Patch series "introduce VMA merge mode to improve brk() performance".
A ~5% performance regression was discovered on the
aim9.brk_test.ops_per_sec by the linux kernel test bot [0].
In the past to satisfy brk() performance we duplicated VMA expansion code
and special-cased do_brk_flags(). This is however horrid and undoes work
to abstract this logic, so in resolving the issue I have endeavoured to
avoid this.
Investigating further I was able to observe that the use of a
vma_iter_next_range() and vma_prev() pair, causing an unnecessary maple
tree walk. In addition there is work that we do that is simply
unnecessary for brk().
Therefore, add a special VMA merge mode VMG_FLAG_JUST_EXPAND to avoid
doing any of this - it assumes the VMA iterator is pointing at the
previous VMA and which skips logic that brk() does not require.
This mostly eliminates the performance regression reducing it to ~2% which
is in the realm of noise. In addition, the will-it-scale test brk2,
written to be more representative of real-world brk() usage, shows a
modest performance improvement - which gives me confidence that we are not
meaningfully regressing real workloads here.
This series includes a test asserting that the 'just expand' mode works as
expected.
With many thanks to Oliver Sang for helping with performance testing of
candidate patch sets!
[0]:https://lore.kernel.org/linux-mm/202409301043.629bea78-oliver.sang@intel.com
This patch (of 2):
We know in advance that do_brk_flags() wants only to perform a VMA
expansion (if the prior VMA is compatible), and that we assume no
mergeable VMA follows it.
These are the semantics of this function prior to the recent rewrite of
the VMA merging logic, however we are now doing more work than necessary -
positioning the VMA iterator at the prior VMA and performing tasks that
are not required.
Add a new field to the vmg struct to permit merge flags and add a new
merge flag VMG_FLAG_JUST_EXPAND which implies this behaviour, and have
do_brk_flags() use this.
This fixes a reported performance regression in a brk() benchmarking suite.
Link: https://lkml.kernel.org/r/cover.1729174352.git.lorenzo.stoakes@oracle.com
Link: https://lkml.kernel.org/r/4e65d4395e5841c5acf8470dbcb714016364fd39.1729174352.git.lorenzo.stoakes@oracle.com
Fixes: cacded5e42 ("mm: avoid using vma_merge() for new VMAs")
Reported-by: kernel test robot <oliver.sang@intel.com>
Closes: https://lore.kernel.org/linux-mm/202409301043.629bea78-oliver.sang@intel.com
Signed-off-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Reviewed-by: Liam R. Howlett <Liam.Howlett@Oracle.com>
Cc: Jann Horn <jannh@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
vms_abort_munmap_vmas() is a recovery path where, on entry, some VMAs have
already been torn down halfway (in a way we can't undo) but are still
present in the maple tree.
At this point, we *must* remove the VMAs from the VMA tree, otherwise we
get UAF.
Because removing VMA tree nodes can require memory allocation, the
existing code has an error path which tries to handle this by reattaching
the VMAs; but that can't be done safely.
A nicer way to fix it would probably be to preallocate enough maple tree
nodes for the removal before the point of no return, or something like
that; but for now, fix it the easy and kinda ugly way, by marking this
allocation __GFP_NOFAIL.
Link: https://lkml.kernel.org/r/20241016-fix-munmap-abort-v1-1-601c94b2240d@google.com
Fixes: 4f87153e82 ("mm: change failure of MAP_FIXED to restoring the gap on failure")
Signed-off-by: Jann Horn <jannh@google.com>
Reviewed-by: Liam R. Howlett <Liam.Howlett@Oracle.com>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Reviewed-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Abstract vma_merge_new_vma() to use vma_merge_struct and rename the
resultant function vma_merge_new_range() to be clear what the purpose of
this function is - a new VMA is desired in the specified range, and we
wish to see if it is possible to 'merge' surrounding VMAs into this range
rather than having to allocate a new VMA.
Note that this function uses vma_extend() exclusively, so adopts its
requirement that the iterator point at or before the gap. We add an
assert to this effect.
This is as opposed to vma_merge_existing_range(), which will be introduced
in a subsequent commit, and provide the same functionality for cases in
which we are modifying an existing VMA.
In mmap_region() and do_brk_flags() we open code scenarios where we prefer
to use vma_expand() rather than invoke a full vma_merge() operation.
Abstract this logic and eliminate all of the open-coding, and also use the
same logic for all cases where we add new VMAs to, rather than ultimately
use vma_merge(), rather use vma_expand().
Doing so removes duplication and simplifies VMA merging in all such cases,
laying the ground for us to eliminate the merging of new VMAs in
vma_merge() altogether.
Also add the ability for the vmg to track state, and able to report
errors, allowing for us to differentiate a failed merge from an inability
to allocate memory in callers.
This makes it far easier to understand what is happening in these cases
avoiding confusion, bugs and allowing for future optimisation.
Also introduce vma_iter_next_rewind() to allow for retrieval of the next,
and (optionally) the prev VMA, rewinding to the start of the previous gap.
Introduce are_anon_vmas_compatible() to abstract individual VMA anon_vma
comparison for the case of merging on both sides where the anon_vma of the
VMA being merged maybe compatible with prev and next, but prev and next's
anon_vma's may not be compatible with each other.
Finally also introduce can_vma_merge_left() / can_vma_merge_right() to
check adjacent VMA compatibility and that they are indeed adjacent.
Link: https://lkml.kernel.org/r/49d37c0769b6b9dc03b27fe4d059173832556392.1725040657.git.lorenzo.stoakes@oracle.com
Signed-off-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Tested-by: Mark Brown <broonie@kernel.org>
Cc: Liam R. Howlett <Liam.Howlett@oracle.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Bert Karwatzki <spasswolf@web.de>
Cc: Jeff Xu <jeffxu@chromium.org>
Cc: Jiri Olsa <olsajiri@gmail.com>
Cc: Kees Cook <kees@kernel.org>
Cc: Lorenzo Stoakes <lstoakes@gmail.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: "Paul E. McKenney" <paulmck@kernel.org>
Cc: Paul Moore <paul@paul-moore.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Rather than passing around huge numbers of parameters to numerous helper
functions, abstract them into a single struct that we thread through the
operation, the vma_merge_struct ('vmg').
Adjust vma_merge() and vma_modify() to accept this parameter, as well as
predicate functions can_vma_merge_before(), can_vma_merge_after(), and the
vma_modify_...() helper functions.
Also introduce VMG_STATE() and VMG_VMA_STATE() helper macros to allow for
easy vmg declaration.
We additionally remove the requirement that vma_merge() is passed a VMA
object representing the candidate new VMA. Previously it used this to
obtain the mm_struct, file and anon_vma properties of the proposed range
(a rather confusing state of affairs), which are now provided by the vmg
directly.
We also remove the pgoff calculation previously performed vma_modify(),
and instead calculate this in VMG_VMA_STATE() via the vma_pgoff_offset()
helper.
Link: https://lkml.kernel.org/r/a955aad09d81329f6fbeb636b2dd10cde7b73dab.1725040657.git.lorenzo.stoakes@oracle.com
Signed-off-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Reviewed-by: Liam R. Howlett <Liam.Howlett@oracle.com>
Cc: Mark Brown <broonie@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Bert Karwatzki <spasswolf@web.de>
Cc: Jeff Xu <jeffxu@chromium.org>
Cc: Jiri Olsa <olsajiri@gmail.com>
Cc: Kees Cook <kees@kernel.org>
Cc: Lorenzo Stoakes <lstoakes@gmail.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: "Paul E. McKenney" <paulmck@kernel.org>
Cc: Paul Moore <paul@paul-moore.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Instead of zeroing the vma tree and then overwriting the area, let the
area be overwritten and then clean up the gathered vmas using
vms_complete_munmap_vmas().
To ensure locking is downgraded correctly, the mm is set regardless of
MAP_FIXED or not (NULL vma).
If a driver is mapping over an existing vma, then clear the ptes before
the call_mmap() invocation. This is done using the vms_clean_up_area()
helper. If there is a close vm_ops, that must also be called to ensure
any cleanup is done before mapping over the area. This also means that
calling open has been added to the abort of an unmap operation, for now.
Since vm_ops->open() and vm_ops->close() are not always undo each other
(state cleanup may exist in ->close() that is lost forever), the code
cannot be left in this way, but that change has been isolated to another
commit to make this point very obvious for traceability.
Temporarily keep track of the number of pages that will be removed and
reduce the charged amount.
This also drops the validate_mm() call in the vma_expand() function. It
is necessary to drop the validate as it would fail since the mm map_count
would be incorrect during a vma expansion, prior to the cleanup from
vms_complete_munmap_vmas().
Clean up the error handing of the vms_gather_munmap_vmas() by calling the
verification within the function.
Link: https://lkml.kernel.org/r/20240830040101.822209-15-Liam.Howlett@oracle.com
Signed-off-by: Liam R. Howlett <Liam.Howlett@Oracle.com>
Reviewed-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Bert Karwatzki <spasswolf@web.de>
Cc: Jeff Xu <jeffxu@chromium.org>
Cc: Jiri Olsa <olsajiri@gmail.com>
Cc: Kees Cook <kees@kernel.org>
Cc: Lorenzo Stoakes <lstoakes@gmail.com>
Cc: Mark Brown <broonie@kernel.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: "Paul E. McKenney" <paulmck@kernel.org>
Cc: Paul Moore <paul@paul-moore.com>
Cc: Sidhartha Kumar <sidhartha.kumar@oracle.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Patch series "mm: Optimize mseal checks", v3.
Optimize mseal checks by removing the separate can_modify_mm() step, and
just doing checks on the individual vmas, when various operations are
themselves iterating through the tree. This provides a nice speedup and
restores performance parity with pre-mseal[3].
will-it-scale mmap1_process[1] -t 1 results:
commit 3450fe2b574b4345e4296ccae395149e1a357fee:
min:277605 max:277605 total:277605
min:281784 max:281784 total:281784
min:277238 max:277238 total:277238
min:281761 max:281761 total:281761
min:274279 max:274279 total:274279
min:254854 max:254854 total:254854
measurement
min:269143 max:269143 total:269143
min:270454 max:270454 total:270454
min:243523 max:243523 total:243523
min:251148 max:251148 total:251148
min:209669 max:209669 total:209669
min:190426 max:190426 total:190426
min:231219 max:231219 total:231219
min:275364 max:275364 total:275364
min:266540 max:266540 total:266540
min:242572 max:242572 total:242572
min:284469 max:284469 total:284469
min:278882 max:278882 total:278882
min:283269 max:283269 total:283269
min:281204 max:281204 total:281204
After this patch set:
min:280580 max:280580 total:280580
min:290514 max:290514 total:290514
min:291006 max:291006 total:291006
min:290352 max:290352 total:290352
min:294582 max:294582 total:294582
min:293075 max:293075 total:293075
measurement
min:295613 max:295613 total:295613
min:294070 max:294070 total:294070
min:293193 max:293193 total:293193
min:291631 max:291631 total:291631
min:295278 max:295278 total:295278
min:293782 max:293782 total:293782
min:290361 max:290361 total:290361
min:294517 max:294517 total:294517
min:293750 max:293750 total:293750
min:293572 max:293572 total:293572
min:295239 max:295239 total:295239
min:292932 max:292932 total:292932
min:293319 max:293319 total:293319
min:294954 max:294954 total:294954
This was a Completely Unscientific test but seems to show there were
around 5-10% gains on ops per second.
Oliver performed his own tests and showed[3] a similar ~5% gain in them.
[1]: mmap1_process does mmap and munmap in a loop. I didn't bother testing multithreading cases.
[2]: https://lore.kernel.org/all/20240807124103.85644-1-mpe@ellerman.id.au/
[3]: https://lore.kernel.org/all/ZrMMJfe9aXSWxJz6@xsang-OptiPlex-9020/
Link: https://lore.kernel.org/all/202408041602.caa0372-oliver.sang@intel.com/
This patch (of 7):
Move can_modify_vma to vma.h so it can be inlined properly (with the
intent to remove can_modify_mm callsites).
Link: https://lkml.kernel.org/r/20240817-mseal-depessimize-v3-1-d8d2e037df30@gmail.com
Signed-off-by: Pedro Falcato <pedro.falcato@gmail.com>
Reviewed-by: Liam R. Howlett <Liam.Howlett@Oracle.com>
Reviewed-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Jeff Xu <jeffxu@chromium.org>
Cc: Kees Cook <kees@kernel.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Pedro Falcato <pedro.falcato@gmail.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
