Save non volatile registers, XER, CTR, MSR and NIP in exception prolog.
Also assign proper value to r2 and r3 there.
For now, recalculate thread pointer in prepare_transfer_to_handler.
It will disappear once KUAP is ported to C.
And remove the comment which is now completely wrong.
Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/56f0cde9dd0362edf2ddba4d887552013eee7329.1615552867.git.christophe.leroy@csgroup.eu
Exception prologs all do the same at the end:
- Save trapno in stack
- Mark stack with exception marker
- Save r0
- Save r3 to r8
Refactor that into a COMMON_EXCEPTION_PROLOG_END macro.
At the same time use r1 instead of r11.
Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/e1c45d2e895e0693c42d2a6840df1105a148efea.1615552867.git.christophe.leroy@csgroup.eu
In order to get more control in exception prolog, dismantle
all non standard exception macros, finishing with EXC_XFER_STD
and EXC_XFER_LITE and EXC_XFER_TEMPLATE.
Also remove transfer_to_handler_full and ret_from_except and
ret_from_except_full as they are not used anymore.
Last parameter of EXCEPTION() is now ignored, will be removed
in a later patch to avoid too much churn.
Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/ca5795d04a220586b7037dbbbe6951dfa9e768eb.1615552867.git.christophe.leroy@csgroup.eu
Until now, non volatile registers were restored everytime they
were saved, ie using EXC_XFER_STD meant saving and restoring
them while EXC_XFER_LITE meant neither saving not restoring them.
Now that they are always saved, EXC_XFER_STD means to restore
them and EXC_XFER_LITE means to not restore them.
Most of the users of EXC_XFER_STD only need to retrieve the
non volatile registers. For them there is no need to restore
the non volatile registers as they have not been modified.
Only very few exceptions require non volatile registers restore.
Opencode the few places which require saving of non volatile
registers.
Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/d1cb12d8023cc6afc1f07150565571373c04945c.1615552867.git.christophe.leroy@csgroup.eu
In order to increase flexibility, add a macro that will for now
call transfer_to_handler.
As transfer_to_handler doesn't do the actual transfer anymore,
also name it prepare_transfer_to_handler. The following patches
will progressively remove the use of transfer_to_handler label.
Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/7f757c52518ab1d7b27ad5113b10f860e803f467.1615552867.git.christophe.leroy@csgroup.eu
Since commit 06d67d5474 ("powerpc: make process.c suitable for both
32-bit and 64-bit"), thread.regs is set on task creation, no need to
set it again and again at each interrupt entry as it never change.
Suggested-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20d52c627303d63e461797df13e6890fc04017d0.1615552867.git.christophe.leroy@csgroup.eu
Now that the MMU is re-enabled before calling the transfer function,
we don't need anymore that hack with the address of the handler and
the return function sitting just after the 'bl' to the transfer
fonction, that function is retrieving via a read relative to 'lr'.
Do a regular call to the transfer function, then to the handler,
then branch to the return function.
Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/73c00f3361ca280ef8fd7814c291bd1f5b6e2081.1615552867.git.christophe.leroy@csgroup.eu
Refactor booke critical registers saving into a few macros
and move it into the exception prolog directly.
Keep the dedicated transfert_to_handler entry point for the
moment allthough they are empty. They will be removed in a
later patch to reduce churn.
Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/269171496f1f5f22afa621695bded22976c9d48d.1615552867.git.christophe.leroy@csgroup.eu
The space in the head section is rather constrained by the fact that
exception vectors are spread every 0x100 bytes and sometimes we
need to have "out of line" code because it doesn't fit.
Now that we are enabling MMU early in the prolog, take that opportunity
to jump somewhere else in the .text section where we don't have any
space constraint.
Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/38b31ca4bc782a4985bc7952a675404d7ff27c24.1615552867.git.christophe.leroy@csgroup.eu
For consistency, add in the macro the label used by exception prolog
to branch to stack overflow processing.
While at it, enclose the macro in #ifdef CONFIG_VMAP_STACK on the 8xx
as already done on book3s/32.
Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/cf80056f5b946572ad98aea9d915dd25b23beda6.1615552867.git.christophe.leroy@csgroup.eu
The check of the emergency context initialisation in
vmap_stack_overflow is buggy for the SMP case, as it
compares r1 with 0 while in the SMP case r1 is offseted
by the CPU id.
Instead of fixing it, just perform static initialisation
of the first emergency context.
Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/4a67ba422be75713286dca0c86ee0d3df2eb6dfa.1615552867.git.christophe.leroy@csgroup.eu
On 40x and 8xx, kernel text is pinned.
On book3s/32, kernel text is mapped by BATs.
Enable instruction translation at the same time as data translation, it
makes things simpler.
In syscall handler, MSR_RI can also be set at the same time because
srr0/srr1 are already saved and r1 is set properly.
On booke, translation is always on, so at the end all PPC32
have translation on early. Just update msr.
Also update comment in power_save_ppc32_restore().
Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/5269c7e5f5d2117358af3a89744d75a116be27b0.1615552867.git.christophe.leroy@csgroup.eu
8xx requires to tag the DAR with a magic value in order to
fixup DAR on faults generated by 'dcbX', as the 8xx
forgets to update the DAR for those faults.
Do the tagging as early as possible, that is before enabling MMU.
Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/853a2e28ca7c5fc85617037030f99fe6070c9536.1615552867.git.christophe.leroy@csgroup.eu
If the code can use a stack in vm area, it can also use a
stack in linear space.
Simplify code by removing old non VMAP stack code on PPC32.
That means the data translation is now re-enabled early in
exception prolog in all cases, not only when using VMAP stacks.
While we are touching EXCEPTION_PROLOG macros, remove the
unused for_rtas parameter in EXCEPTION_PROLOG_1.
Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/7cd6440c60a7e8f4f035b245c57720f51e225aae.1615552866.git.christophe.leroy@csgroup.eu
ksp_limit is there to help detect stack overflows.
That is specific to ppc32 as it was removed from ppc64 in
commit cbc9565ee8 ("powerpc: Remove ksp_limit on ppc64").
There are other means for detecting stack overflows.
As ppc64 has proven to not need it, ppc32 should be able to do
without it too.
Lets remove it and simplify exception handling.
Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/d789c3385b22e07bedc997613c0d26074cb513e7.1615552866.git.christophe.leroy@csgroup.eu
Ensure normal exception handler are able to manage stuff with
MMU enabled. For that we use CONFIG_VMAP_STACK related code
allthough there is no intention to really activate CONFIG_VMAP_STACK
on powerpc 40x for the moment.
40x uses SPRN_DEAR instead of SPRN_DAR and SPRN_ESR instead of
SPRN_DSISR. Take it into account in common macros.
40x MSR value doesn't fit on 15 bits, use LOAD_REG_IMMEDIATE() in
common macros that will be used also with 40x.
Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/01963af2b83037bca270d7bf1336ffcf35da8282.1615552866.git.christophe.leroy@csgroup.eu
In order the enable MMU early in exception prolog, implement
CONFIG_VMAP_STACK principles in critical exception prolog.
There is no intention to use CONFIG_VMAP_STACK on 40x,
but related code will be used to enable MMU early in exception
in a later patch.
Also address (critirq_ctx - PAGE_OFFSET) directly instead of
using tophys() in order to win one instruction.
Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/3fd75ee54c48307119acdbf66cfea966c1463bbd.1615552866.git.christophe.leroy@csgroup.eu
SPRN_SPRG_SCRATCH5 is used to save SPRN_PID.
SPRN_SPRG_SCRATCH6 is already available.
SPRN_PID is only 8 bits. We have r12 that contains CR.
We only need to preserve CR0, so we have space available in r12
to save PID.
Keep PID in r12 and free up SPRN_SPRG_SCRATCH5.
Then In TLB miss handlers, instead of using SPRN_SPRG_SCRATCH0 and
SPRN_SPRG_SCRATCH1, use SPRN_SPRG_SCRATCH5 and SPRN_SPRG_SCRATCH6
to avoid future conflicts with normal exception prologs.
Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/4cdaa85d38e14d594ba902424060ec55babf2c42.1615552866.git.christophe.leroy@csgroup.eu
It is better to rely on the API provided by the MM layer instead of
directly manipulating the mm_users field.
Signed-off-by: Laurent Dufour <ldufour@linux.ibm.com>
Acked-by: Frederic Barrat <fbarrat@linux.ibm.com>
Acked-by: Andrew Donnellan <ajd@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20210310174405.51044-1-ldufour@linux.ibm.com
As per ISA 3.1, prefixed instruction should not cross 64-byte
boundary. So don't allow Uprobe on such prefixed instruction.
There are two ways probed instruction is changed in mapped pages.
First, when Uprobe is activated, it searches for all the relevant
pages and replace instruction in them. In this case, if that probe
is on the 64-byte unaligned prefixed instruction, error out
directly. Second, when Uprobe is already active and user maps a
relevant page via mmap(), instruction is replaced via mmap() code
path. But because Uprobe is invalid, entire mmap() operation can
not be stopped. In this case just print an error and continue.
Signed-off-by: Ravi Bangoria <ravi.bangoria@linux.ibm.com>
Acked-by: Naveen N. Rao <naveen.n.rao@linux.vnet.ibm.com>
Acked-by: Sandipan Das <sandipan@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20210311091538.368590-1-ravi.bangoria@linux.ibm.com
Usually sigset_t is exactly 8B which is a "trivial" size and does not
warrant using __copy_from_user(). Use __get_user() directly in
anticipation of future work to remove the trivial size optimizations
from __copy_from_user().
The ppc32 implementation of get_sigset_t() previously called
copy_from_user() which, unlike __copy_from_user(), calls access_ok().
Replacing this w/ __get_user() (no access_ok()) is fine here since both
callsites in signal_32.c are preceded by an earlier access_ok().
Signed-off-by: Christopher M. Riedl <cmr@codefail.de>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20210227011259.11992-11-cmr@codefail.de
Add uaccess blocks and use the 'unsafe' versions of functions doing user
access where possible to reduce the number of times uaccess has to be
opened/closed.
Co-developed-by: Christopher M. Riedl <cmr@codefail.de>
Signed-off-by: Daniel Axtens <dja@axtens.net>
Signed-off-by: Christopher M. Riedl <cmr@codefail.de>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20210227011259.11992-10-cmr@codefail.de
Add uaccess blocks and use the 'unsafe' versions of functions doing user
access where possible to reduce the number of times uaccess has to be
opened/closed.
There is no 'unsafe' version of copy_siginfo_to_user, so move it
slightly to allow for a "longer" uaccess block.
Co-developed-by: Christopher M. Riedl <cmr@codefail.de>
Signed-off-by: Daniel Axtens <dja@axtens.net>
Signed-off-by: Christopher M. Riedl <cmr@codefail.de>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20210227011259.11992-9-cmr@codefail.de
Previously restore_sigcontext() performed a costly KUAP switch on every
uaccess operation. These repeated uaccess switches cause a significant
drop in signal handling performance.
Rewrite restore_sigcontext() to assume that a userspace read access
window is open by replacing all uaccess functions with their 'unsafe'
versions. Modify the callers to first open, call
unsafe_restore_sigcontext(), and then close the uaccess window.
Signed-off-by: Christopher M. Riedl <cmr@codefail.de>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20210227011259.11992-8-cmr@codefail.de
Previously setup_sigcontext() performed a costly KUAP switch on every
uaccess operation. These repeated uaccess switches cause a significant
drop in signal handling performance.
Rewrite setup_sigcontext() to assume that a userspace write access window
is open by replacing all uaccess functions with their 'unsafe' versions.
Modify the callers to first open, call unsafe_setup_sigcontext() and
then close the uaccess window.
Signed-off-by: Christopher M. Riedl <cmr@codefail.de>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20210227011259.11992-7-cmr@codefail.de
Both rt_sigreturn() and handle_rt_signal_64() contain TM-related ifdefs
which break-up an if/else block. Provide stubs for the ifdef-guarded TM
functions and remove the need for an ifdef in rt_sigreturn().
Rework the remaining TM ifdef in handle_rt_signal64() similar to
commit f1cf4f93de ("powerpc/signal32: Remove ifdefery in middle of if/else").
Unlike in the commit for ppc32, the ifdef can't be removed entirely
since uc_transact in sigframe depends on CONFIG_PPC_TRANSACTIONAL_MEM.
Signed-off-by: Christopher M. Riedl <cmr@codefail.de>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20210227011259.11992-6-cmr@codefail.de
Unlike the other MSR_TM_* macros, MSR_TM_ACTIVE does not reference or
use its parameter unless CONFIG_PPC_TRANSACTIONAL_MEM is defined. This
causes an 'unused variable' compile warning unless the variable is also
guarded with CONFIG_PPC_TRANSACTIONAL_MEM.
Reference but do nothing with the argument in the macro to avoid a
potential compile warning.
Signed-off-by: Christopher M. Riedl <cmr@codefail.de>
Reviewed-by: Daniel Axtens <dja@axtens.net>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20210227011259.11992-5-cmr@codefail.de
The majority of setup_sigcontext() can be refactored to execute in an
"unsafe" context assuming an open uaccess window except for some
non-inline function calls. Move these out into a separate
prepare_setup_sigcontext() function which must be called first and
before opening up a uaccess window. Non-inline function calls should be
avoided during a uaccess window for a few reasons:
- KUAP should be enabled for as much kernel code as possible.
Opening a uaccess window disables KUAP which means any code
executed during this time contributes to a potential attack
surface.
- Non-inline functions default to traceable which means they are
instrumented for ftrace. This adds more code which could run
with KUAP disabled.
- Powerpc does not currently support the objtool UACCESS checks.
All code running with uaccess must be audited manually which
means: less code -> less work -> fewer problems (in theory).
A follow-up commit converts setup_sigcontext() to be "unsafe".
Signed-off-by: Christopher M. Riedl <cmr@codefail.de>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20210227011259.11992-4-cmr@codefail.de
Reuse the "safe" implementation from signal.c but call unsafe_get_user()
directly in a loop to avoid the intermediate copy into a local buffer.
Signed-off-by: Christopher M. Riedl <cmr@codefail.de>
Reviewed-by: Daniel Axtens <dja@axtens.net>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20210227011259.11992-3-cmr@codefail.de
Use the same approach as unsafe_copy_to_user() but instead call
unsafe_get_user() in a loop.
Signed-off-by: Christopher M. Riedl <cmr@codefail.de>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20210227011259.11992-2-cmr@codefail.de
49a7d46a06 (powerpc: Implement smp_cond_load_relaxed()) added
busy-waiting pausing with a preferred SMT priority pattern, lowering
the priority (reducing decode cycles) during the whole loop slowpath.
However, data shows that while this pattern works well with simple
spinlocks, queued spinlocks benefit more being kept in medium priority,
with a cpu_relax() instead, being a low+medium combo on powerpc.
Data is from three benchmarks on a Power9: 9008-22L 64 CPUs with
2 sockets and 8 threads per core.
1. locktorture.
This is data for the lowest and most artificial/pathological level,
with increasing thread counts pounding on the lock. Metrics are total
ops/minute. Despite some small hits in the 4-8 range, scenarios are
either neutral or favorable to this patch.
+=========+==========+==========+=======+
| # tasks | vanilla | dirty | %diff |
+=========+==========+==========+=======+
| 2 | 46718565 | 48751350 | 4.35 |
+---------+----------+----------+-------+
| 4 | 51740198 | 50369082 | -2.65 |
+---------+----------+----------+-------+
| 8 | 63756510 | 62568821 | -1.86 |
+---------+----------+----------+-------+
| 16 | 67824531 | 70966546 | 4.63 |
+---------+----------+----------+-------+
| 32 | 53843519 | 61155508 | 13.58 |
+---------+----------+----------+-------+
| 64 | 53005778 | 53104412 | 0.18 |
+---------+----------+----------+-------+
| 128 | 53331980 | 54606910 | 2.39 |
+=========+==========+==========+=======+
2. sockperf (tcp throughput)
Here a client will do one-way throughput tests to a localhost server, with
increasing message sizes, dealing with the sk_lock. This patch shows to put
the performance of the qspinlock back to par with that of the simple lock:
simple-spinlock vanilla dirty
Hmean 14 73.50 ( 0.00%) 54.44 * -25.93%* 73.45 * -0.07%*
Hmean 100 654.47 ( 0.00%) 385.61 * -41.08%* 771.43 * 17.87%*
Hmean 300 2719.39 ( 0.00%) 2181.67 * -19.77%* 2666.50 * -1.94%*
Hmean 500 4400.59 ( 0.00%) 3390.77 * -22.95%* 4322.14 * -1.78%*
Hmean 850 6726.21 ( 0.00%) 5264.03 * -21.74%* 6863.12 * 2.04%*
3. dbench (tmpfs)
Configured to run with up to ncpusx8 clients, it shows both latency and
throughput metrics. For the latency, with the exception of the 64 case,
there is really nothing to go by:
vanilla dirty
Amean latency-1 1.67 ( 0.00%) 1.67 * 0.09%*
Amean latency-2 2.15 ( 0.00%) 2.08 * 3.36%*
Amean latency-4 2.50 ( 0.00%) 2.56 * -2.27%*
Amean latency-8 2.49 ( 0.00%) 2.48 * 0.31%*
Amean latency-16 2.69 ( 0.00%) 2.72 * -1.37%*
Amean latency-32 2.96 ( 0.00%) 3.04 * -2.60%*
Amean latency-64 7.78 ( 0.00%) 8.17 * -5.07%*
Amean latency-512 186.91 ( 0.00%) 186.41 * 0.27%*
For the dbench4 Throughput (misleading but traditional) there's a small
but rather constant improvement:
vanilla dirty
Hmean 1 849.13 ( 0.00%) 851.51 * 0.28%*
Hmean 2 1664.03 ( 0.00%) 1663.94 * -0.01%*
Hmean 4 3073.70 ( 0.00%) 3104.29 * 1.00%*
Hmean 8 5624.02 ( 0.00%) 5694.16 * 1.25%*
Hmean 16 9169.49 ( 0.00%) 9324.43 * 1.69%*
Hmean 32 11969.37 ( 0.00%) 12127.09 * 1.32%*
Hmean 64 15021.12 ( 0.00%) 15243.14 * 1.48%*
Hmean 512 14891.27 ( 0.00%) 15162.11 * 1.82%*
Measuring the dbench4 Per-VFS Operation latency, shows some very minor
differences within the noise level, around the 0-1% ranges.
Fixes: 49a7d46a06 ("powerpc: Implement smp_cond_load_relaxed()")
Acked-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Davidlohr Bueso <dbueso@suse.de>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20210318204702.71417-1-dave@stgolabs.net
