linux/arch/x86/mm/tlb.c

725 lines
21 KiB
C
Raw Normal View History

#include <linux/init.h>
#include <linux/mm.h>
#include <linux/spinlock.h>
#include <linux/smp.h>
#include <linux/interrupt.h>
#include <linux/export.h>
x86: Spread tlb flush vector between nodes Currently flush tlb vector allocation is based on below equation: sender = smp_processor_id() % 8 This isn't optimal, CPUs from different node can have the same vector, this causes a lot of lock contention. Instead, we can assign the same vectors to CPUs from the same node, while different node has different vectors. This has below advantages: a. if there is lock contention, the lock contention is between CPUs from one node. This should be much cheaper than the contention between nodes. b. completely avoid lock contention between nodes. This especially benefits kswapd, which is the biggest user of tlb flush, since kswapd sets its affinity to specific node. In my test, this could reduce > 20% CPU overhead in extreme case.The test machine has 4 nodes and each node has 16 CPUs. I then bind each node's kswapd to the first CPU of the node. I run a workload with 4 sequential mmap file read thread. The files are empty sparse file. This workload will trigger a lot of page reclaim and tlbflush. The kswapd bind is to easy trigger the extreme tlb flush lock contention because otherwise kswapd keeps migrating between CPUs of a node and I can't get stable result. Sure in real workload, we can't always see so big tlb flush lock contention, but it's possible. [ hpa: folded in fix from Eric Dumazet to use this_cpu_read() ] Signed-off-by: Shaohua Li <shaohua.li@intel.com> LKML-Reference: <1287544023.4571.8.camel@sli10-conroe.sh.intel.com> Cc: Eric Dumazet <eric.dumazet@gmail.com> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2010-10-20 11:07:03 +08:00
#include <linux/cpu.h>
#include <asm/tlbflush.h>
#include <asm/mmu_context.h>
#include <asm/cache.h>
#include <asm/apic.h>
#include <asm/uv/uv.h>
#include <linux/debugfs.h>
/*
* TLB flushing, formerly SMP-only
* c/o Linus Torvalds.
*
* These mean you can really definitely utterly forget about
* writing to user space from interrupts. (Its not allowed anyway).
*
* Optimizations Manfred Spraul <manfred@colorfullife.com>
*
* More scalable flush, from Andi Kleen
*
* Implement flush IPI by CALL_FUNCTION_VECTOR, Alex Shi
*/
x86/mm: Allow flushing for future ASID switches If changing the page tables in such a way that an invalidation of all contexts (aka. PCIDs / ASIDs) is required, they can be actively invalidated by: 1. INVPCID for each PCID (works for single pages too). 2. Load CR3 with each PCID without the NOFLUSH bit set 3. Load CR3 with the NOFLUSH bit set for each and do INVLPG for each address. But, none of these are really feasible since there are ~6 ASIDs (12 with PAGE_TABLE_ISOLATION) at the time that invalidation is required. Instead of actively invalidating them, invalidate the *current* context and also mark the cpu_tlbstate _quickly_ to indicate future invalidation to be required. At the next context-switch, look for this indicator ('invalidate_other' being set) invalidate all of the cpu_tlbstate.ctxs[] entries. This ensures that any future context switches will do a full flush of the TLB, picking up the previous changes. [ tglx: Folded more fixups from Peter ] Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andy Lutomirski <luto@kernel.org> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: David Laight <David.Laight@aculab.com> Cc: Denys Vlasenko <dvlasenk@redhat.com> Cc: Eduardo Valentin <eduval@amazon.com> Cc: Greg KH <gregkh@linuxfoundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Josh Poimboeuf <jpoimboe@redhat.com> Cc: Juergen Gross <jgross@suse.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Will Deacon <will.deacon@arm.com> Cc: aliguori@amazon.com Cc: daniel.gruss@iaik.tugraz.at Cc: hughd@google.com Cc: keescook@google.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-12-04 22:07:57 +08:00
/*
* We get here when we do something requiring a TLB invalidation
* but could not go invalidate all of the contexts. We do the
* necessary invalidation by clearing out the 'ctx_id' which
* forces a TLB flush when the context is loaded.
*/
void clear_asid_other(void)
{
u16 asid;
/*
* This is only expected to be set if we have disabled
* kernel _PAGE_GLOBAL pages.
*/
if (!static_cpu_has(X86_FEATURE_PTI)) {
WARN_ON_ONCE(1);
return;
}
for (asid = 0; asid < TLB_NR_DYN_ASIDS; asid++) {
/* Do not need to flush the current asid */
if (asid == this_cpu_read(cpu_tlbstate.loaded_mm_asid))
continue;
/*
* Make sure the next time we go to switch to
* this asid, we do a flush:
*/
this_cpu_write(cpu_tlbstate.ctxs[asid].ctx_id, 0);
}
this_cpu_write(cpu_tlbstate.invalidate_other, false);
}
atomic64_t last_mm_ctx_id = ATOMIC64_INIT(1);
x86/mm: Flush more aggressively in lazy TLB mode Since commit: 94b1b03b519b ("x86/mm: Rework lazy TLB mode and TLB freshness tracking") x86's lazy TLB mode has been all the way lazy: when running a kernel thread (including the idle thread), the kernel keeps using the last user mm's page tables without attempting to maintain user TLB coherence at all. From a pure semantic perspective, this is fine -- kernel threads won't attempt to access user pages, so having stale TLB entries doesn't matter. Unfortunately, I forgot about a subtlety. By skipping TLB flushes, we also allow any paging-structure caches that may exist on the CPU to become incoherent. This means that we can have a paging-structure cache entry that references a freed page table, and the CPU is within its rights to do a speculative page walk starting at the freed page table. I can imagine this causing two different problems: - A speculative page walk starting from a bogus page table could read IO addresses. I haven't seen any reports of this causing problems. - A speculative page walk that involves a bogus page table can install garbage in the TLB. Such garbage would always be at a user VA, but some AMD CPUs have logic that triggers a machine check when it notices these bogus entries. I've seen a couple reports of this. Boris further explains the failure mode: > It is actually more of an optimization which assumes that paging-structure > entries are in WB DRAM: > > "TlbCacheDis: cacheable memory disable. Read-write. 0=Enables > performance optimization that assumes PML4, PDP, PDE, and PTE entries > are in cacheable WB-DRAM; memory type checks may be bypassed, and > addresses outside of WB-DRAM may result in undefined behavior or NB > protocol errors. 1=Disables performance optimization and allows PML4, > PDP, PDE and PTE entries to be in any memory type. Operating systems > that maintain page tables in memory types other than WB- DRAM must set > TlbCacheDis to insure proper operation." > > The MCE generated is an NB protocol error to signal that > > "Link: A specific coherent-only packet from a CPU was issued to an > IO link. This may be caused by software which addresses page table > structures in a memory type other than cacheable WB-DRAM without > properly configuring MSRC001_0015[TlbCacheDis]. This may occur, for > example, when page table structure addresses are above top of memory. In > such cases, the NB will generate an MCE if it sees a mismatch between > the memory operation generated by the core and the link type." > > I'm assuming coherent-only packets don't go out on IO links, thus the > error. To fix this, reinstate TLB coherence in lazy mode. With this patch applied, we do it in one of two ways: - If we have PCID, we simply switch back to init_mm's page tables when we enter a kernel thread -- this seems to be quite cheap except for the cost of serializing the CPU. - If we don't have PCID, then we set a flag and switch to init_mm the first time we would otherwise need to flush the TLB. The /sys/kernel/debug/x86/tlb_use_lazy_mode debug switch can be changed to override the default mode for benchmarking. In theory, we could optimize this better by only flushing the TLB in lazy CPUs when a page table is freed. Doing that would require auditing the mm code to make sure that all page table freeing goes through tlb_remove_page() as well as reworking some data structures to implement the improved flush logic. Reported-by: Markus Trippelsdorf <markus@trippelsdorf.de> Reported-by: Adam Borowski <kilobyte@angband.pl> Signed-off-by: Andy Lutomirski <luto@kernel.org> Signed-off-by: Borislav Petkov <bp@suse.de> Cc: Borislav Petkov <bp@alien8.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: Daniel Borkmann <daniel@iogearbox.net> Cc: Eric Biggers <ebiggers@google.com> Cc: Johannes Hirte <johannes.hirte@datenkhaos.de> Cc: Kees Cook <keescook@chromium.org> Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Roman Kagan <rkagan@virtuozzo.com> Cc: Thomas Gleixner <tglx@linutronix.de> Fixes: 94b1b03b519b ("x86/mm: Rework lazy TLB mode and TLB freshness tracking") Link: http://lkml.kernel.org/r/20171009170231.fkpraqokz6e4zeco@pd.tnic Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-10-10 00:50:49 +08:00
x86/mm: Implement PCID based optimization: try to preserve old TLB entries using PCID PCID is a "process context ID" -- it's what other architectures call an address space ID. Every non-global TLB entry is tagged with a PCID, only TLB entries that match the currently selected PCID are used, and we can switch PGDs without flushing the TLB. x86's PCID is 12 bits. This is an unorthodox approach to using PCID. x86's PCID is far too short to uniquely identify a process, and we can't even really uniquely identify a running process because there are monster systems with over 4096 CPUs. To make matters worse, past attempts to use all 12 PCID bits have resulted in slowdowns instead of speedups. This patch uses PCID differently. We use a PCID to identify a recently-used mm on a per-cpu basis. An mm has no fixed PCID binding at all; instead, we give it a fresh PCID each time it's loaded except in cases where we want to preserve the TLB, in which case we reuse a recent value. Here are some benchmark results, done on a Skylake laptop at 2.3 GHz (turbo off, intel_pstate requesting max performance) under KVM with the guest using idle=poll (to avoid artifacts when bouncing between CPUs). I haven't done any real statistics here -- I just ran them in a loop and picked the fastest results that didn't look like outliers. Unpatched means commit a4eb8b993554, so all the bookkeeping overhead is gone. ping-pong between two mms on the same CPU using eventfd: patched: 1.22µs patched, nopcid: 1.33µs unpatched: 1.34µs Same ping-pong, but now touch 512 pages (all zero-page to minimize cache misses) each iteration. dTLB misses are measured by dtlb_load_misses.miss_causes_a_walk: patched: 1.8µs 11M dTLB misses patched, nopcid: 6.2µs, 207M dTLB misses unpatched: 6.1µs, 190M dTLB misses Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/9ee75f17a81770feed616358e6860d98a2a5b1e7.1500957502.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-07-25 12:41:38 +08:00
static void choose_new_asid(struct mm_struct *next, u64 next_tlb_gen,
u16 *new_asid, bool *need_flush)
{
u16 asid;
if (!static_cpu_has(X86_FEATURE_PCID)) {
*new_asid = 0;
*need_flush = true;
return;
}
x86/mm: Allow flushing for future ASID switches If changing the page tables in such a way that an invalidation of all contexts (aka. PCIDs / ASIDs) is required, they can be actively invalidated by: 1. INVPCID for each PCID (works for single pages too). 2. Load CR3 with each PCID without the NOFLUSH bit set 3. Load CR3 with the NOFLUSH bit set for each and do INVLPG for each address. But, none of these are really feasible since there are ~6 ASIDs (12 with PAGE_TABLE_ISOLATION) at the time that invalidation is required. Instead of actively invalidating them, invalidate the *current* context and also mark the cpu_tlbstate _quickly_ to indicate future invalidation to be required. At the next context-switch, look for this indicator ('invalidate_other' being set) invalidate all of the cpu_tlbstate.ctxs[] entries. This ensures that any future context switches will do a full flush of the TLB, picking up the previous changes. [ tglx: Folded more fixups from Peter ] Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andy Lutomirski <luto@kernel.org> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: David Laight <David.Laight@aculab.com> Cc: Denys Vlasenko <dvlasenk@redhat.com> Cc: Eduardo Valentin <eduval@amazon.com> Cc: Greg KH <gregkh@linuxfoundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Josh Poimboeuf <jpoimboe@redhat.com> Cc: Juergen Gross <jgross@suse.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Will Deacon <will.deacon@arm.com> Cc: aliguori@amazon.com Cc: daniel.gruss@iaik.tugraz.at Cc: hughd@google.com Cc: keescook@google.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-12-04 22:07:57 +08:00
if (this_cpu_read(cpu_tlbstate.invalidate_other))
clear_asid_other();
x86/mm: Implement PCID based optimization: try to preserve old TLB entries using PCID PCID is a "process context ID" -- it's what other architectures call an address space ID. Every non-global TLB entry is tagged with a PCID, only TLB entries that match the currently selected PCID are used, and we can switch PGDs without flushing the TLB. x86's PCID is 12 bits. This is an unorthodox approach to using PCID. x86's PCID is far too short to uniquely identify a process, and we can't even really uniquely identify a running process because there are monster systems with over 4096 CPUs. To make matters worse, past attempts to use all 12 PCID bits have resulted in slowdowns instead of speedups. This patch uses PCID differently. We use a PCID to identify a recently-used mm on a per-cpu basis. An mm has no fixed PCID binding at all; instead, we give it a fresh PCID each time it's loaded except in cases where we want to preserve the TLB, in which case we reuse a recent value. Here are some benchmark results, done on a Skylake laptop at 2.3 GHz (turbo off, intel_pstate requesting max performance) under KVM with the guest using idle=poll (to avoid artifacts when bouncing between CPUs). I haven't done any real statistics here -- I just ran them in a loop and picked the fastest results that didn't look like outliers. Unpatched means commit a4eb8b993554, so all the bookkeeping overhead is gone. ping-pong between two mms on the same CPU using eventfd: patched: 1.22µs patched, nopcid: 1.33µs unpatched: 1.34µs Same ping-pong, but now touch 512 pages (all zero-page to minimize cache misses) each iteration. dTLB misses are measured by dtlb_load_misses.miss_causes_a_walk: patched: 1.8µs 11M dTLB misses patched, nopcid: 6.2µs, 207M dTLB misses unpatched: 6.1µs, 190M dTLB misses Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/9ee75f17a81770feed616358e6860d98a2a5b1e7.1500957502.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-07-25 12:41:38 +08:00
for (asid = 0; asid < TLB_NR_DYN_ASIDS; asid++) {
if (this_cpu_read(cpu_tlbstate.ctxs[asid].ctx_id) !=
next->context.ctx_id)
continue;
*new_asid = asid;
*need_flush = (this_cpu_read(cpu_tlbstate.ctxs[asid].tlb_gen) <
next_tlb_gen);
return;
}
/*
* We don't currently own an ASID slot on this CPU.
* Allocate a slot.
*/
*new_asid = this_cpu_add_return(cpu_tlbstate.next_asid, 1) - 1;
if (*new_asid >= TLB_NR_DYN_ASIDS) {
*new_asid = 0;
this_cpu_write(cpu_tlbstate.next_asid, 1);
}
*need_flush = true;
}
static void load_new_mm_cr3(pgd_t *pgdir, u16 new_asid, bool need_flush)
{
unsigned long new_mm_cr3;
if (need_flush) {
x86/mm: Use/Fix PCID to optimize user/kernel switches We can use PCID to retain the TLBs across CR3 switches; including those now part of the user/kernel switch. This increases performance of kernel entry/exit at the cost of more expensive/complicated TLB flushing. Now that we have two address spaces, one for kernel and one for user space, we need two PCIDs per mm. We use the top PCID bit to indicate a user PCID (just like we use the PFN LSB for the PGD). Since we do TLB invalidation from kernel space, the existing code will only invalidate the kernel PCID, we augment that by marking the corresponding user PCID invalid, and upon switching back to userspace, use a flushing CR3 write for the switch. In order to access the user_pcid_flush_mask we use PER_CPU storage, which means the previously established SWAPGS vs CR3 ordering is now mandatory and required. Having to do this memory access does require additional registers, most sites have a functioning stack and we can spill one (RAX), sites without functional stack need to otherwise provide the second scratch register. Note: PCID is generally available on Intel Sandybridge and later CPUs. Note: Up until this point TLB flushing was broken in this series. Based-on-code-from: Dave Hansen <dave.hansen@linux.intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andy Lutomirski <luto@kernel.org> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: Dave Hansen <dave.hansen@linux.intel.com> Cc: David Laight <David.Laight@aculab.com> Cc: Denys Vlasenko <dvlasenk@redhat.com> Cc: Eduardo Valentin <eduval@amazon.com> Cc: Greg KH <gregkh@linuxfoundation.org> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Josh Poimboeuf <jpoimboe@redhat.com> Cc: Juergen Gross <jgross@suse.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Will Deacon <will.deacon@arm.com> Cc: aliguori@amazon.com Cc: daniel.gruss@iaik.tugraz.at Cc: hughd@google.com Cc: keescook@google.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-12-04 22:07:59 +08:00
invalidate_user_asid(new_asid);
new_mm_cr3 = build_cr3(pgdir, new_asid);
} else {
new_mm_cr3 = build_cr3_noflush(pgdir, new_asid);
}
/*
* Caution: many callers of this function expect
* that load_cr3() is serializing and orders TLB
* fills with respect to the mm_cpumask writes.
*/
write_cr3(new_mm_cr3);
}
void leave_mm(int cpu)
{
x86/mm: Rework lazy TLB to track the actual loaded mm Lazy TLB state is currently managed in a rather baroque manner. AFAICT, there are three possible states: - Non-lazy. This means that we're running a user thread or a kernel thread that has called use_mm(). current->mm == current->active_mm == cpu_tlbstate.active_mm and cpu_tlbstate.state == TLBSTATE_OK. - Lazy with user mm. We're running a kernel thread without an mm and we're borrowing an mm_struct. We have current->mm == NULL, current->active_mm == cpu_tlbstate.active_mm, cpu_tlbstate.state != TLBSTATE_OK (i.e. TLBSTATE_LAZY or 0). The current cpu is set in mm_cpumask(current->active_mm). CR3 points to current->active_mm->pgd. The TLB is up to date. - Lazy with init_mm. This happens when we call leave_mm(). We have current->mm == NULL, current->active_mm == cpu_tlbstate.active_mm, but that mm is only relelvant insofar as the scheduler is tracking it for refcounting. cpu_tlbstate.state != TLBSTATE_OK. The current cpu is clear in mm_cpumask(current->active_mm). CR3 points to swapper_pg_dir, i.e. init_mm->pgd. This patch simplifies the situation. Other than perf, x86 stops caring about current->active_mm at all. We have cpu_tlbstate.loaded_mm pointing to the mm that CR3 references. The TLB is always up to date for that mm. leave_mm() just switches us to init_mm. There are no longer any special cases for mm_cpumask, and switch_mm() switches mms without worrying about laziness. After this patch, cpu_tlbstate.state serves only to tell the TLB flush code whether it may switch to init_mm instead of doing a normal flush. This makes fairly extensive changes to xen_exit_mmap(), which used to look a bit like black magic. Perf is unchanged. With or without this change, perf may behave a bit erratically if it tries to read user memory in kernel thread context. We should build on this patch to teach perf to never look at user memory when cpu_tlbstate.loaded_mm != current->mm. Signed-off-by: Andy Lutomirski <luto@kernel.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bpetkov@suse.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Michal Hocko <mhocko@suse.com> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Nadav Amit <namit@vmware.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mm@kvack.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-05-29 01:00:15 +08:00
struct mm_struct *loaded_mm = this_cpu_read(cpu_tlbstate.loaded_mm);
/*
* It's plausible that we're in lazy TLB mode while our mm is init_mm.
* If so, our callers still expect us to flush the TLB, but there
* aren't any user TLB entries in init_mm to worry about.
*
* This needs to happen before any other sanity checks due to
* intel_idle's shenanigans.
*/
if (loaded_mm == &init_mm)
return;
x86/mm: Rework lazy TLB mode and TLB freshness tracking x86's lazy TLB mode used to be fairly weak -- it would switch to init_mm the first time it tried to flush a lazy TLB. This meant an unnecessary CR3 write and, if the flush was remote, an unnecessary IPI. Rewrite it entirely. When we enter lazy mode, we simply remove the CPU from mm_cpumask. This means that we need a way to figure out whether we've missed a flush when we switch back out of lazy mode. I use the tlb_gen machinery to track whether a context is up to date. Note to reviewers: this patch, my itself, looks a bit odd. I'm using an array of length 1 containing (ctx_id, tlb_gen) rather than just storing tlb_gen, and making it at array isn't necessary yet. I'm doing this because the next few patches add PCID support, and, with PCID, we need ctx_id, and the array will end up with a length greater than 1. Making it an array now means that there will be less churn and therefore less stress on your eyeballs. NB: This is dubious but, AFAICT, still correct on Xen and UV. xen_exit_mmap() uses mm_cpumask() for nefarious purposes and this patch changes the way that mm_cpumask() works. This should be okay, since Xen *also* iterates all online CPUs to find all the CPUs it needs to twiddle. The UV tlbflush code is rather dated and should be changed. Here are some benchmark results, done on a Skylake laptop at 2.3 GHz (turbo off, intel_pstate requesting max performance) under KVM with the guest using idle=poll (to avoid artifacts when bouncing between CPUs). I haven't done any real statistics here -- I just ran them in a loop and picked the fastest results that didn't look like outliers. Unpatched means commit a4eb8b993554, so all the bookkeeping overhead is gone. MADV_DONTNEED; touch the page; switch CPUs using sched_setaffinity. In an unpatched kernel, MADV_DONTNEED will send an IPI to the previous CPU. This is intended to be a nearly worst-case test. patched: 13.4µs unpatched: 21.6µs Vitaly's pthread_mmap microbenchmark with 8 threads (on four cores), nrounds = 100, 256M data patched: 1.1 seconds or so unpatched: 1.9 seconds or so The sleepup on Vitaly's test appearss to be because it spends a lot of time blocked on mmap_sem, and this patch avoids sending IPIs to blocked CPUs. Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andrew Banman <abanman@sgi.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Dimitri Sivanich <sivanich@sgi.com> Cc: Juergen Gross <jgross@suse.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Travis <travis@sgi.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/ddf2c92962339f4ba39d8fc41b853936ec0b44f1.1498751203.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-06-29 23:53:17 +08:00
/* Warn if we're not lazy. */
x86/mm: Flush more aggressively in lazy TLB mode Since commit: 94b1b03b519b ("x86/mm: Rework lazy TLB mode and TLB freshness tracking") x86's lazy TLB mode has been all the way lazy: when running a kernel thread (including the idle thread), the kernel keeps using the last user mm's page tables without attempting to maintain user TLB coherence at all. From a pure semantic perspective, this is fine -- kernel threads won't attempt to access user pages, so having stale TLB entries doesn't matter. Unfortunately, I forgot about a subtlety. By skipping TLB flushes, we also allow any paging-structure caches that may exist on the CPU to become incoherent. This means that we can have a paging-structure cache entry that references a freed page table, and the CPU is within its rights to do a speculative page walk starting at the freed page table. I can imagine this causing two different problems: - A speculative page walk starting from a bogus page table could read IO addresses. I haven't seen any reports of this causing problems. - A speculative page walk that involves a bogus page table can install garbage in the TLB. Such garbage would always be at a user VA, but some AMD CPUs have logic that triggers a machine check when it notices these bogus entries. I've seen a couple reports of this. Boris further explains the failure mode: > It is actually more of an optimization which assumes that paging-structure > entries are in WB DRAM: > > "TlbCacheDis: cacheable memory disable. Read-write. 0=Enables > performance optimization that assumes PML4, PDP, PDE, and PTE entries > are in cacheable WB-DRAM; memory type checks may be bypassed, and > addresses outside of WB-DRAM may result in undefined behavior or NB > protocol errors. 1=Disables performance optimization and allows PML4, > PDP, PDE and PTE entries to be in any memory type. Operating systems > that maintain page tables in memory types other than WB- DRAM must set > TlbCacheDis to insure proper operation." > > The MCE generated is an NB protocol error to signal that > > "Link: A specific coherent-only packet from a CPU was issued to an > IO link. This may be caused by software which addresses page table > structures in a memory type other than cacheable WB-DRAM without > properly configuring MSRC001_0015[TlbCacheDis]. This may occur, for > example, when page table structure addresses are above top of memory. In > such cases, the NB will generate an MCE if it sees a mismatch between > the memory operation generated by the core and the link type." > > I'm assuming coherent-only packets don't go out on IO links, thus the > error. To fix this, reinstate TLB coherence in lazy mode. With this patch applied, we do it in one of two ways: - If we have PCID, we simply switch back to init_mm's page tables when we enter a kernel thread -- this seems to be quite cheap except for the cost of serializing the CPU. - If we don't have PCID, then we set a flag and switch to init_mm the first time we would otherwise need to flush the TLB. The /sys/kernel/debug/x86/tlb_use_lazy_mode debug switch can be changed to override the default mode for benchmarking. In theory, we could optimize this better by only flushing the TLB in lazy CPUs when a page table is freed. Doing that would require auditing the mm code to make sure that all page table freeing goes through tlb_remove_page() as well as reworking some data structures to implement the improved flush logic. Reported-by: Markus Trippelsdorf <markus@trippelsdorf.de> Reported-by: Adam Borowski <kilobyte@angband.pl> Signed-off-by: Andy Lutomirski <luto@kernel.org> Signed-off-by: Borislav Petkov <bp@suse.de> Cc: Borislav Petkov <bp@alien8.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: Daniel Borkmann <daniel@iogearbox.net> Cc: Eric Biggers <ebiggers@google.com> Cc: Johannes Hirte <johannes.hirte@datenkhaos.de> Cc: Kees Cook <keescook@chromium.org> Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Roman Kagan <rkagan@virtuozzo.com> Cc: Thomas Gleixner <tglx@linutronix.de> Fixes: 94b1b03b519b ("x86/mm: Rework lazy TLB mode and TLB freshness tracking") Link: http://lkml.kernel.org/r/20171009170231.fkpraqokz6e4zeco@pd.tnic Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-10-10 00:50:49 +08:00
WARN_ON(!this_cpu_read(cpu_tlbstate.is_lazy));
x86/mm: Rework lazy TLB to track the actual loaded mm Lazy TLB state is currently managed in a rather baroque manner. AFAICT, there are three possible states: - Non-lazy. This means that we're running a user thread or a kernel thread that has called use_mm(). current->mm == current->active_mm == cpu_tlbstate.active_mm and cpu_tlbstate.state == TLBSTATE_OK. - Lazy with user mm. We're running a kernel thread without an mm and we're borrowing an mm_struct. We have current->mm == NULL, current->active_mm == cpu_tlbstate.active_mm, cpu_tlbstate.state != TLBSTATE_OK (i.e. TLBSTATE_LAZY or 0). The current cpu is set in mm_cpumask(current->active_mm). CR3 points to current->active_mm->pgd. The TLB is up to date. - Lazy with init_mm. This happens when we call leave_mm(). We have current->mm == NULL, current->active_mm == cpu_tlbstate.active_mm, but that mm is only relelvant insofar as the scheduler is tracking it for refcounting. cpu_tlbstate.state != TLBSTATE_OK. The current cpu is clear in mm_cpumask(current->active_mm). CR3 points to swapper_pg_dir, i.e. init_mm->pgd. This patch simplifies the situation. Other than perf, x86 stops caring about current->active_mm at all. We have cpu_tlbstate.loaded_mm pointing to the mm that CR3 references. The TLB is always up to date for that mm. leave_mm() just switches us to init_mm. There are no longer any special cases for mm_cpumask, and switch_mm() switches mms without worrying about laziness. After this patch, cpu_tlbstate.state serves only to tell the TLB flush code whether it may switch to init_mm instead of doing a normal flush. This makes fairly extensive changes to xen_exit_mmap(), which used to look a bit like black magic. Perf is unchanged. With or without this change, perf may behave a bit erratically if it tries to read user memory in kernel thread context. We should build on this patch to teach perf to never look at user memory when cpu_tlbstate.loaded_mm != current->mm. Signed-off-by: Andy Lutomirski <luto@kernel.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bpetkov@suse.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Michal Hocko <mhocko@suse.com> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Nadav Amit <namit@vmware.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mm@kvack.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-05-29 01:00:15 +08:00
switch_mm(NULL, &init_mm, NULL);
}
Revert "x86/mm: Stop calling leave_mm() in idle code" This reverts commit 43858b4f25cf0adc5c2ca9cf5ce5fdf2532941e5. The reason I removed the leave_mm() calls in question is because the heuristic wasn't needed after that patch. With the original version of my PCID series, we never flushed a "lazy cpu" (i.e. a CPU running kernel thread) due a flush on the loaded mm. Unfortunately, that caused architectural issues, so now I've reinstated these flushes on non-PCID systems in: commit b956575bed91 ("x86/mm: Flush more aggressively in lazy TLB mode"). That, in turn, gives us a power management and occasionally performance regression as compared to old kernels: a process that goes into a deep idle state on a given CPU and gets its mm flushed due to activity on a different CPU will wake the idle CPU. Reinstate the old ugly heuristic: if a CPU goes into ACPI C3 or an intel_idle state that is likely to cause a TLB flush gets its mm switched to init_mm before going idle. FWIW, this heuristic is lousy. Whether we should change CR3 before idle isn't a good hint except insofar as the performance hit is a bit lower if the TLB is getting flushed by the idle code anyway. What we really want to know is whether we anticipate being idle long enough that the mm is likely to be flushed before we wake up. This is more a matter of the expected latency than the idle state that gets chosen. This heuristic also completely fails on systems that don't know whether the TLB will be flushed (e.g. AMD systems?). OTOH it may be a bit obsolete anyway -- PCID systems don't presently benefit from this heuristic at all. We also shouldn't do this callback from innermost bit of the idle code due to the RCU nastiness it causes. All the information need is available before rcu_idle_enter() needs to happen. Signed-off-by: Andy Lutomirski <luto@kernel.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Borislav Petkov <bpetkov@suse.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: Denys Vlasenko <dvlasenk@redhat.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Josh Poimboeuf <jpoimboe@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Fixes: 43858b4f25cf "x86/mm: Stop calling leave_mm() in idle code" Link: http://lkml.kernel.org/r/c513bbd4e653747213e05bc7062de000bf0202a5.1509793738.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-11-04 19:16:12 +08:00
EXPORT_SYMBOL_GPL(leave_mm);
void switch_mm(struct mm_struct *prev, struct mm_struct *next,
struct task_struct *tsk)
{
unsigned long flags;
local_irq_save(flags);
switch_mm_irqs_off(prev, next, tsk);
local_irq_restore(flags);
}
x86/mm/64: Fix vmapped stack syncing on very-large-memory 4-level systems Neil Berrington reported a double-fault on a VM with 768GB of RAM that uses large amounts of vmalloc space with PTI enabled. The cause is that load_new_mm_cr3() was never fixed to take the 5-level pgd folding code into account, so, on a 4-level kernel, the pgd synchronization logic compiles away to exactly nothing. Interestingly, the problem doesn't trigger with nopti. I assume this is because the kernel is mapped with global pages if we boot with nopti. The sequence of operations when we create a new task is that we first load its mm while still running on the old stack (which crashes if the old stack is unmapped in the new mm unless the TLB saves us), then we call prepare_switch_to(), and then we switch to the new stack. prepare_switch_to() pokes the new stack directly, which will populate the mapping through vmalloc_fault(). I assume that we're getting lucky on non-PTI systems -- the old stack's TLB entry stays alive long enough to make it all the way through prepare_switch_to() and switch_to() so that we make it to a valid stack. Fixes: b50858ce3e2a ("x86/mm/vmalloc: Add 5-level paging support") Reported-and-tested-by: Neil Berrington <neil.berrington@datacore.com> Signed-off-by: Andy Lutomirski <luto@kernel.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Cc: Konstantin Khlebnikov <khlebnikov@yandex-team.ru> Cc: stable@vger.kernel.org Cc: Dave Hansen <dave.hansen@intel.com> Cc: Borislav Petkov <bp@alien8.de> Link: https://lkml.kernel.org/r/346541c56caed61abbe693d7d2742b4a380c5001.1516914529.git.luto@kernel.org
2018-01-26 05:12:14 +08:00
static void sync_current_stack_to_mm(struct mm_struct *mm)
{
unsigned long sp = current_stack_pointer;
pgd_t *pgd = pgd_offset(mm, sp);
if (CONFIG_PGTABLE_LEVELS > 4) {
if (unlikely(pgd_none(*pgd))) {
pgd_t *pgd_ref = pgd_offset_k(sp);
set_pgd(pgd, *pgd_ref);
}
} else {
/*
* "pgd" is faked. The top level entries are "p4d"s, so sync
* the p4d. This compiles to approximately the same code as
* the 5-level case.
*/
p4d_t *p4d = p4d_offset(pgd, sp);
if (unlikely(p4d_none(*p4d))) {
pgd_t *pgd_ref = pgd_offset_k(sp);
p4d_t *p4d_ref = p4d_offset(pgd_ref, sp);
set_p4d(p4d, *p4d_ref);
}
}
}
void switch_mm_irqs_off(struct mm_struct *prev, struct mm_struct *next,
struct task_struct *tsk)
{
x86/mm: Rework lazy TLB to track the actual loaded mm Lazy TLB state is currently managed in a rather baroque manner. AFAICT, there are three possible states: - Non-lazy. This means that we're running a user thread or a kernel thread that has called use_mm(). current->mm == current->active_mm == cpu_tlbstate.active_mm and cpu_tlbstate.state == TLBSTATE_OK. - Lazy with user mm. We're running a kernel thread without an mm and we're borrowing an mm_struct. We have current->mm == NULL, current->active_mm == cpu_tlbstate.active_mm, cpu_tlbstate.state != TLBSTATE_OK (i.e. TLBSTATE_LAZY or 0). The current cpu is set in mm_cpumask(current->active_mm). CR3 points to current->active_mm->pgd. The TLB is up to date. - Lazy with init_mm. This happens when we call leave_mm(). We have current->mm == NULL, current->active_mm == cpu_tlbstate.active_mm, but that mm is only relelvant insofar as the scheduler is tracking it for refcounting. cpu_tlbstate.state != TLBSTATE_OK. The current cpu is clear in mm_cpumask(current->active_mm). CR3 points to swapper_pg_dir, i.e. init_mm->pgd. This patch simplifies the situation. Other than perf, x86 stops caring about current->active_mm at all. We have cpu_tlbstate.loaded_mm pointing to the mm that CR3 references. The TLB is always up to date for that mm. leave_mm() just switches us to init_mm. There are no longer any special cases for mm_cpumask, and switch_mm() switches mms without worrying about laziness. After this patch, cpu_tlbstate.state serves only to tell the TLB flush code whether it may switch to init_mm instead of doing a normal flush. This makes fairly extensive changes to xen_exit_mmap(), which used to look a bit like black magic. Perf is unchanged. With or without this change, perf may behave a bit erratically if it tries to read user memory in kernel thread context. We should build on this patch to teach perf to never look at user memory when cpu_tlbstate.loaded_mm != current->mm. Signed-off-by: Andy Lutomirski <luto@kernel.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bpetkov@suse.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Michal Hocko <mhocko@suse.com> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Nadav Amit <namit@vmware.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mm@kvack.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-05-29 01:00:15 +08:00
struct mm_struct *real_prev = this_cpu_read(cpu_tlbstate.loaded_mm);
x86/mm: Implement PCID based optimization: try to preserve old TLB entries using PCID PCID is a "process context ID" -- it's what other architectures call an address space ID. Every non-global TLB entry is tagged with a PCID, only TLB entries that match the currently selected PCID are used, and we can switch PGDs without flushing the TLB. x86's PCID is 12 bits. This is an unorthodox approach to using PCID. x86's PCID is far too short to uniquely identify a process, and we can't even really uniquely identify a running process because there are monster systems with over 4096 CPUs. To make matters worse, past attempts to use all 12 PCID bits have resulted in slowdowns instead of speedups. This patch uses PCID differently. We use a PCID to identify a recently-used mm on a per-cpu basis. An mm has no fixed PCID binding at all; instead, we give it a fresh PCID each time it's loaded except in cases where we want to preserve the TLB, in which case we reuse a recent value. Here are some benchmark results, done on a Skylake laptop at 2.3 GHz (turbo off, intel_pstate requesting max performance) under KVM with the guest using idle=poll (to avoid artifacts when bouncing between CPUs). I haven't done any real statistics here -- I just ran them in a loop and picked the fastest results that didn't look like outliers. Unpatched means commit a4eb8b993554, so all the bookkeeping overhead is gone. ping-pong between two mms on the same CPU using eventfd: patched: 1.22µs patched, nopcid: 1.33µs unpatched: 1.34µs Same ping-pong, but now touch 512 pages (all zero-page to minimize cache misses) each iteration. dTLB misses are measured by dtlb_load_misses.miss_causes_a_walk: patched: 1.8µs 11M dTLB misses patched, nopcid: 6.2µs, 207M dTLB misses unpatched: 6.1µs, 190M dTLB misses Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/9ee75f17a81770feed616358e6860d98a2a5b1e7.1500957502.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-07-25 12:41:38 +08:00
u16 prev_asid = this_cpu_read(cpu_tlbstate.loaded_mm_asid);
x86/mm: Rework lazy TLB mode and TLB freshness tracking x86's lazy TLB mode used to be fairly weak -- it would switch to init_mm the first time it tried to flush a lazy TLB. This meant an unnecessary CR3 write and, if the flush was remote, an unnecessary IPI. Rewrite it entirely. When we enter lazy mode, we simply remove the CPU from mm_cpumask. This means that we need a way to figure out whether we've missed a flush when we switch back out of lazy mode. I use the tlb_gen machinery to track whether a context is up to date. Note to reviewers: this patch, my itself, looks a bit odd. I'm using an array of length 1 containing (ctx_id, tlb_gen) rather than just storing tlb_gen, and making it at array isn't necessary yet. I'm doing this because the next few patches add PCID support, and, with PCID, we need ctx_id, and the array will end up with a length greater than 1. Making it an array now means that there will be less churn and therefore less stress on your eyeballs. NB: This is dubious but, AFAICT, still correct on Xen and UV. xen_exit_mmap() uses mm_cpumask() for nefarious purposes and this patch changes the way that mm_cpumask() works. This should be okay, since Xen *also* iterates all online CPUs to find all the CPUs it needs to twiddle. The UV tlbflush code is rather dated and should be changed. Here are some benchmark results, done on a Skylake laptop at 2.3 GHz (turbo off, intel_pstate requesting max performance) under KVM with the guest using idle=poll (to avoid artifacts when bouncing between CPUs). I haven't done any real statistics here -- I just ran them in a loop and picked the fastest results that didn't look like outliers. Unpatched means commit a4eb8b993554, so all the bookkeeping overhead is gone. MADV_DONTNEED; touch the page; switch CPUs using sched_setaffinity. In an unpatched kernel, MADV_DONTNEED will send an IPI to the previous CPU. This is intended to be a nearly worst-case test. patched: 13.4µs unpatched: 21.6µs Vitaly's pthread_mmap microbenchmark with 8 threads (on four cores), nrounds = 100, 256M data patched: 1.1 seconds or so unpatched: 1.9 seconds or so The sleepup on Vitaly's test appearss to be because it spends a lot of time blocked on mmap_sem, and this patch avoids sending IPIs to blocked CPUs. Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andrew Banman <abanman@sgi.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Dimitri Sivanich <sivanich@sgi.com> Cc: Juergen Gross <jgross@suse.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Travis <travis@sgi.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/ddf2c92962339f4ba39d8fc41b853936ec0b44f1.1498751203.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-06-29 23:53:17 +08:00
unsigned cpu = smp_processor_id();
u64 next_tlb_gen;
x86/mm: Rework lazy TLB to track the actual loaded mm Lazy TLB state is currently managed in a rather baroque manner. AFAICT, there are three possible states: - Non-lazy. This means that we're running a user thread or a kernel thread that has called use_mm(). current->mm == current->active_mm == cpu_tlbstate.active_mm and cpu_tlbstate.state == TLBSTATE_OK. - Lazy with user mm. We're running a kernel thread without an mm and we're borrowing an mm_struct. We have current->mm == NULL, current->active_mm == cpu_tlbstate.active_mm, cpu_tlbstate.state != TLBSTATE_OK (i.e. TLBSTATE_LAZY or 0). The current cpu is set in mm_cpumask(current->active_mm). CR3 points to current->active_mm->pgd. The TLB is up to date. - Lazy with init_mm. This happens when we call leave_mm(). We have current->mm == NULL, current->active_mm == cpu_tlbstate.active_mm, but that mm is only relelvant insofar as the scheduler is tracking it for refcounting. cpu_tlbstate.state != TLBSTATE_OK. The current cpu is clear in mm_cpumask(current->active_mm). CR3 points to swapper_pg_dir, i.e. init_mm->pgd. This patch simplifies the situation. Other than perf, x86 stops caring about current->active_mm at all. We have cpu_tlbstate.loaded_mm pointing to the mm that CR3 references. The TLB is always up to date for that mm. leave_mm() just switches us to init_mm. There are no longer any special cases for mm_cpumask, and switch_mm() switches mms without worrying about laziness. After this patch, cpu_tlbstate.state serves only to tell the TLB flush code whether it may switch to init_mm instead of doing a normal flush. This makes fairly extensive changes to xen_exit_mmap(), which used to look a bit like black magic. Perf is unchanged. With or without this change, perf may behave a bit erratically if it tries to read user memory in kernel thread context. We should build on this patch to teach perf to never look at user memory when cpu_tlbstate.loaded_mm != current->mm. Signed-off-by: Andy Lutomirski <luto@kernel.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bpetkov@suse.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Michal Hocko <mhocko@suse.com> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Nadav Amit <namit@vmware.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mm@kvack.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-05-29 01:00:15 +08:00
/*
x86/mm: Rework lazy TLB mode and TLB freshness tracking x86's lazy TLB mode used to be fairly weak -- it would switch to init_mm the first time it tried to flush a lazy TLB. This meant an unnecessary CR3 write and, if the flush was remote, an unnecessary IPI. Rewrite it entirely. When we enter lazy mode, we simply remove the CPU from mm_cpumask. This means that we need a way to figure out whether we've missed a flush when we switch back out of lazy mode. I use the tlb_gen machinery to track whether a context is up to date. Note to reviewers: this patch, my itself, looks a bit odd. I'm using an array of length 1 containing (ctx_id, tlb_gen) rather than just storing tlb_gen, and making it at array isn't necessary yet. I'm doing this because the next few patches add PCID support, and, with PCID, we need ctx_id, and the array will end up with a length greater than 1. Making it an array now means that there will be less churn and therefore less stress on your eyeballs. NB: This is dubious but, AFAICT, still correct on Xen and UV. xen_exit_mmap() uses mm_cpumask() for nefarious purposes and this patch changes the way that mm_cpumask() works. This should be okay, since Xen *also* iterates all online CPUs to find all the CPUs it needs to twiddle. The UV tlbflush code is rather dated and should be changed. Here are some benchmark results, done on a Skylake laptop at 2.3 GHz (turbo off, intel_pstate requesting max performance) under KVM with the guest using idle=poll (to avoid artifacts when bouncing between CPUs). I haven't done any real statistics here -- I just ran them in a loop and picked the fastest results that didn't look like outliers. Unpatched means commit a4eb8b993554, so all the bookkeeping overhead is gone. MADV_DONTNEED; touch the page; switch CPUs using sched_setaffinity. In an unpatched kernel, MADV_DONTNEED will send an IPI to the previous CPU. This is intended to be a nearly worst-case test. patched: 13.4µs unpatched: 21.6µs Vitaly's pthread_mmap microbenchmark with 8 threads (on four cores), nrounds = 100, 256M data patched: 1.1 seconds or so unpatched: 1.9 seconds or so The sleepup on Vitaly's test appearss to be because it spends a lot of time blocked on mmap_sem, and this patch avoids sending IPIs to blocked CPUs. Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andrew Banman <abanman@sgi.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Dimitri Sivanich <sivanich@sgi.com> Cc: Juergen Gross <jgross@suse.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Travis <travis@sgi.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/ddf2c92962339f4ba39d8fc41b853936ec0b44f1.1498751203.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-06-29 23:53:17 +08:00
* NB: The scheduler will call us with prev == next when switching
* from lazy TLB mode to normal mode if active_mm isn't changing.
* When this happens, we don't assume that CR3 (and hence
* cpu_tlbstate.loaded_mm) matches next.
x86/mm: Rework lazy TLB to track the actual loaded mm Lazy TLB state is currently managed in a rather baroque manner. AFAICT, there are three possible states: - Non-lazy. This means that we're running a user thread or a kernel thread that has called use_mm(). current->mm == current->active_mm == cpu_tlbstate.active_mm and cpu_tlbstate.state == TLBSTATE_OK. - Lazy with user mm. We're running a kernel thread without an mm and we're borrowing an mm_struct. We have current->mm == NULL, current->active_mm == cpu_tlbstate.active_mm, cpu_tlbstate.state != TLBSTATE_OK (i.e. TLBSTATE_LAZY or 0). The current cpu is set in mm_cpumask(current->active_mm). CR3 points to current->active_mm->pgd. The TLB is up to date. - Lazy with init_mm. This happens when we call leave_mm(). We have current->mm == NULL, current->active_mm == cpu_tlbstate.active_mm, but that mm is only relelvant insofar as the scheduler is tracking it for refcounting. cpu_tlbstate.state != TLBSTATE_OK. The current cpu is clear in mm_cpumask(current->active_mm). CR3 points to swapper_pg_dir, i.e. init_mm->pgd. This patch simplifies the situation. Other than perf, x86 stops caring about current->active_mm at all. We have cpu_tlbstate.loaded_mm pointing to the mm that CR3 references. The TLB is always up to date for that mm. leave_mm() just switches us to init_mm. There are no longer any special cases for mm_cpumask, and switch_mm() switches mms without worrying about laziness. After this patch, cpu_tlbstate.state serves only to tell the TLB flush code whether it may switch to init_mm instead of doing a normal flush. This makes fairly extensive changes to xen_exit_mmap(), which used to look a bit like black magic. Perf is unchanged. With or without this change, perf may behave a bit erratically if it tries to read user memory in kernel thread context. We should build on this patch to teach perf to never look at user memory when cpu_tlbstate.loaded_mm != current->mm. Signed-off-by: Andy Lutomirski <luto@kernel.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bpetkov@suse.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Michal Hocko <mhocko@suse.com> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Nadav Amit <namit@vmware.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mm@kvack.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-05-29 01:00:15 +08:00
*
* NB: leave_mm() calls us with prev == NULL and tsk == NULL.
*/
x86/mm/64: Enable vmapped stacks (CONFIG_HAVE_ARCH_VMAP_STACK=y) This allows x86_64 kernels to enable vmapped stacks by setting HAVE_ARCH_VMAP_STACK=y - which enables the CONFIG_VMAP_STACK=y high level Kconfig option. There are a couple of interesting bits: First, x86 lazily faults in top-level paging entries for the vmalloc area. This won't work if we get a page fault while trying to access the stack: the CPU will promote it to a double-fault and we'll die. To avoid this problem, probe the new stack when switching stacks and forcibly populate the pgd entry for the stack when switching mms. Second, once we have guard pages around the stack, we'll want to detect and handle stack overflow. I didn't enable it on x86_32. We'd need to rework the double-fault code a bit and I'm concerned about running out of vmalloc virtual addresses under some workloads. This patch, by itself, will behave somewhat erratically when the stack overflows while RSP is still more than a few tens of bytes above the bottom of the stack. Specifically, we'll get #PF and make it to no_context and them oops without reliably triggering a double-fault, and no_context doesn't know about stack overflows. The next patch will improve that case. Thank you to Nadav and Brian for helping me pay enough attention to the SDM to hopefully get this right. Signed-off-by: Andy Lutomirski <luto@kernel.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: Denys Vlasenko <dvlasenk@redhat.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Josh Poimboeuf <jpoimboe@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/c88f3e2920b18e6cc621d772a04a62c06869037e.1470907718.git.luto@kernel.org [ Minor edits. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-11 17:35:23 +08:00
x86/mm: Rework lazy TLB mode and TLB freshness tracking x86's lazy TLB mode used to be fairly weak -- it would switch to init_mm the first time it tried to flush a lazy TLB. This meant an unnecessary CR3 write and, if the flush was remote, an unnecessary IPI. Rewrite it entirely. When we enter lazy mode, we simply remove the CPU from mm_cpumask. This means that we need a way to figure out whether we've missed a flush when we switch back out of lazy mode. I use the tlb_gen machinery to track whether a context is up to date. Note to reviewers: this patch, my itself, looks a bit odd. I'm using an array of length 1 containing (ctx_id, tlb_gen) rather than just storing tlb_gen, and making it at array isn't necessary yet. I'm doing this because the next few patches add PCID support, and, with PCID, we need ctx_id, and the array will end up with a length greater than 1. Making it an array now means that there will be less churn and therefore less stress on your eyeballs. NB: This is dubious but, AFAICT, still correct on Xen and UV. xen_exit_mmap() uses mm_cpumask() for nefarious purposes and this patch changes the way that mm_cpumask() works. This should be okay, since Xen *also* iterates all online CPUs to find all the CPUs it needs to twiddle. The UV tlbflush code is rather dated and should be changed. Here are some benchmark results, done on a Skylake laptop at 2.3 GHz (turbo off, intel_pstate requesting max performance) under KVM with the guest using idle=poll (to avoid artifacts when bouncing between CPUs). I haven't done any real statistics here -- I just ran them in a loop and picked the fastest results that didn't look like outliers. Unpatched means commit a4eb8b993554, so all the bookkeeping overhead is gone. MADV_DONTNEED; touch the page; switch CPUs using sched_setaffinity. In an unpatched kernel, MADV_DONTNEED will send an IPI to the previous CPU. This is intended to be a nearly worst-case test. patched: 13.4µs unpatched: 21.6µs Vitaly's pthread_mmap microbenchmark with 8 threads (on four cores), nrounds = 100, 256M data patched: 1.1 seconds or so unpatched: 1.9 seconds or so The sleepup on Vitaly's test appearss to be because it spends a lot of time blocked on mmap_sem, and this patch avoids sending IPIs to blocked CPUs. Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andrew Banman <abanman@sgi.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Dimitri Sivanich <sivanich@sgi.com> Cc: Juergen Gross <jgross@suse.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Travis <travis@sgi.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/ddf2c92962339f4ba39d8fc41b853936ec0b44f1.1498751203.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-06-29 23:53:17 +08:00
/* We don't want flush_tlb_func_* to run concurrently with us. */
if (IS_ENABLED(CONFIG_PROVE_LOCKING))
WARN_ON_ONCE(!irqs_disabled());
/*
* Verify that CR3 is what we think it is. This will catch
* hypothetical buggy code that directly switches to swapper_pg_dir
x86/mm: Implement PCID based optimization: try to preserve old TLB entries using PCID PCID is a "process context ID" -- it's what other architectures call an address space ID. Every non-global TLB entry is tagged with a PCID, only TLB entries that match the currently selected PCID are used, and we can switch PGDs without flushing the TLB. x86's PCID is 12 bits. This is an unorthodox approach to using PCID. x86's PCID is far too short to uniquely identify a process, and we can't even really uniquely identify a running process because there are monster systems with over 4096 CPUs. To make matters worse, past attempts to use all 12 PCID bits have resulted in slowdowns instead of speedups. This patch uses PCID differently. We use a PCID to identify a recently-used mm on a per-cpu basis. An mm has no fixed PCID binding at all; instead, we give it a fresh PCID each time it's loaded except in cases where we want to preserve the TLB, in which case we reuse a recent value. Here are some benchmark results, done on a Skylake laptop at 2.3 GHz (turbo off, intel_pstate requesting max performance) under KVM with the guest using idle=poll (to avoid artifacts when bouncing between CPUs). I haven't done any real statistics here -- I just ran them in a loop and picked the fastest results that didn't look like outliers. Unpatched means commit a4eb8b993554, so all the bookkeeping overhead is gone. ping-pong between two mms on the same CPU using eventfd: patched: 1.22µs patched, nopcid: 1.33µs unpatched: 1.34µs Same ping-pong, but now touch 512 pages (all zero-page to minimize cache misses) each iteration. dTLB misses are measured by dtlb_load_misses.miss_causes_a_walk: patched: 1.8µs 11M dTLB misses patched, nopcid: 6.2µs, 207M dTLB misses unpatched: 6.1µs, 190M dTLB misses Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/9ee75f17a81770feed616358e6860d98a2a5b1e7.1500957502.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-07-25 12:41:38 +08:00
* without going through leave_mm() / switch_mm_irqs_off() or that
* does something like write_cr3(read_cr3_pa()).
*
* Only do this check if CONFIG_DEBUG_VM=y because __read_cr3()
* isn't free.
x86/mm: Rework lazy TLB mode and TLB freshness tracking x86's lazy TLB mode used to be fairly weak -- it would switch to init_mm the first time it tried to flush a lazy TLB. This meant an unnecessary CR3 write and, if the flush was remote, an unnecessary IPI. Rewrite it entirely. When we enter lazy mode, we simply remove the CPU from mm_cpumask. This means that we need a way to figure out whether we've missed a flush when we switch back out of lazy mode. I use the tlb_gen machinery to track whether a context is up to date. Note to reviewers: this patch, my itself, looks a bit odd. I'm using an array of length 1 containing (ctx_id, tlb_gen) rather than just storing tlb_gen, and making it at array isn't necessary yet. I'm doing this because the next few patches add PCID support, and, with PCID, we need ctx_id, and the array will end up with a length greater than 1. Making it an array now means that there will be less churn and therefore less stress on your eyeballs. NB: This is dubious but, AFAICT, still correct on Xen and UV. xen_exit_mmap() uses mm_cpumask() for nefarious purposes and this patch changes the way that mm_cpumask() works. This should be okay, since Xen *also* iterates all online CPUs to find all the CPUs it needs to twiddle. The UV tlbflush code is rather dated and should be changed. Here are some benchmark results, done on a Skylake laptop at 2.3 GHz (turbo off, intel_pstate requesting max performance) under KVM with the guest using idle=poll (to avoid artifacts when bouncing between CPUs). I haven't done any real statistics here -- I just ran them in a loop and picked the fastest results that didn't look like outliers. Unpatched means commit a4eb8b993554, so all the bookkeeping overhead is gone. MADV_DONTNEED; touch the page; switch CPUs using sched_setaffinity. In an unpatched kernel, MADV_DONTNEED will send an IPI to the previous CPU. This is intended to be a nearly worst-case test. patched: 13.4µs unpatched: 21.6µs Vitaly's pthread_mmap microbenchmark with 8 threads (on four cores), nrounds = 100, 256M data patched: 1.1 seconds or so unpatched: 1.9 seconds or so The sleepup on Vitaly's test appearss to be because it spends a lot of time blocked on mmap_sem, and this patch avoids sending IPIs to blocked CPUs. Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andrew Banman <abanman@sgi.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Dimitri Sivanich <sivanich@sgi.com> Cc: Juergen Gross <jgross@suse.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Travis <travis@sgi.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/ddf2c92962339f4ba39d8fc41b853936ec0b44f1.1498751203.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-06-29 23:53:17 +08:00
*/
#ifdef CONFIG_DEBUG_VM
if (WARN_ON_ONCE(__read_cr3() != build_cr3(real_prev->pgd, prev_asid))) {
/*
* If we were to BUG here, we'd be very likely to kill
* the system so hard that we don't see the call trace.
* Try to recover instead by ignoring the error and doing
* a global flush to minimize the chance of corruption.
*
* (This is far from being a fully correct recovery.
* Architecturally, the CPU could prefetch something
* back into an incorrect ASID slot and leave it there
* to cause trouble down the road. It's better than
* nothing, though.)
*/
__flush_tlb_all();
}
#endif
x86/mm: Flush more aggressively in lazy TLB mode Since commit: 94b1b03b519b ("x86/mm: Rework lazy TLB mode and TLB freshness tracking") x86's lazy TLB mode has been all the way lazy: when running a kernel thread (including the idle thread), the kernel keeps using the last user mm's page tables without attempting to maintain user TLB coherence at all. From a pure semantic perspective, this is fine -- kernel threads won't attempt to access user pages, so having stale TLB entries doesn't matter. Unfortunately, I forgot about a subtlety. By skipping TLB flushes, we also allow any paging-structure caches that may exist on the CPU to become incoherent. This means that we can have a paging-structure cache entry that references a freed page table, and the CPU is within its rights to do a speculative page walk starting at the freed page table. I can imagine this causing two different problems: - A speculative page walk starting from a bogus page table could read IO addresses. I haven't seen any reports of this causing problems. - A speculative page walk that involves a bogus page table can install garbage in the TLB. Such garbage would always be at a user VA, but some AMD CPUs have logic that triggers a machine check when it notices these bogus entries. I've seen a couple reports of this. Boris further explains the failure mode: > It is actually more of an optimization which assumes that paging-structure > entries are in WB DRAM: > > "TlbCacheDis: cacheable memory disable. Read-write. 0=Enables > performance optimization that assumes PML4, PDP, PDE, and PTE entries > are in cacheable WB-DRAM; memory type checks may be bypassed, and > addresses outside of WB-DRAM may result in undefined behavior or NB > protocol errors. 1=Disables performance optimization and allows PML4, > PDP, PDE and PTE entries to be in any memory type. Operating systems > that maintain page tables in memory types other than WB- DRAM must set > TlbCacheDis to insure proper operation." > > The MCE generated is an NB protocol error to signal that > > "Link: A specific coherent-only packet from a CPU was issued to an > IO link. This may be caused by software which addresses page table > structures in a memory type other than cacheable WB-DRAM without > properly configuring MSRC001_0015[TlbCacheDis]. This may occur, for > example, when page table structure addresses are above top of memory. In > such cases, the NB will generate an MCE if it sees a mismatch between > the memory operation generated by the core and the link type." > > I'm assuming coherent-only packets don't go out on IO links, thus the > error. To fix this, reinstate TLB coherence in lazy mode. With this patch applied, we do it in one of two ways: - If we have PCID, we simply switch back to init_mm's page tables when we enter a kernel thread -- this seems to be quite cheap except for the cost of serializing the CPU. - If we don't have PCID, then we set a flag and switch to init_mm the first time we would otherwise need to flush the TLB. The /sys/kernel/debug/x86/tlb_use_lazy_mode debug switch can be changed to override the default mode for benchmarking. In theory, we could optimize this better by only flushing the TLB in lazy CPUs when a page table is freed. Doing that would require auditing the mm code to make sure that all page table freeing goes through tlb_remove_page() as well as reworking some data structures to implement the improved flush logic. Reported-by: Markus Trippelsdorf <markus@trippelsdorf.de> Reported-by: Adam Borowski <kilobyte@angband.pl> Signed-off-by: Andy Lutomirski <luto@kernel.org> Signed-off-by: Borislav Petkov <bp@suse.de> Cc: Borislav Petkov <bp@alien8.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: Daniel Borkmann <daniel@iogearbox.net> Cc: Eric Biggers <ebiggers@google.com> Cc: Johannes Hirte <johannes.hirte@datenkhaos.de> Cc: Kees Cook <keescook@chromium.org> Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Roman Kagan <rkagan@virtuozzo.com> Cc: Thomas Gleixner <tglx@linutronix.de> Fixes: 94b1b03b519b ("x86/mm: Rework lazy TLB mode and TLB freshness tracking") Link: http://lkml.kernel.org/r/20171009170231.fkpraqokz6e4zeco@pd.tnic Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-10-10 00:50:49 +08:00
this_cpu_write(cpu_tlbstate.is_lazy, false);
x86/mm/64: Enable vmapped stacks (CONFIG_HAVE_ARCH_VMAP_STACK=y) This allows x86_64 kernels to enable vmapped stacks by setting HAVE_ARCH_VMAP_STACK=y - which enables the CONFIG_VMAP_STACK=y high level Kconfig option. There are a couple of interesting bits: First, x86 lazily faults in top-level paging entries for the vmalloc area. This won't work if we get a page fault while trying to access the stack: the CPU will promote it to a double-fault and we'll die. To avoid this problem, probe the new stack when switching stacks and forcibly populate the pgd entry for the stack when switching mms. Second, once we have guard pages around the stack, we'll want to detect and handle stack overflow. I didn't enable it on x86_32. We'd need to rework the double-fault code a bit and I'm concerned about running out of vmalloc virtual addresses under some workloads. This patch, by itself, will behave somewhat erratically when the stack overflows while RSP is still more than a few tens of bytes above the bottom of the stack. Specifically, we'll get #PF and make it to no_context and them oops without reliably triggering a double-fault, and no_context doesn't know about stack overflows. The next patch will improve that case. Thank you to Nadav and Brian for helping me pay enough attention to the SDM to hopefully get this right. Signed-off-by: Andy Lutomirski <luto@kernel.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: Denys Vlasenko <dvlasenk@redhat.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Josh Poimboeuf <jpoimboe@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/c88f3e2920b18e6cc621d772a04a62c06869037e.1470907718.git.luto@kernel.org [ Minor edits. ] Signed-off-by: Ingo Molnar <mingo@kernel.org>
2016-08-11 17:35:23 +08:00
membarrier: Document scheduler barrier requirements Document the membarrier requirement on having a full memory barrier in __schedule() after coming from user-space, before storing to rq->curr. It is provided by smp_mb__after_spinlock() in __schedule(). Document that membarrier requires a full barrier on transition from kernel thread to userspace thread. We currently have an implicit barrier from atomic_dec_and_test() in mmdrop() that ensures this. The x86 switch_mm_irqs_off() full barrier is currently provided by many cpumask update operations as well as write_cr3(). Document that write_cr3() provides this barrier. Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com> Acked-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Andrea Parri <parri.andrea@gmail.com> Cc: Andrew Hunter <ahh@google.com> Cc: Andy Lutomirski <luto@kernel.org> Cc: Avi Kivity <avi@scylladb.com> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Boqun Feng <boqun.feng@gmail.com> Cc: Dave Watson <davejwatson@fb.com> Cc: David Sehr <sehr@google.com> Cc: Greg Hackmann <ghackmann@google.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Maged Michael <maged.michael@gmail.com> Cc: Michael Ellerman <mpe@ellerman.id.au> Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com> Cc: Paul Mackerras <paulus@samba.org> Cc: Russell King <linux@armlinux.org.uk> Cc: Will Deacon <will.deacon@arm.com> Cc: linux-api@vger.kernel.org Link: http://lkml.kernel.org/r/20180129202020.8515-4-mathieu.desnoyers@efficios.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-01-30 04:20:12 +08:00
/*
* The membarrier system call requires a full memory barrier
* before returning to user-space, after storing to rq->curr.
* Writing to CR3 provides that full memory barrier.
*/
x86/mm: Rework lazy TLB to track the actual loaded mm Lazy TLB state is currently managed in a rather baroque manner. AFAICT, there are three possible states: - Non-lazy. This means that we're running a user thread or a kernel thread that has called use_mm(). current->mm == current->active_mm == cpu_tlbstate.active_mm and cpu_tlbstate.state == TLBSTATE_OK. - Lazy with user mm. We're running a kernel thread without an mm and we're borrowing an mm_struct. We have current->mm == NULL, current->active_mm == cpu_tlbstate.active_mm, cpu_tlbstate.state != TLBSTATE_OK (i.e. TLBSTATE_LAZY or 0). The current cpu is set in mm_cpumask(current->active_mm). CR3 points to current->active_mm->pgd. The TLB is up to date. - Lazy with init_mm. This happens when we call leave_mm(). We have current->mm == NULL, current->active_mm == cpu_tlbstate.active_mm, but that mm is only relelvant insofar as the scheduler is tracking it for refcounting. cpu_tlbstate.state != TLBSTATE_OK. The current cpu is clear in mm_cpumask(current->active_mm). CR3 points to swapper_pg_dir, i.e. init_mm->pgd. This patch simplifies the situation. Other than perf, x86 stops caring about current->active_mm at all. We have cpu_tlbstate.loaded_mm pointing to the mm that CR3 references. The TLB is always up to date for that mm. leave_mm() just switches us to init_mm. There are no longer any special cases for mm_cpumask, and switch_mm() switches mms without worrying about laziness. After this patch, cpu_tlbstate.state serves only to tell the TLB flush code whether it may switch to init_mm instead of doing a normal flush. This makes fairly extensive changes to xen_exit_mmap(), which used to look a bit like black magic. Perf is unchanged. With or without this change, perf may behave a bit erratically if it tries to read user memory in kernel thread context. We should build on this patch to teach perf to never look at user memory when cpu_tlbstate.loaded_mm != current->mm. Signed-off-by: Andy Lutomirski <luto@kernel.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bpetkov@suse.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Michal Hocko <mhocko@suse.com> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Nadav Amit <namit@vmware.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mm@kvack.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-05-29 01:00:15 +08:00
if (real_prev == next) {
VM_WARN_ON(this_cpu_read(cpu_tlbstate.ctxs[prev_asid].ctx_id) !=
next->context.ctx_id);
x86/mm: Rework lazy TLB mode and TLB freshness tracking x86's lazy TLB mode used to be fairly weak -- it would switch to init_mm the first time it tried to flush a lazy TLB. This meant an unnecessary CR3 write and, if the flush was remote, an unnecessary IPI. Rewrite it entirely. When we enter lazy mode, we simply remove the CPU from mm_cpumask. This means that we need a way to figure out whether we've missed a flush when we switch back out of lazy mode. I use the tlb_gen machinery to track whether a context is up to date. Note to reviewers: this patch, my itself, looks a bit odd. I'm using an array of length 1 containing (ctx_id, tlb_gen) rather than just storing tlb_gen, and making it at array isn't necessary yet. I'm doing this because the next few patches add PCID support, and, with PCID, we need ctx_id, and the array will end up with a length greater than 1. Making it an array now means that there will be less churn and therefore less stress on your eyeballs. NB: This is dubious but, AFAICT, still correct on Xen and UV. xen_exit_mmap() uses mm_cpumask() for nefarious purposes and this patch changes the way that mm_cpumask() works. This should be okay, since Xen *also* iterates all online CPUs to find all the CPUs it needs to twiddle. The UV tlbflush code is rather dated and should be changed. Here are some benchmark results, done on a Skylake laptop at 2.3 GHz (turbo off, intel_pstate requesting max performance) under KVM with the guest using idle=poll (to avoid artifacts when bouncing between CPUs). I haven't done any real statistics here -- I just ran them in a loop and picked the fastest results that didn't look like outliers. Unpatched means commit a4eb8b993554, so all the bookkeeping overhead is gone. MADV_DONTNEED; touch the page; switch CPUs using sched_setaffinity. In an unpatched kernel, MADV_DONTNEED will send an IPI to the previous CPU. This is intended to be a nearly worst-case test. patched: 13.4µs unpatched: 21.6µs Vitaly's pthread_mmap microbenchmark with 8 threads (on four cores), nrounds = 100, 256M data patched: 1.1 seconds or so unpatched: 1.9 seconds or so The sleepup on Vitaly's test appearss to be because it spends a lot of time blocked on mmap_sem, and this patch avoids sending IPIs to blocked CPUs. Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andrew Banman <abanman@sgi.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Dimitri Sivanich <sivanich@sgi.com> Cc: Juergen Gross <jgross@suse.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Travis <travis@sgi.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/ddf2c92962339f4ba39d8fc41b853936ec0b44f1.1498751203.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-06-29 23:53:17 +08:00
/*
x86/mm: Flush more aggressively in lazy TLB mode Since commit: 94b1b03b519b ("x86/mm: Rework lazy TLB mode and TLB freshness tracking") x86's lazy TLB mode has been all the way lazy: when running a kernel thread (including the idle thread), the kernel keeps using the last user mm's page tables without attempting to maintain user TLB coherence at all. From a pure semantic perspective, this is fine -- kernel threads won't attempt to access user pages, so having stale TLB entries doesn't matter. Unfortunately, I forgot about a subtlety. By skipping TLB flushes, we also allow any paging-structure caches that may exist on the CPU to become incoherent. This means that we can have a paging-structure cache entry that references a freed page table, and the CPU is within its rights to do a speculative page walk starting at the freed page table. I can imagine this causing two different problems: - A speculative page walk starting from a bogus page table could read IO addresses. I haven't seen any reports of this causing problems. - A speculative page walk that involves a bogus page table can install garbage in the TLB. Such garbage would always be at a user VA, but some AMD CPUs have logic that triggers a machine check when it notices these bogus entries. I've seen a couple reports of this. Boris further explains the failure mode: > It is actually more of an optimization which assumes that paging-structure > entries are in WB DRAM: > > "TlbCacheDis: cacheable memory disable. Read-write. 0=Enables > performance optimization that assumes PML4, PDP, PDE, and PTE entries > are in cacheable WB-DRAM; memory type checks may be bypassed, and > addresses outside of WB-DRAM may result in undefined behavior or NB > protocol errors. 1=Disables performance optimization and allows PML4, > PDP, PDE and PTE entries to be in any memory type. Operating systems > that maintain page tables in memory types other than WB- DRAM must set > TlbCacheDis to insure proper operation." > > The MCE generated is an NB protocol error to signal that > > "Link: A specific coherent-only packet from a CPU was issued to an > IO link. This may be caused by software which addresses page table > structures in a memory type other than cacheable WB-DRAM without > properly configuring MSRC001_0015[TlbCacheDis]. This may occur, for > example, when page table structure addresses are above top of memory. In > such cases, the NB will generate an MCE if it sees a mismatch between > the memory operation generated by the core and the link type." > > I'm assuming coherent-only packets don't go out on IO links, thus the > error. To fix this, reinstate TLB coherence in lazy mode. With this patch applied, we do it in one of two ways: - If we have PCID, we simply switch back to init_mm's page tables when we enter a kernel thread -- this seems to be quite cheap except for the cost of serializing the CPU. - If we don't have PCID, then we set a flag and switch to init_mm the first time we would otherwise need to flush the TLB. The /sys/kernel/debug/x86/tlb_use_lazy_mode debug switch can be changed to override the default mode for benchmarking. In theory, we could optimize this better by only flushing the TLB in lazy CPUs when a page table is freed. Doing that would require auditing the mm code to make sure that all page table freeing goes through tlb_remove_page() as well as reworking some data structures to implement the improved flush logic. Reported-by: Markus Trippelsdorf <markus@trippelsdorf.de> Reported-by: Adam Borowski <kilobyte@angband.pl> Signed-off-by: Andy Lutomirski <luto@kernel.org> Signed-off-by: Borislav Petkov <bp@suse.de> Cc: Borislav Petkov <bp@alien8.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: Daniel Borkmann <daniel@iogearbox.net> Cc: Eric Biggers <ebiggers@google.com> Cc: Johannes Hirte <johannes.hirte@datenkhaos.de> Cc: Kees Cook <keescook@chromium.org> Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Roman Kagan <rkagan@virtuozzo.com> Cc: Thomas Gleixner <tglx@linutronix.de> Fixes: 94b1b03b519b ("x86/mm: Rework lazy TLB mode and TLB freshness tracking") Link: http://lkml.kernel.org/r/20171009170231.fkpraqokz6e4zeco@pd.tnic Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-10-10 00:50:49 +08:00
* We don't currently support having a real mm loaded without
* our cpu set in mm_cpumask(). We have all the bookkeeping
* in place to figure out whether we would need to flush
* if our cpu were cleared in mm_cpumask(), but we don't
* currently use it.
*/
x86/mm: Flush more aggressively in lazy TLB mode Since commit: 94b1b03b519b ("x86/mm: Rework lazy TLB mode and TLB freshness tracking") x86's lazy TLB mode has been all the way lazy: when running a kernel thread (including the idle thread), the kernel keeps using the last user mm's page tables without attempting to maintain user TLB coherence at all. From a pure semantic perspective, this is fine -- kernel threads won't attempt to access user pages, so having stale TLB entries doesn't matter. Unfortunately, I forgot about a subtlety. By skipping TLB flushes, we also allow any paging-structure caches that may exist on the CPU to become incoherent. This means that we can have a paging-structure cache entry that references a freed page table, and the CPU is within its rights to do a speculative page walk starting at the freed page table. I can imagine this causing two different problems: - A speculative page walk starting from a bogus page table could read IO addresses. I haven't seen any reports of this causing problems. - A speculative page walk that involves a bogus page table can install garbage in the TLB. Such garbage would always be at a user VA, but some AMD CPUs have logic that triggers a machine check when it notices these bogus entries. I've seen a couple reports of this. Boris further explains the failure mode: > It is actually more of an optimization which assumes that paging-structure > entries are in WB DRAM: > > "TlbCacheDis: cacheable memory disable. Read-write. 0=Enables > performance optimization that assumes PML4, PDP, PDE, and PTE entries > are in cacheable WB-DRAM; memory type checks may be bypassed, and > addresses outside of WB-DRAM may result in undefined behavior or NB > protocol errors. 1=Disables performance optimization and allows PML4, > PDP, PDE and PTE entries to be in any memory type. Operating systems > that maintain page tables in memory types other than WB- DRAM must set > TlbCacheDis to insure proper operation." > > The MCE generated is an NB protocol error to signal that > > "Link: A specific coherent-only packet from a CPU was issued to an > IO link. This may be caused by software which addresses page table > structures in a memory type other than cacheable WB-DRAM without > properly configuring MSRC001_0015[TlbCacheDis]. This may occur, for > example, when page table structure addresses are above top of memory. In > such cases, the NB will generate an MCE if it sees a mismatch between > the memory operation generated by the core and the link type." > > I'm assuming coherent-only packets don't go out on IO links, thus the > error. To fix this, reinstate TLB coherence in lazy mode. With this patch applied, we do it in one of two ways: - If we have PCID, we simply switch back to init_mm's page tables when we enter a kernel thread -- this seems to be quite cheap except for the cost of serializing the CPU. - If we don't have PCID, then we set a flag and switch to init_mm the first time we would otherwise need to flush the TLB. The /sys/kernel/debug/x86/tlb_use_lazy_mode debug switch can be changed to override the default mode for benchmarking. In theory, we could optimize this better by only flushing the TLB in lazy CPUs when a page table is freed. Doing that would require auditing the mm code to make sure that all page table freeing goes through tlb_remove_page() as well as reworking some data structures to implement the improved flush logic. Reported-by: Markus Trippelsdorf <markus@trippelsdorf.de> Reported-by: Adam Borowski <kilobyte@angband.pl> Signed-off-by: Andy Lutomirski <luto@kernel.org> Signed-off-by: Borislav Petkov <bp@suse.de> Cc: Borislav Petkov <bp@alien8.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: Daniel Borkmann <daniel@iogearbox.net> Cc: Eric Biggers <ebiggers@google.com> Cc: Johannes Hirte <johannes.hirte@datenkhaos.de> Cc: Kees Cook <keescook@chromium.org> Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Roman Kagan <rkagan@virtuozzo.com> Cc: Thomas Gleixner <tglx@linutronix.de> Fixes: 94b1b03b519b ("x86/mm: Rework lazy TLB mode and TLB freshness tracking") Link: http://lkml.kernel.org/r/20171009170231.fkpraqokz6e4zeco@pd.tnic Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-10-10 00:50:49 +08:00
if (WARN_ON_ONCE(real_prev != &init_mm &&
!cpumask_test_cpu(cpu, mm_cpumask(next))))
cpumask_set_cpu(cpu, mm_cpumask(next));
return;
x86/mm: Rework lazy TLB mode and TLB freshness tracking x86's lazy TLB mode used to be fairly weak -- it would switch to init_mm the first time it tried to flush a lazy TLB. This meant an unnecessary CR3 write and, if the flush was remote, an unnecessary IPI. Rewrite it entirely. When we enter lazy mode, we simply remove the CPU from mm_cpumask. This means that we need a way to figure out whether we've missed a flush when we switch back out of lazy mode. I use the tlb_gen machinery to track whether a context is up to date. Note to reviewers: this patch, my itself, looks a bit odd. I'm using an array of length 1 containing (ctx_id, tlb_gen) rather than just storing tlb_gen, and making it at array isn't necessary yet. I'm doing this because the next few patches add PCID support, and, with PCID, we need ctx_id, and the array will end up with a length greater than 1. Making it an array now means that there will be less churn and therefore less stress on your eyeballs. NB: This is dubious but, AFAICT, still correct on Xen and UV. xen_exit_mmap() uses mm_cpumask() for nefarious purposes and this patch changes the way that mm_cpumask() works. This should be okay, since Xen *also* iterates all online CPUs to find all the CPUs it needs to twiddle. The UV tlbflush code is rather dated and should be changed. Here are some benchmark results, done on a Skylake laptop at 2.3 GHz (turbo off, intel_pstate requesting max performance) under KVM with the guest using idle=poll (to avoid artifacts when bouncing between CPUs). I haven't done any real statistics here -- I just ran them in a loop and picked the fastest results that didn't look like outliers. Unpatched means commit a4eb8b993554, so all the bookkeeping overhead is gone. MADV_DONTNEED; touch the page; switch CPUs using sched_setaffinity. In an unpatched kernel, MADV_DONTNEED will send an IPI to the previous CPU. This is intended to be a nearly worst-case test. patched: 13.4µs unpatched: 21.6µs Vitaly's pthread_mmap microbenchmark with 8 threads (on four cores), nrounds = 100, 256M data patched: 1.1 seconds or so unpatched: 1.9 seconds or so The sleepup on Vitaly's test appearss to be because it spends a lot of time blocked on mmap_sem, and this patch avoids sending IPIs to blocked CPUs. Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andrew Banman <abanman@sgi.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Dimitri Sivanich <sivanich@sgi.com> Cc: Juergen Gross <jgross@suse.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Travis <travis@sgi.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/ddf2c92962339f4ba39d8fc41b853936ec0b44f1.1498751203.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-06-29 23:53:17 +08:00
} else {
x86/mm: Implement PCID based optimization: try to preserve old TLB entries using PCID PCID is a "process context ID" -- it's what other architectures call an address space ID. Every non-global TLB entry is tagged with a PCID, only TLB entries that match the currently selected PCID are used, and we can switch PGDs without flushing the TLB. x86's PCID is 12 bits. This is an unorthodox approach to using PCID. x86's PCID is far too short to uniquely identify a process, and we can't even really uniquely identify a running process because there are monster systems with over 4096 CPUs. To make matters worse, past attempts to use all 12 PCID bits have resulted in slowdowns instead of speedups. This patch uses PCID differently. We use a PCID to identify a recently-used mm on a per-cpu basis. An mm has no fixed PCID binding at all; instead, we give it a fresh PCID each time it's loaded except in cases where we want to preserve the TLB, in which case we reuse a recent value. Here are some benchmark results, done on a Skylake laptop at 2.3 GHz (turbo off, intel_pstate requesting max performance) under KVM with the guest using idle=poll (to avoid artifacts when bouncing between CPUs). I haven't done any real statistics here -- I just ran them in a loop and picked the fastest results that didn't look like outliers. Unpatched means commit a4eb8b993554, so all the bookkeeping overhead is gone. ping-pong between two mms on the same CPU using eventfd: patched: 1.22µs patched, nopcid: 1.33µs unpatched: 1.34µs Same ping-pong, but now touch 512 pages (all zero-page to minimize cache misses) each iteration. dTLB misses are measured by dtlb_load_misses.miss_causes_a_walk: patched: 1.8µs 11M dTLB misses patched, nopcid: 6.2µs, 207M dTLB misses unpatched: 6.1µs, 190M dTLB misses Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/9ee75f17a81770feed616358e6860d98a2a5b1e7.1500957502.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-07-25 12:41:38 +08:00
u16 new_asid;
bool need_flush;
x86/mm: Rework lazy TLB mode and TLB freshness tracking x86's lazy TLB mode used to be fairly weak -- it would switch to init_mm the first time it tried to flush a lazy TLB. This meant an unnecessary CR3 write and, if the flush was remote, an unnecessary IPI. Rewrite it entirely. When we enter lazy mode, we simply remove the CPU from mm_cpumask. This means that we need a way to figure out whether we've missed a flush when we switch back out of lazy mode. I use the tlb_gen machinery to track whether a context is up to date. Note to reviewers: this patch, my itself, looks a bit odd. I'm using an array of length 1 containing (ctx_id, tlb_gen) rather than just storing tlb_gen, and making it at array isn't necessary yet. I'm doing this because the next few patches add PCID support, and, with PCID, we need ctx_id, and the array will end up with a length greater than 1. Making it an array now means that there will be less churn and therefore less stress on your eyeballs. NB: This is dubious but, AFAICT, still correct on Xen and UV. xen_exit_mmap() uses mm_cpumask() for nefarious purposes and this patch changes the way that mm_cpumask() works. This should be okay, since Xen *also* iterates all online CPUs to find all the CPUs it needs to twiddle. The UV tlbflush code is rather dated and should be changed. Here are some benchmark results, done on a Skylake laptop at 2.3 GHz (turbo off, intel_pstate requesting max performance) under KVM with the guest using idle=poll (to avoid artifacts when bouncing between CPUs). I haven't done any real statistics here -- I just ran them in a loop and picked the fastest results that didn't look like outliers. Unpatched means commit a4eb8b993554, so all the bookkeeping overhead is gone. MADV_DONTNEED; touch the page; switch CPUs using sched_setaffinity. In an unpatched kernel, MADV_DONTNEED will send an IPI to the previous CPU. This is intended to be a nearly worst-case test. patched: 13.4µs unpatched: 21.6µs Vitaly's pthread_mmap microbenchmark with 8 threads (on four cores), nrounds = 100, 256M data patched: 1.1 seconds or so unpatched: 1.9 seconds or so The sleepup on Vitaly's test appearss to be because it spends a lot of time blocked on mmap_sem, and this patch avoids sending IPIs to blocked CPUs. Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andrew Banman <abanman@sgi.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Dimitri Sivanich <sivanich@sgi.com> Cc: Juergen Gross <jgross@suse.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Travis <travis@sgi.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/ddf2c92962339f4ba39d8fc41b853936ec0b44f1.1498751203.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-06-29 23:53:17 +08:00
if (IS_ENABLED(CONFIG_VMAP_STACK)) {
/*
* If our current stack is in vmalloc space and isn't
* mapped in the new pgd, we'll double-fault. Forcibly
* map it.
*/
x86/mm/64: Fix vmapped stack syncing on very-large-memory 4-level systems Neil Berrington reported a double-fault on a VM with 768GB of RAM that uses large amounts of vmalloc space with PTI enabled. The cause is that load_new_mm_cr3() was never fixed to take the 5-level pgd folding code into account, so, on a 4-level kernel, the pgd synchronization logic compiles away to exactly nothing. Interestingly, the problem doesn't trigger with nopti. I assume this is because the kernel is mapped with global pages if we boot with nopti. The sequence of operations when we create a new task is that we first load its mm while still running on the old stack (which crashes if the old stack is unmapped in the new mm unless the TLB saves us), then we call prepare_switch_to(), and then we switch to the new stack. prepare_switch_to() pokes the new stack directly, which will populate the mapping through vmalloc_fault(). I assume that we're getting lucky on non-PTI systems -- the old stack's TLB entry stays alive long enough to make it all the way through prepare_switch_to() and switch_to() so that we make it to a valid stack. Fixes: b50858ce3e2a ("x86/mm/vmalloc: Add 5-level paging support") Reported-and-tested-by: Neil Berrington <neil.berrington@datacore.com> Signed-off-by: Andy Lutomirski <luto@kernel.org> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Cc: Konstantin Khlebnikov <khlebnikov@yandex-team.ru> Cc: stable@vger.kernel.org Cc: Dave Hansen <dave.hansen@intel.com> Cc: Borislav Petkov <bp@alien8.de> Link: https://lkml.kernel.org/r/346541c56caed61abbe693d7d2742b4a380c5001.1516914529.git.luto@kernel.org
2018-01-26 05:12:14 +08:00
sync_current_stack_to_mm(next);
x86/mm: Rework lazy TLB mode and TLB freshness tracking x86's lazy TLB mode used to be fairly weak -- it would switch to init_mm the first time it tried to flush a lazy TLB. This meant an unnecessary CR3 write and, if the flush was remote, an unnecessary IPI. Rewrite it entirely. When we enter lazy mode, we simply remove the CPU from mm_cpumask. This means that we need a way to figure out whether we've missed a flush when we switch back out of lazy mode. I use the tlb_gen machinery to track whether a context is up to date. Note to reviewers: this patch, my itself, looks a bit odd. I'm using an array of length 1 containing (ctx_id, tlb_gen) rather than just storing tlb_gen, and making it at array isn't necessary yet. I'm doing this because the next few patches add PCID support, and, with PCID, we need ctx_id, and the array will end up with a length greater than 1. Making it an array now means that there will be less churn and therefore less stress on your eyeballs. NB: This is dubious but, AFAICT, still correct on Xen and UV. xen_exit_mmap() uses mm_cpumask() for nefarious purposes and this patch changes the way that mm_cpumask() works. This should be okay, since Xen *also* iterates all online CPUs to find all the CPUs it needs to twiddle. The UV tlbflush code is rather dated and should be changed. Here are some benchmark results, done on a Skylake laptop at 2.3 GHz (turbo off, intel_pstate requesting max performance) under KVM with the guest using idle=poll (to avoid artifacts when bouncing between CPUs). I haven't done any real statistics here -- I just ran them in a loop and picked the fastest results that didn't look like outliers. Unpatched means commit a4eb8b993554, so all the bookkeeping overhead is gone. MADV_DONTNEED; touch the page; switch CPUs using sched_setaffinity. In an unpatched kernel, MADV_DONTNEED will send an IPI to the previous CPU. This is intended to be a nearly worst-case test. patched: 13.4µs unpatched: 21.6µs Vitaly's pthread_mmap microbenchmark with 8 threads (on four cores), nrounds = 100, 256M data patched: 1.1 seconds or so unpatched: 1.9 seconds or so The sleepup on Vitaly's test appearss to be because it spends a lot of time blocked on mmap_sem, and this patch avoids sending IPIs to blocked CPUs. Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andrew Banman <abanman@sgi.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Dimitri Sivanich <sivanich@sgi.com> Cc: Juergen Gross <jgross@suse.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Travis <travis@sgi.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/ddf2c92962339f4ba39d8fc41b853936ec0b44f1.1498751203.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-06-29 23:53:17 +08:00
}
x86/mm: Rework lazy TLB mode and TLB freshness tracking x86's lazy TLB mode used to be fairly weak -- it would switch to init_mm the first time it tried to flush a lazy TLB. This meant an unnecessary CR3 write and, if the flush was remote, an unnecessary IPI. Rewrite it entirely. When we enter lazy mode, we simply remove the CPU from mm_cpumask. This means that we need a way to figure out whether we've missed a flush when we switch back out of lazy mode. I use the tlb_gen machinery to track whether a context is up to date. Note to reviewers: this patch, my itself, looks a bit odd. I'm using an array of length 1 containing (ctx_id, tlb_gen) rather than just storing tlb_gen, and making it at array isn't necessary yet. I'm doing this because the next few patches add PCID support, and, with PCID, we need ctx_id, and the array will end up with a length greater than 1. Making it an array now means that there will be less churn and therefore less stress on your eyeballs. NB: This is dubious but, AFAICT, still correct on Xen and UV. xen_exit_mmap() uses mm_cpumask() for nefarious purposes and this patch changes the way that mm_cpumask() works. This should be okay, since Xen *also* iterates all online CPUs to find all the CPUs it needs to twiddle. The UV tlbflush code is rather dated and should be changed. Here are some benchmark results, done on a Skylake laptop at 2.3 GHz (turbo off, intel_pstate requesting max performance) under KVM with the guest using idle=poll (to avoid artifacts when bouncing between CPUs). I haven't done any real statistics here -- I just ran them in a loop and picked the fastest results that didn't look like outliers. Unpatched means commit a4eb8b993554, so all the bookkeeping overhead is gone. MADV_DONTNEED; touch the page; switch CPUs using sched_setaffinity. In an unpatched kernel, MADV_DONTNEED will send an IPI to the previous CPU. This is intended to be a nearly worst-case test. patched: 13.4µs unpatched: 21.6µs Vitaly's pthread_mmap microbenchmark with 8 threads (on four cores), nrounds = 100, 256M data patched: 1.1 seconds or so unpatched: 1.9 seconds or so The sleepup on Vitaly's test appearss to be because it spends a lot of time blocked on mmap_sem, and this patch avoids sending IPIs to blocked CPUs. Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andrew Banman <abanman@sgi.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Dimitri Sivanich <sivanich@sgi.com> Cc: Juergen Gross <jgross@suse.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Travis <travis@sgi.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/ddf2c92962339f4ba39d8fc41b853936ec0b44f1.1498751203.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-06-29 23:53:17 +08:00
/* Stop remote flushes for the previous mm */
x86/mm: Flush more aggressively in lazy TLB mode Since commit: 94b1b03b519b ("x86/mm: Rework lazy TLB mode and TLB freshness tracking") x86's lazy TLB mode has been all the way lazy: when running a kernel thread (including the idle thread), the kernel keeps using the last user mm's page tables without attempting to maintain user TLB coherence at all. From a pure semantic perspective, this is fine -- kernel threads won't attempt to access user pages, so having stale TLB entries doesn't matter. Unfortunately, I forgot about a subtlety. By skipping TLB flushes, we also allow any paging-structure caches that may exist on the CPU to become incoherent. This means that we can have a paging-structure cache entry that references a freed page table, and the CPU is within its rights to do a speculative page walk starting at the freed page table. I can imagine this causing two different problems: - A speculative page walk starting from a bogus page table could read IO addresses. I haven't seen any reports of this causing problems. - A speculative page walk that involves a bogus page table can install garbage in the TLB. Such garbage would always be at a user VA, but some AMD CPUs have logic that triggers a machine check when it notices these bogus entries. I've seen a couple reports of this. Boris further explains the failure mode: > It is actually more of an optimization which assumes that paging-structure > entries are in WB DRAM: > > "TlbCacheDis: cacheable memory disable. Read-write. 0=Enables > performance optimization that assumes PML4, PDP, PDE, and PTE entries > are in cacheable WB-DRAM; memory type checks may be bypassed, and > addresses outside of WB-DRAM may result in undefined behavior or NB > protocol errors. 1=Disables performance optimization and allows PML4, > PDP, PDE and PTE entries to be in any memory type. Operating systems > that maintain page tables in memory types other than WB- DRAM must set > TlbCacheDis to insure proper operation." > > The MCE generated is an NB protocol error to signal that > > "Link: A specific coherent-only packet from a CPU was issued to an > IO link. This may be caused by software which addresses page table > structures in a memory type other than cacheable WB-DRAM without > properly configuring MSRC001_0015[TlbCacheDis]. This may occur, for > example, when page table structure addresses are above top of memory. In > such cases, the NB will generate an MCE if it sees a mismatch between > the memory operation generated by the core and the link type." > > I'm assuming coherent-only packets don't go out on IO links, thus the > error. To fix this, reinstate TLB coherence in lazy mode. With this patch applied, we do it in one of two ways: - If we have PCID, we simply switch back to init_mm's page tables when we enter a kernel thread -- this seems to be quite cheap except for the cost of serializing the CPU. - If we don't have PCID, then we set a flag and switch to init_mm the first time we would otherwise need to flush the TLB. The /sys/kernel/debug/x86/tlb_use_lazy_mode debug switch can be changed to override the default mode for benchmarking. In theory, we could optimize this better by only flushing the TLB in lazy CPUs when a page table is freed. Doing that would require auditing the mm code to make sure that all page table freeing goes through tlb_remove_page() as well as reworking some data structures to implement the improved flush logic. Reported-by: Markus Trippelsdorf <markus@trippelsdorf.de> Reported-by: Adam Borowski <kilobyte@angband.pl> Signed-off-by: Andy Lutomirski <luto@kernel.org> Signed-off-by: Borislav Petkov <bp@suse.de> Cc: Borislav Petkov <bp@alien8.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: Daniel Borkmann <daniel@iogearbox.net> Cc: Eric Biggers <ebiggers@google.com> Cc: Johannes Hirte <johannes.hirte@datenkhaos.de> Cc: Kees Cook <keescook@chromium.org> Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Roman Kagan <rkagan@virtuozzo.com> Cc: Thomas Gleixner <tglx@linutronix.de> Fixes: 94b1b03b519b ("x86/mm: Rework lazy TLB mode and TLB freshness tracking") Link: http://lkml.kernel.org/r/20171009170231.fkpraqokz6e4zeco@pd.tnic Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-10-10 00:50:49 +08:00
VM_WARN_ON_ONCE(!cpumask_test_cpu(cpu, mm_cpumask(real_prev)) &&
real_prev != &init_mm);
cpumask_clear_cpu(cpu, mm_cpumask(real_prev));
x86/mm: Rework lazy TLB to track the actual loaded mm Lazy TLB state is currently managed in a rather baroque manner. AFAICT, there are three possible states: - Non-lazy. This means that we're running a user thread or a kernel thread that has called use_mm(). current->mm == current->active_mm == cpu_tlbstate.active_mm and cpu_tlbstate.state == TLBSTATE_OK. - Lazy with user mm. We're running a kernel thread without an mm and we're borrowing an mm_struct. We have current->mm == NULL, current->active_mm == cpu_tlbstate.active_mm, cpu_tlbstate.state != TLBSTATE_OK (i.e. TLBSTATE_LAZY or 0). The current cpu is set in mm_cpumask(current->active_mm). CR3 points to current->active_mm->pgd. The TLB is up to date. - Lazy with init_mm. This happens when we call leave_mm(). We have current->mm == NULL, current->active_mm == cpu_tlbstate.active_mm, but that mm is only relelvant insofar as the scheduler is tracking it for refcounting. cpu_tlbstate.state != TLBSTATE_OK. The current cpu is clear in mm_cpumask(current->active_mm). CR3 points to swapper_pg_dir, i.e. init_mm->pgd. This patch simplifies the situation. Other than perf, x86 stops caring about current->active_mm at all. We have cpu_tlbstate.loaded_mm pointing to the mm that CR3 references. The TLB is always up to date for that mm. leave_mm() just switches us to init_mm. There are no longer any special cases for mm_cpumask, and switch_mm() switches mms without worrying about laziness. After this patch, cpu_tlbstate.state serves only to tell the TLB flush code whether it may switch to init_mm instead of doing a normal flush. This makes fairly extensive changes to xen_exit_mmap(), which used to look a bit like black magic. Perf is unchanged. With or without this change, perf may behave a bit erratically if it tries to read user memory in kernel thread context. We should build on this patch to teach perf to never look at user memory when cpu_tlbstate.loaded_mm != current->mm. Signed-off-by: Andy Lutomirski <luto@kernel.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bpetkov@suse.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Michal Hocko <mhocko@suse.com> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Nadav Amit <namit@vmware.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mm@kvack.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-05-29 01:00:15 +08:00
x86/mm: Rework lazy TLB mode and TLB freshness tracking x86's lazy TLB mode used to be fairly weak -- it would switch to init_mm the first time it tried to flush a lazy TLB. This meant an unnecessary CR3 write and, if the flush was remote, an unnecessary IPI. Rewrite it entirely. When we enter lazy mode, we simply remove the CPU from mm_cpumask. This means that we need a way to figure out whether we've missed a flush when we switch back out of lazy mode. I use the tlb_gen machinery to track whether a context is up to date. Note to reviewers: this patch, my itself, looks a bit odd. I'm using an array of length 1 containing (ctx_id, tlb_gen) rather than just storing tlb_gen, and making it at array isn't necessary yet. I'm doing this because the next few patches add PCID support, and, with PCID, we need ctx_id, and the array will end up with a length greater than 1. Making it an array now means that there will be less churn and therefore less stress on your eyeballs. NB: This is dubious but, AFAICT, still correct on Xen and UV. xen_exit_mmap() uses mm_cpumask() for nefarious purposes and this patch changes the way that mm_cpumask() works. This should be okay, since Xen *also* iterates all online CPUs to find all the CPUs it needs to twiddle. The UV tlbflush code is rather dated and should be changed. Here are some benchmark results, done on a Skylake laptop at 2.3 GHz (turbo off, intel_pstate requesting max performance) under KVM with the guest using idle=poll (to avoid artifacts when bouncing between CPUs). I haven't done any real statistics here -- I just ran them in a loop and picked the fastest results that didn't look like outliers. Unpatched means commit a4eb8b993554, so all the bookkeeping overhead is gone. MADV_DONTNEED; touch the page; switch CPUs using sched_setaffinity. In an unpatched kernel, MADV_DONTNEED will send an IPI to the previous CPU. This is intended to be a nearly worst-case test. patched: 13.4µs unpatched: 21.6µs Vitaly's pthread_mmap microbenchmark with 8 threads (on four cores), nrounds = 100, 256M data patched: 1.1 seconds or so unpatched: 1.9 seconds or so The sleepup on Vitaly's test appearss to be because it spends a lot of time blocked on mmap_sem, and this patch avoids sending IPIs to blocked CPUs. Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andrew Banman <abanman@sgi.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Dimitri Sivanich <sivanich@sgi.com> Cc: Juergen Gross <jgross@suse.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Travis <travis@sgi.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/ddf2c92962339f4ba39d8fc41b853936ec0b44f1.1498751203.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-06-29 23:53:17 +08:00
/*
* Start remote flushes and then read tlb_gen.
*/
cpumask_set_cpu(cpu, mm_cpumask(next));
next_tlb_gen = atomic64_read(&next->context.tlb_gen);
x86/mm: Rework lazy TLB to track the actual loaded mm Lazy TLB state is currently managed in a rather baroque manner. AFAICT, there are three possible states: - Non-lazy. This means that we're running a user thread or a kernel thread that has called use_mm(). current->mm == current->active_mm == cpu_tlbstate.active_mm and cpu_tlbstate.state == TLBSTATE_OK. - Lazy with user mm. We're running a kernel thread without an mm and we're borrowing an mm_struct. We have current->mm == NULL, current->active_mm == cpu_tlbstate.active_mm, cpu_tlbstate.state != TLBSTATE_OK (i.e. TLBSTATE_LAZY or 0). The current cpu is set in mm_cpumask(current->active_mm). CR3 points to current->active_mm->pgd. The TLB is up to date. - Lazy with init_mm. This happens when we call leave_mm(). We have current->mm == NULL, current->active_mm == cpu_tlbstate.active_mm, but that mm is only relelvant insofar as the scheduler is tracking it for refcounting. cpu_tlbstate.state != TLBSTATE_OK. The current cpu is clear in mm_cpumask(current->active_mm). CR3 points to swapper_pg_dir, i.e. init_mm->pgd. This patch simplifies the situation. Other than perf, x86 stops caring about current->active_mm at all. We have cpu_tlbstate.loaded_mm pointing to the mm that CR3 references. The TLB is always up to date for that mm. leave_mm() just switches us to init_mm. There are no longer any special cases for mm_cpumask, and switch_mm() switches mms without worrying about laziness. After this patch, cpu_tlbstate.state serves only to tell the TLB flush code whether it may switch to init_mm instead of doing a normal flush. This makes fairly extensive changes to xen_exit_mmap(), which used to look a bit like black magic. Perf is unchanged. With or without this change, perf may behave a bit erratically if it tries to read user memory in kernel thread context. We should build on this patch to teach perf to never look at user memory when cpu_tlbstate.loaded_mm != current->mm. Signed-off-by: Andy Lutomirski <luto@kernel.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bpetkov@suse.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Michal Hocko <mhocko@suse.com> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Nadav Amit <namit@vmware.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mm@kvack.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-05-29 01:00:15 +08:00
x86/mm: Implement PCID based optimization: try to preserve old TLB entries using PCID PCID is a "process context ID" -- it's what other architectures call an address space ID. Every non-global TLB entry is tagged with a PCID, only TLB entries that match the currently selected PCID are used, and we can switch PGDs without flushing the TLB. x86's PCID is 12 bits. This is an unorthodox approach to using PCID. x86's PCID is far too short to uniquely identify a process, and we can't even really uniquely identify a running process because there are monster systems with over 4096 CPUs. To make matters worse, past attempts to use all 12 PCID bits have resulted in slowdowns instead of speedups. This patch uses PCID differently. We use a PCID to identify a recently-used mm on a per-cpu basis. An mm has no fixed PCID binding at all; instead, we give it a fresh PCID each time it's loaded except in cases where we want to preserve the TLB, in which case we reuse a recent value. Here are some benchmark results, done on a Skylake laptop at 2.3 GHz (turbo off, intel_pstate requesting max performance) under KVM with the guest using idle=poll (to avoid artifacts when bouncing between CPUs). I haven't done any real statistics here -- I just ran them in a loop and picked the fastest results that didn't look like outliers. Unpatched means commit a4eb8b993554, so all the bookkeeping overhead is gone. ping-pong between two mms on the same CPU using eventfd: patched: 1.22µs patched, nopcid: 1.33µs unpatched: 1.34µs Same ping-pong, but now touch 512 pages (all zero-page to minimize cache misses) each iteration. dTLB misses are measured by dtlb_load_misses.miss_causes_a_walk: patched: 1.8µs 11M dTLB misses patched, nopcid: 6.2µs, 207M dTLB misses unpatched: 6.1µs, 190M dTLB misses Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/9ee75f17a81770feed616358e6860d98a2a5b1e7.1500957502.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-07-25 12:41:38 +08:00
choose_new_asid(next, next_tlb_gen, &new_asid, &need_flush);
x86/mm: Rework lazy TLB to track the actual loaded mm Lazy TLB state is currently managed in a rather baroque manner. AFAICT, there are three possible states: - Non-lazy. This means that we're running a user thread or a kernel thread that has called use_mm(). current->mm == current->active_mm == cpu_tlbstate.active_mm and cpu_tlbstate.state == TLBSTATE_OK. - Lazy with user mm. We're running a kernel thread without an mm and we're borrowing an mm_struct. We have current->mm == NULL, current->active_mm == cpu_tlbstate.active_mm, cpu_tlbstate.state != TLBSTATE_OK (i.e. TLBSTATE_LAZY or 0). The current cpu is set in mm_cpumask(current->active_mm). CR3 points to current->active_mm->pgd. The TLB is up to date. - Lazy with init_mm. This happens when we call leave_mm(). We have current->mm == NULL, current->active_mm == cpu_tlbstate.active_mm, but that mm is only relelvant insofar as the scheduler is tracking it for refcounting. cpu_tlbstate.state != TLBSTATE_OK. The current cpu is clear in mm_cpumask(current->active_mm). CR3 points to swapper_pg_dir, i.e. init_mm->pgd. This patch simplifies the situation. Other than perf, x86 stops caring about current->active_mm at all. We have cpu_tlbstate.loaded_mm pointing to the mm that CR3 references. The TLB is always up to date for that mm. leave_mm() just switches us to init_mm. There are no longer any special cases for mm_cpumask, and switch_mm() switches mms without worrying about laziness. After this patch, cpu_tlbstate.state serves only to tell the TLB flush code whether it may switch to init_mm instead of doing a normal flush. This makes fairly extensive changes to xen_exit_mmap(), which used to look a bit like black magic. Perf is unchanged. With or without this change, perf may behave a bit erratically if it tries to read user memory in kernel thread context. We should build on this patch to teach perf to never look at user memory when cpu_tlbstate.loaded_mm != current->mm. Signed-off-by: Andy Lutomirski <luto@kernel.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bpetkov@suse.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Michal Hocko <mhocko@suse.com> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Nadav Amit <namit@vmware.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mm@kvack.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-05-29 01:00:15 +08:00
x86/mm: Implement PCID based optimization: try to preserve old TLB entries using PCID PCID is a "process context ID" -- it's what other architectures call an address space ID. Every non-global TLB entry is tagged with a PCID, only TLB entries that match the currently selected PCID are used, and we can switch PGDs without flushing the TLB. x86's PCID is 12 bits. This is an unorthodox approach to using PCID. x86's PCID is far too short to uniquely identify a process, and we can't even really uniquely identify a running process because there are monster systems with over 4096 CPUs. To make matters worse, past attempts to use all 12 PCID bits have resulted in slowdowns instead of speedups. This patch uses PCID differently. We use a PCID to identify a recently-used mm on a per-cpu basis. An mm has no fixed PCID binding at all; instead, we give it a fresh PCID each time it's loaded except in cases where we want to preserve the TLB, in which case we reuse a recent value. Here are some benchmark results, done on a Skylake laptop at 2.3 GHz (turbo off, intel_pstate requesting max performance) under KVM with the guest using idle=poll (to avoid artifacts when bouncing between CPUs). I haven't done any real statistics here -- I just ran them in a loop and picked the fastest results that didn't look like outliers. Unpatched means commit a4eb8b993554, so all the bookkeeping overhead is gone. ping-pong between two mms on the same CPU using eventfd: patched: 1.22µs patched, nopcid: 1.33µs unpatched: 1.34µs Same ping-pong, but now touch 512 pages (all zero-page to minimize cache misses) each iteration. dTLB misses are measured by dtlb_load_misses.miss_causes_a_walk: patched: 1.8µs 11M dTLB misses patched, nopcid: 6.2µs, 207M dTLB misses unpatched: 6.1µs, 190M dTLB misses Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/9ee75f17a81770feed616358e6860d98a2a5b1e7.1500957502.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-07-25 12:41:38 +08:00
if (need_flush) {
this_cpu_write(cpu_tlbstate.ctxs[new_asid].ctx_id, next->context.ctx_id);
this_cpu_write(cpu_tlbstate.ctxs[new_asid].tlb_gen, next_tlb_gen);
load_new_mm_cr3(next->pgd, new_asid, true);
Revert "x86/mm: Stop calling leave_mm() in idle code" This reverts commit 43858b4f25cf0adc5c2ca9cf5ce5fdf2532941e5. The reason I removed the leave_mm() calls in question is because the heuristic wasn't needed after that patch. With the original version of my PCID series, we never flushed a "lazy cpu" (i.e. a CPU running kernel thread) due a flush on the loaded mm. Unfortunately, that caused architectural issues, so now I've reinstated these flushes on non-PCID systems in: commit b956575bed91 ("x86/mm: Flush more aggressively in lazy TLB mode"). That, in turn, gives us a power management and occasionally performance regression as compared to old kernels: a process that goes into a deep idle state on a given CPU and gets its mm flushed due to activity on a different CPU will wake the idle CPU. Reinstate the old ugly heuristic: if a CPU goes into ACPI C3 or an intel_idle state that is likely to cause a TLB flush gets its mm switched to init_mm before going idle. FWIW, this heuristic is lousy. Whether we should change CR3 before idle isn't a good hint except insofar as the performance hit is a bit lower if the TLB is getting flushed by the idle code anyway. What we really want to know is whether we anticipate being idle long enough that the mm is likely to be flushed before we wake up. This is more a matter of the expected latency than the idle state that gets chosen. This heuristic also completely fails on systems that don't know whether the TLB will be flushed (e.g. AMD systems?). OTOH it may be a bit obsolete anyway -- PCID systems don't presently benefit from this heuristic at all. We also shouldn't do this callback from innermost bit of the idle code due to the RCU nastiness it causes. All the information need is available before rcu_idle_enter() needs to happen. Signed-off-by: Andy Lutomirski <luto@kernel.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Borislav Petkov <bpetkov@suse.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: Denys Vlasenko <dvlasenk@redhat.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Josh Poimboeuf <jpoimboe@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Fixes: 43858b4f25cf "x86/mm: Stop calling leave_mm() in idle code" Link: http://lkml.kernel.org/r/c513bbd4e653747213e05bc7062de000bf0202a5.1509793738.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-11-04 19:16:12 +08:00
/*
* NB: This gets called via leave_mm() in the idle path
* where RCU functions differently. Tracing normally
* uses RCU, so we need to use the _rcuidle variant.
*
* (There is no good reason for this. The idle code should
* be rearranged to call this before rcu_idle_enter().)
*/
trace_tlb_flush_rcuidle(TLB_FLUSH_ON_TASK_SWITCH, TLB_FLUSH_ALL);
x86/mm: Implement PCID based optimization: try to preserve old TLB entries using PCID PCID is a "process context ID" -- it's what other architectures call an address space ID. Every non-global TLB entry is tagged with a PCID, only TLB entries that match the currently selected PCID are used, and we can switch PGDs without flushing the TLB. x86's PCID is 12 bits. This is an unorthodox approach to using PCID. x86's PCID is far too short to uniquely identify a process, and we can't even really uniquely identify a running process because there are monster systems with over 4096 CPUs. To make matters worse, past attempts to use all 12 PCID bits have resulted in slowdowns instead of speedups. This patch uses PCID differently. We use a PCID to identify a recently-used mm on a per-cpu basis. An mm has no fixed PCID binding at all; instead, we give it a fresh PCID each time it's loaded except in cases where we want to preserve the TLB, in which case we reuse a recent value. Here are some benchmark results, done on a Skylake laptop at 2.3 GHz (turbo off, intel_pstate requesting max performance) under KVM with the guest using idle=poll (to avoid artifacts when bouncing between CPUs). I haven't done any real statistics here -- I just ran them in a loop and picked the fastest results that didn't look like outliers. Unpatched means commit a4eb8b993554, so all the bookkeeping overhead is gone. ping-pong between two mms on the same CPU using eventfd: patched: 1.22µs patched, nopcid: 1.33µs unpatched: 1.34µs Same ping-pong, but now touch 512 pages (all zero-page to minimize cache misses) each iteration. dTLB misses are measured by dtlb_load_misses.miss_causes_a_walk: patched: 1.8µs 11M dTLB misses patched, nopcid: 6.2µs, 207M dTLB misses unpatched: 6.1µs, 190M dTLB misses Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/9ee75f17a81770feed616358e6860d98a2a5b1e7.1500957502.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-07-25 12:41:38 +08:00
} else {
/* The new ASID is already up to date. */
load_new_mm_cr3(next->pgd, new_asid, false);
Revert "x86/mm: Stop calling leave_mm() in idle code" This reverts commit 43858b4f25cf0adc5c2ca9cf5ce5fdf2532941e5. The reason I removed the leave_mm() calls in question is because the heuristic wasn't needed after that patch. With the original version of my PCID series, we never flushed a "lazy cpu" (i.e. a CPU running kernel thread) due a flush on the loaded mm. Unfortunately, that caused architectural issues, so now I've reinstated these flushes on non-PCID systems in: commit b956575bed91 ("x86/mm: Flush more aggressively in lazy TLB mode"). That, in turn, gives us a power management and occasionally performance regression as compared to old kernels: a process that goes into a deep idle state on a given CPU and gets its mm flushed due to activity on a different CPU will wake the idle CPU. Reinstate the old ugly heuristic: if a CPU goes into ACPI C3 or an intel_idle state that is likely to cause a TLB flush gets its mm switched to init_mm before going idle. FWIW, this heuristic is lousy. Whether we should change CR3 before idle isn't a good hint except insofar as the performance hit is a bit lower if the TLB is getting flushed by the idle code anyway. What we really want to know is whether we anticipate being idle long enough that the mm is likely to be flushed before we wake up. This is more a matter of the expected latency than the idle state that gets chosen. This heuristic also completely fails on systems that don't know whether the TLB will be flushed (e.g. AMD systems?). OTOH it may be a bit obsolete anyway -- PCID systems don't presently benefit from this heuristic at all. We also shouldn't do this callback from innermost bit of the idle code due to the RCU nastiness it causes. All the information need is available before rcu_idle_enter() needs to happen. Signed-off-by: Andy Lutomirski <luto@kernel.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Borislav Petkov <bpetkov@suse.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: Denys Vlasenko <dvlasenk@redhat.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Josh Poimboeuf <jpoimboe@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Fixes: 43858b4f25cf "x86/mm: Stop calling leave_mm() in idle code" Link: http://lkml.kernel.org/r/c513bbd4e653747213e05bc7062de000bf0202a5.1509793738.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-11-04 19:16:12 +08:00
/* See above wrt _rcuidle. */
trace_tlb_flush_rcuidle(TLB_FLUSH_ON_TASK_SWITCH, 0);
x86/mm: Implement PCID based optimization: try to preserve old TLB entries using PCID PCID is a "process context ID" -- it's what other architectures call an address space ID. Every non-global TLB entry is tagged with a PCID, only TLB entries that match the currently selected PCID are used, and we can switch PGDs without flushing the TLB. x86's PCID is 12 bits. This is an unorthodox approach to using PCID. x86's PCID is far too short to uniquely identify a process, and we can't even really uniquely identify a running process because there are monster systems with over 4096 CPUs. To make matters worse, past attempts to use all 12 PCID bits have resulted in slowdowns instead of speedups. This patch uses PCID differently. We use a PCID to identify a recently-used mm on a per-cpu basis. An mm has no fixed PCID binding at all; instead, we give it a fresh PCID each time it's loaded except in cases where we want to preserve the TLB, in which case we reuse a recent value. Here are some benchmark results, done on a Skylake laptop at 2.3 GHz (turbo off, intel_pstate requesting max performance) under KVM with the guest using idle=poll (to avoid artifacts when bouncing between CPUs). I haven't done any real statistics here -- I just ran them in a loop and picked the fastest results that didn't look like outliers. Unpatched means commit a4eb8b993554, so all the bookkeeping overhead is gone. ping-pong between two mms on the same CPU using eventfd: patched: 1.22µs patched, nopcid: 1.33µs unpatched: 1.34µs Same ping-pong, but now touch 512 pages (all zero-page to minimize cache misses) each iteration. dTLB misses are measured by dtlb_load_misses.miss_causes_a_walk: patched: 1.8µs 11M dTLB misses patched, nopcid: 6.2µs, 207M dTLB misses unpatched: 6.1µs, 190M dTLB misses Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/9ee75f17a81770feed616358e6860d98a2a5b1e7.1500957502.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-07-25 12:41:38 +08:00
}
this_cpu_write(cpu_tlbstate.loaded_mm, next);
this_cpu_write(cpu_tlbstate.loaded_mm_asid, new_asid);
x86/mm: Rework lazy TLB mode and TLB freshness tracking x86's lazy TLB mode used to be fairly weak -- it would switch to init_mm the first time it tried to flush a lazy TLB. This meant an unnecessary CR3 write and, if the flush was remote, an unnecessary IPI. Rewrite it entirely. When we enter lazy mode, we simply remove the CPU from mm_cpumask. This means that we need a way to figure out whether we've missed a flush when we switch back out of lazy mode. I use the tlb_gen machinery to track whether a context is up to date. Note to reviewers: this patch, my itself, looks a bit odd. I'm using an array of length 1 containing (ctx_id, tlb_gen) rather than just storing tlb_gen, and making it at array isn't necessary yet. I'm doing this because the next few patches add PCID support, and, with PCID, we need ctx_id, and the array will end up with a length greater than 1. Making it an array now means that there will be less churn and therefore less stress on your eyeballs. NB: This is dubious but, AFAICT, still correct on Xen and UV. xen_exit_mmap() uses mm_cpumask() for nefarious purposes and this patch changes the way that mm_cpumask() works. This should be okay, since Xen *also* iterates all online CPUs to find all the CPUs it needs to twiddle. The UV tlbflush code is rather dated and should be changed. Here are some benchmark results, done on a Skylake laptop at 2.3 GHz (turbo off, intel_pstate requesting max performance) under KVM with the guest using idle=poll (to avoid artifacts when bouncing between CPUs). I haven't done any real statistics here -- I just ran them in a loop and picked the fastest results that didn't look like outliers. Unpatched means commit a4eb8b993554, so all the bookkeeping overhead is gone. MADV_DONTNEED; touch the page; switch CPUs using sched_setaffinity. In an unpatched kernel, MADV_DONTNEED will send an IPI to the previous CPU. This is intended to be a nearly worst-case test. patched: 13.4µs unpatched: 21.6µs Vitaly's pthread_mmap microbenchmark with 8 threads (on four cores), nrounds = 100, 256M data patched: 1.1 seconds or so unpatched: 1.9 seconds or so The sleepup on Vitaly's test appearss to be because it spends a lot of time blocked on mmap_sem, and this patch avoids sending IPIs to blocked CPUs. Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andrew Banman <abanman@sgi.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Dimitri Sivanich <sivanich@sgi.com> Cc: Juergen Gross <jgross@suse.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Travis <travis@sgi.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/ddf2c92962339f4ba39d8fc41b853936ec0b44f1.1498751203.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-06-29 23:53:17 +08:00
}
x86/mm: Rework lazy TLB to track the actual loaded mm Lazy TLB state is currently managed in a rather baroque manner. AFAICT, there are three possible states: - Non-lazy. This means that we're running a user thread or a kernel thread that has called use_mm(). current->mm == current->active_mm == cpu_tlbstate.active_mm and cpu_tlbstate.state == TLBSTATE_OK. - Lazy with user mm. We're running a kernel thread without an mm and we're borrowing an mm_struct. We have current->mm == NULL, current->active_mm == cpu_tlbstate.active_mm, cpu_tlbstate.state != TLBSTATE_OK (i.e. TLBSTATE_LAZY or 0). The current cpu is set in mm_cpumask(current->active_mm). CR3 points to current->active_mm->pgd. The TLB is up to date. - Lazy with init_mm. This happens when we call leave_mm(). We have current->mm == NULL, current->active_mm == cpu_tlbstate.active_mm, but that mm is only relelvant insofar as the scheduler is tracking it for refcounting. cpu_tlbstate.state != TLBSTATE_OK. The current cpu is clear in mm_cpumask(current->active_mm). CR3 points to swapper_pg_dir, i.e. init_mm->pgd. This patch simplifies the situation. Other than perf, x86 stops caring about current->active_mm at all. We have cpu_tlbstate.loaded_mm pointing to the mm that CR3 references. The TLB is always up to date for that mm. leave_mm() just switches us to init_mm. There are no longer any special cases for mm_cpumask, and switch_mm() switches mms without worrying about laziness. After this patch, cpu_tlbstate.state serves only to tell the TLB flush code whether it may switch to init_mm instead of doing a normal flush. This makes fairly extensive changes to xen_exit_mmap(), which used to look a bit like black magic. Perf is unchanged. With or without this change, perf may behave a bit erratically if it tries to read user memory in kernel thread context. We should build on this patch to teach perf to never look at user memory when cpu_tlbstate.loaded_mm != current->mm. Signed-off-by: Andy Lutomirski <luto@kernel.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bpetkov@suse.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Michal Hocko <mhocko@suse.com> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Nadav Amit <namit@vmware.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mm@kvack.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-05-29 01:00:15 +08:00
load_mm_cr4(next);
x86/ldt: Simplify the LDT switching logic Originally, Linux reloaded the LDT whenever the prev mm or the next mm had an LDT. It was changed in 2002 in: 0bbed3beb4f2 ("[PATCH] Thread-Local Storage (TLS) support") (commit from the historical tree), like this: - /* load_LDT, if either the previous or next thread - * has a non-default LDT. + /* + * load the LDT, if the LDT is different: */ - if (next->context.size+prev->context.size) + if (unlikely(prev->context.ldt != next->context.ldt)) load_LDT(&next->context); The current code is unlikely to avoid any LDT reloads, since different mms won't share an LDT. When we redo lazy mode to stop flush IPIs without switching to init_mm, though, the current logic would become incorrect: it will be possible to have real_prev == next but nonetheless have a stale LDT descriptor. Simplify the code to update LDTR if either the previous or the next mm has an LDT, i.e. effectively restore the historical logic.. While we're at it, clean up the code by moving all the ifdeffery to a header where it belongs. Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Reviewed-by: Borislav Petkov <bp@suse.de> Acked-by: Rik van Riel <riel@redhat.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Denys Vlasenko <dvlasenk@redhat.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Josh Poimboeuf <jpoimboe@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/2a859ac01245f9594c58f9d0a8b2ed8a7cd2507e.1498022414.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-06-21 13:22:08 +08:00
switch_ldt(real_prev, next);
}
x86/mm: Flush more aggressively in lazy TLB mode Since commit: 94b1b03b519b ("x86/mm: Rework lazy TLB mode and TLB freshness tracking") x86's lazy TLB mode has been all the way lazy: when running a kernel thread (including the idle thread), the kernel keeps using the last user mm's page tables without attempting to maintain user TLB coherence at all. From a pure semantic perspective, this is fine -- kernel threads won't attempt to access user pages, so having stale TLB entries doesn't matter. Unfortunately, I forgot about a subtlety. By skipping TLB flushes, we also allow any paging-structure caches that may exist on the CPU to become incoherent. This means that we can have a paging-structure cache entry that references a freed page table, and the CPU is within its rights to do a speculative page walk starting at the freed page table. I can imagine this causing two different problems: - A speculative page walk starting from a bogus page table could read IO addresses. I haven't seen any reports of this causing problems. - A speculative page walk that involves a bogus page table can install garbage in the TLB. Such garbage would always be at a user VA, but some AMD CPUs have logic that triggers a machine check when it notices these bogus entries. I've seen a couple reports of this. Boris further explains the failure mode: > It is actually more of an optimization which assumes that paging-structure > entries are in WB DRAM: > > "TlbCacheDis: cacheable memory disable. Read-write. 0=Enables > performance optimization that assumes PML4, PDP, PDE, and PTE entries > are in cacheable WB-DRAM; memory type checks may be bypassed, and > addresses outside of WB-DRAM may result in undefined behavior or NB > protocol errors. 1=Disables performance optimization and allows PML4, > PDP, PDE and PTE entries to be in any memory type. Operating systems > that maintain page tables in memory types other than WB- DRAM must set > TlbCacheDis to insure proper operation." > > The MCE generated is an NB protocol error to signal that > > "Link: A specific coherent-only packet from a CPU was issued to an > IO link. This may be caused by software which addresses page table > structures in a memory type other than cacheable WB-DRAM without > properly configuring MSRC001_0015[TlbCacheDis]. This may occur, for > example, when page table structure addresses are above top of memory. In > such cases, the NB will generate an MCE if it sees a mismatch between > the memory operation generated by the core and the link type." > > I'm assuming coherent-only packets don't go out on IO links, thus the > error. To fix this, reinstate TLB coherence in lazy mode. With this patch applied, we do it in one of two ways: - If we have PCID, we simply switch back to init_mm's page tables when we enter a kernel thread -- this seems to be quite cheap except for the cost of serializing the CPU. - If we don't have PCID, then we set a flag and switch to init_mm the first time we would otherwise need to flush the TLB. The /sys/kernel/debug/x86/tlb_use_lazy_mode debug switch can be changed to override the default mode for benchmarking. In theory, we could optimize this better by only flushing the TLB in lazy CPUs when a page table is freed. Doing that would require auditing the mm code to make sure that all page table freeing goes through tlb_remove_page() as well as reworking some data structures to implement the improved flush logic. Reported-by: Markus Trippelsdorf <markus@trippelsdorf.de> Reported-by: Adam Borowski <kilobyte@angband.pl> Signed-off-by: Andy Lutomirski <luto@kernel.org> Signed-off-by: Borislav Petkov <bp@suse.de> Cc: Borislav Petkov <bp@alien8.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: Daniel Borkmann <daniel@iogearbox.net> Cc: Eric Biggers <ebiggers@google.com> Cc: Johannes Hirte <johannes.hirte@datenkhaos.de> Cc: Kees Cook <keescook@chromium.org> Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Roman Kagan <rkagan@virtuozzo.com> Cc: Thomas Gleixner <tglx@linutronix.de> Fixes: 94b1b03b519b ("x86/mm: Rework lazy TLB mode and TLB freshness tracking") Link: http://lkml.kernel.org/r/20171009170231.fkpraqokz6e4zeco@pd.tnic Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-10-10 00:50:49 +08:00
/*
* Please ignore the name of this function. It should be called
* switch_to_kernel_thread().
*
x86/mm: Flush more aggressively in lazy TLB mode Since commit: 94b1b03b519b ("x86/mm: Rework lazy TLB mode and TLB freshness tracking") x86's lazy TLB mode has been all the way lazy: when running a kernel thread (including the idle thread), the kernel keeps using the last user mm's page tables without attempting to maintain user TLB coherence at all. From a pure semantic perspective, this is fine -- kernel threads won't attempt to access user pages, so having stale TLB entries doesn't matter. Unfortunately, I forgot about a subtlety. By skipping TLB flushes, we also allow any paging-structure caches that may exist on the CPU to become incoherent. This means that we can have a paging-structure cache entry that references a freed page table, and the CPU is within its rights to do a speculative page walk starting at the freed page table. I can imagine this causing two different problems: - A speculative page walk starting from a bogus page table could read IO addresses. I haven't seen any reports of this causing problems. - A speculative page walk that involves a bogus page table can install garbage in the TLB. Such garbage would always be at a user VA, but some AMD CPUs have logic that triggers a machine check when it notices these bogus entries. I've seen a couple reports of this. Boris further explains the failure mode: > It is actually more of an optimization which assumes that paging-structure > entries are in WB DRAM: > > "TlbCacheDis: cacheable memory disable. Read-write. 0=Enables > performance optimization that assumes PML4, PDP, PDE, and PTE entries > are in cacheable WB-DRAM; memory type checks may be bypassed, and > addresses outside of WB-DRAM may result in undefined behavior or NB > protocol errors. 1=Disables performance optimization and allows PML4, > PDP, PDE and PTE entries to be in any memory type. Operating systems > that maintain page tables in memory types other than WB- DRAM must set > TlbCacheDis to insure proper operation." > > The MCE generated is an NB protocol error to signal that > > "Link: A specific coherent-only packet from a CPU was issued to an > IO link. This may be caused by software which addresses page table > structures in a memory type other than cacheable WB-DRAM without > properly configuring MSRC001_0015[TlbCacheDis]. This may occur, for > example, when page table structure addresses are above top of memory. In > such cases, the NB will generate an MCE if it sees a mismatch between > the memory operation generated by the core and the link type." > > I'm assuming coherent-only packets don't go out on IO links, thus the > error. To fix this, reinstate TLB coherence in lazy mode. With this patch applied, we do it in one of two ways: - If we have PCID, we simply switch back to init_mm's page tables when we enter a kernel thread -- this seems to be quite cheap except for the cost of serializing the CPU. - If we don't have PCID, then we set a flag and switch to init_mm the first time we would otherwise need to flush the TLB. The /sys/kernel/debug/x86/tlb_use_lazy_mode debug switch can be changed to override the default mode for benchmarking. In theory, we could optimize this better by only flushing the TLB in lazy CPUs when a page table is freed. Doing that would require auditing the mm code to make sure that all page table freeing goes through tlb_remove_page() as well as reworking some data structures to implement the improved flush logic. Reported-by: Markus Trippelsdorf <markus@trippelsdorf.de> Reported-by: Adam Borowski <kilobyte@angband.pl> Signed-off-by: Andy Lutomirski <luto@kernel.org> Signed-off-by: Borislav Petkov <bp@suse.de> Cc: Borislav Petkov <bp@alien8.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: Daniel Borkmann <daniel@iogearbox.net> Cc: Eric Biggers <ebiggers@google.com> Cc: Johannes Hirte <johannes.hirte@datenkhaos.de> Cc: Kees Cook <keescook@chromium.org> Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Roman Kagan <rkagan@virtuozzo.com> Cc: Thomas Gleixner <tglx@linutronix.de> Fixes: 94b1b03b519b ("x86/mm: Rework lazy TLB mode and TLB freshness tracking") Link: http://lkml.kernel.org/r/20171009170231.fkpraqokz6e4zeco@pd.tnic Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-10-10 00:50:49 +08:00
* enter_lazy_tlb() is a hint from the scheduler that we are entering a
* kernel thread or other context without an mm. Acceptable implementations
* include doing nothing whatsoever, switching to init_mm, or various clever
* lazy tricks to try to minimize TLB flushes.
*
* The scheduler reserves the right to call enter_lazy_tlb() several times
* in a row. It will notify us that we're going back to a real mm by
* calling switch_mm_irqs_off().
*/
void enter_lazy_tlb(struct mm_struct *mm, struct task_struct *tsk)
{
if (this_cpu_read(cpu_tlbstate.loaded_mm) == &init_mm)
return;
if (tlb_defer_switch_to_init_mm()) {
x86/mm: Flush more aggressively in lazy TLB mode Since commit: 94b1b03b519b ("x86/mm: Rework lazy TLB mode and TLB freshness tracking") x86's lazy TLB mode has been all the way lazy: when running a kernel thread (including the idle thread), the kernel keeps using the last user mm's page tables without attempting to maintain user TLB coherence at all. From a pure semantic perspective, this is fine -- kernel threads won't attempt to access user pages, so having stale TLB entries doesn't matter. Unfortunately, I forgot about a subtlety. By skipping TLB flushes, we also allow any paging-structure caches that may exist on the CPU to become incoherent. This means that we can have a paging-structure cache entry that references a freed page table, and the CPU is within its rights to do a speculative page walk starting at the freed page table. I can imagine this causing two different problems: - A speculative page walk starting from a bogus page table could read IO addresses. I haven't seen any reports of this causing problems. - A speculative page walk that involves a bogus page table can install garbage in the TLB. Such garbage would always be at a user VA, but some AMD CPUs have logic that triggers a machine check when it notices these bogus entries. I've seen a couple reports of this. Boris further explains the failure mode: > It is actually more of an optimization which assumes that paging-structure > entries are in WB DRAM: > > "TlbCacheDis: cacheable memory disable. Read-write. 0=Enables > performance optimization that assumes PML4, PDP, PDE, and PTE entries > are in cacheable WB-DRAM; memory type checks may be bypassed, and > addresses outside of WB-DRAM may result in undefined behavior or NB > protocol errors. 1=Disables performance optimization and allows PML4, > PDP, PDE and PTE entries to be in any memory type. Operating systems > that maintain page tables in memory types other than WB- DRAM must set > TlbCacheDis to insure proper operation." > > The MCE generated is an NB protocol error to signal that > > "Link: A specific coherent-only packet from a CPU was issued to an > IO link. This may be caused by software which addresses page table > structures in a memory type other than cacheable WB-DRAM without > properly configuring MSRC001_0015[TlbCacheDis]. This may occur, for > example, when page table structure addresses are above top of memory. In > such cases, the NB will generate an MCE if it sees a mismatch between > the memory operation generated by the core and the link type." > > I'm assuming coherent-only packets don't go out on IO links, thus the > error. To fix this, reinstate TLB coherence in lazy mode. With this patch applied, we do it in one of two ways: - If we have PCID, we simply switch back to init_mm's page tables when we enter a kernel thread -- this seems to be quite cheap except for the cost of serializing the CPU. - If we don't have PCID, then we set a flag and switch to init_mm the first time we would otherwise need to flush the TLB. The /sys/kernel/debug/x86/tlb_use_lazy_mode debug switch can be changed to override the default mode for benchmarking. In theory, we could optimize this better by only flushing the TLB in lazy CPUs when a page table is freed. Doing that would require auditing the mm code to make sure that all page table freeing goes through tlb_remove_page() as well as reworking some data structures to implement the improved flush logic. Reported-by: Markus Trippelsdorf <markus@trippelsdorf.de> Reported-by: Adam Borowski <kilobyte@angband.pl> Signed-off-by: Andy Lutomirski <luto@kernel.org> Signed-off-by: Borislav Petkov <bp@suse.de> Cc: Borislav Petkov <bp@alien8.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: Daniel Borkmann <daniel@iogearbox.net> Cc: Eric Biggers <ebiggers@google.com> Cc: Johannes Hirte <johannes.hirte@datenkhaos.de> Cc: Kees Cook <keescook@chromium.org> Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Roman Kagan <rkagan@virtuozzo.com> Cc: Thomas Gleixner <tglx@linutronix.de> Fixes: 94b1b03b519b ("x86/mm: Rework lazy TLB mode and TLB freshness tracking") Link: http://lkml.kernel.org/r/20171009170231.fkpraqokz6e4zeco@pd.tnic Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-10-10 00:50:49 +08:00
/*
* There's a significant optimization that may be possible
* here. We have accurate enough TLB flush tracking that we
* don't need to maintain coherence of TLB per se when we're
* lazy. We do, however, need to maintain coherence of
* paging-structure caches. We could, in principle, leave our
* old mm loaded and only switch to init_mm when
* tlb_remove_page() happens.
*/
this_cpu_write(cpu_tlbstate.is_lazy, true);
} else {
switch_mm(NULL, &init_mm, NULL);
}
}
/*
* Call this when reinitializing a CPU. It fixes the following potential
* problems:
*
* - The ASID changed from what cpu_tlbstate thinks it is (most likely
* because the CPU was taken down and came back up with CR3's PCID
* bits clear. CPU hotplug can do this.
*
* - The TLB contains junk in slots corresponding to inactive ASIDs.
*
* - The CPU went so far out to lunch that it may have missed a TLB
* flush.
*/
void initialize_tlbstate_and_flush(void)
{
int i;
struct mm_struct *mm = this_cpu_read(cpu_tlbstate.loaded_mm);
u64 tlb_gen = atomic64_read(&init_mm.context.tlb_gen);
unsigned long cr3 = __read_cr3();
/* Assert that CR3 already references the right mm. */
WARN_ON((cr3 & CR3_ADDR_MASK) != __pa(mm->pgd));
/*
* Assert that CR4.PCIDE is set if needed. (CR4.PCIDE initialization
* doesn't work like other CR4 bits because it can only be set from
* long mode.)
*/
WARN_ON(boot_cpu_has(X86_FEATURE_PCID) &&
!(cr4_read_shadow() & X86_CR4_PCIDE));
/* Force ASID 0 and force a TLB flush. */
write_cr3(build_cr3(mm->pgd, 0));
/* Reinitialize tlbstate. */
this_cpu_write(cpu_tlbstate.loaded_mm_asid, 0);
this_cpu_write(cpu_tlbstate.next_asid, 1);
this_cpu_write(cpu_tlbstate.ctxs[0].ctx_id, mm->context.ctx_id);
this_cpu_write(cpu_tlbstate.ctxs[0].tlb_gen, tlb_gen);
for (i = 1; i < TLB_NR_DYN_ASIDS; i++)
this_cpu_write(cpu_tlbstate.ctxs[i].ctx_id, 0);
}
x86/mm: Track the TLB's tlb_gen and update the flushing algorithm There are two kernel features that would benefit from tracking how up-to-date each CPU's TLB is in the case where IPIs aren't keeping it up to date in real time: - Lazy mm switching currently works by switching to init_mm when it would otherwise flush. This is wasteful: there isn't fundamentally any need to update CR3 at all when going lazy or when returning from lazy mode, nor is there any need to receive flush IPIs at all. Instead, we should just stop trying to keep the TLB coherent when we go lazy and, when unlazying, check whether we missed any flushes. - PCID will let us keep recent user contexts alive in the TLB. If we start doing this, we need a way to decide whether those contexts are up to date. On some paravirt systems, remote TLBs can be flushed without IPIs. This won't update the target CPUs' tlb_gens, which may cause unnecessary local flushes later on. We can address this if it becomes a problem by carefully updating the target CPU's tlb_gen directly. By itself, this patch is a very minor optimization that avoids unnecessary flushes when multiple TLB flushes targetting the same CPU race. The complexity in this patch would not be worth it on its own, but it will enable improved lazy TLB tracking and PCID. Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/1210fb244bc9cbe7677f7f0b72db4d359675f24b.1498751203.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-06-29 23:53:16 +08:00
/*
* flush_tlb_func_common()'s memory ordering requirement is that any
* TLB fills that happen after we flush the TLB are ordered after we
* read active_mm's tlb_gen. We don't need any explicit barriers
* because all x86 flush operations are serializing and the
* atomic64_read operation won't be reordered by the compiler.
*/
x86/mm: Refactor flush_tlb_mm_range() to merge local and remote cases The local flush path is very similar to the remote flush path. Merge them. This is intended to make no difference to behavior whatsoever. It removes some code and will make future changes to the flushing mechanics simpler. This patch does remove one small optimization: flush_tlb_mm_range() now has an unconditional smp_mb() instead of using MOV to CR3 or INVLPG as a full barrier when applicable. I think this is okay for a few reasons. First, smp_mb() is quite cheap compared to the cost of a TLB flush. Second, this rearrangement makes a bigger optimization available: with some work on the SMP function call code, we could do the local and remote flushes in parallel. Third, I'm planning a rework of the TLB flush algorithm that will require an atomic operation at the beginning of each flush, and that operation will replace the smp_mb(). Signed-off-by: Andy Lutomirski <luto@kernel.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bpetkov@suse.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Michal Hocko <mhocko@suse.com> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Nadav Amit <namit@vmware.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mm@kvack.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-05-29 01:00:12 +08:00
static void flush_tlb_func_common(const struct flush_tlb_info *f,
bool local, enum tlb_flush_reason reason)
{
x86/mm: Track the TLB's tlb_gen and update the flushing algorithm There are two kernel features that would benefit from tracking how up-to-date each CPU's TLB is in the case where IPIs aren't keeping it up to date in real time: - Lazy mm switching currently works by switching to init_mm when it would otherwise flush. This is wasteful: there isn't fundamentally any need to update CR3 at all when going lazy or when returning from lazy mode, nor is there any need to receive flush IPIs at all. Instead, we should just stop trying to keep the TLB coherent when we go lazy and, when unlazying, check whether we missed any flushes. - PCID will let us keep recent user contexts alive in the TLB. If we start doing this, we need a way to decide whether those contexts are up to date. On some paravirt systems, remote TLBs can be flushed without IPIs. This won't update the target CPUs' tlb_gens, which may cause unnecessary local flushes later on. We can address this if it becomes a problem by carefully updating the target CPU's tlb_gen directly. By itself, this patch is a very minor optimization that avoids unnecessary flushes when multiple TLB flushes targetting the same CPU race. The complexity in this patch would not be worth it on its own, but it will enable improved lazy TLB tracking and PCID. Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/1210fb244bc9cbe7677f7f0b72db4d359675f24b.1498751203.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-06-29 23:53:16 +08:00
/*
* We have three different tlb_gen values in here. They are:
*
* - mm_tlb_gen: the latest generation.
* - local_tlb_gen: the generation that this CPU has already caught
* up to.
* - f->new_tlb_gen: the generation that the requester of the flush
* wants us to catch up to.
*/
struct mm_struct *loaded_mm = this_cpu_read(cpu_tlbstate.loaded_mm);
x86/mm: Implement PCID based optimization: try to preserve old TLB entries using PCID PCID is a "process context ID" -- it's what other architectures call an address space ID. Every non-global TLB entry is tagged with a PCID, only TLB entries that match the currently selected PCID are used, and we can switch PGDs without flushing the TLB. x86's PCID is 12 bits. This is an unorthodox approach to using PCID. x86's PCID is far too short to uniquely identify a process, and we can't even really uniquely identify a running process because there are monster systems with over 4096 CPUs. To make matters worse, past attempts to use all 12 PCID bits have resulted in slowdowns instead of speedups. This patch uses PCID differently. We use a PCID to identify a recently-used mm on a per-cpu basis. An mm has no fixed PCID binding at all; instead, we give it a fresh PCID each time it's loaded except in cases where we want to preserve the TLB, in which case we reuse a recent value. Here are some benchmark results, done on a Skylake laptop at 2.3 GHz (turbo off, intel_pstate requesting max performance) under KVM with the guest using idle=poll (to avoid artifacts when bouncing between CPUs). I haven't done any real statistics here -- I just ran them in a loop and picked the fastest results that didn't look like outliers. Unpatched means commit a4eb8b993554, so all the bookkeeping overhead is gone. ping-pong between two mms on the same CPU using eventfd: patched: 1.22µs patched, nopcid: 1.33µs unpatched: 1.34µs Same ping-pong, but now touch 512 pages (all zero-page to minimize cache misses) each iteration. dTLB misses are measured by dtlb_load_misses.miss_causes_a_walk: patched: 1.8µs 11M dTLB misses patched, nopcid: 6.2µs, 207M dTLB misses unpatched: 6.1µs, 190M dTLB misses Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/9ee75f17a81770feed616358e6860d98a2a5b1e7.1500957502.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-07-25 12:41:38 +08:00
u32 loaded_mm_asid = this_cpu_read(cpu_tlbstate.loaded_mm_asid);
x86/mm: Track the TLB's tlb_gen and update the flushing algorithm There are two kernel features that would benefit from tracking how up-to-date each CPU's TLB is in the case where IPIs aren't keeping it up to date in real time: - Lazy mm switching currently works by switching to init_mm when it would otherwise flush. This is wasteful: there isn't fundamentally any need to update CR3 at all when going lazy or when returning from lazy mode, nor is there any need to receive flush IPIs at all. Instead, we should just stop trying to keep the TLB coherent when we go lazy and, when unlazying, check whether we missed any flushes. - PCID will let us keep recent user contexts alive in the TLB. If we start doing this, we need a way to decide whether those contexts are up to date. On some paravirt systems, remote TLBs can be flushed without IPIs. This won't update the target CPUs' tlb_gens, which may cause unnecessary local flushes later on. We can address this if it becomes a problem by carefully updating the target CPU's tlb_gen directly. By itself, this patch is a very minor optimization that avoids unnecessary flushes when multiple TLB flushes targetting the same CPU race. The complexity in this patch would not be worth it on its own, but it will enable improved lazy TLB tracking and PCID. Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/1210fb244bc9cbe7677f7f0b72db4d359675f24b.1498751203.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-06-29 23:53:16 +08:00
u64 mm_tlb_gen = atomic64_read(&loaded_mm->context.tlb_gen);
x86/mm: Implement PCID based optimization: try to preserve old TLB entries using PCID PCID is a "process context ID" -- it's what other architectures call an address space ID. Every non-global TLB entry is tagged with a PCID, only TLB entries that match the currently selected PCID are used, and we can switch PGDs without flushing the TLB. x86's PCID is 12 bits. This is an unorthodox approach to using PCID. x86's PCID is far too short to uniquely identify a process, and we can't even really uniquely identify a running process because there are monster systems with over 4096 CPUs. To make matters worse, past attempts to use all 12 PCID bits have resulted in slowdowns instead of speedups. This patch uses PCID differently. We use a PCID to identify a recently-used mm on a per-cpu basis. An mm has no fixed PCID binding at all; instead, we give it a fresh PCID each time it's loaded except in cases where we want to preserve the TLB, in which case we reuse a recent value. Here are some benchmark results, done on a Skylake laptop at 2.3 GHz (turbo off, intel_pstate requesting max performance) under KVM with the guest using idle=poll (to avoid artifacts when bouncing between CPUs). I haven't done any real statistics here -- I just ran them in a loop and picked the fastest results that didn't look like outliers. Unpatched means commit a4eb8b993554, so all the bookkeeping overhead is gone. ping-pong between two mms on the same CPU using eventfd: patched: 1.22µs patched, nopcid: 1.33µs unpatched: 1.34µs Same ping-pong, but now touch 512 pages (all zero-page to minimize cache misses) each iteration. dTLB misses are measured by dtlb_load_misses.miss_causes_a_walk: patched: 1.8µs 11M dTLB misses patched, nopcid: 6.2µs, 207M dTLB misses unpatched: 6.1µs, 190M dTLB misses Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/9ee75f17a81770feed616358e6860d98a2a5b1e7.1500957502.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-07-25 12:41:38 +08:00
u64 local_tlb_gen = this_cpu_read(cpu_tlbstate.ctxs[loaded_mm_asid].tlb_gen);
x86/mm: Track the TLB's tlb_gen and update the flushing algorithm There are two kernel features that would benefit from tracking how up-to-date each CPU's TLB is in the case where IPIs aren't keeping it up to date in real time: - Lazy mm switching currently works by switching to init_mm when it would otherwise flush. This is wasteful: there isn't fundamentally any need to update CR3 at all when going lazy or when returning from lazy mode, nor is there any need to receive flush IPIs at all. Instead, we should just stop trying to keep the TLB coherent when we go lazy and, when unlazying, check whether we missed any flushes. - PCID will let us keep recent user contexts alive in the TLB. If we start doing this, we need a way to decide whether those contexts are up to date. On some paravirt systems, remote TLBs can be flushed without IPIs. This won't update the target CPUs' tlb_gens, which may cause unnecessary local flushes later on. We can address this if it becomes a problem by carefully updating the target CPU's tlb_gen directly. By itself, this patch is a very minor optimization that avoids unnecessary flushes when multiple TLB flushes targetting the same CPU race. The complexity in this patch would not be worth it on its own, but it will enable improved lazy TLB tracking and PCID. Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/1210fb244bc9cbe7677f7f0b72db4d359675f24b.1498751203.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-06-29 23:53:16 +08:00
x86/mm: Don't reenter flush_tlb_func_common() It was historically possible to have two concurrent TLB flushes targetting the same CPU: one initiated locally and one initiated remotely. This can now cause an OOPS in leave_mm() at arch/x86/mm/tlb.c:47: if (this_cpu_read(cpu_tlbstate.state) == TLBSTATE_OK) BUG(); with this call trace: flush_tlb_func_local arch/x86/mm/tlb.c:239 [inline] flush_tlb_mm_range+0x26d/0x370 arch/x86/mm/tlb.c:317 Without reentrancy, this OOPS is impossible: leave_mm() is only called if we're not in TLBSTATE_OK, but then we're unexpectedly in TLBSTATE_OK in leave_mm(). This can be caused by flush_tlb_func_remote() happening between the two checks and calling leave_mm(), resulting in two consecutive leave_mm() calls on the same CPU with no intervening switch_mm() calls. We never saw this OOPS before because the old leave_mm() implementation didn't put us back in TLBSTATE_OK, so the assertion didn't fire. Nadav noticed the reentrancy issue in a different context, but neither of us realized that it caused a problem yet. Reported-by: Levin, Alexander (Sasha Levin) <alexander.levin@verizon.com> Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: linux-mm@kvack.org Fixes: 3d28ebceaffa ("x86/mm: Rework lazy TLB to track the actual loaded mm") Link: http://lkml.kernel.org/r/855acf733268d521c9f2e191faee2dcc23a29729.1498751203.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-06-29 23:53:13 +08:00
/* This code cannot presently handle being reentered. */
VM_WARN_ON(!irqs_disabled());
x86/mm: Flush more aggressively in lazy TLB mode Since commit: 94b1b03b519b ("x86/mm: Rework lazy TLB mode and TLB freshness tracking") x86's lazy TLB mode has been all the way lazy: when running a kernel thread (including the idle thread), the kernel keeps using the last user mm's page tables without attempting to maintain user TLB coherence at all. From a pure semantic perspective, this is fine -- kernel threads won't attempt to access user pages, so having stale TLB entries doesn't matter. Unfortunately, I forgot about a subtlety. By skipping TLB flushes, we also allow any paging-structure caches that may exist on the CPU to become incoherent. This means that we can have a paging-structure cache entry that references a freed page table, and the CPU is within its rights to do a speculative page walk starting at the freed page table. I can imagine this causing two different problems: - A speculative page walk starting from a bogus page table could read IO addresses. I haven't seen any reports of this causing problems. - A speculative page walk that involves a bogus page table can install garbage in the TLB. Such garbage would always be at a user VA, but some AMD CPUs have logic that triggers a machine check when it notices these bogus entries. I've seen a couple reports of this. Boris further explains the failure mode: > It is actually more of an optimization which assumes that paging-structure > entries are in WB DRAM: > > "TlbCacheDis: cacheable memory disable. Read-write. 0=Enables > performance optimization that assumes PML4, PDP, PDE, and PTE entries > are in cacheable WB-DRAM; memory type checks may be bypassed, and > addresses outside of WB-DRAM may result in undefined behavior or NB > protocol errors. 1=Disables performance optimization and allows PML4, > PDP, PDE and PTE entries to be in any memory type. Operating systems > that maintain page tables in memory types other than WB- DRAM must set > TlbCacheDis to insure proper operation." > > The MCE generated is an NB protocol error to signal that > > "Link: A specific coherent-only packet from a CPU was issued to an > IO link. This may be caused by software which addresses page table > structures in a memory type other than cacheable WB-DRAM without > properly configuring MSRC001_0015[TlbCacheDis]. This may occur, for > example, when page table structure addresses are above top of memory. In > such cases, the NB will generate an MCE if it sees a mismatch between > the memory operation generated by the core and the link type." > > I'm assuming coherent-only packets don't go out on IO links, thus the > error. To fix this, reinstate TLB coherence in lazy mode. With this patch applied, we do it in one of two ways: - If we have PCID, we simply switch back to init_mm's page tables when we enter a kernel thread -- this seems to be quite cheap except for the cost of serializing the CPU. - If we don't have PCID, then we set a flag and switch to init_mm the first time we would otherwise need to flush the TLB. The /sys/kernel/debug/x86/tlb_use_lazy_mode debug switch can be changed to override the default mode for benchmarking. In theory, we could optimize this better by only flushing the TLB in lazy CPUs when a page table is freed. Doing that would require auditing the mm code to make sure that all page table freeing goes through tlb_remove_page() as well as reworking some data structures to implement the improved flush logic. Reported-by: Markus Trippelsdorf <markus@trippelsdorf.de> Reported-by: Adam Borowski <kilobyte@angband.pl> Signed-off-by: Andy Lutomirski <luto@kernel.org> Signed-off-by: Borislav Petkov <bp@suse.de> Cc: Borislav Petkov <bp@alien8.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: Daniel Borkmann <daniel@iogearbox.net> Cc: Eric Biggers <ebiggers@google.com> Cc: Johannes Hirte <johannes.hirte@datenkhaos.de> Cc: Kees Cook <keescook@chromium.org> Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Roman Kagan <rkagan@virtuozzo.com> Cc: Thomas Gleixner <tglx@linutronix.de> Fixes: 94b1b03b519b ("x86/mm: Rework lazy TLB mode and TLB freshness tracking") Link: http://lkml.kernel.org/r/20171009170231.fkpraqokz6e4zeco@pd.tnic Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-10-10 00:50:49 +08:00
if (unlikely(loaded_mm == &init_mm))
return;
x86/mm: Implement PCID based optimization: try to preserve old TLB entries using PCID PCID is a "process context ID" -- it's what other architectures call an address space ID. Every non-global TLB entry is tagged with a PCID, only TLB entries that match the currently selected PCID are used, and we can switch PGDs without flushing the TLB. x86's PCID is 12 bits. This is an unorthodox approach to using PCID. x86's PCID is far too short to uniquely identify a process, and we can't even really uniquely identify a running process because there are monster systems with over 4096 CPUs. To make matters worse, past attempts to use all 12 PCID bits have resulted in slowdowns instead of speedups. This patch uses PCID differently. We use a PCID to identify a recently-used mm on a per-cpu basis. An mm has no fixed PCID binding at all; instead, we give it a fresh PCID each time it's loaded except in cases where we want to preserve the TLB, in which case we reuse a recent value. Here are some benchmark results, done on a Skylake laptop at 2.3 GHz (turbo off, intel_pstate requesting max performance) under KVM with the guest using idle=poll (to avoid artifacts when bouncing between CPUs). I haven't done any real statistics here -- I just ran them in a loop and picked the fastest results that didn't look like outliers. Unpatched means commit a4eb8b993554, so all the bookkeeping overhead is gone. ping-pong between two mms on the same CPU using eventfd: patched: 1.22µs patched, nopcid: 1.33µs unpatched: 1.34µs Same ping-pong, but now touch 512 pages (all zero-page to minimize cache misses) each iteration. dTLB misses are measured by dtlb_load_misses.miss_causes_a_walk: patched: 1.8µs 11M dTLB misses patched, nopcid: 6.2µs, 207M dTLB misses unpatched: 6.1µs, 190M dTLB misses Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/9ee75f17a81770feed616358e6860d98a2a5b1e7.1500957502.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-07-25 12:41:38 +08:00
VM_WARN_ON(this_cpu_read(cpu_tlbstate.ctxs[loaded_mm_asid].ctx_id) !=
x86/mm: Track the TLB's tlb_gen and update the flushing algorithm There are two kernel features that would benefit from tracking how up-to-date each CPU's TLB is in the case where IPIs aren't keeping it up to date in real time: - Lazy mm switching currently works by switching to init_mm when it would otherwise flush. This is wasteful: there isn't fundamentally any need to update CR3 at all when going lazy or when returning from lazy mode, nor is there any need to receive flush IPIs at all. Instead, we should just stop trying to keep the TLB coherent when we go lazy and, when unlazying, check whether we missed any flushes. - PCID will let us keep recent user contexts alive in the TLB. If we start doing this, we need a way to decide whether those contexts are up to date. On some paravirt systems, remote TLBs can be flushed without IPIs. This won't update the target CPUs' tlb_gens, which may cause unnecessary local flushes later on. We can address this if it becomes a problem by carefully updating the target CPU's tlb_gen directly. By itself, this patch is a very minor optimization that avoids unnecessary flushes when multiple TLB flushes targetting the same CPU race. The complexity in this patch would not be worth it on its own, but it will enable improved lazy TLB tracking and PCID. Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/1210fb244bc9cbe7677f7f0b72db4d359675f24b.1498751203.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-06-29 23:53:16 +08:00
loaded_mm->context.ctx_id);
x86/mm: Flush more aggressively in lazy TLB mode Since commit: 94b1b03b519b ("x86/mm: Rework lazy TLB mode and TLB freshness tracking") x86's lazy TLB mode has been all the way lazy: when running a kernel thread (including the idle thread), the kernel keeps using the last user mm's page tables without attempting to maintain user TLB coherence at all. From a pure semantic perspective, this is fine -- kernel threads won't attempt to access user pages, so having stale TLB entries doesn't matter. Unfortunately, I forgot about a subtlety. By skipping TLB flushes, we also allow any paging-structure caches that may exist on the CPU to become incoherent. This means that we can have a paging-structure cache entry that references a freed page table, and the CPU is within its rights to do a speculative page walk starting at the freed page table. I can imagine this causing two different problems: - A speculative page walk starting from a bogus page table could read IO addresses. I haven't seen any reports of this causing problems. - A speculative page walk that involves a bogus page table can install garbage in the TLB. Such garbage would always be at a user VA, but some AMD CPUs have logic that triggers a machine check when it notices these bogus entries. I've seen a couple reports of this. Boris further explains the failure mode: > It is actually more of an optimization which assumes that paging-structure > entries are in WB DRAM: > > "TlbCacheDis: cacheable memory disable. Read-write. 0=Enables > performance optimization that assumes PML4, PDP, PDE, and PTE entries > are in cacheable WB-DRAM; memory type checks may be bypassed, and > addresses outside of WB-DRAM may result in undefined behavior or NB > protocol errors. 1=Disables performance optimization and allows PML4, > PDP, PDE and PTE entries to be in any memory type. Operating systems > that maintain page tables in memory types other than WB- DRAM must set > TlbCacheDis to insure proper operation." > > The MCE generated is an NB protocol error to signal that > > "Link: A specific coherent-only packet from a CPU was issued to an > IO link. This may be caused by software which addresses page table > structures in a memory type other than cacheable WB-DRAM without > properly configuring MSRC001_0015[TlbCacheDis]. This may occur, for > example, when page table structure addresses are above top of memory. In > such cases, the NB will generate an MCE if it sees a mismatch between > the memory operation generated by the core and the link type." > > I'm assuming coherent-only packets don't go out on IO links, thus the > error. To fix this, reinstate TLB coherence in lazy mode. With this patch applied, we do it in one of two ways: - If we have PCID, we simply switch back to init_mm's page tables when we enter a kernel thread -- this seems to be quite cheap except for the cost of serializing the CPU. - If we don't have PCID, then we set a flag and switch to init_mm the first time we would otherwise need to flush the TLB. The /sys/kernel/debug/x86/tlb_use_lazy_mode debug switch can be changed to override the default mode for benchmarking. In theory, we could optimize this better by only flushing the TLB in lazy CPUs when a page table is freed. Doing that would require auditing the mm code to make sure that all page table freeing goes through tlb_remove_page() as well as reworking some data structures to implement the improved flush logic. Reported-by: Markus Trippelsdorf <markus@trippelsdorf.de> Reported-by: Adam Borowski <kilobyte@angband.pl> Signed-off-by: Andy Lutomirski <luto@kernel.org> Signed-off-by: Borislav Petkov <bp@suse.de> Cc: Borislav Petkov <bp@alien8.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: Daniel Borkmann <daniel@iogearbox.net> Cc: Eric Biggers <ebiggers@google.com> Cc: Johannes Hirte <johannes.hirte@datenkhaos.de> Cc: Kees Cook <keescook@chromium.org> Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Roman Kagan <rkagan@virtuozzo.com> Cc: Thomas Gleixner <tglx@linutronix.de> Fixes: 94b1b03b519b ("x86/mm: Rework lazy TLB mode and TLB freshness tracking") Link: http://lkml.kernel.org/r/20171009170231.fkpraqokz6e4zeco@pd.tnic Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-10-10 00:50:49 +08:00
if (this_cpu_read(cpu_tlbstate.is_lazy)) {
x86/mm: Track the TLB's tlb_gen and update the flushing algorithm There are two kernel features that would benefit from tracking how up-to-date each CPU's TLB is in the case where IPIs aren't keeping it up to date in real time: - Lazy mm switching currently works by switching to init_mm when it would otherwise flush. This is wasteful: there isn't fundamentally any need to update CR3 at all when going lazy or when returning from lazy mode, nor is there any need to receive flush IPIs at all. Instead, we should just stop trying to keep the TLB coherent when we go lazy and, when unlazying, check whether we missed any flushes. - PCID will let us keep recent user contexts alive in the TLB. If we start doing this, we need a way to decide whether those contexts are up to date. On some paravirt systems, remote TLBs can be flushed without IPIs. This won't update the target CPUs' tlb_gens, which may cause unnecessary local flushes later on. We can address this if it becomes a problem by carefully updating the target CPU's tlb_gen directly. By itself, this patch is a very minor optimization that avoids unnecessary flushes when multiple TLB flushes targetting the same CPU race. The complexity in this patch would not be worth it on its own, but it will enable improved lazy TLB tracking and PCID. Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/1210fb244bc9cbe7677f7f0b72db4d359675f24b.1498751203.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-06-29 23:53:16 +08:00
/*
x86/mm: Flush more aggressively in lazy TLB mode Since commit: 94b1b03b519b ("x86/mm: Rework lazy TLB mode and TLB freshness tracking") x86's lazy TLB mode has been all the way lazy: when running a kernel thread (including the idle thread), the kernel keeps using the last user mm's page tables without attempting to maintain user TLB coherence at all. From a pure semantic perspective, this is fine -- kernel threads won't attempt to access user pages, so having stale TLB entries doesn't matter. Unfortunately, I forgot about a subtlety. By skipping TLB flushes, we also allow any paging-structure caches that may exist on the CPU to become incoherent. This means that we can have a paging-structure cache entry that references a freed page table, and the CPU is within its rights to do a speculative page walk starting at the freed page table. I can imagine this causing two different problems: - A speculative page walk starting from a bogus page table could read IO addresses. I haven't seen any reports of this causing problems. - A speculative page walk that involves a bogus page table can install garbage in the TLB. Such garbage would always be at a user VA, but some AMD CPUs have logic that triggers a machine check when it notices these bogus entries. I've seen a couple reports of this. Boris further explains the failure mode: > It is actually more of an optimization which assumes that paging-structure > entries are in WB DRAM: > > "TlbCacheDis: cacheable memory disable. Read-write. 0=Enables > performance optimization that assumes PML4, PDP, PDE, and PTE entries > are in cacheable WB-DRAM; memory type checks may be bypassed, and > addresses outside of WB-DRAM may result in undefined behavior or NB > protocol errors. 1=Disables performance optimization and allows PML4, > PDP, PDE and PTE entries to be in any memory type. Operating systems > that maintain page tables in memory types other than WB- DRAM must set > TlbCacheDis to insure proper operation." > > The MCE generated is an NB protocol error to signal that > > "Link: A specific coherent-only packet from a CPU was issued to an > IO link. This may be caused by software which addresses page table > structures in a memory type other than cacheable WB-DRAM without > properly configuring MSRC001_0015[TlbCacheDis]. This may occur, for > example, when page table structure addresses are above top of memory. In > such cases, the NB will generate an MCE if it sees a mismatch between > the memory operation generated by the core and the link type." > > I'm assuming coherent-only packets don't go out on IO links, thus the > error. To fix this, reinstate TLB coherence in lazy mode. With this patch applied, we do it in one of two ways: - If we have PCID, we simply switch back to init_mm's page tables when we enter a kernel thread -- this seems to be quite cheap except for the cost of serializing the CPU. - If we don't have PCID, then we set a flag and switch to init_mm the first time we would otherwise need to flush the TLB. The /sys/kernel/debug/x86/tlb_use_lazy_mode debug switch can be changed to override the default mode for benchmarking. In theory, we could optimize this better by only flushing the TLB in lazy CPUs when a page table is freed. Doing that would require auditing the mm code to make sure that all page table freeing goes through tlb_remove_page() as well as reworking some data structures to implement the improved flush logic. Reported-by: Markus Trippelsdorf <markus@trippelsdorf.de> Reported-by: Adam Borowski <kilobyte@angband.pl> Signed-off-by: Andy Lutomirski <luto@kernel.org> Signed-off-by: Borislav Petkov <bp@suse.de> Cc: Borislav Petkov <bp@alien8.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: Daniel Borkmann <daniel@iogearbox.net> Cc: Eric Biggers <ebiggers@google.com> Cc: Johannes Hirte <johannes.hirte@datenkhaos.de> Cc: Kees Cook <keescook@chromium.org> Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Roman Kagan <rkagan@virtuozzo.com> Cc: Thomas Gleixner <tglx@linutronix.de> Fixes: 94b1b03b519b ("x86/mm: Rework lazy TLB mode and TLB freshness tracking") Link: http://lkml.kernel.org/r/20171009170231.fkpraqokz6e4zeco@pd.tnic Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-10-10 00:50:49 +08:00
* We're in lazy mode. We need to at least flush our
* paging-structure cache to avoid speculatively reading
* garbage into our TLB. Since switching to init_mm is barely
* slower than a minimal flush, just switch to init_mm.
x86/mm: Track the TLB's tlb_gen and update the flushing algorithm There are two kernel features that would benefit from tracking how up-to-date each CPU's TLB is in the case where IPIs aren't keeping it up to date in real time: - Lazy mm switching currently works by switching to init_mm when it would otherwise flush. This is wasteful: there isn't fundamentally any need to update CR3 at all when going lazy or when returning from lazy mode, nor is there any need to receive flush IPIs at all. Instead, we should just stop trying to keep the TLB coherent when we go lazy and, when unlazying, check whether we missed any flushes. - PCID will let us keep recent user contexts alive in the TLB. If we start doing this, we need a way to decide whether those contexts are up to date. On some paravirt systems, remote TLBs can be flushed without IPIs. This won't update the target CPUs' tlb_gens, which may cause unnecessary local flushes later on. We can address this if it becomes a problem by carefully updating the target CPU's tlb_gen directly. By itself, this patch is a very minor optimization that avoids unnecessary flushes when multiple TLB flushes targetting the same CPU race. The complexity in this patch would not be worth it on its own, but it will enable improved lazy TLB tracking and PCID. Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/1210fb244bc9cbe7677f7f0b72db4d359675f24b.1498751203.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-06-29 23:53:16 +08:00
*/
x86/mm: Flush more aggressively in lazy TLB mode Since commit: 94b1b03b519b ("x86/mm: Rework lazy TLB mode and TLB freshness tracking") x86's lazy TLB mode has been all the way lazy: when running a kernel thread (including the idle thread), the kernel keeps using the last user mm's page tables without attempting to maintain user TLB coherence at all. From a pure semantic perspective, this is fine -- kernel threads won't attempt to access user pages, so having stale TLB entries doesn't matter. Unfortunately, I forgot about a subtlety. By skipping TLB flushes, we also allow any paging-structure caches that may exist on the CPU to become incoherent. This means that we can have a paging-structure cache entry that references a freed page table, and the CPU is within its rights to do a speculative page walk starting at the freed page table. I can imagine this causing two different problems: - A speculative page walk starting from a bogus page table could read IO addresses. I haven't seen any reports of this causing problems. - A speculative page walk that involves a bogus page table can install garbage in the TLB. Such garbage would always be at a user VA, but some AMD CPUs have logic that triggers a machine check when it notices these bogus entries. I've seen a couple reports of this. Boris further explains the failure mode: > It is actually more of an optimization which assumes that paging-structure > entries are in WB DRAM: > > "TlbCacheDis: cacheable memory disable. Read-write. 0=Enables > performance optimization that assumes PML4, PDP, PDE, and PTE entries > are in cacheable WB-DRAM; memory type checks may be bypassed, and > addresses outside of WB-DRAM may result in undefined behavior or NB > protocol errors. 1=Disables performance optimization and allows PML4, > PDP, PDE and PTE entries to be in any memory type. Operating systems > that maintain page tables in memory types other than WB- DRAM must set > TlbCacheDis to insure proper operation." > > The MCE generated is an NB protocol error to signal that > > "Link: A specific coherent-only packet from a CPU was issued to an > IO link. This may be caused by software which addresses page table > structures in a memory type other than cacheable WB-DRAM without > properly configuring MSRC001_0015[TlbCacheDis]. This may occur, for > example, when page table structure addresses are above top of memory. In > such cases, the NB will generate an MCE if it sees a mismatch between > the memory operation generated by the core and the link type." > > I'm assuming coherent-only packets don't go out on IO links, thus the > error. To fix this, reinstate TLB coherence in lazy mode. With this patch applied, we do it in one of two ways: - If we have PCID, we simply switch back to init_mm's page tables when we enter a kernel thread -- this seems to be quite cheap except for the cost of serializing the CPU. - If we don't have PCID, then we set a flag and switch to init_mm the first time we would otherwise need to flush the TLB. The /sys/kernel/debug/x86/tlb_use_lazy_mode debug switch can be changed to override the default mode for benchmarking. In theory, we could optimize this better by only flushing the TLB in lazy CPUs when a page table is freed. Doing that would require auditing the mm code to make sure that all page table freeing goes through tlb_remove_page() as well as reworking some data structures to implement the improved flush logic. Reported-by: Markus Trippelsdorf <markus@trippelsdorf.de> Reported-by: Adam Borowski <kilobyte@angband.pl> Signed-off-by: Andy Lutomirski <luto@kernel.org> Signed-off-by: Borislav Petkov <bp@suse.de> Cc: Borislav Petkov <bp@alien8.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: Daniel Borkmann <daniel@iogearbox.net> Cc: Eric Biggers <ebiggers@google.com> Cc: Johannes Hirte <johannes.hirte@datenkhaos.de> Cc: Kees Cook <keescook@chromium.org> Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Roman Kagan <rkagan@virtuozzo.com> Cc: Thomas Gleixner <tglx@linutronix.de> Fixes: 94b1b03b519b ("x86/mm: Rework lazy TLB mode and TLB freshness tracking") Link: http://lkml.kernel.org/r/20171009170231.fkpraqokz6e4zeco@pd.tnic Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-10-10 00:50:49 +08:00
switch_mm_irqs_off(NULL, &init_mm, NULL);
return;
}
x86/mm: Track the TLB's tlb_gen and update the flushing algorithm There are two kernel features that would benefit from tracking how up-to-date each CPU's TLB is in the case where IPIs aren't keeping it up to date in real time: - Lazy mm switching currently works by switching to init_mm when it would otherwise flush. This is wasteful: there isn't fundamentally any need to update CR3 at all when going lazy or when returning from lazy mode, nor is there any need to receive flush IPIs at all. Instead, we should just stop trying to keep the TLB coherent when we go lazy and, when unlazying, check whether we missed any flushes. - PCID will let us keep recent user contexts alive in the TLB. If we start doing this, we need a way to decide whether those contexts are up to date. On some paravirt systems, remote TLBs can be flushed without IPIs. This won't update the target CPUs' tlb_gens, which may cause unnecessary local flushes later on. We can address this if it becomes a problem by carefully updating the target CPU's tlb_gen directly. By itself, this patch is a very minor optimization that avoids unnecessary flushes when multiple TLB flushes targetting the same CPU race. The complexity in this patch would not be worth it on its own, but it will enable improved lazy TLB tracking and PCID. Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/1210fb244bc9cbe7677f7f0b72db4d359675f24b.1498751203.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-06-29 23:53:16 +08:00
if (unlikely(local_tlb_gen == mm_tlb_gen)) {
/*
* There's nothing to do: we're already up to date. This can
* happen if two concurrent flushes happen -- the first flush to
* be handled can catch us all the way up, leaving no work for
* the second flush.
*/
x86/mm: Rework lazy TLB mode and TLB freshness tracking x86's lazy TLB mode used to be fairly weak -- it would switch to init_mm the first time it tried to flush a lazy TLB. This meant an unnecessary CR3 write and, if the flush was remote, an unnecessary IPI. Rewrite it entirely. When we enter lazy mode, we simply remove the CPU from mm_cpumask. This means that we need a way to figure out whether we've missed a flush when we switch back out of lazy mode. I use the tlb_gen machinery to track whether a context is up to date. Note to reviewers: this patch, my itself, looks a bit odd. I'm using an array of length 1 containing (ctx_id, tlb_gen) rather than just storing tlb_gen, and making it at array isn't necessary yet. I'm doing this because the next few patches add PCID support, and, with PCID, we need ctx_id, and the array will end up with a length greater than 1. Making it an array now means that there will be less churn and therefore less stress on your eyeballs. NB: This is dubious but, AFAICT, still correct on Xen and UV. xen_exit_mmap() uses mm_cpumask() for nefarious purposes and this patch changes the way that mm_cpumask() works. This should be okay, since Xen *also* iterates all online CPUs to find all the CPUs it needs to twiddle. The UV tlbflush code is rather dated and should be changed. Here are some benchmark results, done on a Skylake laptop at 2.3 GHz (turbo off, intel_pstate requesting max performance) under KVM with the guest using idle=poll (to avoid artifacts when bouncing between CPUs). I haven't done any real statistics here -- I just ran them in a loop and picked the fastest results that didn't look like outliers. Unpatched means commit a4eb8b993554, so all the bookkeeping overhead is gone. MADV_DONTNEED; touch the page; switch CPUs using sched_setaffinity. In an unpatched kernel, MADV_DONTNEED will send an IPI to the previous CPU. This is intended to be a nearly worst-case test. patched: 13.4µs unpatched: 21.6µs Vitaly's pthread_mmap microbenchmark with 8 threads (on four cores), nrounds = 100, 256M data patched: 1.1 seconds or so unpatched: 1.9 seconds or so The sleepup on Vitaly's test appearss to be because it spends a lot of time blocked on mmap_sem, and this patch avoids sending IPIs to blocked CPUs. Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andrew Banman <abanman@sgi.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Dimitri Sivanich <sivanich@sgi.com> Cc: Juergen Gross <jgross@suse.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Travis <travis@sgi.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/ddf2c92962339f4ba39d8fc41b853936ec0b44f1.1498751203.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-06-29 23:53:17 +08:00
trace_tlb_flush(reason, 0);
x86/mm: Track the TLB's tlb_gen and update the flushing algorithm There are two kernel features that would benefit from tracking how up-to-date each CPU's TLB is in the case where IPIs aren't keeping it up to date in real time: - Lazy mm switching currently works by switching to init_mm when it would otherwise flush. This is wasteful: there isn't fundamentally any need to update CR3 at all when going lazy or when returning from lazy mode, nor is there any need to receive flush IPIs at all. Instead, we should just stop trying to keep the TLB coherent when we go lazy and, when unlazying, check whether we missed any flushes. - PCID will let us keep recent user contexts alive in the TLB. If we start doing this, we need a way to decide whether those contexts are up to date. On some paravirt systems, remote TLBs can be flushed without IPIs. This won't update the target CPUs' tlb_gens, which may cause unnecessary local flushes later on. We can address this if it becomes a problem by carefully updating the target CPU's tlb_gen directly. By itself, this patch is a very minor optimization that avoids unnecessary flushes when multiple TLB flushes targetting the same CPU race. The complexity in this patch would not be worth it on its own, but it will enable improved lazy TLB tracking and PCID. Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/1210fb244bc9cbe7677f7f0b72db4d359675f24b.1498751203.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-06-29 23:53:16 +08:00
return;
}
WARN_ON_ONCE(local_tlb_gen > mm_tlb_gen);
WARN_ON_ONCE(f->new_tlb_gen > mm_tlb_gen);
/*
* If we get to this point, we know that our TLB is out of date.
* This does not strictly imply that we need to flush (it's
* possible that f->new_tlb_gen <= local_tlb_gen), but we're
* going to need to flush in the very near future, so we might
* as well get it over with.
*
* The only question is whether to do a full or partial flush.
*
* We do a partial flush if requested and two extra conditions
* are met:
*
* 1. f->new_tlb_gen == local_tlb_gen + 1. We have an invariant that
* we've always done all needed flushes to catch up to
* local_tlb_gen. If, for example, local_tlb_gen == 2 and
* f->new_tlb_gen == 3, then we know that the flush needed to bring
* us up to date for tlb_gen 3 is the partial flush we're
* processing.
*
* As an example of why this check is needed, suppose that there
* are two concurrent flushes. The first is a full flush that
* changes context.tlb_gen from 1 to 2. The second is a partial
* flush that changes context.tlb_gen from 2 to 3. If they get
* processed on this CPU in reverse order, we'll see
* local_tlb_gen == 1, mm_tlb_gen == 3, and end != TLB_FLUSH_ALL.
* If we were to use __flush_tlb_single() and set local_tlb_gen to
* 3, we'd be break the invariant: we'd update local_tlb_gen above
* 1 without the full flush that's needed for tlb_gen 2.
*
* 2. f->new_tlb_gen == mm_tlb_gen. This is purely an optimiation.
* Partial TLB flushes are not all that much cheaper than full TLB
* flushes, so it seems unlikely that it would be a performance win
* to do a partial flush if that won't bring our TLB fully up to
* date. By doing a full flush instead, we can increase
* local_tlb_gen all the way to mm_tlb_gen and we can probably
* avoid another flush in the very near future.
*/
if (f->end != TLB_FLUSH_ALL &&
f->new_tlb_gen == local_tlb_gen + 1 &&
f->new_tlb_gen == mm_tlb_gen) {
/* Partial flush */
unsigned long addr;
unsigned long nr_pages = (f->end - f->start) >> PAGE_SHIFT;
x86/mm: Track the TLB's tlb_gen and update the flushing algorithm There are two kernel features that would benefit from tracking how up-to-date each CPU's TLB is in the case where IPIs aren't keeping it up to date in real time: - Lazy mm switching currently works by switching to init_mm when it would otherwise flush. This is wasteful: there isn't fundamentally any need to update CR3 at all when going lazy or when returning from lazy mode, nor is there any need to receive flush IPIs at all. Instead, we should just stop trying to keep the TLB coherent when we go lazy and, when unlazying, check whether we missed any flushes. - PCID will let us keep recent user contexts alive in the TLB. If we start doing this, we need a way to decide whether those contexts are up to date. On some paravirt systems, remote TLBs can be flushed without IPIs. This won't update the target CPUs' tlb_gens, which may cause unnecessary local flushes later on. We can address this if it becomes a problem by carefully updating the target CPU's tlb_gen directly. By itself, this patch is a very minor optimization that avoids unnecessary flushes when multiple TLB flushes targetting the same CPU race. The complexity in this patch would not be worth it on its own, but it will enable improved lazy TLB tracking and PCID. Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/1210fb244bc9cbe7677f7f0b72db4d359675f24b.1498751203.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-06-29 23:53:16 +08:00
addr = f->start;
while (addr < f->end) {
__flush_tlb_single(addr);
addr += PAGE_SIZE;
}
x86/mm: Refactor flush_tlb_mm_range() to merge local and remote cases The local flush path is very similar to the remote flush path. Merge them. This is intended to make no difference to behavior whatsoever. It removes some code and will make future changes to the flushing mechanics simpler. This patch does remove one small optimization: flush_tlb_mm_range() now has an unconditional smp_mb() instead of using MOV to CR3 or INVLPG as a full barrier when applicable. I think this is okay for a few reasons. First, smp_mb() is quite cheap compared to the cost of a TLB flush. Second, this rearrangement makes a bigger optimization available: with some work on the SMP function call code, we could do the local and remote flushes in parallel. Third, I'm planning a rework of the TLB flush algorithm that will require an atomic operation at the beginning of each flush, and that operation will replace the smp_mb(). Signed-off-by: Andy Lutomirski <luto@kernel.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bpetkov@suse.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Michal Hocko <mhocko@suse.com> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Nadav Amit <namit@vmware.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mm@kvack.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-05-29 01:00:12 +08:00
if (local)
count_vm_tlb_events(NR_TLB_LOCAL_FLUSH_ONE, nr_pages);
trace_tlb_flush(reason, nr_pages);
x86/mm: Track the TLB's tlb_gen and update the flushing algorithm There are two kernel features that would benefit from tracking how up-to-date each CPU's TLB is in the case where IPIs aren't keeping it up to date in real time: - Lazy mm switching currently works by switching to init_mm when it would otherwise flush. This is wasteful: there isn't fundamentally any need to update CR3 at all when going lazy or when returning from lazy mode, nor is there any need to receive flush IPIs at all. Instead, we should just stop trying to keep the TLB coherent when we go lazy and, when unlazying, check whether we missed any flushes. - PCID will let us keep recent user contexts alive in the TLB. If we start doing this, we need a way to decide whether those contexts are up to date. On some paravirt systems, remote TLBs can be flushed without IPIs. This won't update the target CPUs' tlb_gens, which may cause unnecessary local flushes later on. We can address this if it becomes a problem by carefully updating the target CPU's tlb_gen directly. By itself, this patch is a very minor optimization that avoids unnecessary flushes when multiple TLB flushes targetting the same CPU race. The complexity in this patch would not be worth it on its own, but it will enable improved lazy TLB tracking and PCID. Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/1210fb244bc9cbe7677f7f0b72db4d359675f24b.1498751203.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-06-29 23:53:16 +08:00
} else {
/* Full flush. */
local_flush_tlb();
if (local)
count_vm_tlb_event(NR_TLB_LOCAL_FLUSH_ALL);
trace_tlb_flush(reason, TLB_FLUSH_ALL);
}
x86/mm: Track the TLB's tlb_gen and update the flushing algorithm There are two kernel features that would benefit from tracking how up-to-date each CPU's TLB is in the case where IPIs aren't keeping it up to date in real time: - Lazy mm switching currently works by switching to init_mm when it would otherwise flush. This is wasteful: there isn't fundamentally any need to update CR3 at all when going lazy or when returning from lazy mode, nor is there any need to receive flush IPIs at all. Instead, we should just stop trying to keep the TLB coherent when we go lazy and, when unlazying, check whether we missed any flushes. - PCID will let us keep recent user contexts alive in the TLB. If we start doing this, we need a way to decide whether those contexts are up to date. On some paravirt systems, remote TLBs can be flushed without IPIs. This won't update the target CPUs' tlb_gens, which may cause unnecessary local flushes later on. We can address this if it becomes a problem by carefully updating the target CPU's tlb_gen directly. By itself, this patch is a very minor optimization that avoids unnecessary flushes when multiple TLB flushes targetting the same CPU race. The complexity in this patch would not be worth it on its own, but it will enable improved lazy TLB tracking and PCID. Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/1210fb244bc9cbe7677f7f0b72db4d359675f24b.1498751203.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-06-29 23:53:16 +08:00
/* Both paths above update our state to mm_tlb_gen. */
x86/mm: Implement PCID based optimization: try to preserve old TLB entries using PCID PCID is a "process context ID" -- it's what other architectures call an address space ID. Every non-global TLB entry is tagged with a PCID, only TLB entries that match the currently selected PCID are used, and we can switch PGDs without flushing the TLB. x86's PCID is 12 bits. This is an unorthodox approach to using PCID. x86's PCID is far too short to uniquely identify a process, and we can't even really uniquely identify a running process because there are monster systems with over 4096 CPUs. To make matters worse, past attempts to use all 12 PCID bits have resulted in slowdowns instead of speedups. This patch uses PCID differently. We use a PCID to identify a recently-used mm on a per-cpu basis. An mm has no fixed PCID binding at all; instead, we give it a fresh PCID each time it's loaded except in cases where we want to preserve the TLB, in which case we reuse a recent value. Here are some benchmark results, done on a Skylake laptop at 2.3 GHz (turbo off, intel_pstate requesting max performance) under KVM with the guest using idle=poll (to avoid artifacts when bouncing between CPUs). I haven't done any real statistics here -- I just ran them in a loop and picked the fastest results that didn't look like outliers. Unpatched means commit a4eb8b993554, so all the bookkeeping overhead is gone. ping-pong between two mms on the same CPU using eventfd: patched: 1.22µs patched, nopcid: 1.33µs unpatched: 1.34µs Same ping-pong, but now touch 512 pages (all zero-page to minimize cache misses) each iteration. dTLB misses are measured by dtlb_load_misses.miss_causes_a_walk: patched: 1.8µs 11M dTLB misses patched, nopcid: 6.2µs, 207M dTLB misses unpatched: 6.1µs, 190M dTLB misses Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/9ee75f17a81770feed616358e6860d98a2a5b1e7.1500957502.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-07-25 12:41:38 +08:00
this_cpu_write(cpu_tlbstate.ctxs[loaded_mm_asid].tlb_gen, mm_tlb_gen);
}
x86/mm: Refactor flush_tlb_mm_range() to merge local and remote cases The local flush path is very similar to the remote flush path. Merge them. This is intended to make no difference to behavior whatsoever. It removes some code and will make future changes to the flushing mechanics simpler. This patch does remove one small optimization: flush_tlb_mm_range() now has an unconditional smp_mb() instead of using MOV to CR3 or INVLPG as a full barrier when applicable. I think this is okay for a few reasons. First, smp_mb() is quite cheap compared to the cost of a TLB flush. Second, this rearrangement makes a bigger optimization available: with some work on the SMP function call code, we could do the local and remote flushes in parallel. Third, I'm planning a rework of the TLB flush algorithm that will require an atomic operation at the beginning of each flush, and that operation will replace the smp_mb(). Signed-off-by: Andy Lutomirski <luto@kernel.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bpetkov@suse.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Michal Hocko <mhocko@suse.com> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Nadav Amit <namit@vmware.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mm@kvack.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-05-29 01:00:12 +08:00
static void flush_tlb_func_local(void *info, enum tlb_flush_reason reason)
{
const struct flush_tlb_info *f = info;
flush_tlb_func_common(f, true, reason);
}
static void flush_tlb_func_remote(void *info)
{
const struct flush_tlb_info *f = info;
inc_irq_stat(irq_tlb_count);
x86/mm: Rework lazy TLB to track the actual loaded mm Lazy TLB state is currently managed in a rather baroque manner. AFAICT, there are three possible states: - Non-lazy. This means that we're running a user thread or a kernel thread that has called use_mm(). current->mm == current->active_mm == cpu_tlbstate.active_mm and cpu_tlbstate.state == TLBSTATE_OK. - Lazy with user mm. We're running a kernel thread without an mm and we're borrowing an mm_struct. We have current->mm == NULL, current->active_mm == cpu_tlbstate.active_mm, cpu_tlbstate.state != TLBSTATE_OK (i.e. TLBSTATE_LAZY or 0). The current cpu is set in mm_cpumask(current->active_mm). CR3 points to current->active_mm->pgd. The TLB is up to date. - Lazy with init_mm. This happens when we call leave_mm(). We have current->mm == NULL, current->active_mm == cpu_tlbstate.active_mm, but that mm is only relelvant insofar as the scheduler is tracking it for refcounting. cpu_tlbstate.state != TLBSTATE_OK. The current cpu is clear in mm_cpumask(current->active_mm). CR3 points to swapper_pg_dir, i.e. init_mm->pgd. This patch simplifies the situation. Other than perf, x86 stops caring about current->active_mm at all. We have cpu_tlbstate.loaded_mm pointing to the mm that CR3 references. The TLB is always up to date for that mm. leave_mm() just switches us to init_mm. There are no longer any special cases for mm_cpumask, and switch_mm() switches mms without worrying about laziness. After this patch, cpu_tlbstate.state serves only to tell the TLB flush code whether it may switch to init_mm instead of doing a normal flush. This makes fairly extensive changes to xen_exit_mmap(), which used to look a bit like black magic. Perf is unchanged. With or without this change, perf may behave a bit erratically if it tries to read user memory in kernel thread context. We should build on this patch to teach perf to never look at user memory when cpu_tlbstate.loaded_mm != current->mm. Signed-off-by: Andy Lutomirski <luto@kernel.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bpetkov@suse.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Michal Hocko <mhocko@suse.com> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Nadav Amit <namit@vmware.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mm@kvack.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-05-29 01:00:15 +08:00
if (f->mm && f->mm != this_cpu_read(cpu_tlbstate.loaded_mm))
x86/mm: Refactor flush_tlb_mm_range() to merge local and remote cases The local flush path is very similar to the remote flush path. Merge them. This is intended to make no difference to behavior whatsoever. It removes some code and will make future changes to the flushing mechanics simpler. This patch does remove one small optimization: flush_tlb_mm_range() now has an unconditional smp_mb() instead of using MOV to CR3 or INVLPG as a full barrier when applicable. I think this is okay for a few reasons. First, smp_mb() is quite cheap compared to the cost of a TLB flush. Second, this rearrangement makes a bigger optimization available: with some work on the SMP function call code, we could do the local and remote flushes in parallel. Third, I'm planning a rework of the TLB flush algorithm that will require an atomic operation at the beginning of each flush, and that operation will replace the smp_mb(). Signed-off-by: Andy Lutomirski <luto@kernel.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bpetkov@suse.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Michal Hocko <mhocko@suse.com> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Nadav Amit <namit@vmware.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mm@kvack.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-05-29 01:00:12 +08:00
return;
count_vm_tlb_event(NR_TLB_REMOTE_FLUSH_RECEIVED);
flush_tlb_func_common(f, false, TLB_REMOTE_SHOOTDOWN);
}
void native_flush_tlb_others(const struct cpumask *cpumask,
const struct flush_tlb_info *info)
{
count_vm_tlb_event(NR_TLB_REMOTE_FLUSH);
if (info->end == TLB_FLUSH_ALL)
trace_tlb_flush(TLB_REMOTE_SEND_IPI, TLB_FLUSH_ALL);
else
trace_tlb_flush(TLB_REMOTE_SEND_IPI,
(info->end - info->start) >> PAGE_SHIFT);
if (is_uv_system()) {
x86/mm: Rework lazy TLB mode and TLB freshness tracking x86's lazy TLB mode used to be fairly weak -- it would switch to init_mm the first time it tried to flush a lazy TLB. This meant an unnecessary CR3 write and, if the flush was remote, an unnecessary IPI. Rewrite it entirely. When we enter lazy mode, we simply remove the CPU from mm_cpumask. This means that we need a way to figure out whether we've missed a flush when we switch back out of lazy mode. I use the tlb_gen machinery to track whether a context is up to date. Note to reviewers: this patch, my itself, looks a bit odd. I'm using an array of length 1 containing (ctx_id, tlb_gen) rather than just storing tlb_gen, and making it at array isn't necessary yet. I'm doing this because the next few patches add PCID support, and, with PCID, we need ctx_id, and the array will end up with a length greater than 1. Making it an array now means that there will be less churn and therefore less stress on your eyeballs. NB: This is dubious but, AFAICT, still correct on Xen and UV. xen_exit_mmap() uses mm_cpumask() for nefarious purposes and this patch changes the way that mm_cpumask() works. This should be okay, since Xen *also* iterates all online CPUs to find all the CPUs it needs to twiddle. The UV tlbflush code is rather dated and should be changed. Here are some benchmark results, done on a Skylake laptop at 2.3 GHz (turbo off, intel_pstate requesting max performance) under KVM with the guest using idle=poll (to avoid artifacts when bouncing between CPUs). I haven't done any real statistics here -- I just ran them in a loop and picked the fastest results that didn't look like outliers. Unpatched means commit a4eb8b993554, so all the bookkeeping overhead is gone. MADV_DONTNEED; touch the page; switch CPUs using sched_setaffinity. In an unpatched kernel, MADV_DONTNEED will send an IPI to the previous CPU. This is intended to be a nearly worst-case test. patched: 13.4µs unpatched: 21.6µs Vitaly's pthread_mmap microbenchmark with 8 threads (on four cores), nrounds = 100, 256M data patched: 1.1 seconds or so unpatched: 1.9 seconds or so The sleepup on Vitaly's test appearss to be because it spends a lot of time blocked on mmap_sem, and this patch avoids sending IPIs to blocked CPUs. Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andrew Banman <abanman@sgi.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Dimitri Sivanich <sivanich@sgi.com> Cc: Juergen Gross <jgross@suse.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Travis <travis@sgi.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/ddf2c92962339f4ba39d8fc41b853936ec0b44f1.1498751203.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-06-29 23:53:17 +08:00
/*
* This whole special case is confused. UV has a "Broadcast
* Assist Unit", which seems to be a fancy way to send IPIs.
* Back when x86 used an explicit TLB flush IPI, UV was
* optimized to use its own mechanism. These days, x86 uses
* smp_call_function_many(), but UV still uses a manual IPI,
* and that IPI's action is out of date -- it does a manual
* flush instead of calling flush_tlb_func_remote(). This
* means that the percpu tlb_gen variables won't be updated
* and we'll do pointless flushes on future context switches.
*
* Rather than hooking native_flush_tlb_others() here, I think
* that UV should be updated so that smp_call_function_many(),
* etc, are optimal on UV.
*/
unsigned int cpu;
cpu = smp_processor_id();
cpumask = uv_flush_tlb_others(cpumask, info);
if (cpumask)
x86/mm: Refactor flush_tlb_mm_range() to merge local and remote cases The local flush path is very similar to the remote flush path. Merge them. This is intended to make no difference to behavior whatsoever. It removes some code and will make future changes to the flushing mechanics simpler. This patch does remove one small optimization: flush_tlb_mm_range() now has an unconditional smp_mb() instead of using MOV to CR3 or INVLPG as a full barrier when applicable. I think this is okay for a few reasons. First, smp_mb() is quite cheap compared to the cost of a TLB flush. Second, this rearrangement makes a bigger optimization available: with some work on the SMP function call code, we could do the local and remote flushes in parallel. Third, I'm planning a rework of the TLB flush algorithm that will require an atomic operation at the beginning of each flush, and that operation will replace the smp_mb(). Signed-off-by: Andy Lutomirski <luto@kernel.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bpetkov@suse.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Michal Hocko <mhocko@suse.com> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Nadav Amit <namit@vmware.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mm@kvack.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-05-29 01:00:12 +08:00
smp_call_function_many(cpumask, flush_tlb_func_remote,
(void *)info, 1);
return;
}
x86/mm: Refactor flush_tlb_mm_range() to merge local and remote cases The local flush path is very similar to the remote flush path. Merge them. This is intended to make no difference to behavior whatsoever. It removes some code and will make future changes to the flushing mechanics simpler. This patch does remove one small optimization: flush_tlb_mm_range() now has an unconditional smp_mb() instead of using MOV to CR3 or INVLPG as a full barrier when applicable. I think this is okay for a few reasons. First, smp_mb() is quite cheap compared to the cost of a TLB flush. Second, this rearrangement makes a bigger optimization available: with some work on the SMP function call code, we could do the local and remote flushes in parallel. Third, I'm planning a rework of the TLB flush algorithm that will require an atomic operation at the beginning of each flush, and that operation will replace the smp_mb(). Signed-off-by: Andy Lutomirski <luto@kernel.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bpetkov@suse.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Michal Hocko <mhocko@suse.com> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Nadav Amit <namit@vmware.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mm@kvack.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-05-29 01:00:12 +08:00
smp_call_function_many(cpumask, flush_tlb_func_remote,
(void *)info, 1);
}
x86/mm: Set TLB flush tunable to sane value (33) This has been run through Intel's LKP tests across a wide range of modern sytems and workloads and it wasn't shown to make a measurable performance difference positive or negative. Now that we have some shiny new tracepoints, we can actually figure out what the heck is going on. During a kernel compile, 60% of the flush_tlb_mm_range() calls are for a single page. It breaks down like this: size percent percent<= V V V GLOBAL: 2.20% 2.20% avg cycles: 2283 1: 56.92% 59.12% avg cycles: 1276 2: 13.78% 72.90% avg cycles: 1505 3: 8.26% 81.16% avg cycles: 1880 4: 7.41% 88.58% avg cycles: 2447 5: 1.73% 90.31% avg cycles: 2358 6: 1.32% 91.63% avg cycles: 2563 7: 1.14% 92.77% avg cycles: 2862 8: 0.62% 93.39% avg cycles: 3542 9: 0.08% 93.47% avg cycles: 3289 10: 0.43% 93.90% avg cycles: 3570 11: 0.20% 94.10% avg cycles: 3767 12: 0.08% 94.18% avg cycles: 3996 13: 0.03% 94.20% avg cycles: 4077 14: 0.02% 94.23% avg cycles: 4836 15: 0.04% 94.26% avg cycles: 5699 16: 0.06% 94.32% avg cycles: 5041 17: 0.57% 94.89% avg cycles: 5473 18: 0.02% 94.91% avg cycles: 5396 19: 0.03% 94.95% avg cycles: 5296 20: 0.02% 94.96% avg cycles: 6749 21: 0.18% 95.14% avg cycles: 6225 22: 0.01% 95.15% avg cycles: 6393 23: 0.01% 95.16% avg cycles: 6861 24: 0.12% 95.28% avg cycles: 6912 25: 0.05% 95.32% avg cycles: 7190 26: 0.01% 95.33% avg cycles: 7793 27: 0.01% 95.34% avg cycles: 7833 28: 0.01% 95.35% avg cycles: 8253 29: 0.08% 95.42% avg cycles: 8024 30: 0.03% 95.45% avg cycles: 9670 31: 0.01% 95.46% avg cycles: 8949 32: 0.01% 95.46% avg cycles: 9350 33: 3.11% 98.57% avg cycles: 8534 34: 0.02% 98.60% avg cycles: 10977 35: 0.02% 98.62% avg cycles: 11400 We get in to dimishing returns pretty quickly. On pre-IvyBridge CPUs, we used to set the limit at 8 pages, and it was set at 128 on IvyBrige. That 128 number looks pretty silly considering that less than 0.5% of the flushes are that large. The previous code tried to size this number based on the size of the TLB. Good idea, but it's error-prone, needs maintenance (which it didn't get up to now), and probably would not matter in practice much. Settting it to 33 means that we cover the mallopt M_TRIM_THRESHOLD, which is the most universally common size to do flushes. That's the short version. Here's the long one for why I chose 33: 1. These numbers have a constant bias in the timestamps from the tracing. Probably counts for a couple hundred cycles in each of these tests, but it should be fairly _even_ across all of them. The smallest delta between the tracepoints I have ever seen is 335 cycles. This is one reason the cycles/page cost goes down in general as the flushes get larger. The true cost is nearer to 100 cycles. 2. A full flush is more expensive than a single invlpg, but not by much (single percentages). 3. A dtlb miss is 17.1ns (~45 cycles) and a itlb miss is 13.0ns (~34 cycles). At those rates, refilling the 512-entry dTLB takes 22,000 cycles. 4. 22,000 cycles is approximately the equivalent of doing 85 invlpg operations. But, the odds are that the TLB can actually be filled up faster than that because TLB misses that are close in time also tend to leverage the same caches. 6. ~98% of flushes are <=33 pages. There are a lot of flushes of 33 pages, probably because libc's M_TRIM_THRESHOLD is set to 128k (32 pages) 7. I've found no consistent data to support changing the IvyBridge vs. SandyBridge tunable by a factor of 16 I used the performance counters on this hardware (IvyBridge i5-3320M) to figure out the tlb miss costs: ocperf.py stat -e dtlb_load_misses.walk_duration,dtlb_load_misses.walk_completed,dtlb_store_misses.walk_duration,dtlb_store_misses.walk_completed,itlb_misses.walk_duration,itlb_misses.walk_completed,itlb.itlb_flush 7,720,030,970 dtlb_load_misses_walk_duration [57.13%] 169,856,353 dtlb_load_misses_walk_completed [57.15%] 708,832,859 dtlb_store_misses_walk_duration [57.17%] 19,346,823 dtlb_store_misses_walk_completed [57.17%] 2,779,687,402 itlb_misses_walk_duration [57.15%] 82,241,148 itlb_misses_walk_completed [57.13%] 770,717 itlb_itlb_flush [57.11%] Show that a dtlb miss is 17.1ns (~45 cycles) and a itlb miss is 13.0ns (~34 cycles). At those rates, refilling the 512-entry dTLB takes 22,000 cycles. On a SandyBridge system with more cores and larger caches, those are dtlb=13.4ns and itlb=9.5ns. cat perf.stat.txt | perl -pe 's/,//g' | awk '/itlb_misses_walk_duration/ { icyc+=$1 } /itlb_misses_walk_completed/ { imiss+=$1 } /dtlb_.*_walk_duration/ { dcyc+=$1 } /dtlb_.*.*completed/ { dmiss+=$1 } END {print "itlb cyc/miss: ", icyc/imiss, " dtlb cyc/miss: ", dcyc/dmiss, " ----- ", icyc,imiss, dcyc,dmiss } On Westmere CPUs, the counters to use are: itlb_flush,itlb_misses.walk_cycles,itlb_misses.any,dtlb_misses.walk_cycles,dtlb_misses.any The assumptions that this code went in under: https://lkml.org/lkml/2012/6/12/119 say that a flush and a refill are about 100ns. Being generous, that is over by a factor of 6 on the refill side, although it is fairly close on the cost of an invlpg. An increase of a single invlpg operation seems to lengthen the flush range operation by about 200 cycles. Here is one example of the data collected for flushing 10 and 11 pages (full data are below): 10: 0.43% 93.90% avg cycles: 3570 cycles/page: 357 samples: 4714 11: 0.20% 94.10% avg cycles: 3767 cycles/page: 342 samples: 2145 How to generate this table: echo 10000 > /sys/kernel/debug/tracing/buffer_size_kb echo x86-tsc > /sys/kernel/debug/tracing/trace_clock echo 'reason != 0' > /sys/kernel/debug/tracing/events/tlb/tlb_flush/filter echo 1 > /sys/kernel/debug/tracing/events/tlb/tlb_flush/enable Pipe the trace output in to this script: http://sr71.net/~dave/intel/201402-tlb/trace-time-diff-process.pl.txt Note that these data were gathered with the invlpg threshold set to 150 pages. Only data points with >=50 of samples were printed: Flush % of %<= in flush this pages es size ------------------------------------------------------------------------------ -1: 2.20% 2.20% avg cycles: 2283 cycles/page: xxxx samples: 23960 1: 56.92% 59.12% avg cycles: 1276 cycles/page: 1276 samples: 620895 2: 13.78% 72.90% avg cycles: 1505 cycles/page: 752 samples: 150335 3: 8.26% 81.16% avg cycles: 1880 cycles/page: 626 samples: 90131 4: 7.41% 88.58% avg cycles: 2447 cycles/page: 611 samples: 80877 5: 1.73% 90.31% avg cycles: 2358 cycles/page: 471 samples: 18885 6: 1.32% 91.63% avg cycles: 2563 cycles/page: 427 samples: 14397 7: 1.14% 92.77% avg cycles: 2862 cycles/page: 408 samples: 12441 8: 0.62% 93.39% avg cycles: 3542 cycles/page: 442 samples: 6721 9: 0.08% 93.47% avg cycles: 3289 cycles/page: 365 samples: 917 10: 0.43% 93.90% avg cycles: 3570 cycles/page: 357 samples: 4714 11: 0.20% 94.10% avg cycles: 3767 cycles/page: 342 samples: 2145 12: 0.08% 94.18% avg cycles: 3996 cycles/page: 333 samples: 864 13: 0.03% 94.20% avg cycles: 4077 cycles/page: 313 samples: 289 14: 0.02% 94.23% avg cycles: 4836 cycles/page: 345 samples: 236 15: 0.04% 94.26% avg cycles: 5699 cycles/page: 379 samples: 390 16: 0.06% 94.32% avg cycles: 5041 cycles/page: 315 samples: 643 17: 0.57% 94.89% avg cycles: 5473 cycles/page: 321 samples: 6229 18: 0.02% 94.91% avg cycles: 5396 cycles/page: 299 samples: 224 19: 0.03% 94.95% avg cycles: 5296 cycles/page: 278 samples: 367 20: 0.02% 94.96% avg cycles: 6749 cycles/page: 337 samples: 185 21: 0.18% 95.14% avg cycles: 6225 cycles/page: 296 samples: 1964 22: 0.01% 95.15% avg cycles: 6393 cycles/page: 290 samples: 83 23: 0.01% 95.16% avg cycles: 6861 cycles/page: 298 samples: 61 24: 0.12% 95.28% avg cycles: 6912 cycles/page: 288 samples: 1307 25: 0.05% 95.32% avg cycles: 7190 cycles/page: 287 samples: 533 26: 0.01% 95.33% avg cycles: 7793 cycles/page: 299 samples: 94 27: 0.01% 95.34% avg cycles: 7833 cycles/page: 290 samples: 66 28: 0.01% 95.35% avg cycles: 8253 cycles/page: 294 samples: 73 29: 0.08% 95.42% avg cycles: 8024 cycles/page: 276 samples: 846 30: 0.03% 95.45% avg cycles: 9670 cycles/page: 322 samples: 296 31: 0.01% 95.46% avg cycles: 8949 cycles/page: 288 samples: 79 32: 0.01% 95.46% avg cycles: 9350 cycles/page: 292 samples: 60 33: 3.11% 98.57% avg cycles: 8534 cycles/page: 258 samples: 33936 34: 0.02% 98.60% avg cycles: 10977 cycles/page: 322 samples: 268 35: 0.02% 98.62% avg cycles: 11400 cycles/page: 325 samples: 177 36: 0.01% 98.63% avg cycles: 11504 cycles/page: 319 samples: 161 37: 0.02% 98.65% avg cycles: 11596 cycles/page: 313 samples: 182 38: 0.02% 98.66% avg cycles: 11850 cycles/page: 311 samples: 195 39: 0.01% 98.68% avg cycles: 12158 cycles/page: 311 samples: 128 40: 0.01% 98.68% avg cycles: 11626 cycles/page: 290 samples: 78 41: 0.04% 98.73% avg cycles: 11435 cycles/page: 278 samples: 477 42: 0.01% 98.73% avg cycles: 12571 cycles/page: 299 samples: 74 43: 0.01% 98.74% avg cycles: 12562 cycles/page: 292 samples: 78 44: 0.01% 98.75% avg cycles: 12991 cycles/page: 295 samples: 108 45: 0.01% 98.76% avg cycles: 13169 cycles/page: 292 samples: 78 46: 0.02% 98.78% avg cycles: 12891 cycles/page: 280 samples: 261 47: 0.01% 98.79% avg cycles: 13099 cycles/page: 278 samples: 67 48: 0.01% 98.80% avg cycles: 13851 cycles/page: 288 samples: 77 49: 0.01% 98.80% avg cycles: 13749 cycles/page: 280 samples: 66 50: 0.01% 98.81% avg cycles: 13949 cycles/page: 278 samples: 73 52: 0.00% 98.82% avg cycles: 14243 cycles/page: 273 samples: 52 54: 0.01% 98.83% avg cycles: 15312 cycles/page: 283 samples: 87 55: 0.01% 98.84% avg cycles: 15197 cycles/page: 276 samples: 109 56: 0.02% 98.86% avg cycles: 15234 cycles/page: 272 samples: 208 57: 0.00% 98.86% avg cycles: 14888 cycles/page: 261 samples: 53 58: 0.01% 98.87% avg cycles: 15037 cycles/page: 259 samples: 59 59: 0.01% 98.87% avg cycles: 15752 cycles/page: 266 samples: 63 62: 0.00% 98.89% avg cycles: 16222 cycles/page: 261 samples: 54 64: 0.02% 98.91% avg cycles: 17179 cycles/page: 268 samples: 248 65: 0.12% 99.03% avg cycles: 18762 cycles/page: 288 samples: 1324 85: 0.00% 99.10% avg cycles: 21649 cycles/page: 254 samples: 50 127: 0.01% 99.18% avg cycles: 32397 cycles/page: 255 samples: 75 128: 0.13% 99.31% avg cycles: 31711 cycles/page: 247 samples: 1466 129: 0.18% 99.49% avg cycles: 33017 cycles/page: 255 samples: 1927 181: 0.33% 99.84% avg cycles: 2489 cycles/page: 13 samples: 3547 256: 0.05% 99.91% avg cycles: 2305 cycles/page: 9 samples: 550 512: 0.03% 99.95% avg cycles: 2133 cycles/page: 4 samples: 304 1512: 0.01% 99.99% avg cycles: 3038 cycles/page: 2 samples: 65 Here are the tlb counters during a 10-second slice of a kernel compile for a SandyBridge system. It's better than IvyBridge, but probably due to the larger caches since this was one of the 'X' extreme parts. 10,873,007,282 dtlb_load_misses_walk_duration 250,711,333 dtlb_load_misses_walk_completed 1,212,395,865 dtlb_store_misses_walk_duration 31,615,772 dtlb_store_misses_walk_completed 5,091,010,274 itlb_misses_walk_duration 163,193,511 itlb_misses_walk_completed 1,321,980 itlb_itlb_flush 10.008045158 seconds time elapsed # cat perf.stat.1392743721.txt | perl -pe 's/,//g' | awk '/itlb_misses_walk_duration/ { icyc+=$1 } /itlb_misses_walk_completed/ { imiss+=$1 } /dtlb_.*_walk_duration/ { dcyc+=$1 } /dtlb_.*.*completed/ { dmiss+=$1 } END {print "itlb cyc/miss: ", icyc/imiss/3.3, " dtlb cyc/miss: ", dcyc/dmiss/3.3, " ----- ", icyc,imiss, dcyc,dmiss }' itlb ns/miss: 9.45338 dtlb ns/miss: 12.9716 Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Link: http://lkml.kernel.org/r/20140731154103.10C1115E@viggo.jf.intel.com Acked-by: Rik van Riel <riel@redhat.com> Acked-by: Mel Gorman <mgorman@suse.de> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2014-07-31 23:41:03 +08:00
/*
* See Documentation/x86/tlb.txt for details. We choose 33
* because it is large enough to cover the vast majority (at
* least 95%) of allocations, and is small enough that we are
* confident it will not cause too much overhead. Each single
* flush is about 100 ns, so this caps the maximum overhead at
* _about_ 3,000 ns.
*
* This is in units of pages.
*/
static unsigned long tlb_single_page_flush_ceiling __read_mostly = 33;
x86/mm: Rip out complicated, out-of-date, buggy TLB flushing I think the flush_tlb_mm_range() code that tries to tune the flush sizes based on the CPU needs to get ripped out for several reasons: 1. It is obviously buggy. It uses mm->total_vm to judge the task's footprint in the TLB. It should certainly be using some measure of RSS, *NOT* ->total_vm since only resident memory can populate the TLB. 2. Haswell, and several other CPUs are missing from the intel_tlb_flushall_shift_set() function. Thus, it has been demonstrated to bitrot quickly in practice. 3. It is plain wrong in my vm: [ 0.037444] Last level iTLB entries: 4KB 0, 2MB 0, 4MB 0 [ 0.037444] Last level dTLB entries: 4KB 0, 2MB 0, 4MB 0 [ 0.037444] tlb_flushall_shift: 6 Which leads to it to never use invlpg. 4. The assumptions about TLB refill costs are wrong: http://lkml.kernel.org/r/1337782555-8088-3-git-send-email-alex.shi@intel.com (more on this in later patches) 5. I can not reproduce the original data: https://lkml.org/lkml/2012/5/17/59 I believe the sample times were too short. Running the benchmark in a loop yields times that vary quite a bit. Note that this leaves us with a static ceiling of 1 page. This is a conservative, dumb setting, and will be revised in a later patch. This also removes the code which attempts to predict whether we are flushing data or instructions. We expect instruction flushes to be relatively rare and not worth tuning for explicitly. Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Link: http://lkml.kernel.org/r/20140731154055.ABC88E89@viggo.jf.intel.com Acked-by: Rik van Riel <riel@redhat.com> Acked-by: Mel Gorman <mgorman@suse.de> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2014-07-31 23:40:55 +08:00
x86/tlb: enable tlb flush range support for x86 Not every tlb_flush execution moment is really need to evacuate all TLB entries, like in munmap, just few 'invlpg' is better for whole process performance, since it leaves most of TLB entries for later accessing. This patch also rewrite flush_tlb_range for 2 purposes: 1, split it out to get flush_blt_mm_range function. 2, clean up to reduce line breaking, thanks for Borislav's input. My micro benchmark 'mummap' http://lkml.org/lkml/2012/5/17/59 show that the random memory access on other CPU has 0~50% speed up on a 2P * 4cores * HT NHM EP while do 'munmap'. Thanks Yongjie's testing on this patch: ------------- I used Linux 3.4-RC6 w/ and w/o his patches as Xen dom0 and guest kernel. After running two benchmarks in Xen HVM guest, I found his patches brought about 1%~3% performance gain in 'kernel build' and 'netperf' testing, though the performance gain was not very stable in 'kernel build' testing. Some detailed testing results are below. Testing Environment: Hardware: Romley-EP platform Xen version: latest upstream Linux kernel: 3.4-RC6 Guest vCPU number: 8 NIC: Intel 82599 (10GB bandwidth) In 'kernel build' testing in guest: Command line | performance gain make -j 4 | 3.81% make -j 8 | 0.37% make -j 16 | -0.52% In 'netperf' testing, we tested TCP_STREAM with default socket size 16384 byte as large packet and 64 byte as small packet. I used several clients to add networking pressure, then 'netperf' server automatically generated several threads to response them. I also used large-size packet and small-size packet in the testing. Packet size | Thread number | performance gain 16384 bytes | 4 | 0.02% 16384 bytes | 8 | 2.21% 16384 bytes | 16 | 2.04% 64 bytes | 4 | 1.07% 64 bytes | 8 | 3.31% 64 bytes | 16 | 0.71% Signed-off-by: Alex Shi <alex.shi@intel.com> Link: http://lkml.kernel.org/r/1340845344-27557-8-git-send-email-alex.shi@intel.com Tested-by: Ren, Yongjie <yongjie.ren@intel.com> Signed-off-by: H. Peter Anvin <hpa@zytor.com>
2012-06-28 09:02:22 +08:00
void flush_tlb_mm_range(struct mm_struct *mm, unsigned long start,
unsigned long end, unsigned long vmflag)
{
x86/mm: Refactor flush_tlb_mm_range() to merge local and remote cases The local flush path is very similar to the remote flush path. Merge them. This is intended to make no difference to behavior whatsoever. It removes some code and will make future changes to the flushing mechanics simpler. This patch does remove one small optimization: flush_tlb_mm_range() now has an unconditional smp_mb() instead of using MOV to CR3 or INVLPG as a full barrier when applicable. I think this is okay for a few reasons. First, smp_mb() is quite cheap compared to the cost of a TLB flush. Second, this rearrangement makes a bigger optimization available: with some work on the SMP function call code, we could do the local and remote flushes in parallel. Third, I'm planning a rework of the TLB flush algorithm that will require an atomic operation at the beginning of each flush, and that operation will replace the smp_mb(). Signed-off-by: Andy Lutomirski <luto@kernel.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bpetkov@suse.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Michal Hocko <mhocko@suse.com> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Nadav Amit <namit@vmware.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mm@kvack.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-05-29 01:00:12 +08:00
int cpu;
x86/mm: Make flush_tlb_mm_range() more predictable I'm about to rewrite the function almost completely, but first I want to get a functional change out of the way. Currently, if flush_tlb_mm_range() does not flush the local TLB at all, it will never do individual page flushes on remote CPUs. This seems to be an accident, and preserving it will be awkward. Let's change it first so that any regressions in the rewrite will be easier to bisect and so that the rewrite can attempt to change no visible behavior at all. The fix is simple: we can simply avoid short-circuiting the calculation of base_pages_to_flush. As a side effect, this also eliminates a potential corner case: if tlb_single_page_flush_ceiling == TLB_FLUSH_ALL, flush_tlb_mm_range() could have ended up flushing the entire address space one page at a time. Signed-off-by: Andy Lutomirski <luto@kernel.org> Acked-by: Dave Hansen <dave.hansen@intel.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: Denys Vlasenko <dvlasenk@redhat.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Josh Poimboeuf <jpoimboe@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Nadav Amit <namit@vmware.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/4b29b771d9975aad7154c314534fec235618175a.1492844372.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-04-22 15:01:21 +08:00
x86/mm: Refactor flush_tlb_mm_range() to merge local and remote cases The local flush path is very similar to the remote flush path. Merge them. This is intended to make no difference to behavior whatsoever. It removes some code and will make future changes to the flushing mechanics simpler. This patch does remove one small optimization: flush_tlb_mm_range() now has an unconditional smp_mb() instead of using MOV to CR3 or INVLPG as a full barrier when applicable. I think this is okay for a few reasons. First, smp_mb() is quite cheap compared to the cost of a TLB flush. Second, this rearrangement makes a bigger optimization available: with some work on the SMP function call code, we could do the local and remote flushes in parallel. Third, I'm planning a rework of the TLB flush algorithm that will require an atomic operation at the beginning of each flush, and that operation will replace the smp_mb(). Signed-off-by: Andy Lutomirski <luto@kernel.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bpetkov@suse.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Michal Hocko <mhocko@suse.com> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Nadav Amit <namit@vmware.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mm@kvack.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-05-29 01:00:12 +08:00
struct flush_tlb_info info = {
.mm = mm,
};
x86/mm: Make flush_tlb_mm_range() more predictable I'm about to rewrite the function almost completely, but first I want to get a functional change out of the way. Currently, if flush_tlb_mm_range() does not flush the local TLB at all, it will never do individual page flushes on remote CPUs. This seems to be an accident, and preserving it will be awkward. Let's change it first so that any regressions in the rewrite will be easier to bisect and so that the rewrite can attempt to change no visible behavior at all. The fix is simple: we can simply avoid short-circuiting the calculation of base_pages_to_flush. As a side effect, this also eliminates a potential corner case: if tlb_single_page_flush_ceiling == TLB_FLUSH_ALL, flush_tlb_mm_range() could have ended up flushing the entire address space one page at a time. Signed-off-by: Andy Lutomirski <luto@kernel.org> Acked-by: Dave Hansen <dave.hansen@intel.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Borislav Petkov <bp@alien8.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: Denys Vlasenko <dvlasenk@redhat.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Josh Poimboeuf <jpoimboe@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Michal Hocko <mhocko@suse.com> Cc: Nadav Amit <namit@vmware.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/4b29b771d9975aad7154c314534fec235618175a.1492844372.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-04-22 15:01:21 +08:00
x86/mm: Refactor flush_tlb_mm_range() to merge local and remote cases The local flush path is very similar to the remote flush path. Merge them. This is intended to make no difference to behavior whatsoever. It removes some code and will make future changes to the flushing mechanics simpler. This patch does remove one small optimization: flush_tlb_mm_range() now has an unconditional smp_mb() instead of using MOV to CR3 or INVLPG as a full barrier when applicable. I think this is okay for a few reasons. First, smp_mb() is quite cheap compared to the cost of a TLB flush. Second, this rearrangement makes a bigger optimization available: with some work on the SMP function call code, we could do the local and remote flushes in parallel. Third, I'm planning a rework of the TLB flush algorithm that will require an atomic operation at the beginning of each flush, and that operation will replace the smp_mb(). Signed-off-by: Andy Lutomirski <luto@kernel.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bpetkov@suse.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Michal Hocko <mhocko@suse.com> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Nadav Amit <namit@vmware.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mm@kvack.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-05-29 01:00:12 +08:00
cpu = get_cpu();
/* This is also a barrier that synchronizes with switch_mm(). */
x86/mm: Track the TLB's tlb_gen and update the flushing algorithm There are two kernel features that would benefit from tracking how up-to-date each CPU's TLB is in the case where IPIs aren't keeping it up to date in real time: - Lazy mm switching currently works by switching to init_mm when it would otherwise flush. This is wasteful: there isn't fundamentally any need to update CR3 at all when going lazy or when returning from lazy mode, nor is there any need to receive flush IPIs at all. Instead, we should just stop trying to keep the TLB coherent when we go lazy and, when unlazying, check whether we missed any flushes. - PCID will let us keep recent user contexts alive in the TLB. If we start doing this, we need a way to decide whether those contexts are up to date. On some paravirt systems, remote TLBs can be flushed without IPIs. This won't update the target CPUs' tlb_gens, which may cause unnecessary local flushes later on. We can address this if it becomes a problem by carefully updating the target CPU's tlb_gen directly. By itself, this patch is a very minor optimization that avoids unnecessary flushes when multiple TLB flushes targetting the same CPU race. The complexity in this patch would not be worth it on its own, but it will enable improved lazy TLB tracking and PCID. Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/1210fb244bc9cbe7677f7f0b72db4d359675f24b.1498751203.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-06-29 23:53:16 +08:00
info.new_tlb_gen = inc_mm_tlb_gen(mm);
x86/mm: Refactor flush_tlb_mm_range() to merge local and remote cases The local flush path is very similar to the remote flush path. Merge them. This is intended to make no difference to behavior whatsoever. It removes some code and will make future changes to the flushing mechanics simpler. This patch does remove one small optimization: flush_tlb_mm_range() now has an unconditional smp_mb() instead of using MOV to CR3 or INVLPG as a full barrier when applicable. I think this is okay for a few reasons. First, smp_mb() is quite cheap compared to the cost of a TLB flush. Second, this rearrangement makes a bigger optimization available: with some work on the SMP function call code, we could do the local and remote flushes in parallel. Third, I'm planning a rework of the TLB flush algorithm that will require an atomic operation at the beginning of each flush, and that operation will replace the smp_mb(). Signed-off-by: Andy Lutomirski <luto@kernel.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bpetkov@suse.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Michal Hocko <mhocko@suse.com> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Nadav Amit <namit@vmware.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mm@kvack.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-05-29 01:00:12 +08:00
/* Should we flush just the requested range? */
if ((end != TLB_FLUSH_ALL) &&
!(vmflag & VM_HUGETLB) &&
((end - start) >> PAGE_SHIFT) <= tlb_single_page_flush_ceiling) {
info.start = start;
info.end = end;
} else {
info.start = 0UL;
info.end = TLB_FLUSH_ALL;
}
x86/mm: Refactor flush_tlb_mm_range() to merge local and remote cases The local flush path is very similar to the remote flush path. Merge them. This is intended to make no difference to behavior whatsoever. It removes some code and will make future changes to the flushing mechanics simpler. This patch does remove one small optimization: flush_tlb_mm_range() now has an unconditional smp_mb() instead of using MOV to CR3 or INVLPG as a full barrier when applicable. I think this is okay for a few reasons. First, smp_mb() is quite cheap compared to the cost of a TLB flush. Second, this rearrangement makes a bigger optimization available: with some work on the SMP function call code, we could do the local and remote flushes in parallel. Third, I'm planning a rework of the TLB flush algorithm that will require an atomic operation at the beginning of each flush, and that operation will replace the smp_mb(). Signed-off-by: Andy Lutomirski <luto@kernel.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bpetkov@suse.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Michal Hocko <mhocko@suse.com> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Nadav Amit <namit@vmware.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mm@kvack.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-05-29 01:00:12 +08:00
x86/mm: Don't reenter flush_tlb_func_common() It was historically possible to have two concurrent TLB flushes targetting the same CPU: one initiated locally and one initiated remotely. This can now cause an OOPS in leave_mm() at arch/x86/mm/tlb.c:47: if (this_cpu_read(cpu_tlbstate.state) == TLBSTATE_OK) BUG(); with this call trace: flush_tlb_func_local arch/x86/mm/tlb.c:239 [inline] flush_tlb_mm_range+0x26d/0x370 arch/x86/mm/tlb.c:317 Without reentrancy, this OOPS is impossible: leave_mm() is only called if we're not in TLBSTATE_OK, but then we're unexpectedly in TLBSTATE_OK in leave_mm(). This can be caused by flush_tlb_func_remote() happening between the two checks and calling leave_mm(), resulting in two consecutive leave_mm() calls on the same CPU with no intervening switch_mm() calls. We never saw this OOPS before because the old leave_mm() implementation didn't put us back in TLBSTATE_OK, so the assertion didn't fire. Nadav noticed the reentrancy issue in a different context, but neither of us realized that it caused a problem yet. Reported-by: Levin, Alexander (Sasha Levin) <alexander.levin@verizon.com> Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: linux-mm@kvack.org Fixes: 3d28ebceaffa ("x86/mm: Rework lazy TLB to track the actual loaded mm") Link: http://lkml.kernel.org/r/855acf733268d521c9f2e191faee2dcc23a29729.1498751203.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-06-29 23:53:13 +08:00
if (mm == this_cpu_read(cpu_tlbstate.loaded_mm)) {
VM_WARN_ON(irqs_disabled());
local_irq_disable();
x86/mm: Refactor flush_tlb_mm_range() to merge local and remote cases The local flush path is very similar to the remote flush path. Merge them. This is intended to make no difference to behavior whatsoever. It removes some code and will make future changes to the flushing mechanics simpler. This patch does remove one small optimization: flush_tlb_mm_range() now has an unconditional smp_mb() instead of using MOV to CR3 or INVLPG as a full barrier when applicable. I think this is okay for a few reasons. First, smp_mb() is quite cheap compared to the cost of a TLB flush. Second, this rearrangement makes a bigger optimization available: with some work on the SMP function call code, we could do the local and remote flushes in parallel. Third, I'm planning a rework of the TLB flush algorithm that will require an atomic operation at the beginning of each flush, and that operation will replace the smp_mb(). Signed-off-by: Andy Lutomirski <luto@kernel.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bpetkov@suse.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Michal Hocko <mhocko@suse.com> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Nadav Amit <namit@vmware.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mm@kvack.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-05-29 01:00:12 +08:00
flush_tlb_func_local(&info, TLB_LOCAL_MM_SHOOTDOWN);
x86/mm: Don't reenter flush_tlb_func_common() It was historically possible to have two concurrent TLB flushes targetting the same CPU: one initiated locally and one initiated remotely. This can now cause an OOPS in leave_mm() at arch/x86/mm/tlb.c:47: if (this_cpu_read(cpu_tlbstate.state) == TLBSTATE_OK) BUG(); with this call trace: flush_tlb_func_local arch/x86/mm/tlb.c:239 [inline] flush_tlb_mm_range+0x26d/0x370 arch/x86/mm/tlb.c:317 Without reentrancy, this OOPS is impossible: leave_mm() is only called if we're not in TLBSTATE_OK, but then we're unexpectedly in TLBSTATE_OK in leave_mm(). This can be caused by flush_tlb_func_remote() happening between the two checks and calling leave_mm(), resulting in two consecutive leave_mm() calls on the same CPU with no intervening switch_mm() calls. We never saw this OOPS before because the old leave_mm() implementation didn't put us back in TLBSTATE_OK, so the assertion didn't fire. Nadav noticed the reentrancy issue in a different context, but neither of us realized that it caused a problem yet. Reported-by: Levin, Alexander (Sasha Levin) <alexander.levin@verizon.com> Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: linux-mm@kvack.org Fixes: 3d28ebceaffa ("x86/mm: Rework lazy TLB to track the actual loaded mm") Link: http://lkml.kernel.org/r/855acf733268d521c9f2e191faee2dcc23a29729.1498751203.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-06-29 23:53:13 +08:00
local_irq_enable();
}
x86/mm: Refactor flush_tlb_mm_range() to merge local and remote cases The local flush path is very similar to the remote flush path. Merge them. This is intended to make no difference to behavior whatsoever. It removes some code and will make future changes to the flushing mechanics simpler. This patch does remove one small optimization: flush_tlb_mm_range() now has an unconditional smp_mb() instead of using MOV to CR3 or INVLPG as a full barrier when applicable. I think this is okay for a few reasons. First, smp_mb() is quite cheap compared to the cost of a TLB flush. Second, this rearrangement makes a bigger optimization available: with some work on the SMP function call code, we could do the local and remote flushes in parallel. Third, I'm planning a rework of the TLB flush algorithm that will require an atomic operation at the beginning of each flush, and that operation will replace the smp_mb(). Signed-off-by: Andy Lutomirski <luto@kernel.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bpetkov@suse.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Michal Hocko <mhocko@suse.com> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Nadav Amit <namit@vmware.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mm@kvack.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-05-29 01:00:12 +08:00
if (cpumask_any_but(mm_cpumask(mm), cpu) < nr_cpu_ids)
flush_tlb_others(mm_cpumask(mm), &info);
x86/mm: Rework lazy TLB mode and TLB freshness tracking x86's lazy TLB mode used to be fairly weak -- it would switch to init_mm the first time it tried to flush a lazy TLB. This meant an unnecessary CR3 write and, if the flush was remote, an unnecessary IPI. Rewrite it entirely. When we enter lazy mode, we simply remove the CPU from mm_cpumask. This means that we need a way to figure out whether we've missed a flush when we switch back out of lazy mode. I use the tlb_gen machinery to track whether a context is up to date. Note to reviewers: this patch, my itself, looks a bit odd. I'm using an array of length 1 containing (ctx_id, tlb_gen) rather than just storing tlb_gen, and making it at array isn't necessary yet. I'm doing this because the next few patches add PCID support, and, with PCID, we need ctx_id, and the array will end up with a length greater than 1. Making it an array now means that there will be less churn and therefore less stress on your eyeballs. NB: This is dubious but, AFAICT, still correct on Xen and UV. xen_exit_mmap() uses mm_cpumask() for nefarious purposes and this patch changes the way that mm_cpumask() works. This should be okay, since Xen *also* iterates all online CPUs to find all the CPUs it needs to twiddle. The UV tlbflush code is rather dated and should be changed. Here are some benchmark results, done on a Skylake laptop at 2.3 GHz (turbo off, intel_pstate requesting max performance) under KVM with the guest using idle=poll (to avoid artifacts when bouncing between CPUs). I haven't done any real statistics here -- I just ran them in a loop and picked the fastest results that didn't look like outliers. Unpatched means commit a4eb8b993554, so all the bookkeeping overhead is gone. MADV_DONTNEED; touch the page; switch CPUs using sched_setaffinity. In an unpatched kernel, MADV_DONTNEED will send an IPI to the previous CPU. This is intended to be a nearly worst-case test. patched: 13.4µs unpatched: 21.6µs Vitaly's pthread_mmap microbenchmark with 8 threads (on four cores), nrounds = 100, 256M data patched: 1.1 seconds or so unpatched: 1.9 seconds or so The sleepup on Vitaly's test appearss to be because it spends a lot of time blocked on mmap_sem, and this patch avoids sending IPIs to blocked CPUs. Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andrew Banman <abanman@sgi.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Dimitri Sivanich <sivanich@sgi.com> Cc: Juergen Gross <jgross@suse.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Travis <travis@sgi.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/ddf2c92962339f4ba39d8fc41b853936ec0b44f1.1498751203.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-06-29 23:53:17 +08:00
x86/mm: Refactor flush_tlb_mm_range() to merge local and remote cases The local flush path is very similar to the remote flush path. Merge them. This is intended to make no difference to behavior whatsoever. It removes some code and will make future changes to the flushing mechanics simpler. This patch does remove one small optimization: flush_tlb_mm_range() now has an unconditional smp_mb() instead of using MOV to CR3 or INVLPG as a full barrier when applicable. I think this is okay for a few reasons. First, smp_mb() is quite cheap compared to the cost of a TLB flush. Second, this rearrangement makes a bigger optimization available: with some work on the SMP function call code, we could do the local and remote flushes in parallel. Third, I'm planning a rework of the TLB flush algorithm that will require an atomic operation at the beginning of each flush, and that operation will replace the smp_mb(). Signed-off-by: Andy Lutomirski <luto@kernel.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bpetkov@suse.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Michal Hocko <mhocko@suse.com> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Nadav Amit <namit@vmware.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mm@kvack.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-05-29 01:00:12 +08:00
put_cpu();
}
static void do_flush_tlb_all(void *info)
{
count_vm_tlb_event(NR_TLB_REMOTE_FLUSH_RECEIVED);
__flush_tlb_all();
}
void flush_tlb_all(void)
{
count_vm_tlb_event(NR_TLB_REMOTE_FLUSH);
on_each_cpu(do_flush_tlb_all, NULL, 1);
}
static void do_kernel_range_flush(void *info)
{
struct flush_tlb_info *f = info;
unsigned long addr;
/* flush range by one by one 'invlpg' */
for (addr = f->start; addr < f->end; addr += PAGE_SIZE)
__flush_tlb_one(addr);
}
void flush_tlb_kernel_range(unsigned long start, unsigned long end)
{
/* Balance as user space task's flush, a bit conservative */
x86/mm: Rip out complicated, out-of-date, buggy TLB flushing I think the flush_tlb_mm_range() code that tries to tune the flush sizes based on the CPU needs to get ripped out for several reasons: 1. It is obviously buggy. It uses mm->total_vm to judge the task's footprint in the TLB. It should certainly be using some measure of RSS, *NOT* ->total_vm since only resident memory can populate the TLB. 2. Haswell, and several other CPUs are missing from the intel_tlb_flushall_shift_set() function. Thus, it has been demonstrated to bitrot quickly in practice. 3. It is plain wrong in my vm: [ 0.037444] Last level iTLB entries: 4KB 0, 2MB 0, 4MB 0 [ 0.037444] Last level dTLB entries: 4KB 0, 2MB 0, 4MB 0 [ 0.037444] tlb_flushall_shift: 6 Which leads to it to never use invlpg. 4. The assumptions about TLB refill costs are wrong: http://lkml.kernel.org/r/1337782555-8088-3-git-send-email-alex.shi@intel.com (more on this in later patches) 5. I can not reproduce the original data: https://lkml.org/lkml/2012/5/17/59 I believe the sample times were too short. Running the benchmark in a loop yields times that vary quite a bit. Note that this leaves us with a static ceiling of 1 page. This is a conservative, dumb setting, and will be revised in a later patch. This also removes the code which attempts to predict whether we are flushing data or instructions. We expect instruction flushes to be relatively rare and not worth tuning for explicitly. Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Link: http://lkml.kernel.org/r/20140731154055.ABC88E89@viggo.jf.intel.com Acked-by: Rik van Riel <riel@redhat.com> Acked-by: Mel Gorman <mgorman@suse.de> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2014-07-31 23:40:55 +08:00
if (end == TLB_FLUSH_ALL ||
(end - start) > tlb_single_page_flush_ceiling << PAGE_SHIFT) {
on_each_cpu(do_flush_tlb_all, NULL, 1);
x86/mm: Rip out complicated, out-of-date, buggy TLB flushing I think the flush_tlb_mm_range() code that tries to tune the flush sizes based on the CPU needs to get ripped out for several reasons: 1. It is obviously buggy. It uses mm->total_vm to judge the task's footprint in the TLB. It should certainly be using some measure of RSS, *NOT* ->total_vm since only resident memory can populate the TLB. 2. Haswell, and several other CPUs are missing from the intel_tlb_flushall_shift_set() function. Thus, it has been demonstrated to bitrot quickly in practice. 3. It is plain wrong in my vm: [ 0.037444] Last level iTLB entries: 4KB 0, 2MB 0, 4MB 0 [ 0.037444] Last level dTLB entries: 4KB 0, 2MB 0, 4MB 0 [ 0.037444] tlb_flushall_shift: 6 Which leads to it to never use invlpg. 4. The assumptions about TLB refill costs are wrong: http://lkml.kernel.org/r/1337782555-8088-3-git-send-email-alex.shi@intel.com (more on this in later patches) 5. I can not reproduce the original data: https://lkml.org/lkml/2012/5/17/59 I believe the sample times were too short. Running the benchmark in a loop yields times that vary quite a bit. Note that this leaves us with a static ceiling of 1 page. This is a conservative, dumb setting, and will be revised in a later patch. This also removes the code which attempts to predict whether we are flushing data or instructions. We expect instruction flushes to be relatively rare and not worth tuning for explicitly. Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Link: http://lkml.kernel.org/r/20140731154055.ABC88E89@viggo.jf.intel.com Acked-by: Rik van Riel <riel@redhat.com> Acked-by: Mel Gorman <mgorman@suse.de> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
2014-07-31 23:40:55 +08:00
} else {
struct flush_tlb_info info;
info.start = start;
info.end = end;
on_each_cpu(do_kernel_range_flush, &info, 1);
}
}
mm, x86/mm: Make the batched unmap TLB flush API more generic try_to_unmap_flush() used to open-code a rather x86-centric flush sequence: local_flush_tlb() + flush_tlb_others(). Rearrange the code so that the arch (only x86 for now) provides arch_tlbbatch_add_mm() and arch_tlbbatch_flush() and the core code calls those functions instead. I'll want this for x86 because, to enable address space ids, I can't support the flush_tlb_others() mode used by exising try_to_unmap_flush() implementation with good performance. I can support the new API fairly easily, though. I imagine that other architectures may be in a similar position. Architectures with strong remote flush primitives (arm64?) may have even worse performance problems with flush_tlb_others() the way that try_to_unmap_flush() uses it. Signed-off-by: Andy Lutomirski <luto@kernel.org> Acked-by: Kees Cook <keescook@chromium.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Borislav Petkov <bpetkov@suse.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Michal Hocko <mhocko@suse.com> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Nadav Amit <namit@vmware.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Sasha Levin <sasha.levin@oracle.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/19f25a8581f9fb77876b7ff3b001f89835e34ea3.1495492063.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-05-23 06:30:03 +08:00
void arch_tlbbatch_flush(struct arch_tlbflush_unmap_batch *batch)
{
struct flush_tlb_info info = {
.mm = NULL,
.start = 0UL,
.end = TLB_FLUSH_ALL,
};
mm, x86/mm: Make the batched unmap TLB flush API more generic try_to_unmap_flush() used to open-code a rather x86-centric flush sequence: local_flush_tlb() + flush_tlb_others(). Rearrange the code so that the arch (only x86 for now) provides arch_tlbbatch_add_mm() and arch_tlbbatch_flush() and the core code calls those functions instead. I'll want this for x86 because, to enable address space ids, I can't support the flush_tlb_others() mode used by exising try_to_unmap_flush() implementation with good performance. I can support the new API fairly easily, though. I imagine that other architectures may be in a similar position. Architectures with strong remote flush primitives (arm64?) may have even worse performance problems with flush_tlb_others() the way that try_to_unmap_flush() uses it. Signed-off-by: Andy Lutomirski <luto@kernel.org> Acked-by: Kees Cook <keescook@chromium.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Borislav Petkov <bpetkov@suse.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Michal Hocko <mhocko@suse.com> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Nadav Amit <namit@vmware.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Sasha Levin <sasha.levin@oracle.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/19f25a8581f9fb77876b7ff3b001f89835e34ea3.1495492063.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-05-23 06:30:03 +08:00
int cpu = get_cpu();
x86/mm: Don't reenter flush_tlb_func_common() It was historically possible to have two concurrent TLB flushes targetting the same CPU: one initiated locally and one initiated remotely. This can now cause an OOPS in leave_mm() at arch/x86/mm/tlb.c:47: if (this_cpu_read(cpu_tlbstate.state) == TLBSTATE_OK) BUG(); with this call trace: flush_tlb_func_local arch/x86/mm/tlb.c:239 [inline] flush_tlb_mm_range+0x26d/0x370 arch/x86/mm/tlb.c:317 Without reentrancy, this OOPS is impossible: leave_mm() is only called if we're not in TLBSTATE_OK, but then we're unexpectedly in TLBSTATE_OK in leave_mm(). This can be caused by flush_tlb_func_remote() happening between the two checks and calling leave_mm(), resulting in two consecutive leave_mm() calls on the same CPU with no intervening switch_mm() calls. We never saw this OOPS before because the old leave_mm() implementation didn't put us back in TLBSTATE_OK, so the assertion didn't fire. Nadav noticed the reentrancy issue in a different context, but neither of us realized that it caused a problem yet. Reported-by: Levin, Alexander (Sasha Levin) <alexander.levin@verizon.com> Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: linux-mm@kvack.org Fixes: 3d28ebceaffa ("x86/mm: Rework lazy TLB to track the actual loaded mm") Link: http://lkml.kernel.org/r/855acf733268d521c9f2e191faee2dcc23a29729.1498751203.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-06-29 23:53:13 +08:00
if (cpumask_test_cpu(cpu, &batch->cpumask)) {
VM_WARN_ON(irqs_disabled());
local_irq_disable();
flush_tlb_func_local(&info, TLB_LOCAL_SHOOTDOWN);
x86/mm: Don't reenter flush_tlb_func_common() It was historically possible to have two concurrent TLB flushes targetting the same CPU: one initiated locally and one initiated remotely. This can now cause an OOPS in leave_mm() at arch/x86/mm/tlb.c:47: if (this_cpu_read(cpu_tlbstate.state) == TLBSTATE_OK) BUG(); with this call trace: flush_tlb_func_local arch/x86/mm/tlb.c:239 [inline] flush_tlb_mm_range+0x26d/0x370 arch/x86/mm/tlb.c:317 Without reentrancy, this OOPS is impossible: leave_mm() is only called if we're not in TLBSTATE_OK, but then we're unexpectedly in TLBSTATE_OK in leave_mm(). This can be caused by flush_tlb_func_remote() happening between the two checks and calling leave_mm(), resulting in two consecutive leave_mm() calls on the same CPU with no intervening switch_mm() calls. We never saw this OOPS before because the old leave_mm() implementation didn't put us back in TLBSTATE_OK, so the assertion didn't fire. Nadav noticed the reentrancy issue in a different context, but neither of us realized that it caused a problem yet. Reported-by: Levin, Alexander (Sasha Levin) <alexander.levin@verizon.com> Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: linux-mm@kvack.org Fixes: 3d28ebceaffa ("x86/mm: Rework lazy TLB to track the actual loaded mm") Link: http://lkml.kernel.org/r/855acf733268d521c9f2e191faee2dcc23a29729.1498751203.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-06-29 23:53:13 +08:00
local_irq_enable();
}
mm, x86/mm: Make the batched unmap TLB flush API more generic try_to_unmap_flush() used to open-code a rather x86-centric flush sequence: local_flush_tlb() + flush_tlb_others(). Rearrange the code so that the arch (only x86 for now) provides arch_tlbbatch_add_mm() and arch_tlbbatch_flush() and the core code calls those functions instead. I'll want this for x86 because, to enable address space ids, I can't support the flush_tlb_others() mode used by exising try_to_unmap_flush() implementation with good performance. I can support the new API fairly easily, though. I imagine that other architectures may be in a similar position. Architectures with strong remote flush primitives (arm64?) may have even worse performance problems with flush_tlb_others() the way that try_to_unmap_flush() uses it. Signed-off-by: Andy Lutomirski <luto@kernel.org> Acked-by: Kees Cook <keescook@chromium.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Borislav Petkov <bpetkov@suse.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Michal Hocko <mhocko@suse.com> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Nadav Amit <namit@vmware.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Sasha Levin <sasha.levin@oracle.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/19f25a8581f9fb77876b7ff3b001f89835e34ea3.1495492063.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-05-23 06:30:03 +08:00
if (cpumask_any_but(&batch->cpumask, cpu) < nr_cpu_ids)
flush_tlb_others(&batch->cpumask, &info);
x86/mm: Rework lazy TLB mode and TLB freshness tracking x86's lazy TLB mode used to be fairly weak -- it would switch to init_mm the first time it tried to flush a lazy TLB. This meant an unnecessary CR3 write and, if the flush was remote, an unnecessary IPI. Rewrite it entirely. When we enter lazy mode, we simply remove the CPU from mm_cpumask. This means that we need a way to figure out whether we've missed a flush when we switch back out of lazy mode. I use the tlb_gen machinery to track whether a context is up to date. Note to reviewers: this patch, my itself, looks a bit odd. I'm using an array of length 1 containing (ctx_id, tlb_gen) rather than just storing tlb_gen, and making it at array isn't necessary yet. I'm doing this because the next few patches add PCID support, and, with PCID, we need ctx_id, and the array will end up with a length greater than 1. Making it an array now means that there will be less churn and therefore less stress on your eyeballs. NB: This is dubious but, AFAICT, still correct on Xen and UV. xen_exit_mmap() uses mm_cpumask() for nefarious purposes and this patch changes the way that mm_cpumask() works. This should be okay, since Xen *also* iterates all online CPUs to find all the CPUs it needs to twiddle. The UV tlbflush code is rather dated and should be changed. Here are some benchmark results, done on a Skylake laptop at 2.3 GHz (turbo off, intel_pstate requesting max performance) under KVM with the guest using idle=poll (to avoid artifacts when bouncing between CPUs). I haven't done any real statistics here -- I just ran them in a loop and picked the fastest results that didn't look like outliers. Unpatched means commit a4eb8b993554, so all the bookkeeping overhead is gone. MADV_DONTNEED; touch the page; switch CPUs using sched_setaffinity. In an unpatched kernel, MADV_DONTNEED will send an IPI to the previous CPU. This is intended to be a nearly worst-case test. patched: 13.4µs unpatched: 21.6µs Vitaly's pthread_mmap microbenchmark with 8 threads (on four cores), nrounds = 100, 256M data patched: 1.1 seconds or so unpatched: 1.9 seconds or so The sleepup on Vitaly's test appearss to be because it spends a lot of time blocked on mmap_sem, and this patch avoids sending IPIs to blocked CPUs. Signed-off-by: Andy Lutomirski <luto@kernel.org> Reviewed-by: Nadav Amit <nadav.amit@gmail.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Cc: Andrew Banman <abanman@sgi.com> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Arjan van de Ven <arjan@linux.intel.com> Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com> Cc: Borislav Petkov <bp@alien8.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Dimitri Sivanich <sivanich@sgi.com> Cc: Juergen Gross <jgross@suse.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Mike Travis <travis@sgi.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/ddf2c92962339f4ba39d8fc41b853936ec0b44f1.1498751203.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-06-29 23:53:17 +08:00
mm, x86/mm: Make the batched unmap TLB flush API more generic try_to_unmap_flush() used to open-code a rather x86-centric flush sequence: local_flush_tlb() + flush_tlb_others(). Rearrange the code so that the arch (only x86 for now) provides arch_tlbbatch_add_mm() and arch_tlbbatch_flush() and the core code calls those functions instead. I'll want this for x86 because, to enable address space ids, I can't support the flush_tlb_others() mode used by exising try_to_unmap_flush() implementation with good performance. I can support the new API fairly easily, though. I imagine that other architectures may be in a similar position. Architectures with strong remote flush primitives (arm64?) may have even worse performance problems with flush_tlb_others() the way that try_to_unmap_flush() uses it. Signed-off-by: Andy Lutomirski <luto@kernel.org> Acked-by: Kees Cook <keescook@chromium.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Borislav Petkov <bpetkov@suse.de> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Michal Hocko <mhocko@suse.com> Cc: Nadav Amit <nadav.amit@gmail.com> Cc: Nadav Amit <namit@vmware.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Sasha Levin <sasha.levin@oracle.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: linux-mm@kvack.org Link: http://lkml.kernel.org/r/19f25a8581f9fb77876b7ff3b001f89835e34ea3.1495492063.git.luto@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-05-23 06:30:03 +08:00
cpumask_clear(&batch->cpumask);
put_cpu();
}
static ssize_t tlbflush_read_file(struct file *file, char __user *user_buf,
size_t count, loff_t *ppos)
{
char buf[32];
unsigned int len;
len = sprintf(buf, "%ld\n", tlb_single_page_flush_ceiling);
return simple_read_from_buffer(user_buf, count, ppos, buf, len);
}
static ssize_t tlbflush_write_file(struct file *file,
const char __user *user_buf, size_t count, loff_t *ppos)
{
char buf[32];
ssize_t len;
int ceiling;
len = min(count, sizeof(buf) - 1);
if (copy_from_user(buf, user_buf, len))
return -EFAULT;
buf[len] = '\0';
if (kstrtoint(buf, 0, &ceiling))
return -EINVAL;
if (ceiling < 0)
return -EINVAL;
tlb_single_page_flush_ceiling = ceiling;
return count;
}
static const struct file_operations fops_tlbflush = {
.read = tlbflush_read_file,
.write = tlbflush_write_file,
.llseek = default_llseek,
};
static int __init create_tlb_single_page_flush_ceiling(void)
{
debugfs_create_file("tlb_single_page_flush_ceiling", S_IRUSR | S_IWUSR,
arch_debugfs_dir, NULL, &fops_tlbflush);
return 0;
}
late_initcall(create_tlb_single_page_flush_ceiling);