diff options
author | Linus Torvalds <torvalds@linux-foundation.org> | 2018-08-03 10:38:21 -0700 |
---|---|---|
committer | Linus Torvalds <torvalds@linux-foundation.org> | 2018-08-03 10:38:21 -0700 |
commit | 2ed7533cb7d657410ac29adf08d608b634e58eb0 (patch) | |
tree | 0c155334c9d9a208801c8d8c46b231bb44ecc3f8 /arch | |
parent | 7827fc7d24043522726c58612d2c7f39843721cf (diff) | |
parent | cca19f0b684f4ed6aabf6ad07ae3e15e77bfd78a (diff) |
Merge tag 'powerpc-4.18-5' of git://git.kernel.org/pub/scm/linux/kernel/git/powerpc/linux
Pull powerpc fixes from Michael Ellerman:
"One fix for a regression in a recent TLB flush optimisation, which
caused us to incorrectly not send TLB invalidations to coprocessors.
Thanks to Frederic Barrat, Nicholas Piggin, Vaibhav Jain"
* tag 'powerpc-4.18-5' of git://git.kernel.org/pub/scm/linux/kernel/git/powerpc/linux:
powerpc/64s/radix: Fix missing global invalidations when removing copro
Diffstat (limited to 'arch')
-rw-r--r-- | arch/powerpc/include/asm/mmu_context.h | 33 |
1 files changed, 21 insertions, 12 deletions
diff --git a/arch/powerpc/include/asm/mmu_context.h b/arch/powerpc/include/asm/mmu_context.h index 79d570cbf332..b2f89b621b15 100644 --- a/arch/powerpc/include/asm/mmu_context.h +++ b/arch/powerpc/include/asm/mmu_context.h @@ -143,24 +143,33 @@ static inline void mm_context_remove_copro(struct mm_struct *mm) { int c; - c = atomic_dec_if_positive(&mm->context.copros); - - /* Detect imbalance between add and remove */ - WARN_ON(c < 0); - /* - * Need to broadcast a global flush of the full mm before - * decrementing active_cpus count, as the next TLBI may be - * local and the nMMU and/or PSL need to be cleaned up. - * Should be rare enough so that it's acceptable. + * When removing the last copro, we need to broadcast a global + * flush of the full mm, as the next TLBI may be local and the + * nMMU and/or PSL need to be cleaned up. + * + * Both the 'copros' and 'active_cpus' counts are looked at in + * flush_all_mm() to determine the scope (local/global) of the + * TLBIs, so we need to flush first before decrementing + * 'copros'. If this API is used by several callers for the + * same context, it can lead to over-flushing. It's hopefully + * not common enough to be a problem. * * Skip on hash, as we don't know how to do the proper flush * for the time being. Invalidations will remain global if - * used on hash. + * used on hash. Note that we can't drop 'copros' either, as + * it could make some invalidations local with no flush + * in-between. */ - if (c == 0 && radix_enabled()) { + if (radix_enabled()) { flush_all_mm(mm); - dec_mm_active_cpus(mm); + + c = atomic_dec_if_positive(&mm->context.copros); + /* Detect imbalance between add and remove */ + WARN_ON(c < 0); + + if (c == 0) + dec_mm_active_cpus(mm); } } #else |