[linux-nvidia-6.18] Backport TLBI optimization patches#337
Open
nvmochs wants to merge 2 commits intoNVIDIA:linux-nvidia-6.18from
Open
[linux-nvidia-6.18] Backport TLBI optimization patches#337nvmochs wants to merge 2 commits intoNVIDIA:linux-nvidia-6.18from
nvmochs wants to merge 2 commits intoNVIDIA:linux-nvidia-6.18from
Conversation
The TLBI instruction accepts XZR as a register argument, and for TLBI operations with a register argument, there is no functional difference between using XZR or another GPR which contains zeroes. Operations without a register argument are encoded as if XZR were used. Allow the __TLBI_1() macro to use XZR when a register argument is all zeroes. Today this only results in a trivial code saving in __do_compat_cache_op()'s workaround for Neoverse-N1 erratum #1542419. In subsequent patches this pattern will be used more generally. There should be no functional change as a result of this patch. Signed-off-by: Mark Rutland <mark.rutland@arm.com> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Marc Zyngier <maz@kernel.org> Cc: Oliver Upton <oupton@kernel.org> Cc: Ryan Roberts <ryan.roberts@arm.com> Cc: Will Deacon <will@kernel.org> Signed-off-by: Will Deacon <will@kernel.org> (cherry picked from commit bfd9c93) Signed-off-by: Matthew R. Ochs <mochs@nvidia.com>
The ARM64_WORKAROUND_REPEAT_TLBI workaround is used to mitigate several errata where broadcast TLBI;DSB sequences don't provide all the architecturally required synchronization. The workaround performs more work than necessary, and can have significant overhead. This patch optimizes the workaround, as explained below. The workaround was originally added for Qualcomm Falkor erratum 1009 in commit: d9ff80f ("arm64: Work around Falkor erratum 1009") As noted in the message for that commit, the workaround is applied even in cases where it is not strictly necessary. The workaround was later reused without changes for: * Arm Cortex-A76 erratum #1286807 SDEN v33: https://developer.arm.com/documentation/SDEN-885749/33-0/ * Arm Cortex-A55 erratum #2441007 SDEN v16: https://developer.arm.com/documentation/SDEN-859338/1600/ * Arm Cortex-A510 erratum #2441009 SDEN v19: https://developer.arm.com/documentation/SDEN-1873351/1900/ The important details to note are as follows: 1. All relevant errata only affect the ordering and/or completion of memory accesses which have been translated by an invalidated TLB entry. The actual invalidation of TLB entries is unaffected. 2. The existing workaround is applied to both broadcast and local TLB invalidation, whereas for all relevant errata it is only necessary to apply a workaround for broadcast invalidation. 3. The existing workaround replaces every TLBI with a TLBI;DSB;TLBI sequence, whereas for all relevant errata it is only necessary to execute a single additional TLBI;DSB sequence after any number of TLBIs are completed by a DSB. For example, for a sequence of batched TLBIs: TLBI <op1>[, <arg1>] TLBI <op2>[, <arg2>] TLBI <op3>[, <arg3>] DSB ISH ... the existing workaround will expand this to: TLBI <op1>[, <arg1>] DSB ISH // additional TLBI <op1>[, <arg1>] // additional TLBI <op2>[, <arg2>] DSB ISH // additional TLBI <op2>[, <arg2>] // additional TLBI <op3>[, <arg3>] DSB ISH // additional TLBI <op3>[, <arg3>] // additional DSB ISH ... whereas it is sufficient to have: TLBI <op1>[, <arg1>] TLBI <op2>[, <arg2>] TLBI <op3>[, <arg3>] DSB ISH TLBI <opX>[, <argX>] // additional DSB ISH // additional Using a single additional TBLI and DSB at the end of the sequence can have significantly lower overhead as each DSB which completes a TLBI must synchronize with other PEs in the system, with potential performance effects both locally and system-wide. 4. The existing workaround repeats each specific TLBI operation, whereas for all relevant errata it is sufficient for the additional TLBI to use *any* operation which will be broadcast, regardless of which translation regime or stage of translation the operation applies to. For example, for a single TLBI: TLBI ALLE2IS DSB ISH ... the existing workaround will expand this to: TLBI ALLE2IS DSB ISH TLBI ALLE2IS // additional DSB ISH // additional ... whereas it is sufficient to have: TLBI ALLE2IS DSB ISH TLBI VALE1IS, XZR // additional DSB ISH // additional As the additional TLBI doesn't have to match a specific earlier TLBI, the additional TLBI can be implemented in separate code, with no memory of the earlier TLBIs. The additional TLBI can also use a cheaper TLBI operation. 5. The existing workaround is applied to both Stage-1 and Stage-2 TLB invalidation, whereas for all relevant errata it is only necessary to apply a workaround for Stage-1 invalidation. Architecturally, TLBI operations which invalidate only Stage-2 information (e.g. IPAS2E1IS) are not required to invalidate TLB entries which combine information from Stage-1 and Stage-2 translation table entries, and consequently may not complete memory accesses translated by those combined entries. In these cases, completion of memory accesses is only guaranteed after subsequent invalidation of Stage-1 information (e.g. VMALLE1IS). Taking the above points into account, this patch reworks the workaround logic to reduce overhead: * New __tlbi_sync_s1ish() and __tlbi_sync_s1ish_hyp() functions are added and used in place of any dsb(ish) which is used to complete broadcast Stage-1 TLB maintenance. When the ARM64_WORKAROUND_REPEAT_TLBI workaround is enabled, these helpers will execute an additional TLBI;DSB sequence. For consistency, it might make sense to add __tlbi_sync_*() helpers for local and stage 2 maintenance. For now I've left those with open-coded dsb() to keep the diff small. * The duplication of TLBIs in __TLBI_0() and __TLBI_1() is removed. This is no longer needed as the necessary synchronization will happen in __tlbi_sync_s1ish() or __tlbi_sync_s1ish_hyp(). * The additional TLBI operation is chosen to have minimal impact: - __tlbi_sync_s1ish() uses "TLBI VALE1IS, XZR". This is only used at EL1 or at EL2 with {E2H,TGE}=={1,1}, where it will target an unused entry for the reserved ASID in the kernel's own translation regime, and have no adverse affect. - __tlbi_sync_s1ish_hyp() uses "TLBI VALE2IS, XZR". This is only used in hyp code, where it will target an unused entry in the hyp code's TTBR0 mapping, and should have no adverse effect. * As __TLBI_0() and __TLBI_1() no longer replace each TLBI with a TLBI;DSB;TLBI sequence, batching TLBIs is worthwhile, and there's no need for arch_tlbbatch_should_defer() to consider ARM64_WORKAROUND_REPEAT_TLBI. When building defconfig with GCC 15.1.0, compared to v6.19-rc1, this patch saves ~1KiB of text, makes the vmlinux ~42KiB smaller, and makes the resulting Image 64KiB smaller: | [mark@lakrids:~/src/linux]% size vmlinux-* | text data bss dec hex filename | 21179831 19660919 708216 4154896 279fca6 vmlinux-after | 21181075 19660903 708216 41550194 27a0172 vmlinux-before | [mark@lakrids:~/src/linux]% ls -l vmlinux-* | -rwxr-xr-x 1 mark mark 157771472 Feb 4 12:05 vmlinux-after | -rwxr-xr-x 1 mark mark 157815432 Feb 4 12:05 vmlinux-before | [mark@lakrids:~/src/linux]% ls -l Image-* | -rw-r--r-- 1 mark mark 41007616 Feb 4 12:05 Image-after | -rw-r--r-- 1 mark mark 41073152 Feb 4 12:05 Image-before Signed-off-by: Mark Rutland <mark.rutland@arm.com> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Marc Zyngier <maz@kernel.org> Cc: Oliver Upton <oupton@kernel.org> Cc: Ryan Roberts <ryan.roberts@arm.com> Cc: Will Deacon <will@kernel.org> Signed-off-by: Will Deacon <will@kernel.org> (cherry picked from commit a8f7868) Signed-off-by: Matthew R. Ochs <mochs@nvidia.com>
Collaborator
|
|
clsotog
approved these changes
Mar 10, 2026
Collaborator
clsotog
left a comment
clsotog
left a comment
There was a problem hiding this comment.
Acked-by: Carol L Soto <csoto@nvidia.com>
jamieNguyenNVIDIA
approved these changes
Mar 11, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
These two upstream patches provide some minor TLBI-related optimizations and may also be dependencies for some future patches.
LKML: https://lore.kernel.org/all/20260218164348.2022831-1-mark.rutland@arm.com/
v7.0:
a8f7868 arm64: tlb: Optimize ARM64_WORKAROUND_REPEAT_TLBI
bfd9c93 arm64: tlb: Allow XZR argument to TLBI ops
These patches picked clean and were tested by running
stress-ng --vm 8 --vm-bytes 75% --fork 8 --timeout 600s.