x86/tlb: fix assisted flush usage
PATCH v4 1/1 x86/tlb: fix assisted flush usage
Commit e9aca9470ed86 introduced a regression when avoiding sending IPIs for certain flush operations. Xen page fault handler (spurious_page_fault) relies on blocking interrupts in order to prevent handling TLB flush IPIs and thus preventing other CPUs from removing page tables pages. Switching to assisted flushing avoided such IPIs, and thus can result in pages belonging to the page tables being removed (and possibly re-used) while __page_fault_type is being executed.
Force some of the TLB flushes to use IPIs, thus avoiding the assisted TLB flush. Those selected flushes are the page type change (when switching from a page table type to a different one, ie: a page that has been removed as a page table) and page allocation. This sadly has a negative performance impact on the pvshim, as less assisted flushes can be used. Note the flush in grant-table code is also switched to use an IPI even when not strictly needed. This is done so that a common arch_flush_tlb_mask can be introduced and always used in common code.
Introduce a new flag (FLUSH_FORCE_IPI) and helper to force a TLB flush using an IPI (flush_tlb_mask_sync, x86 only). Note that the flag is only meaningfully defined when the hypervisor supports PV or shadow paging mode, as otherwise hardware assisted paging domains are in charge of their page tables and won't share page tables with Xen, thus not influencing the result of page walks performed by the spurious fault handler.
Just passing this new flag when calling flush_area_mask prevents the usage of the assisted flush without any other side effects.
Note the flag is not defined on Arm.
---
Changes since v3:
- Introduce arch_flush_tlb_mask.
Changes since v2:
- Always do a physical IPI triggered flush in filtered_flush_tlb_mask, since it's always required by the current callers of the function.
Changes since v1:
- Add a comment describing the usage of FLUSH_FORCE_IPI (and why no modifications to flush_area_mask are required).
- Use PGT_root_page_table instead of PGT_l4_page_table.
- Also perform IPI flushes if configured with shadow paging support.
- Use ifdef instead of if.
--- xen/arch/arm/smp.c | 2 +- xen/arch/x86/mm.c | 12 +++++++++++- xen/common/grant_table.c | 2 +- xen/include/asm-arm/flushtlb.h | 2 +- xen/include/asm-x86/flushtlb.h | 25 +++++++++++++++++++++++++ xen/include/xen/mm.h | 2 +- 6 files changed, 40 insertions(+), 5 deletions(-)
xen/arch/arm/smp.c +1 -1
@@ -5,7 +5,7 @@#include <asm/gic.h> #include <asm/flushtlb.h> -void flush_tlb_mask (const cpumask_t *mask) +void arch_flush_tlb_mask (const cpumask_t *mask) {/* No need to IPI other processors on ARM, the processor takes care of it. */ flush_all_guests_tlb ();
xen/arch/x86/mm.c +11 -1
@@ -2894,7 +2894,17 @@ static int _get_page_type(struct page_info *page, unsigned long type, ((nx & PGT_type_mask) == PGT_writable_page)) ) {perfc_incr (need_flush_tlb_flush); -flush_tlb_mask (mask); +if ( (x & PGT_type_mask) && + (x & PGT_type_mask) <= PGT_root_page_table ) +/* +* If page was a page table make sure the flush is +* performed using an IPI in order to avoid changing +* the type of a page table page under the feet of +* spurious_page_fault. +*/ +flush_tlb_mask_sync (mask); +else +flush_tlb_mask (mask); }/* We lose existing type and validity. */
xen/common/grant_table.c +1 -1
@@ -393,7 +393,7 @@ static inline void grant_write_unlock(struct grant_table *gt)static inline void gnttab_flush_tlb (const struct domain *d) {if ( !paging_mode_external (d) ) -flush_tlb_mask (d->dirty_cpumask); +arch_flush_tlb_mask (d->dirty_cpumask); } static inline unsigned int
xen/include/asm-arm/flushtlb.h +1 -1
@@ -26,7 +26,7 @@ static inline void page_set_tlbflush_timestamp(struct page_info *page)#endif /* Flush specified CPUs' TLBs */ -void flush_tlb_mask (const cpumask_t *mask); +void arch_flush_tlb_mask (const cpumask_t *mask);/* * Flush a range of VA's hypervisor mappings from the TLB of the local
xen/include/asm-x86/flushtlb.h +25 -0
@@ -126,6 +126,16 @@ void switch_cr3_cr4(unsigned long cr3, unsigned long cr4);#else #define FLUSH_HVM_ASID_CORE 0#endif +#if defined(CONFIG_PV ) || defined(CONFIG_SHADOW_PAGING ) +/* +* Force an IPI to be sent. Note that adding this to the flags passed to +* flush_area_mask will prevent using the assisted flush without having any +* other side effect. +*/ +# define FLUSH_FORCE_IPI 0x8000 +#else +# define FLUSH_FORCE_IPI 0 +#endif /* Flush local TLBs/caches. */ unsigned int flush_area_local (const void *va,unsigned int flags);
@@ -148,9 +158,24 @@ void flush_area_mask(const cpumask_t *, const void *va, unsigned int flags);/* Flush specified CPUs' TLBs */ #define flush_tlb_mask (mask) \ flush_mask(mask, FLUSH_TLB) +/* +* Flush specified CPUs' TLBs and force the usage of an IPI to do so. This is +* required for certain operations that rely on page tables themselves not +* being freed and reused when interrupts are blocked, as the flush IPI won't +* be fulfilled until exiting from that critical region. +*/ +#define flush_tlb_mask_sync (mask) \ + flush_mask(mask, FLUSH_TLB | FLUSH_FORCE_IPI)#define flush_tlb_one_mask (mask,v) \ flush_area_mask(mask, (const void *)(v), FLUSH_TLB|FLUSH_ORDER(0)) +/* +* Alias the common code TLB flush helper to the sync one in order to be on the +* safe side. Note that not all calls from common code strictly require the +* _sync variant. +*/ +#define arch_flush_tlb_mask flush_tlb_mask_sync +/* Flush all CPUs' TLBs */ #define flush_tlb_all () \ flush_tlb_mask(&cpu_online_map)
xen/include/xen/mm.h +1 -1
@@ -648,7 +648,7 @@ static inline void filtered_flush_tlb_mask(uint32_t tlbflush_timestamp)if ( !cpumask_empty (&mask) ) {perfc_incr (need_flush_tlb_flush); -flush_tlb_mask (&mask); +arch_flush_tlb_mask (&mask); } }
Signature · 1 line
2.26.2
Quoted 27 lines · Roger Pau Monne
Commit
e9aca9470ed86introduced a regression when avoiding sending IPIs for certain flush operations. Xen page fault handler (spurious_page_fault) relies on blocking interrupts in order to prevent handling TLB flush IPIs and thus preventing other CPUs from removing page tables pages. Switching to assisted flushing avoided such IPIs, and thus can result in pages belonging to the page tables being removed (and possibly re-used) while__page_fault_typeis being executed.Force some of the TLB flushes to use IPIs, thus avoiding the assisted TLB flush. Those selected flushes are the page type change (when switching from a page table type to a different one, ie: a page that has been removed as a page table) and page allocation. This sadly has a negative performance impact on the pvshim, as less assisted flushes can be used. Note the flush in grant-table code is also switched to use an IPI even when not strictly needed. This is done so that a common
arch_flush_tlb_maskcan be introduced and always used in common code.Introduce a new flag (
FLUSH_FORCE_IPI) and helper to force a TLB flush using an IPI (flush_tlb_mask_sync, x86 only). Note that the flag is only meaningfully defined when the hypervisor supports PV or shadow paging mode, as otherwise hardware assisted paging domains are in charge of their page tables and won't share page tables with Xen, thus not influencing the result of page walks performed by the spurious fault handler.
Quoted 8 lines
Just passing this new flag when calling
flush_area_maskprevents the usage of the assisted flush without any other side effects.Note the flag is not defined on Arm.
Fixes e9aca9470ed86 ('x86/tlb: use Xen L0 assisted TLB flush when available')Reported-by Andrew Cooper <[email protected]>Signed-off-by Roger Pau Monné <[email protected]>
In principle
A few cosmetic remarks though:
Quoted 8 lines
xen/arch/x86/mm.c +1
-1@@ -2894,7 +2894,17 @@ static int _get_page_type(struct page_info *page, unsigned long type, ((nx & PGT_type_mask) == PGT_writable_page)) ) {perfc_incr (need_flush_tlb_flush); -flush_tlb_mask (mask); + if ( (x & PGT_type_mask) &&
xen/arch/x86/mm.c static int _get_page_type(struct page_info *page, unsigned long type, · line 2907
Quoted 10 lines
xen/arch/x86/mm.c +10
-0+ (x & PGT_type_mask) <= PGT_root_page_table ) +/* +* If page was a page table make sure the flush is +* performed using an IPI in order to avoid changing +* the type of a page table page under the feet of +* spurious_page_fault. +*/ +flush_tlb_mask_sync (mask); + else +flush_tlb_mask (mask );
Effectively this now is the only user of the new macro. I'd prefer avoiding its introduction (and hence avoiding the questionable "_sync" suffix), doing
flush_mask(mask, FLUSH_TLB | (... ? FLUSH_FORCE_IPI : 0));
here and ...
Quoted 11 lines
xen/arch/x86/mm.c +7
-0@@ -148,9 +158,24 @@ void flush_area_mask(const cpumask_t *, const void *va, unsigned int flags);/* Flush specified CPUs' TLBs */ #define flush_tlb_mask (mask) \ flush_mask(mask, FLUSH_TLB) +/* +* Flush specified CPUs' TLBs and force the usage of an IPI to do so. This is +* required for certain operations that rely on page tables themselves not +* being freed and reused when interrupts are blocked, as the flush IPI won't +* be fulfilled until exiting from that critical region. +*/ +#define flush_tlb_mask_sync (mask) \
xen/arch/x86/mm.c void flush_area_mask(const cpumask_t *, const void *va, unsigned int flags); · line 177
Quoted 10 lines
xen/arch/x86/mm.c +7
-0+ flush_mask(mask, FLUSH_TLB | FLUSH_FORCE_IPI)#define flush_tlb_one_mask (mask,v) \ flush_area_mask(mask, (const void *)(v), FLUSH_TLB|FLUSH_ORDER(0)) +/* +* Alias the common code TLB flush helper to the sync one in order to be on the +* safe side. Note that not all calls from common code strictly require the +* _sync variant. +*/ +#define arch_flush_tlb_mask flush_tlb_mask_sync
...
#define arch_flush_tlb_mask(mask) \
flush_mask(mask, FLUSH_TLB | FLUSH_FORCE_IPI)
here. I'd be okay making these adjustments while committing, if you and others don't object.
Jan
Quoted 64 lines · Jan Beulich
Quoted 27 lines · Roger Pau Monne
Commit
e9aca9470ed86introduced a regression when avoiding sending IPIs for certain flush operations. Xen page fault handler (spurious_page_fault) relies on blocking interrupts in order to prevent handling TLB flush IPIs and thus preventing other CPUs from removing page tables pages. Switching to assisted flushing avoided such IPIs, and thus can result in pages belonging to the page tables being removed (and possibly re-used) while__page_fault_typeis being executed.Force some of the TLB flushes to use IPIs, thus avoiding the assisted TLB flush. Those selected flushes are the page type change (when switching from a page table type to a different one, ie: a page that has been removed as a page table) and page allocation. This sadly has a negative performance impact on the pvshim, as less assisted flushes can be used. Note the flush in grant-table code is also switched to use an IPI even when not strictly needed. This is done so that a common
arch_flush_tlb_maskcan be introduced and always used in common code.Introduce a new flag (
FLUSH_FORCE_IPI) and helper to force a TLB flush using an IPI (flush_tlb_mask_sync, x86 only). Note that the flag is only meaningfully defined when the hypervisor supports PV or shadow paging mode, as otherwise hardware assisted paging domains are in charge of their page tables and won't share page tables with Xen, thus not influencing the result of page walks performed by the spurious fault handler.Quoted 8 lines
Just passing this new flag when calling
flush_area_maskprevents the usage of the assisted flush without any other side effects.Note the flag is not defined on Arm.
Fixes e9aca9470ed86 ('x86/tlb: use Xen L0 assisted TLB flush when available')Reported-by Andrew Cooper <[email protected]>Signed-off-by Roger Pau Monné <[email protected]>In principle
Reviewed-by Jan Beulich <[email protected]>A few cosmetic remarks though:
Quoted 8 lines
xen/arch/x86/mm.c +1
-1@@ -2894,7 +2894,17 @@ static int _get_page_type(struct page_info *page, unsigned long type, ((nx & PGT_type_mask) == PGT_writable_page)) ) {perfc_incr (need_flush_tlb_flush); -flush_tlb_mask (mask); + if ( (x & PGT_type_mask) &&xen/arch/x86/mm.c static int _get_page_type(struct page_info *page, unsigned long type, · line 2907
Quoted 10 lines
xen/arch/x86/mm.c +10
-0+ (x & PGT_type_mask) <= PGT_root_page_table ) +/* +* If page was a page table make sure the flush is +* performed using an IPI in order to avoid changing +* the type of a page table page under the feet of +* spurious_page_fault. +*/ +flush_tlb_mask_sync (mask); + else +flush_tlb_mask (mask );Effectively this now is the only user of the new macro. I'd prefer avoiding its introduction (and hence avoiding the questionable "
_sync" suffix), doing
Quoted 1 line
flush_mask(mask, FLUSH_TLB | (... ? FLUSH_FORCE_IPI : 0));
Right, maybe placing the '(x & PGT_type_mask) && (x & PGT_type_mask) <= PGT_root_page_table' condition inside the parameter list of flush_mask will make the code hard to read, so it might be worth to keep the if?
Quoted 26 lines
here and ...
Quoted 21 lines
xen/arch/x86/mm.c +14
-0@@ -148,9 +158,24 @@ void flush_area_mask(const cpumask_t *, const void *va, unsigned int flags);/* Flush specified CPUs' TLBs */ #define flush_tlb_mask (mask) \ flush_mask(mask, FLUSH_TLB) +/* +* Flush specified CPUs' TLBs and force the usage of an IPI to do so. This is +* required for certain operations that rely on page tables themselves not +* being freed and reused when interrupts are blocked, as the flush IPI won't +* be fulfilled until exiting from that critical region. +*/ +#define flush_tlb_mask_sync (mask) \ + flush_mask(mask, FLUSH_TLB | FLUSH_FORCE_IPI)#define flush_tlb_one_mask (mask,v) \ flush_area_mask(mask, (const void *)(v), FLUSH_TLB|FLUSH_ORDER(0)) +/* +* Alias the common code TLB flush helper to the sync one in order to be on the +* safe side. Note that not all calls from common code strictly require the +* _sync variant. +*/ +#define arch_flush_tlb_mask flush_tlb_mask_sync...
Quoted 2 lines
#define
arch_flush_tlb_mask(mask)\
flush_mask(mask, FLUSH_TLB | FLUSH_FORCE_IPI)
Sure. Feel free to slightly adjust the comment, I think doing s/Alias/Force/ would be enough.
Quoted 2 lines
here. I'd be okay making these adjustments while committing, if you and others don't object.
That's fine, I leave to your judgment whether to use the ternary operator in the _get_page_type case.
Roger.
Hi Roger,
Quoted 27 lines · Roger Pau Monne
Commit
e9aca9470ed86introduced a regression when avoiding sending IPIs for certain flush operations. Xen page fault handler (spurious_page_fault) relies on blocking interrupts in order to prevent handling TLB flush IPIs and thus preventing other CPUs from removing page tables pages. Switching to assisted flushing avoided such IPIs, and thus can result in pages belonging to the page tables being removed (and possibly re-used) while__page_fault_typeis being executed.Force some of the TLB flushes to use IPIs, thus avoiding the assisted TLB flush. Those selected flushes are the page type change (when switching from a page table type to a different one, ie: a page that has been removed as a page table) and page allocation. This sadly has a negative performance impact on the pvshim, as less assisted flushes can be used. Note the flush in grant-table code is also switched to use an IPI even when not strictly needed. This is done so that a common
arch_flush_tlb_maskcan be introduced and always used in common code.Introduce a new flag (
FLUSH_FORCE_IPI) and helper to force a TLB flush using an IPI (flush_tlb_mask_sync, x86 only). Note that the flag is only meaningfully defined when the hypervisor supports PV or shadow paging mode, as otherwise hardware assisted paging domains are in charge of their page tables and won't share page tables with Xen, thus not influencing the result of page walks performed by the spurious fault handler.
Quoted 8 lines
Just passing this new flag when calling
flush_area_maskprevents the usage of the assisted flush without any other side effects.Note the flag is not defined on Arm.
Fixes e9aca9470ed86 ('x86/tlb: use Xen L0 assisted TLB flush when available')Reported-by Andrew Cooper <[email protected]>Signed-off-by Roger Pau Monné <[email protected]>
Cheers,
Signature · 1 line
Julien Grall
Quoted 43 lines
-----Original Message-----
From: Julien Grall <[email protected]>
Sent: 30 June 2020 13:48
To: Roger Pau Monne <[email protected]>; [email protected]
Cc: [email protected]; Stefano Stabellini <[email protected]>; Volodymyr Babchuk <[email protected]>; Andrew Cooper <[email protected]>; George Dunlap <[email protected]>; Ian Jackson <[email protected]>; Jan Beulich <[email protected]>; Wei Liu <[email protected]>
Subject: Re: [PATCH for-4.14 v4] x86/tlb: fix assisted flush usageHi Roger,
Quoted 30 lines · Roger Pau Monne
Commit
e9aca9470ed86introduced a regression when avoiding sending IPIs for certain flush operations. Xen page fault handler (spurious_page_fault) relies on blocking interrupts in order to prevent handling TLB flush IPIs and thus preventing other CPUs from removing page tables pages. Switching to assisted flushing avoided such IPIs, and thus can result in pages belonging to the page tables being removed (and possibly re-used) while__page_fault_typeis being executed.Force some of the TLB flushes to use IPIs, thus avoiding the assisted TLB flush. Those selected flushes are the page type change (when switching from a page table type to a different one, ie: a page that has been removed as a page table) and page allocation. This sadly has a negative performance impact on the pvshim, as less assisted flushes can be used. Note the flush in grant-table code is also switched to use an IPI even when not strictly needed. This is done so that a common
arch_flush_tlb_maskcan be introduced and always used in common code.Introduce a new flag (
FLUSH_FORCE_IPI) and helper to force a TLB flush using an IPI (flush_tlb_mask_sync, x86 only). Note that the flag is only meaningfully defined when the hypervisor supports PV or shadow paging mode, as otherwise hardware assisted paging domains are in charge of their page tables and won't share page tables with Xen, thus not influencing the result of page walks performed by the spurious fault handler.Just passing this new flag when calling
flush_area_maskprevents the usage of the assisted flush without any other side effects.
Quoted 8 lines
Quoted 5 lines
Note the flag is not defined on Arm.
Fixes e9aca9470ed86 ('x86/tlb: use Xen L0 assisted TLB flush when available')Reported-by Andrew Cooper <[email protected]>Signed-off-by Roger Pau Monné <[email protected]>Acked-by Julien Grall <[email protected]>
And...