get_cpu_ptr() disabled preemption and returns the ->fq object of the
current CPU. raw_cpu_ptr() does the same except that it not disable
preemption which means the scheduler can move it to another CPU after it
obtained the per-CPU object.
In this case this is not bad because the data structure itself is
protected with a spin_lock. This change shouldn't matter however on RT
it does because the sleeping lock can't be accessed with disabled
preemption.
Cc: Joerg Roedel <joro@8bytes.org>
Cc: iommu@lists.linux-foundation.org
Reported-by: vinadhy@gmail.com
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
Since IOVA allocation failure is not unusual case we need to flush
CPUs' rcache in hope we will succeed in next round.
However, it is useful to decide whether we need rcache flush step because
of two reasons:
- Not scalability. On large system with ~100 CPUs iterating and flushing
rcache for each CPU becomes serious bottleneck so we may want to defer it.
- free_cpu_cached_iovas() does not care about max PFN we are interested in.
Thus we may flush our rcaches and still get no new IOVA like in the
commonly used scenario:
if (dma_limit > DMA_BIT_MASK(32) && dev_is_pci(dev))
iova = alloc_iova_fast(iovad, iova_len, DMA_BIT_MASK(32) >> shift);
if (!iova)
iova = alloc_iova_fast(iovad, iova_len, dma_limit >> shift);
1. First alloc_iova_fast() call is limited to DMA_BIT_MASK(32) to get
PCI devices a SAC address
2. alloc_iova() fails due to full 32-bit space
3. rcaches contain PFNs out of 32-bit space so free_cpu_cached_iovas()
throws entries away for nothing and alloc_iova() fails again
4. Next alloc_iova_fast() call cannot take advantage of rcache since we
have just defeated caches. In this case we pick the slowest option
to proceed.
This patch reworks flushed_rcache local flag to be additional function
argument instead and control rcache flush step. Also, it updates all users
to do the flush as the last chance.
Signed-off-by: Tomasz Nowicki <Tomasz.Nowicki@caviumnetworks.com>
Reviewed-by: Robin Murphy <robin.murphy@arm.com>
Tested-by: Nate Watterson <nwatters@codeaurora.org>
Signed-off-by: Joerg Roedel <jroedel@suse.de>
Anchor nodes are not reserved IOVAs in the way that copy_reserved_iova()
cares about - while the failure from reserve_iova() is benign since the
target domain will already have its own anchor, we still don't want to
be triggering spurious warnings.
Reported-by: kernel test robot <fengguang.wu@intel.com>
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Fixes: bb68b2fbfb ('iommu/iova: Add rbtree anchor node')
Signed-off-by: Joerg Roedel <jroedel@suse.de>
When devices with different DMA masks are using the same domain, or for
PCI devices where we usually try a speculative 32-bit allocation first,
there is a fair possibility that the top PFN of the rcache stack at any
given time may be unsuitable for the lower limit, prompting a fallback
to allocating anew from the rbtree. Consequently, we may end up
artifically increasing pressure on the 32-bit IOVA space as unused IOVAs
accumulate lower down in the rcache stacks, while callers with 32-bit
masks also impose unnecessary rbtree overhead.
In such cases, let's try a bit harder to satisfy the allocation locally
first - scanning the whole stack should still be relatively inexpensive.
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Signed-off-by: Joerg Roedel <jroedel@suse.de>
When popping a pfn from an rcache, we are currently checking it directly
against limit_pfn for viability. Since this represents iova->pfn_lo, it
is technically possible for the corresponding iova->pfn_hi to be greater
than limit_pfn. Although we generally get away with it in practice since
limit_pfn is typically a power-of-two boundary and the IOVAs are
size-aligned, it's pretty trivial to make the iova_rcache_get() path
take the allocation size into account for complete safety.
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Signed-off-by: Joerg Roedel <jroedel@suse.de>
All put_iova_domain() should have to worry about is freeing memory - by
that point the domain must no longer be live, so the act of cleaning up
doesn't need to be concurrency-safe or maintain the rbtree in a
self-consistent state. There's no need to waste time with locking or
emptying the rcache magazines, and we can just use the postorder
traversal helper to clear out the remaining rbtree entries in-place.
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Signed-off-by: Joerg Roedel <jroedel@suse.de>
The logic of __get_cached_rbnode() is a little obtuse, but then
__get_prev_node_of_cached_rbnode_or_last_node_and_update_limit_pfn()
wouldn't exactly roll off the tongue...
Now that we have the invariant that there is always a valid node to
start searching downwards from, everything gets a bit easier to follow
if we simplify that function to do what it says on the tin and return
the cached node (or anchor node as appropriate) directly. In turn, we
can then deduplicate the rb_prev() and limit_pfn logic into the main
loop itself, further reduce the amount of code under the lock, and
generally make the inner workings a bit less subtle.
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Signed-off-by: Joerg Roedel <jroedel@suse.de>
Add a permanent dummy IOVA reservation to the rbtree, such that we can
always access the top of the address space instantly. The immediate
benefit is that we remove the overhead of the rb_last() traversal when
not using the cached node, but it also paves the way for further
simplifications.
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Signed-off-by: Joerg Roedel <jroedel@suse.de>
Now that the cached node optimisation can apply to all allocations, the
couple of users which were playing tricks with dma_32bit_pfn in order to
benefit from it can stop doing so. Conversely, there is also no need for
all the other users to explicitly calculate a 'real' 32-bit PFN, when
init_iova_domain() can happily do that itself from the page granularity.
CC: Thierry Reding <thierry.reding@gmail.com>
CC: Jonathan Hunter <jonathanh@nvidia.com>
CC: David Airlie <airlied@linux.ie>
CC: Sudeep Dutt <sudeep.dutt@intel.com>
CC: Ashutosh Dixit <ashutosh.dixit@intel.com>
Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com>
Tested-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Tested-by: Zhen Lei <thunder.leizhen@huawei.com>
Tested-by: Nate Watterson <nwatters@codeaurora.org>
[rm: use iova_shift(), rewrote commit message]
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Signed-off-by: Joerg Roedel <jroedel@suse.de>
The cached node mechanism provides a significant performance benefit for
allocations using a 32-bit DMA mask, but in the case of non-PCI devices
or where the 32-bit space is full, the loss of this benefit can be
significant - on large systems there can be many thousands of entries in
the tree, such that walking all the way down to find free space every
time becomes increasingly awful.
Maintain a similar cached node for the whole IOVA space as a superset of
the 32-bit space so that performance can remain much more consistent.
Inspired by work by Zhen Lei <thunder.leizhen@huawei.com>.
Tested-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Tested-by: Zhen Lei <thunder.leizhen@huawei.com>
Tested-by: Nate Watterson <nwatters@codeaurora.org>
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Signed-off-by: Joerg Roedel <jroedel@suse.de>
The mask for calculating the padding size doesn't change, so there's no
need to recalculate it every loop iteration. Furthermore, Once we've
done that, it becomes clear that we don't actually need to calculate a
padding size at all - by flipping the arithmetic around, we can just
combine the upper limit, size, and mask directly to check against the
lower limit.
For an arm64 build, this alone knocks 20% off the object code size of
the entire alloc_iova() function!
Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com>
Tested-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Tested-by: Zhen Lei <thunder.leizhen@huawei.com>
Tested-by: Nate Watterson <nwatters@codeaurora.org>
[rm: simplified more of the arithmetic, rewrote commit message]
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Signed-off-by: Joerg Roedel <jroedel@suse.de>
Checking the IOVA bounds separately before deciding which direction to
continue the search (if necessary) results in redundantly comparing both
pfns twice each. GCC can already determine that the final comparison op
is redundant and optimise it down to 3 in total, but we can go one
further with a little tweak of the ordering (which makes the intent of
the code that much cleaner as a bonus).
Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com>
Tested-by: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Tested-by: Zhen Lei <thunder.leizhen@huawei.com>
Tested-by: Nate Watterson <nwatters@codeaurora.org>
[rm: rewrote commit message to clarify]
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Signed-off-by: Joerg Roedel <jroedel@suse.de>
Add a timer to flush entries from the Flush-Queues every
10ms. This makes sure that no stale TLB entries remain for
too long after an IOVA has been unmapped.
Signed-off-by: Joerg Roedel <jroedel@suse.de>
The lock is taken from the same CPU most of the time. But
having it allows to flush the queue also from another CPU if
necessary.
This will be used by a timer to regularily flush any pending
IOVAs from the Flush-Queues.
Signed-off-by: Joerg Roedel <jroedel@suse.de>
There are two counters:
* fq_flush_start_cnt - Increased when a TLB flush
is started.
* fq_flush_finish_cnt - Increased when a TLB flush
is finished.
The fq_flush_start_cnt is assigned to every Flush-Queue
entry on its creation. When freeing entries from the
Flush-Queue, the value in the entry is compared to the
fq_flush_finish_cnt. The entry can only be freed when its
value is less than the value of fq_flush_finish_cnt.
The reason for these counters it to take advantage of IOMMU
TLB flushes that happened on other CPUs. These already
flushed the TLB for Flush-Queue entries on other CPUs so
that they can already be freed without flushing the TLB
again.
This makes it less likely that the Flush-Queue is full and
saves IOMMU TLB flushes.
Signed-off-by: Joerg Roedel <jroedel@suse.de>
Add a function to add entries to the Flush-Queue ring
buffer. If the buffer is full, call the flush-callback and
free the entries.
Signed-off-by: Joerg Roedel <jroedel@suse.de>
This patch adds the basic data-structures to implement
flush-queues in the generic IOVA code. It also adds the
initialization and destroy routines for these data
structures.
The initialization routine is designed so that the use of
this feature is optional for the users of IOVA code.
Signed-off-by: Joerg Roedel <jroedel@suse.de>
Commit 583248e662 ("iommu/iova: Disable preemption around use of
this_cpu_ptr()") disables preemption while accessing a per-CPU variable.
This does keep lockdep quiet. However I don't see the point why it is
bad if we get migrated after its access to another CPU.
__iova_rcache_insert() and __iova_rcache_get() immediately locks the
variable after obtaining it - before accessing its members.
_If_ we get migrated away after retrieving the address of cpu_rcache
before taking the lock then the *other* task on the same CPU will
retrieve the same address of cpu_rcache and will spin on the lock.
alloc_iova_fast() disables preemption while invoking
free_cpu_cached_iovas() on each CPU. The function itself uses
per_cpu_ptr() which does not trigger a warning (like this_cpu_ptr()
does). It _could_ make sense to use get_online_cpus() instead but the we
have a hotplug notifier for CPU down (and none for up) so we are good.
Cc: Joerg Roedel <joro@8bytes.org>
Cc: iommu@lists.linux-foundation.org
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Signed-off-by: Joerg Roedel <jroedel@suse.de>
When walking the rbtree, the fact that iovad->start_pfn and limit_pfn
are both inclusive limits creates an ambiguity once limit_pfn reaches
the bottom of the address space and they overlap. Commit 5016bdb796
("iommu/iova: Fix underflow bug in __alloc_and_insert_iova_range") fixed
the worst side-effect of this, that of underflow wraparound leading to
bogus allocations, but the remaining fallout is that any attempt to
allocate start_pfn itself erroneously fails.
The cleanest way to resolve the ambiguity is to simply make limit_pfn an
exclusive limit when inside the guts of the rbtree. Since we're working
with PFNs, representing one past the top of the address space is always
possible without fear of overflow, and elsewhere it just makes life a
little more straightforward.
Reported-by: Aaron Sierra <asierra@xes-inc.com>
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Signed-off-by: Joerg Roedel <jroedel@suse.de>
Normally, calling alloc_iova() using an iova_domain with insufficient
pfns remaining between start_pfn and dma_limit will fail and return a
NULL pointer. Unexpectedly, if such a "full" iova_domain contains an
iova with pfn_lo == 0, the alloc_iova() call will instead succeed and
return an iova containing invalid pfns.
This is caused by an underflow bug in __alloc_and_insert_iova_range()
that occurs after walking the "full" iova tree when the search ends
at the iova with pfn_lo == 0 and limit_pfn is then adjusted to be just
below that (-1). This (now huge) limit_pfn gives the impression that a
vast amount of space is available between it and start_pfn and thus
a new iova is allocated with the invalid pfn_hi value, 0xFFF.... .
To rememdy this, a check is introduced to ensure that adjustments to
limit_pfn will not underflow.
This issue has been observed in the wild, and is easily reproduced with
the following sample code.
struct iova_domain *iovad = kzalloc(sizeof(*iovad), GFP_KERNEL);
struct iova *rsvd_iova, *good_iova, *bad_iova;
unsigned long limit_pfn = 3;
unsigned long start_pfn = 1;
unsigned long va_size = 2;
init_iova_domain(iovad, SZ_4K, start_pfn, limit_pfn);
rsvd_iova = reserve_iova(iovad, 0, 0);
good_iova = alloc_iova(iovad, va_size, limit_pfn, true);
bad_iova = alloc_iova(iovad, va_size, limit_pfn, true);
Prior to the patch, this yielded:
*rsvd_iova == {0, 0} /* Expected */
*good_iova == {2, 3} /* Expected */
*bad_iova == {-2, -1} /* Oh no... */
After the patch, bad_iova is NULL as expected since inadequate
space remains between limit_pfn and start_pfn after allocating
good_iova.
Signed-off-by: Nate Watterson <nwatters@codeaurora.org>
Signed-off-by: Joerg Roedel <jroedel@suse.de>
This patch consolidates almost the same code used in iova_insert_rbtree()
and __alloc_and_insert_iova_range() functions. While touching this code,
replace BUG() with WARN_ON(1) to avoid taking down the whole system in
case of corrupted iova tree or incorrect calls.
Signed-off-by: Marek Szyprowski <m.szyprowski@samsung.com>
Signed-off-by: Joerg Roedel <jroedel@suse.de>
To make the code clearer, use rb_entry() instead of container_of() to
deal with rbtree.
Signed-off-by: Geliang Tang <geliangtang@gmail.com>
Signed-off-by: Joerg Roedel <jroedel@suse.de>
When searching for a free IOVA range, we optimise the tree traversal
by starting from the cached32_node, instead of the last node, when
limit_pfn is equal to dma_32bit_pfn. However, if limit_pfn happens to
be smaller, then we'll go ahead and start from the top even though
dma_32bit_pfn is still a more suitable upper bound. Since this is
clearly a silly thing to do, adjust the lookup condition appropriately.
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Signed-off-by: Joerg Roedel <jroedel@suse.de>
IOVA allocation has two problems that impede high-throughput I/O.
First, it can do a linear search over the allocated IOVA ranges.
Second, the rbtree spinlock that serializes IOVA allocations becomes
contended.
Address these problems by creating an API for caching allocated IOVA
ranges, so that the IOVA allocator isn't accessed frequently. This
patch adds a per-CPU cache, from which CPUs can alloc/free IOVAs
without taking the rbtree spinlock. The per-CPU caches are backed by
a global cache, to avoid invoking the (linear-time) IOVA allocator
without needing to make the per-CPU cache size excessive. This design
is based on magazines, as described in "Magazines and Vmem: Extending
the Slab Allocator to Many CPUs and Arbitrary Resources" (currently
available at https://www.usenix.org/legacy/event/usenix01/bonwick.html)
Adding caching on top of the existing rbtree allocator maintains the
property that IOVAs are densely packed in the IO virtual address space,
which is important for keeping IOMMU page table usage low.
To keep the cache size reasonable, we bound the IOVA space a CPU can
cache by 32 MiB (we cache a bounded number of IOVA ranges, and only
ranges of size <= 128 KiB). The shared global cache is bounded at
4 MiB of IOVA space.
Signed-off-by: Omer Peleg <omer@cs.technion.ac.il>
[mad@cs.technion.ac.il: rebased, cleaned up and reworded the commit message]
Signed-off-by: Adam Morrison <mad@cs.technion.ac.il>
Reviewed-by: Shaohua Li <shli@fb.com>
Reviewed-by: Ben Serebrin <serebrin@google.com>
[dwmw2: split out VT-d part into a separate patch]
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
The iova library has use outside the intel-iommu driver, thus make it a
module.
Signed-off-by: Sakari Ailus <sakari.ailus@linux.intel.com>
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
Use EXPORT_SYMBOL_GPL() to export the iova library symbols. The symbols
include:
init_iova_domain();
iova_cache_get();
iova_cache_put();
iova_cache_init();
alloc_iova();
find_iova();
__free_iova();
free_iova();
put_iova_domain();
reserve_iova();
copy_reserved_iova();
Signed-off-by: Sakari Ailus <sakari.ailus@linux.intel.com>
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
This is necessary to separate intel-iommu from the iova library.
Signed-off-by: Sakari Ailus <sakari.ailus@linux.intel.com>
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
Currently, allocating a size-aligned IOVA region quietly adjusts the
actual allocation size in the process, returning a rounded-up
power-of-two-sized allocation. This results in mismatched behaviour in
the IOMMU driver if the original size was not a power of two, where the
original size is mapped, but the rounded-up IOVA size is unmapped.
Whilst some IOMMUs will happily unmap already-unmapped pages, others
consider this an error, so fix it by computing the necessary alignment
padding without altering the actual allocation size. Also clean up by
making pad_size unsigned, since its callers always pass unsigned values
and negative padding makes little sense here anyway.
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Signed-off-by: David Woodhouse <David.Woodhouse@intel.com>
Fixed checkpatch warnings for missing blank line after
declaration of struct.
Signed-off-by: Robert Callicotte <rcallicotte@gmail.com>
Signed-off-by: Joerg Roedel <jroedel@suse.de>
Systems may contain heterogeneous IOMMUs supporting differing minimum
page sizes, which may also not be common with the CPU page size.
Thus it is practical to have an explicit notion of IOVA granularity
to simplify handling of mapping and allocation constraints.
As an initial step, move the IOVA page granularity from an implicit
compile-time constant to a per-domain property so we can make use
of it in IOVA domain context at runtime. To keep the abstraction tidy,
extend the little API of inline iova_* helpers to parallel some of the
equivalent PAGE_* macros.
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Signed-off-by: Joerg Roedel <jroedel@suse.de>
To share the IOVA allocator with other architectures, it needs to
accommodate more general aperture restrictions; move the lower limit
from a compile-time constant to a runtime domain property to allow
IOVA domains with different requirements to co-exist.
Also reword the slightly unclear description of alloc_iova since we're
touching it anyway.
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Signed-off-by: Joerg Roedel <jroedel@suse.de>
In order to share the IOVA allocator with other architectures, break
the unnecssary dependency on the Intel IOMMU driver and move the
remaining IOVA internals to iova.c
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
Signed-off-by: Joerg Roedel <jroedel@suse.de>
This should ease finding similarities with different platforms,
with the intention of solving problems once in a generic framework
which everyone can use.
Note: to move intel-iommu.c, the declaration of pci_find_upstream_pcie_bridge()
has to move from drivers/pci/pci.h to include/linux/pci.h. This is handled
in this patch, too.
As suggested, also drop DMAR's EXPERIMENTAL tag while we're at it.
Compile-tested on x86_64.
Signed-off-by: Ohad Ben-Cohen <ohad@wizery.com>
Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
