mm/mmu_gather: Invalidate TLB correctly on batch allocation failure and flush
Subsystems (5)
RFC PATCH 2/2 mm/mmu_gather: Avoid multiple page walk cache flush
In reply to Aneesh Kumar K.VOn tlb_finish_mmu() kernel does a tlb flush before mmu gather table invalidate. The mmu gather table invalidate depending on kernel config also does another TLBI. Avoid the later on tlb_finish_mmu().
--- mm/mmu_gather.c | 12 +++++++++--- 1 file changed, 9 insertions(+), 3 deletions(-)
mm/mmu_gather.c +9 -3
@@ -143,17 +143,23 @@ static void tlb_remove_table_rcu(struct rcu_head *head)free_page ((unsigned long )batch); } -static void tlb_table_flush (struct mmu_gather *tlb) +static void __tlb_table_flush (struct mmu_gather *tlb,bool table_inval) {struct mmu_table_batch **batch = &tlb->batch;if (*batch) { -tlb_table_invalidate (tlb); +if (table_inval) +tlb_table_invalidate (tlb);call_rcu (&(*batch)->rcu, tlb_remove_table_rcu); *batch =NULL ; } } +static void tlb_table_flush (struct mmu_gather *tlb) +{ +__tlb_table_flush (tlb, true); +} +void tlb_remove_table (struct mmu_gather *tlb,void *table) {struct mmu_table_batch **batch = &tlb->batch;
@@ -178,7 +184,7 @@ void tlb_remove_table(struct mmu_gather *tlb, void *table)static void tlb_flush_mmu_free (struct mmu_gather *tlb) {#ifdef CONFIG_HAVE_RCU_TABLE_FREE -tlb_table_flush (tlb); +__tlb_table_flush (tlb, false);#endif #ifndef CONFIG_HAVE_MMU_GATHER_NO_GATHER tlb_batch_pages_flush (tlb);
Signature · 1 line
2.23.0
RFC PATCH 1/2 mm/mmu_gather: Invalidate TLB correctly on batch allocation failure and flush
Architectures for which we have hardware walkers of Linux page table should flush TLB on mmu gather batch allocation failures and batch flush. Some architectures like POWER supports multiple translation modes (hash and radix) and in the case of POWER only radix translation mode needs the above TLBI. This is because for hash translation mode kernel wants to avoid this extra flush since there are no hardware walkers of linux page table. With radix translation, the hardware also walks linux page table and with that, kernel needs to make sure to TLB invalidate page walk cache before page table pages are freed.
More details in
commit: d86564a2f085 ("mm/tlb, x86/mm: Support invalidating TLB caches for RCU_TABLE_FREE")
Based on changes from Peter Zijlstra <[email protected]>
--- arch/Kconfig | 3 --- arch/powerpc/Kconfig | 1 - arch/powerpc/include/asm/tlb.h | 4 ++++ arch/sparc/Kconfig | 1 - arch/sparc/include/asm/tlb_64.h | 9 +++++++++ include/asm-generic/tlb.h | 22 +++++++++++++++------- mm/mmu_gather.c | 16 ++++++++-------- 7 files changed, 36 insertions(+), 20 deletions(-)
arch/Kconfig +0 -3
@@ -396,9 +396,6 @@ config HAVE_ARCH_JUMP_LABEL_RELATIVE config HAVE_RCU_TABLE_FREE bool -config HAVE_RCU_TABLE_NO_INVALIDATE - bool - config HAVE_MMU_GATHER_PAGE_SIZE bool
arch/powerpc/Kconfig +0 -1
@@ -223,7 +223,6 @@ config PPC select HAVE_PERF_REGS select HAVE_PERF_USER_STACK_DUMP select HAVE_RCU_TABLE_FREE if SMP - select HAVE_RCU_TABLE_NO_INVALIDATE if HAVE_RCU_TABLE_FREE select HAVE_MMU_GATHER_PAGE_SIZE select HAVE_REGS_AND_STACK_ACCESS_API select HAVE_RELIABLE_STACKTRACE if PPC_BOOK3S_64 && CPU_LITTLE_ENDIAN
arch/powerpc/include/asm/tlb.h +4 -0
@@ -27,6 +27,10 @@#define tlb_flush tlb_flushextern void tlb_flush (struct mmu_gather *tlb); +#ifdef CONFIG_HAVE_RCU_TABLE_FREE +#define tlb_needs_table_invalidate () radix_enabled() +#endif +/* Get the generic bits... */ #include <asm-generic/tlb.h>
arch/sparc/Kconfig +0 -1
@@ -65,7 +65,6 @@ config SPARC64 select HAVE_KRETPROBES select HAVE_KPROBES select HAVE_RCU_TABLE_FREE if SMP - select HAVE_RCU_TABLE_NO_INVALIDATE if HAVE_RCU_TABLE_FREE select HAVE_MEMBLOCK_NODE_MAP select HAVE_ARCH_TRANSPARENT_HUGEPAGE select HAVE_DYNAMIC_FTRACE
arch/sparc/include/asm/tlb_64.h +9 -0
@@ -28,6 +28,15 @@ void flush_tlb_pending(void);#define __tlb_remove_tlb_entry (tlb, ptep, address) do { } while (0)#define tlb_flush (tlb) flush_tlb_pending() +/* +* SPARC64's hardware TLB fill does not use the Linux page-tables +* and therefore we don't need a TLBI when freeing page-table pages. +*/ + +#ifdef CONFIG_HAVE_RCU_TABLE_FREE +#define tlb_needs_table_invalidate () (false) +#endif +#include <asm-generic/tlb.h> #endif /* _SPARC64_TLB_H */
include/asm-generic/tlb.h +15 -7
@@ -137,13 +137,6 @@* When used, an architecture is expected to provide __tlb_remove_table() * which does the actual freeing of these pages. * -* HAVE_RCU_TABLE_NO_INVALIDATE -* -* This makes HAVE_RCU_TABLE_FREE avoid calling tlb_flush_mmu_tlbonly() before -* freeing the page-table pages. This can be avoided if you use -* HAVE_RCU_TABLE_FREE and your architecture does _NOT_ use the Linux -* page-tables natively. -* * MMU_GATHER_NO_RANGE * * Use this if your architecture lacks an efficient flush_tlb_range().
@@ -189,8 +182,23 @@ struct mmu_table_batch {extern void tlb_remove_table (struct mmu_gather *tlb,void *table); +/* +* This allows an architecture that does not use the linux page-tables for +* hardware to skip the TLBI when freeing page tables. +*/ +#ifndef tlb_needs_table_invalidate +#define tlb_needs_table_invalidate () (true) +#endif + +#else + +#ifdef tlb_needs_table_invalidate +#error tlb_needs_table_invalidate() requires MMU_GATHER_RCU_TABLE_FREE#endif +#endif /* CONFIG_HAVE_RCU_TABLE_FREE */ + +#ifndef CONFIG_HAVE_MMU_GATHER_NO_GATHER /* * If we can't allocate a page to make a big batch of page pointers
mm/mmu_gather.c +8 -8
@@ -102,14 +102,14 @@ bool __tlb_remove_page_size(struct mmu_gather *tlb, struct page *page, int page_*/ static inline void tlb_table_invalidate (struct mmu_gather *tlb) { -#ifndef CONFIG_HAVE_RCU_TABLE_NO_INVALIDATE -/* -* Invalidate page-table caches used by hardware walkers. Then we still -* need to RCU-sched wait while freeing the pages because software -* walkers can still be in-flight. -*/ -tlb_flush_mmu_tlbonly (tlb); -#endif +if (tlb_needs_table_invalidate ()) { +/* +* Invalidate page-table caches used by hardware walkers. Then +* we still need to RCU-sched wait while freeing the pages +* because software walkers can still be in-flight. +*/ +tlb_flush_mmu_tlbonly (tlb); + } }static void tlb_remove_table_smp_sync (void *arg)
Signature · 1 line
2.23.0
Peter Zijlstra
maintainerRe: [RFC PATCH 2/2] mm/mmu_gather: Avoid multiple page walk cache flush
In reply to Aneesh Kumar K.VQuoted 3 lines · Aneesh Kumar K.V
On
tlb_finish_mmu()kernel does a tlb flush before mmu gather table invalidate. The mmu gather table invalidate depending on kernel config also does another TLBI. Avoid the later ontlb_finish_mmu().
That is already avoided, if you look at tlb_flush_mmu_tlbonly() it does __tlb_range_reset(), which results in ->end = 0, which then triggers the early exit on the next invocation:
if (!tlb->end) return;
Peter Zijlstra
maintainerQuoted 11 lines · Aneesh Kumar K.V
Architectures for which we have hardware walkers of Linux page table should flush TLB on mmu gather batch allocation failures and batch flush. Some architectures like POWER supports multiple translation modes (hash and radix) and in the case of POWER only radix translation mode needs the above TLBI. This is because for hash translation mode kernel wants to avoid this extra flush since there are no hardware walkers of linux page table. With radix translation, the hardware also walks linux page table and with that, kernel needs to make sure to TLB invalidate page walk cache before page table pages are freed.
Based on changes from Peter Zijlstra <[email protected]>
AFAICT it is all my patch ;-)
Anyway, this commit:
Quoted 2 lines
More details in
commit:d86564a2f085("mm/tlb, x86/mm: Support invalidating TLB caches forRCU_TABLE_FREE")
states that you do an explicit invalidate in __p*_free_tlb(), which, if I'm not mistaken is still there:
arch/powerpc/include/asm/nohash/pgalloc.h: tlb_flush_pgtable(tlb, address);
Or am I reading this wrong? I'm thinking you can remove that now.
arch/powerpc/include/asm/tlb.h line 30
Quoted 9 lines
arch/powerpc/include/asm/tlb.h +1
-0@@ -27,6 +27,10 @@#define tlb_flush tlb_flushextern void tlb_flush (struct mmu_gather *tlb); +#ifdef CONFIG_HAVE_RCU_TABLE_FREE
/*
* PPC-Hash does not use the linux page-tables, so we can avoid
* the TLBI for page-table freeing, PPC-Radix otoh does use the
* page-tables and needs the TLBI.
*/
arch/powerpc/include/asm/tlb.h line 32
Quoted 2 lines
arch/powerpc/include/asm/tlb.h +2
-0+#define tlb_needs_table_invalidate () radix_enabled() +#endif
Also, are you really sure about the !SMP case? Esp. on Radix I'm thinking that the PWC (page-walk-cache) can give trouble even on UP, when we get preempted in the middle of mmu_gather. Hmm?
Quoted 2 lines
arch/powerpc/include/asm/tlb.h +0
-0/* Get the generic bits... */ #include <asm-generic/tlb.h>
Re: [RFC PATCH 2/2] mm/mmu_gather: Avoid multiple page walk cache flush
In reply to Peter ZijlstraQuoted 5 lines · Peter Zijlstra
Quoted 3 lines · Aneesh Kumar K.V
On
tlb_finish_mmu()kernel does a tlb flush before mmu gather table invalidate. The mmu gather table invalidate depending on kernel config also does another TLBI. Avoid the later ontlb_finish_mmu().
Quoted 7 lines
That is already avoided, if you look at
tlb_flush_mmu_tlbonly()it does__tlb_range_reset(), which results in ->end = 0, which then triggers the early exit on the next invocation:if (!tlb->end) return;
Is that true for tlb->fulmm flush?
-aneesh
Quoted 11 lines · Peter Zijlstra
Quoted 9 lines · Aneesh Kumar K.V
Architectures for which we have hardware walkers of Linux page table should flush TLB on mmu gather batch allocation failures and batch flush. Some architectures like POWER supports multiple translation modes (hash and radix) and in the case of POWER only radix translation mode needs the above TLBI. This is because for hash translation mode kernel wants to avoid this extra flush since there are no hardware walkers of linux page table. With radix translation, the hardware also walks linux page table and with that, kernel needs to make sure to TLB invalidate page walk cache before page table pages are freed.
Quoted 3 lines
Quoted 1 line
Based on changes from Peter Zijlstra <[email protected]>
AFAICT it is all my patch ;-)
Yes. I moved the changes you had to upstream. I can update the From: in the next version if you are ok with that?
Quoted 11 lines
Anyway, this commit:
Quoted 2 lines
More details in
commit:d86564a2f085("mm/tlb, x86/mm: Support invalidating TLB caches forRCU_TABLE_FREE")states that you do an explicit invalidate in __p*_free_tlb(), which, if I'm not mistaken is still there:
arch/powerpc/include/asm/nohash/pgalloc.h: tlb_flush_pgtable(tlb, address);
nohash is not really radix. So we still do the tlb flush from the pte_free_tlb for nohash and for PPC-radix, we let tlb_table_invalidate to flush that.
Quoted 16 lines
Or am I reading this wrong? I'm thinking you can remove that now.
arch/powerpc/include/asm/tlb.h line 30
Quoted 9 lines
arch/powerpc/include/asm/tlb.h +1
-0@@ -27,6 +27,10 @@#define tlb_flush tlb_flushextern void tlb_flush (struct mmu_gather *tlb); +#ifdef CONFIG_HAVE_RCU_TABLE_FREE /*
* PPC-Hash does not use the linux page-tables, so we can avoid
* the TLBI for page-table freeing, PPC-Radix otoh does use the
* page-tables and needs the TLBI.
*/
Quoted 7 lines
arch/powerpc/include/asm/tlb.h line 32
Quoted 2 lines
arch/powerpc/include/asm/tlb.h +2
-0+#define tlb_needs_table_invalidate () radix_enabled() +#endif Also, are you really sure about the !SMP case? Esp. on Radix I'm thinking that the PWC (page-walk-cache) can give trouble even on UP, when we get preempted in the middle of
mmu_gather. Hmm?
Yes, looking at !SMP I guess we do have issue there. we do free the pagetable pages directly in __p*_free_tlb() with the current code. That will definitely not work. Are you suggesting we enable HAVE_RCU_TABLE_FREE even for !SMP?
Quoted 4 lines
arch/powerpc/include/asm/tlb.h line 35
Quoted 2 lines
arch/powerpc/include/asm/tlb.h +0
-0/* Get the generic bits... */ #include <asm-generic/tlb.h>
-aneesh
Peter Zijlstra
maintainerRe: [RFC PATCH 2/2] mm/mmu_gather: Avoid multiple page walk cache flush
In reply to Aneesh Kumar K.VQuoted 10 lines · Aneesh Kumar K.V
Quoted 9 lines · Peter Zijlstra
Quoted 3 lines · Aneesh Kumar K.V
On
tlb_finish_mmu()kernel does a tlb flush before mmu gather table invalidate. The mmu gather table invalidate depending on kernel config also does another TLBI. Avoid the later ontlb_finish_mmu().That is already avoided, if you look at
tlb_flush_mmu_tlbonly()it does__tlb_range_reset(), which results in ->end = 0, which then triggers the early exit on the next invocation:
Quoted 5 lines
Quoted 3 lines
if (!tlb->end) return;Is that true for
tlb->fulmmflush?
Hmm, no, but I'm thinking you patch is broken, even for that case. We must issue the TLBI before call_rcu().
Perhaps if we replace !tlb->end with something like:
!tlb->freed_tables && !tlb->cleared_p*
(which GCC should be able to do with a single load and mask)
I've not really thought too hard about it yet, I need to run some errands, but I'll look at it more closely when I get back.
Peter Zijlstra
maintainerQuoted 12 lines · Aneesh Kumar K.V
Quoted 11 lines · Peter Zijlstra
Quoted 9 lines · Aneesh Kumar K.V
Architectures for which we have hardware walkers of Linux page table should flush TLB on mmu gather batch allocation failures and batch flush. Some architectures like POWER supports multiple translation modes (hash and radix) and in the case of POWER only radix translation mode needs the above TLBI. This is because for hash translation mode kernel wants to avoid this extra flush since there are no hardware walkers of linux page table. With radix translation, the hardware also walks linux page table and with that, kernel needs to make sure to TLB invalidate page walk cache before page table pages are freed.
Quoted 6 lines
Quoted 3 lines
Quoted 1 line
Based on changes from Peter Zijlstra <[email protected]>
AFAICT it is all my patch ;-)
Yes. I moved the changes you had to upstream. I can update the From: in the next version if you are ok with that?
Well, since PPC isn't broken per finding the invalidate in __p*_free_tlb(), lets do these things on top of the patches I proposed here. Also, you mnight want to run benchmarks to see if the movement of that TLBI actually helps (I'm thinking the cost of the PTESYNC might add up).
Peter Zijlstra
maintainerPATCH 1/1 asm-generic/tlb: Avoid potential double flush
In reply to Peter ZijlstraQuoted 8 lines · Peter Zijlstra
Perhaps if we replace !tlb->end with something like:
!tlb->freed_tables && !tlb->cleared_p*(which GCC should be able to do with a single load and mask)
I've not really thought too hard about it yet, I need to run some errands, but I'll look at it more closely when I get back.
AFAICT this should work.
---
Subject: asm-generic/tlb: Avoid potential double flush
Aneesh reported that:
tlb_flush_mmu() tlb_flush_mmu_tlbonly() tlb_flush() <-- #1 tlb_flush_mmu_free() tlb_table_flush() tlb_table_invalidate() tlb_flush_mmu_tlbonly() tlb_flush() <-- #2
does two TLBIs when tlb->fullmm, because __tlb_reset_range() will not clear tlb->end in that case.
Observe that any caller to __tlb_adjust_range() also sets at least one of the tlb->freed_tables || tlb->cleared_p* bits, and those are unconditionally cleared by __tlb_reset_range().
Change the condition for actually issuing TLBI to having one of those bits set, as opposed to having tlb->end != 0.
--- include/asm-generic/tlb.h | 7 ++++++- 1 file changed, 6 insertions(+), 1 deletion(-)
include/asm-generic/tlb.h +6 -1
@@ -402,7 +402,12 @@ tlb_update_vma_flags(struct mmu_gather *tlb, struct vm_area_struct *vma) { }static inline void tlb_flush_mmu_tlbonly (struct mmu_gather *tlb) { -if (!tlb->end) +/* +* Anything calling __tlb_adjust_range() also sets at least one of +* these bits. +*/ +if (!(tlb->freed_tables || tlb->cleared_ptes || tlb->cleared_pmds || + tlb->cleared_puds || tlb->cleared_p4ds))return ;tlb_flush (tlb);
Quoted 17 lines · Peter Zijlstra
Quoted 16 lines · Aneesh Kumar K.V
Quoted 14 lines · Peter Zijlstra
Quoted 11 lines · Aneesh Kumar K.V
Architectures for which we have hardware walkers of Linux page table should flush TLB on mmu gather batch allocation failures and batch flush. Some architectures like POWER supports multiple translation modes (hash and radix) and in the case of POWER only radix translation mode needs the above TLBI. This is because for hash translation mode kernel wants to avoid this extra flush since there are no hardware walkers of linux page table. With radix translation, the hardware also walks linux page table and with that, kernel needs to make sure to TLB invalidate page walk cache before page table pages are freed.
Based on changes from Peter Zijlstra <[email protected]>
AFAICT it is all my patch ;-)
Quoted 8 lines
Quoted 2 lines
Yes. I moved the changes you had to upstream. I can update the From: in the next version if you are ok with that?
Well, since PPC isn't broken per finding the invalidate in __p*_free_tlb(), lets do these things on top of the patches I proposed here. Also, you mnight want to run benchmarks to see if the movement of that TLBI actually helps (I'm thinking the cost of the PTESYNC might add up).
Upstream ppc64 is broken after the commit: a46cc7a90fd8 ("powerpc/mm/radix: Improve TLB/PWC flushes").
Also the patches are not adding any extra TLBI on either radix or hash.
Considering we need to backport this to stable and other distributions, how about we do this early patches in your series before the Kconfig rename? This should enable stable to pick them up with less dependencies.
-aneesh
Peter Zijlstra
maintainerQuoted 8 lines · Aneesh Kumar K.V
Upstream ppc64 is broken after the commit:
a46cc7a90fd8("powerpc/mm/radix: Improve TLB/PWC flushes").Also the patches are not adding any extra TLBI on either radix or hash.
Considering we need to backport this to stable and other distributions, how about we do this early patches in your series before the
Kconfigrename? This should enable stable to pick them up with less dependencies.
OK I suppose. Will you send a new series?