kasan: Fix ordering between MTE tag colouring and page->flags
Subsystems (6)
Catalin Marinas
maintainerv2 cover kasan: Fix ordering between MTE tag colouring and page->flags
Hi,
That's a second attempt on fixing the race race between setting the allocation (in-memory) tags in a page and the corresponding logical tag in page->flags. Initial version here:
https://lore.kernel.org/r/[email protected]
This new series does not introduce any new GFP flags but instead always skips unpoisoning of the user pages (we already skip the poisoning on free). Any unpoisoned page will have the page->flags tag reset.
For the background:
On a system with MTE and KASAN_HW_TAGS enabled, when a page is allocated kasan_unpoison_pages() sets a random tag and saves it in page->flags so that page_to_virt() re-creates the correct tagged pointer. We need to ensure that the in-memory tags are visible before setting the page->flags:
P0 (__kasan_unpoison_range): P1 (access via virt_to_page):
Wtags=x Rflags=x
| |
| DMB | address dependency
V V
Wflags=x Rtags=xThe first patch changes the order of page unpoisoning with the tag storing in page->flags. page_kasan_tag_set() has the right barriers through try_cmpxchg().
If a page is mapped in user-space with PROT_MTE, the architecture code will set the allocation tag to 0 and a subsequent page_to_virt() dereference will fault. We currently try to fix this by resetting the tag in page->flags so that it is 0xff (match-all, not faulting). However, setting the tags and flags can race with another CPU reading the flags (page_to_virt()) and barriers can't help, e.g.:
P0 (mte_sync_page_tags): P1 (memcpy from virt_to_page):
Rflags!=0xff
Wflags=0xff
DMB (doesn't help)
Wtags=0
Rtags=0 // fault
Since clearing the flags in the arch code doesn't work, to do this at page allocation time when __GFP_SKIP_KASAN_UNPOISON is passed.
Thanks.
Catalin Marinas (4):
mm: kasan: Ensure the tags are visible before the tag in page->flags
mm: kasan: Skip unpoisoning of user pages
mm: kasan: Skip page unpoisoning only if __GFP_SKIP_KASAN_UNPOISON
arm64: kasan: Revert "arm64: mte: reset the page tag in page->flags"
arch/arm64/kernel/hibernate.c | 5 ----- arch/arm64/kernel/mte.c | 9 --------- arch/arm64/mm/copypage.c | 9 --------- arch/arm64/mm/fault.c | 1 - arch/arm64/mm/mteswap.c | 9 --------- include/linux/gfp.h | 2 +- mm/kasan/common.c | 3 ++- mm/page_alloc.c | 19 ++++++++++--------- 8 files changed, 13 insertions(+), 44 deletions(-)
Catalin Marinas
maintainerPATCH v2 1/4 mm: kasan: Ensure the tags are visible before the tag in page->flags
__kasan_unpoison_pages() colours the memory with a random tag and stores it in page->flags in order to re-create the tagged pointer via page_to_virt() later. When the tag from the page->flags is read, ensure that the in-memory tags are already visible by re-ordering the page_kasan_tag_set() after kasan_unpoison(). The former already has barriers in place through try_cmpxchg(). On the reader side, the order is ensured by the address dependency between page->flags and the memory access.
--- mm/kasan/common.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)
mm/kasan/common.c +2 -1
@@ -108,9 +108,10 @@ void __kasan_unpoison_pages(struct page *page, unsigned int order, bool init)return ; tag =kasan_random_tag (); +kasan_unpoison (set_tag (page_address (page), tag), +PAGE_SIZE << order, init);for (i =0 ; i < (1 << order); i++)page_kasan_tag_set (page + i, tag); -kasan_unpoison (page_address (page),PAGE_SIZE << order, init); }void __kasan_poison_pages (struct page *page,unsigned int order,bool init)
Re: [PATCH v2 1/4] mm: kasan: Ensure the tags are visible before the tag in page->flags
Quoted 9 lines · Catalin Marinas
__kasan_unpoison_pages()colours the memory with a random tag and stores it inpage->flagsin order to re-create the tagged pointer viapage_to_virt()later. When the tag from thepage->flagsis read, ensure that the in-memory tags are already visible by re-ordering thepage_kasan_tag_set()afterkasan_unpoison(). The former already has barriers in place throughtry_cmpxchg(). On the reader side, the order is ensured by the address dependency betweenpage->flagsand the memory access.
Quoted 4 lines
Signed-off-by Catalin Marinas <[email protected]>Reviewed-by Andrey Konovalov <[email protected]>Cc Andrey Ryabinin <[email protected]>Cc Vincenzo Frascino <[email protected]>
Quoted 20 lines
--- mm/kasan/common.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)mm/kasan/common.c +2
-1@@ -108,9 +108,10 @@ void __kasan_unpoison_pages(struct page *page, unsigned int order, bool init)return ; tag =kasan_random_tag (); +kasan_unpoison (set_tag (page_address (page), tag), +PAGE_SIZE << order, init);for (i =0 ; i < (1 << order); i++)page_kasan_tag_set (page + i, tag); -kasan_unpoison (page_address (page),PAGE_SIZE << order, init); }void __kasan_poison_pages (struct page *page,unsigned int order,bool init)
Signature · 2 lines
Regards, Vincenzo
Will Deacon
maintainerRe: [PATCH v2 1/4] mm: kasan: Ensure the tags are visible before the tag in page->flags
In reply to Catalin MarinasQuoted 22 lines · Catalin Marinas
__kasan_unpoison_pages()colours the memory with a random tag and stores it inpage->flagsin order to re-create the tagged pointer viapage_to_virt()later. When the tag from thepage->flagsis read, ensure that the in-memory tags are already visible by re-ordering thepage_kasan_tag_set()afterkasan_unpoison(). The former already has barriers in place throughtry_cmpxchg(). On the reader side, the order is ensured by the address dependency betweenpage->flagsand the memory access.Signed-off-by Catalin Marinas <[email protected]>Reviewed-by Andrey Konovalov <[email protected]>Cc Andrey Ryabinin <[email protected]>Cc Vincenzo Frascino <[email protected]>--- mm/kasan/common.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)mm/kasan/common.c +0
-0
mm/kasan/common.c void __kasan_unpoison_pages(struct page *page, unsigned int order, bool init) · line 113
Quoted 8 lines
mm/kasan/common.c +2
-1@@ -108,9 +108,10 @@ void __kasan_unpoison_pages(struct page *page, unsigned int order, bool init)return ; tag =kasan_random_tag (); +kasan_unpoison (set_tag (page_address (page), tag), +PAGE_SIZE << order, init);for (i =0 ; i < (1 << order); i++)page_kasan_tag_set (page + i, tag); -kasan_unpoison (page_address (page),PAGE_SIZE << order, init);
This looks good to me, but after reading the cover letter I'm wondering whether the try_cmpxchg() in page_kasan_tag_set() could be relaxed to try_cmpxchg_release() as a separate optimisation?
Will
Catalin Marinas
maintainerRe: [PATCH v2 1/4] mm: kasan: Ensure the tags are visible before the tag in page->flags
Quoted 25 lines · Will Deacon
Quoted 24 lines · Catalin Marinas
__kasan_unpoison_pages()colours the memory with a random tag and stores it inpage->flagsin order to re-create the tagged pointer viapage_to_virt()later. When the tag from thepage->flagsis read, ensure that the in-memory tags are already visible by re-ordering thepage_kasan_tag_set()afterkasan_unpoison(). The former already has barriers in place throughtry_cmpxchg(). On the reader side, the order is ensured by the address dependency betweenpage->flagsand the memory access.Signed-off-by Catalin Marinas <[email protected]>Reviewed-by Andrey Konovalov <[email protected]>Cc Andrey Ryabinin <[email protected]>Cc Vincenzo Frascino <[email protected]>--- mm/kasan/common.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)mm/kasan/common.c +0
-0@@ -108,9 +108,10 @@ void __kasan_unpoison_pages(struct page *page, unsigned int order, bool init)return ;
Quoted 10 lines
mm/kasan/common.c void __kasan_unpoison_pages(struct page *page, unsigned int order, bool init) · line 113
Quoted 6 lines
mm/kasan/common.c +2
-1tag =kasan_random_tag (); +kasan_unpoison (set_tag (page_address (page), tag), +PAGE_SIZE << order, init);for (i =0 ; i < (1 << order); i++)page_kasan_tag_set (page + i, tag); -kasan_unpoison (page_address (page),PAGE_SIZE << order, init);This looks good to me, but after reading the cover letter I'm wondering whether the
try_cmpxchg()inpage_kasan_tag_set()could be relaxed totry_cmpxchg_release()as a separate optimisation?
I think it can be a try_cmpxchg_release() (I did not realise we have one). I'll post a patch later today.
Signature · 1 line
Catalin
Catalin Marinas
maintainerPATCH v2 2/4 mm: kasan: Skip unpoisoning of user pages
In reply to Catalin MarinasCommit c275c5c6d50a ("kasan: disable freed user page poisoning with HW tags") added __GFP_SKIP_KASAN_POISON to GFP_HIGHUSER_MOVABLE. A similar argument can be made about unpoisoning, so also add __GFP_SKIP_KASAN_UNPOISON to user pages. To ensure the user page is still accessible via page_address() without a kasan fault, reset the page->flags tag.
With the above changes, there is no need for the arm64 tag_clear_highpage() to reset the page->flags tag.
--- arch/arm64/mm/fault.c | 1 - include/linux/gfp.h | 2 +- mm/page_alloc.c | 7 +++++-- 3 files changed, 6 insertions(+), 4 deletions(-)
arch/arm64/mm/fault.c +0 -1
@@ -927,6 +927,5 @@ struct page *alloc_zeroed_user_highpage_movable(struct vm_area_struct *vma,void tag_clear_highpage (struct page *page) {mte_zero_clear_page_tags (page_address (page)); -page_kasan_tag_reset (page);set_bit (PG_mte_tagged, &page->flags); }
include/linux/gfp.h +1 -1
@@ -348,7 +348,7 @@ struct vm_area_struct;#define GFP_DMA32 __GFP_DMA32#define GFP_HIGHUSER (GFP_USER | __GFP_HIGHMEM)#define GFP_HIGHUSER_MOVABLE (GFP_HIGHUSER | __GFP_MOVABLE | \ - __GFP_SKIP_KASAN_POISON) + __GFP_SKIP_KASAN_POISON | __GFP_SKIP_KASAN_UNPOISON)#define GFP_TRANSHUGE_LIGHT ((GFP_HIGHUSER_MOVABLE | __GFP_COMP | \ __GFP_NOMEMALLOC | __GFP_NOWARN) & ~__GFP_RECLAIM)#define GFP_TRANSHUGE (GFP_TRANSHUGE_LIGHT | __GFP_DIRECT_RECLAIM)
mm/page_alloc.c +5 -2
@@ -2397,6 +2397,7 @@ inline void post_alloc_hook(struct page *page, unsigned int order,bool init = !want_init_on_free () &&want_init_on_alloc (gfp_flags) && !should_skip_init (gfp_flags);bool init_tags = init && (gfp_flags & __GFP_ZEROTAGS); +int i;set_page_private (page,0 );set_page_refcounted (page);
@@ -2422,8 +2423,6 @@ inline void post_alloc_hook(struct page *page, unsigned int order,* should be initialized as well). */ if (init_tags) { -int i; -/* Initialize both memory and tags. */ for (i =0 ; i !=1 << order; ++i)tag_clear_highpage (page + i);
@@ -2438,6 +2437,10 @@ inline void post_alloc_hook(struct page *page, unsigned int order,/* Note that memory is already initialized by KASAN. */ if (kasan_has_integrated_init ()) init = false; + }else { +/* Ensure page_address() dereferencing does not fault. */ +for (i =0 ; i !=1 << order; ++i) +page_kasan_tag_reset (page + i); }/* If memory is still not initialized, do it now. */ if (init)
Re: [PATCH v2 2/4] mm: kasan: Skip unpoisoning of user pages
Quoted 68 lines · Catalin Marinas
Commit
c275c5c6d50a("kasan: disable freed user page poisoning with HW tags") added__GFP_SKIP_KASAN_POISONtoGFP_HIGHUSER_MOVABLE. A similar argument can be made about unpoisoning, so also add__GFP_SKIP_KASAN_UNPOISONto user pages. To ensure the user page is still accessible viapage_address()without a kasan fault, reset thepage->flagstag.With the above changes, there is no need for the arm64
tag_clear_highpage()to reset thepage->flagstag.Signed-off-by Catalin Marinas <[email protected]>Cc Andrey Ryabinin <[email protected]>Cc Andrey Konovalov <[email protected]>Cc Peter Collingbourne <[email protected]>Cc Vincenzo Frascino <[email protected]>--- arch/arm64/mm/fault.c | 1 - include/linux/gfp.h | 2 +- mm/page_alloc.c | 7 +++++-- 3 files changed, 6 insertions(+), 4 deletions(-)arch/arm64/mm/fault.c +0
-1@@ -927,6 +927,5 @@ struct page *alloc_zeroed_user_highpage_movable(struct vm_area_struct *vma,void tag_clear_highpage (struct page *page) {mte_zero_clear_page_tags (page_address (page)); -page_kasan_tag_reset (page);set_bit (PG_mte_tagged, &page->flags); }include/linux/gfp.h +1
-1@@ -348,7 +348,7 @@ struct vm_area_struct;#define GFP_DMA32 __GFP_DMA32#define GFP_HIGHUSER (GFP_USER | __GFP_HIGHMEM)#define GFP_HIGHUSER_MOVABLE (GFP_HIGHUSER | __GFP_MOVABLE | \ - __GFP_SKIP_KASAN_POISON) + __GFP_SKIP_KASAN_POISON | __GFP_SKIP_KASAN_UNPOISON)#define GFP_TRANSHUGE_LIGHT ((GFP_HIGHUSER_MOVABLE | __GFP_COMP | \ __GFP_NOMEMALLOC | __GFP_NOWARN) & ~__GFP_RECLAIM)#define GFP_TRANSHUGE (GFP_TRANSHUGE_LIGHT | __GFP_DIRECT_RECLAIM)mm/page_alloc.c +1
-2@@ -2397,6 +2397,7 @@ inline void post_alloc_hook(struct page *page, unsigned int order,bool init = !want_init_on_free () &&want_init_on_alloc (gfp_flags) && !should_skip_init (gfp_flags);bool init_tags = init && (gfp_flags & __GFP_ZEROTAGS); +int i;set_page_private (page,0 );set_page_refcounted (page);@@ -2422,8 +2423,6 @@ inline void post_alloc_hook(struct page *page, unsigned int order,* should be initialized as well). */ if (init_tags) { -int i; -/* Initialize both memory and tags. */ for (i =0 ; i !=1 << order; ++i)tag_clear_highpage (page + i);
mm/page_alloc.c inline void post_alloc_hook(struct page *page, unsigned int order, · line 2446
Quoted 10 lines
mm/page_alloc.c +4
-0@@ -2438,6 +2437,10 @@ inline void post_alloc_hook(struct page *page, unsigned int order,/* Note that memory is already initialized by KASAN. */ if (kasan_has_integrated_init ()) init = false; + }else { +/* Ensure page_address() dereferencing does not fault. */ +for (i =0 ; i !=1 << order; ++i) +page_kasan_tag_reset (page + i); }/* If memory is still not initialized, do it now. */ if (init)
Re: [PATCH v2 2/4] mm: kasan: Skip unpoisoning of user pages
In reply to Catalin MarinasQuoted 50 lines · Catalin Marinas
Commit
c275c5c6d50a("kasan: disable freed user page poisoning with HW tags") added__GFP_SKIP_KASAN_POISONtoGFP_HIGHUSER_MOVABLE. A similar argument can be made about unpoisoning, so also add__GFP_SKIP_KASAN_UNPOISONto user pages. To ensure the user page is still accessible viapage_address()without a kasan fault, reset thepage->flagstag.With the above changes, there is no need for the arm64
tag_clear_highpage()to reset thepage->flagstag.Signed-off-by Catalin Marinas <[email protected]>Cc Andrey Ryabinin <[email protected]>Cc Andrey Konovalov <[email protected]>Cc Peter Collingbourne <[email protected]>Cc Vincenzo Frascino <[email protected]>--- arch/arm64/mm/fault.c | 1 - include/linux/gfp.h | 2 +- mm/page_alloc.c | 7 +++++-- 3 files changed, 6 insertions(+), 4 deletions(-)arch/arm64/mm/fault.c +0
-1@@ -927,6 +927,5 @@ struct page *alloc_zeroed_user_highpage_movable(struct vm_area_struct *vma,void tag_clear_highpage (struct page *page) {mte_zero_clear_page_tags (page_address (page)); -page_kasan_tag_reset (page);set_bit (PG_mte_tagged, &page->flags); }include/linux/gfp.h +1
-1@@ -348,7 +348,7 @@ struct vm_area_struct;#define GFP_DMA32 __GFP_DMA32#define GFP_HIGHUSER (GFP_USER | __GFP_HIGHMEM)#define GFP_HIGHUSER_MOVABLE (GFP_HIGHUSER | __GFP_MOVABLE | \ - __GFP_SKIP_KASAN_POISON) + __GFP_SKIP_KASAN_POISON | __GFP_SKIP_KASAN_UNPOISON)#define GFP_TRANSHUGE_LIGHT ((GFP_HIGHUSER_MOVABLE | __GFP_COMP | \ __GFP_NOMEMALLOC | __GFP_NOWARN) & ~__GFP_RECLAIM)#define GFP_TRANSHUGE (GFP_TRANSHUGE_LIGHT | __GFP_DIRECT_RECLAIM)mm/page_alloc.c +0
-0
mm/page_alloc.c inline void post_alloc_hook(struct page *page, unsigned int order, · line 2401
Quoted 5 lines
mm/page_alloc.c +1
-0@@ -2397,6 +2397,7 @@ inline void post_alloc_hook(struct page *page, unsigned int order,bool init = !want_init_on_free () &&want_init_on_alloc (gfp_flags) && !should_skip_init (gfp_flags);bool init_tags = init && (gfp_flags & __GFP_ZEROTAGS); +int i;
Nit: Since "i" is not used outside of the for loop context we could use the contract form "for (int i = 0; ..." which is allowed by C11.
Quoted 12 lines
mm/page_alloc.c +0
-2set_page_private (page,0 );set_page_refcounted (page);@@ -2422,8 +2423,6 @@ inline void post_alloc_hook(struct page *page, unsigned int order,* should be initialized as well). */ if (init_tags) { -int i; -/* Initialize both memory and tags. */ for (i =0 ; i !=1 << order; ++i)tag_clear_highpage (page + i);
mm/page_alloc.c inline void post_alloc_hook(struct page *page, unsigned int order, · line 2446
Quoted 10 lines
mm/page_alloc.c +4
-0@@ -2438,6 +2437,10 @@ inline void post_alloc_hook(struct page *page, unsigned int order,/* Note that memory is already initialized by KASAN. */ if (kasan_has_integrated_init ()) init = false; + }else { +/* Ensure page_address() dereferencing does not fault. */ +for (i =0 ; i !=1 << order; ++i) +page_kasan_tag_reset (page + i); }/* If memory is still not initialized, do it now. */ if (init)
Either way:
Signature · 2 lines
Regards, Vincenzo
Catalin Marinas
maintainerRe: [PATCH v2 2/4] mm: kasan: Skip unpoisoning of user pages
Quoted 14 lines · Vincenzo Frascino
mm/page_alloc.c inline void post_alloc_hook(struct page *page, unsigned int order, · line 2401
Quoted 10 lines · Catalin Marinas
mm/page_alloc.c +1
-0@@ -2397,6 +2397,7 @@ inline void post_alloc_hook(struct page *page, unsigned int order,bool init = !want_init_on_free () &&want_init_on_alloc (gfp_flags) && !should_skip_init (gfp_flags);bool init_tags = init && (gfp_flags & __GFP_ZEROTAGS); +int i;Nit: Since "i" is not used outside of the for loop context we could use the contract form "for (int i = 0; ..." which is allowed by C11.
Oh yeah, Linux moved to C11 in 5.18. But it looks better to be consistent with the other for loop in this file.
Quoted 1 line
Reviewed-by Vincenzo Frascino <[email protected]>
Thanks.
Signature · 1 line
Catalin
Catalin Marinas
maintainerPATCH v2 3/4 mm: kasan: Skip page unpoisoning only if __GFP_SKIP_KASAN_UNPOISON
In reply to Catalin MarinasCurrently post_alloc_hook() skips the kasan unpoisoning if the tags will be zeroed (__GFP_ZEROTAGS) or __GFP_SKIP_KASAN_UNPOISON is passed. Since __GFP_ZEROTAGS is now accompanied by __GFP_SKIP_KASAN_UNPOISON, remove the extra check.
--- mm/page_alloc.c | 12 +++++------- 1 file changed, 5 insertions(+), 7 deletions(-)
mm/page_alloc.c +5 -7
@@ -2361,7 +2361,7 @@ static inline bool check_new_pcp(struct page *page, unsigned int order) }#endif /* CONFIG_DEBUG_VM */ -static inline bool should_skip_kasan_unpoison (gfp_t flags,bool init_tags) +static inline bool should_skip_kasan_unpoison (gfp_t flags) {/* Don't skip if a software KASAN mode is enabled. */ if (IS_ENABLED(CONFIG_KASAN_GENERIC) ||
@@ -2373,12 +2373,10 @@ static inline bool should_skip_kasan_unpoison(gfp_t flags, bool init_tags)return true;/* -* With hardware tag-based KASAN enabled, skip if either: -* -* 1. Memory tags have already been cleared via tag_clear_highpage(). -* 2. Skipping has been requested via __GFP_SKIP_KASAN_UNPOISON. +* With hardware tag-based KASAN enabled, skip if this has been +* requested via __GFP_SKIP_KASAN_UNPOISON. */ -return init_tags || (flags & __GFP_SKIP_KASAN_UNPOISON); +return flags & __GFP_SKIP_KASAN_UNPOISON; }static inline bool should_skip_init (gfp_t flags)
@@ -2430,7 +2428,7 @@ inline void post_alloc_hook(struct page *page, unsigned int order,/* Note that memory is already initialized by the loop above. */ init = false; } -if (!should_skip_kasan_unpoison (gfp_flags, init_tags)) { +if (!should_skip_kasan_unpoison (gfp_flags)) {/* Unpoison shadow memory or set memory tags. */ kasan_unpoison_pages (page, order, init);
Re: [PATCH v2 3/4] mm: kasan: Skip page unpoisoning only if __GFP_SKIP_KASAN_UNPOISON
Quoted 45 lines · Catalin Marinas
Currently
post_alloc_hook()skips the kasan unpoisoning if the tags will be zeroed (__GFP_ZEROTAGS) or__GFP_SKIP_KASAN_UNPOISONis passed. Since__GFP_ZEROTAGSis now accompanied by__GFP_SKIP_KASAN_UNPOISON, remove the extra check.Signed-off-by Catalin Marinas <[email protected]>Cc Andrey Ryabinin <[email protected]>Cc Andrey Konovalov <[email protected]>Cc Peter Collingbourne <[email protected]>Cc Vincenzo Frascino <[email protected]>--- mm/page_alloc.c | 12 +++++------- 1 file changed, 5 insertions(+), 7 deletions(-)mm/page_alloc.c +4
-6@@ -2361,7 +2361,7 @@ static inline bool check_new_pcp(struct page *page, unsigned int order) }#endif /* CONFIG_DEBUG_VM */ -static inline bool should_skip_kasan_unpoison (gfp_t flags,bool init_tags) +static inline bool should_skip_kasan_unpoison (gfp_t flags) {/* Don't skip if a software KASAN mode is enabled. */ if (IS_ENABLED(CONFIG_KASAN_GENERIC) ||@@ -2373,12 +2373,10 @@ static inline bool should_skip_kasan_unpoison(gfp_t flags, bool init_tags)return true;/* -* With hardware tag-based KASAN enabled, skip if either: -* -* 1. Memory tags have already been cleared via tag_clear_highpage(). -* 2. Skipping has been requested via __GFP_SKIP_KASAN_UNPOISON. +* With hardware tag-based KASAN enabled, skip if this has been +* requested via __GFP_SKIP_KASAN_UNPOISON. */ -return init_tags || (flags & __GFP_SKIP_KASAN_UNPOISON); +return flags & __GFP_SKIP_KASAN_UNPOISON; }static inline bool should_skip_init (gfp_t flags)
mm/page_alloc.c inline void post_alloc_hook(struct page *page, unsigned int order, · line 2434
Quoted 8 lines
mm/page_alloc.c +1
-1@@ -2430,7 +2428,7 @@ inline void post_alloc_hook(struct page *page, unsigned int order,/* Note that memory is already initialized by the loop above. */ init = false; } -if (!should_skip_kasan_unpoison (gfp_flags, init_tags)) { +if (!should_skip_kasan_unpoison (gfp_flags)) {/* Unpoison shadow memory or set memory tags. */ kasan_unpoison_pages (page, order, init);
Re: [PATCH v2 3/4] mm: kasan: Skip page unpoisoning only if __GFP_SKIP_KASAN_UNPOISON
In reply to Catalin MarinasQuoted 10 lines · Catalin Marinas
Currently
post_alloc_hook()skips the kasan unpoisoning if the tags will be zeroed (__GFP_ZEROTAGS) or__GFP_SKIP_KASAN_UNPOISONis passed. Since__GFP_ZEROTAGSis now accompanied by__GFP_SKIP_KASAN_UNPOISON, remove the extra check.Signed-off-by Catalin Marinas <[email protected]>Cc Andrey Ryabinin <[email protected]>Cc Andrey Konovalov <[email protected]>Cc Peter Collingbourne <[email protected]>Cc Vincenzo Frascino <[email protected]>
Quoted 42 lines
--- mm/page_alloc.c | 12 +++++------- 1 file changed, 5 insertions(+), 7 deletions(-)mm/page_alloc.c +5
-7@@ -2361,7 +2361,7 @@ static inline bool check_new_pcp(struct page *page, unsigned int order) }#endif /* CONFIG_DEBUG_VM */ -static inline bool should_skip_kasan_unpoison (gfp_t flags,bool init_tags) +static inline bool should_skip_kasan_unpoison (gfp_t flags) {/* Don't skip if a software KASAN mode is enabled. */ if (IS_ENABLED(CONFIG_KASAN_GENERIC) ||@@ -2373,12 +2373,10 @@ static inline bool should_skip_kasan_unpoison(gfp_t flags, bool init_tags)return true;/* -* With hardware tag-based KASAN enabled, skip if either: -* -* 1. Memory tags have already been cleared via tag_clear_highpage(). -* 2. Skipping has been requested via __GFP_SKIP_KASAN_UNPOISON. +* With hardware tag-based KASAN enabled, skip if this has been +* requested via __GFP_SKIP_KASAN_UNPOISON. */ -return init_tags || (flags & __GFP_SKIP_KASAN_UNPOISON); +return flags & __GFP_SKIP_KASAN_UNPOISON; }static inline bool should_skip_init (gfp_t flags)@@ -2430,7 +2428,7 @@ inline void post_alloc_hook(struct page *page, unsigned int order,/* Note that memory is already initialized by the loop above. */ init = false; } -if (!should_skip_kasan_unpoison (gfp_flags, init_tags)) { +if (!should_skip_kasan_unpoison (gfp_flags)) {/* Unpoison shadow memory or set memory tags. */ kasan_unpoison_pages (page, order, init);
Signature · 2 lines
Regards, Vincenzo
Catalin Marinas
maintainerPATCH v2 4/4 arm64: kasan: Revert "arm64: mte: reset the page tag in page->flags"
In reply to Catalin MarinasThis reverts commit e5b8d9218951e59df986f627ec93569a0d22149b.
Pages mapped in user-space with PROT_MTE have the allocation tags either zeroed or copied/restored to some user values. In order for the kernel to access such pages via page_address(), resetting the tag in page->flags was necessary. This tag resetting was deferred to set_pte_at() -> mte_sync_page_tags() but it can race with another CPU reading the flags (via page_to_virt()):
P0 (mte_sync_page_tags): P1 (memcpy from virt_to_page):
Rflags!=0xff
Wflags=0xff
DMB (doesn't help)
Wtags=0
Rtags=0 // fault
Since now the post_alloc_hook() function resets the page->flags tag when unpoisoning is skipped for user pages (including the __GFP_ZEROTAGS case), revert the arm64 commit calling page_kasan_tag_reset().
--- arch/arm64/kernel/hibernate.c | 5 ----- arch/arm64/kernel/mte.c | 9 --------- arch/arm64/mm/copypage.c | 9 --------- arch/arm64/mm/mteswap.c | 9 --------- 4 files changed, 32 deletions(-)
arch/arm64/kernel/hibernate.c +0 -5
@@ -300,11 +300,6 @@ static void swsusp_mte_restore_tags(void)unsigned long pfn = xa_state.xa_index;struct page *page =pfn_to_online_page (pfn); -/* -* It is not required to invoke page_kasan_tag_reset(page) -* at this point since the tags stored in page->flags are -* already restored. -*/ mte_restore_page_tags (page_address (page), tags);mte_free_tag_storage (tags);
arch/arm64/kernel/mte.c +0 -9
@@ -48,15 +48,6 @@ static void mte_sync_page_tags(struct page *page, pte_t old_pte,if (!pte_is_tagged)return ; -page_kasan_tag_reset (page); -/* -* We need smp_wmb() in between setting the flags and clearing the -* tags because if another thread reads page->flags and builds a -* tagged address out of it, there is an actual dependency to the -* memory access, but on the current thread we do not guarantee that -* the new page->flags are visible before the tags were updated. -*/ -smp_wmb ();mte_clear_page_tags (page_address (page)); }
arch/arm64/mm/copypage.c +0 -9
@@ -23,15 +23,6 @@ void copy_highpage(struct page *to, struct page *from)if (system_supports_mte () &&test_bit (PG_mte_tagged, &from->flags)) {set_bit (PG_mte_tagged, &to->flags); -page_kasan_tag_reset (to); -/* -* We need smp_wmb() in between setting the flags and clearing the -* tags because if another thread reads page->flags and builds a -* tagged address out of it, there is an actual dependency to the -* memory access, but on the current thread we do not guarantee that -* the new page->flags are visible before the tags were updated. -*/ -smp_wmb ();mte_copy_page_tags (kto, kfrom); } }
arch/arm64/mm/mteswap.c +0 -9
@@ -53,15 +53,6 @@ bool mte_restore_tags(swp_entry_t entry, struct page *page)if (!tags)return false; -page_kasan_tag_reset (page); -/* -* We need smp_wmb() in between setting the flags and clearing the -* tags because if another thread reads page->flags and builds a -* tagged address out of it, there is an actual dependency to the -* memory access, but on the current thread we do not guarantee that -* the new page->flags are visible before the tags were updated. -*/ -smp_wmb ();mte_restore_page_tags (page_address (page), tags);return true;
Re: [PATCH v2 4/4] arm64: kasan: Revert "arm64: mte: reset the page tag in page->flags"
Quoted 99 lines · Catalin Marinas
This reverts commit
e5b8d9218951e59df986f627ec93569a0d22149b.Pages mapped in user-space with
PROT_MTEhave the allocation tags either zeroed or copied/restored to some user values. In order for the kernel to access such pages viapage_address(), resetting the tag inpage->flagswas necessary. This tag resetting was deferred toset_pte_at()->mte_sync_page_tags()but it can race with another CPU reading the flags (viapage_to_virt()):P0 (
mte_sync_page_tags): P1 (memcpy fromvirt_to_page):Rflags!=0xffWflags=0xff
DMB (doesn't help)
Wtags=0
Rtags=0 // faultSince now the
post_alloc_hook()function resets thepage->flagstag when unpoisoning is skipped for user pages (including the__GFP_ZEROTAGScase), revert the arm64 commit callingpage_kasan_tag_reset().Signed-off-by Catalin Marinas <[email protected]>Cc Will Deacon <[email protected]>Cc Vincenzo Frascino <[email protected]>Cc Andrey Konovalov <[email protected]>Cc Peter Collingbourne <[email protected]>--- arch/arm64/kernel/hibernate.c | 5 ----- arch/arm64/kernel/mte.c | 9 --------- arch/arm64/mm/copypage.c | 9 --------- arch/arm64/mm/mteswap.c | 9 --------- 4 files changed, 32 deletions(-)arch/arm64/kernel/hibernate.c +0
-5@@ -300,11 +300,6 @@ static void swsusp_mte_restore_tags(void)unsigned long pfn = xa_state.xa_index;struct page *page =pfn_to_online_page (pfn); -/* -* It is not required to invoke page_kasan_tag_reset(page) -* at this point since the tags stored in page->flags are -* already restored. -*/ mte_restore_page_tags (page_address (page), tags);mte_free_tag_storage (tags);arch/arm64/kernel/mte.c +0
-9@@ -48,15 +48,6 @@ static void mte_sync_page_tags(struct page *page, pte_t old_pte,if (!pte_is_tagged)return ; -page_kasan_tag_reset (page); -/* -* We need smp_wmb() in between setting the flags and clearing the -* tags because if another thread reads page->flags and builds a -* tagged address out of it, there is an actual dependency to the -* memory access, but on the current thread we do not guarantee that -* the new page->flags are visible before the tags were updated. -*/ -smp_wmb ();mte_clear_page_tags (page_address (page)); }arch/arm64/mm/copypage.c +0
-9@@ -23,15 +23,6 @@ void copy_highpage(struct page *to, struct page *from)if (system_supports_mte () &&test_bit (PG_mte_tagged, &from->flags)) {set_bit (PG_mte_tagged, &to->flags); -page_kasan_tag_reset (to); -/* -* We need smp_wmb() in between setting the flags and clearing the -* tags because if another thread reads page->flags and builds a -* tagged address out of it, there is an actual dependency to the -* memory access, but on the current thread we do not guarantee that -* the new page->flags are visible before the tags were updated. -*/ -smp_wmb ();mte_copy_page_tags (kto, kfrom); } }arch/arm64/mm/mteswap.c +0
-2@@ -53,15 +53,6 @@ bool mte_restore_tags(swp_entry_t entry, struct page *page)if (!tags)return false; -page_kasan_tag_reset (page); -/*
arch/arm64/mm/mteswap.c bool mte_restore_tags(swp_entry_t entry, struct page *page) · line 58
Quoted 10 lines
arch/arm64/mm/mteswap.c +0
-7-* We need smp_wmb() in between setting the flags and clearing the -* tags because if another thread reads page->flags and builds a -* tagged address out of it, there is an actual dependency to the -* memory access, but on the current thread we do not guarantee that -* the new page->flags are visible before the tags were updated. -*/ -smp_wmb ();mte_restore_page_tags (page_address (page), tags);return true;
Re: [PATCH v2 4/4] arm64: kasan: Revert "arm64: mte: reset the page tag in page->flags"
In reply to Catalin MarinasQuoted 20 lines · Catalin Marinas
This reverts commit
e5b8d9218951e59df986f627ec93569a0d22149b.Pages mapped in user-space with
PROT_MTEhave the allocation tags either zeroed or copied/restored to some user values. In order for the kernel to access such pages viapage_address(), resetting the tag inpage->flagswas necessary. This tag resetting was deferred toset_pte_at()->mte_sync_page_tags()but it can race with another CPU reading the flags (viapage_to_virt()):P0 (
mte_sync_page_tags): P1 (memcpy fromvirt_to_page):Rflags!=0xffWflags=0xff
DMB (doesn't help)
Wtags=0
Rtags=0 // faultSince now the
post_alloc_hook()function resets thepage->flagstag when unpoisoning is skipped for user pages (including the__GFP_ZEROTAGScase), revert the arm64 commit callingpage_kasan_tag_reset().
Quoted 5 lines
Signed-off-by Catalin Marinas <[email protected]>Cc Will Deacon <[email protected]>Cc Vincenzo Frascino <[email protected]>Cc Andrey Konovalov <[email protected]>Cc Peter Collingbourne <[email protected]>
Quoted 83 lines
--- arch/arm64/kernel/hibernate.c | 5 ----- arch/arm64/kernel/mte.c | 9 --------- arch/arm64/mm/copypage.c | 9 --------- arch/arm64/mm/mteswap.c | 9 --------- 4 files changed, 32 deletions(-)arch/arm64/kernel/hibernate.c +0
-5@@ -300,11 +300,6 @@ static void swsusp_mte_restore_tags(void)unsigned long pfn = xa_state.xa_index;struct page *page =pfn_to_online_page (pfn); -/* -* It is not required to invoke page_kasan_tag_reset(page) -* at this point since the tags stored in page->flags are -* already restored. -*/ mte_restore_page_tags (page_address (page), tags);mte_free_tag_storage (tags);arch/arm64/kernel/mte.c +0
-9@@ -48,15 +48,6 @@ static void mte_sync_page_tags(struct page *page, pte_t old_pte,if (!pte_is_tagged)return ; -page_kasan_tag_reset (page); -/* -* We need smp_wmb() in between setting the flags and clearing the -* tags because if another thread reads page->flags and builds a -* tagged address out of it, there is an actual dependency to the -* memory access, but on the current thread we do not guarantee that -* the new page->flags are visible before the tags were updated. -*/ -smp_wmb ();mte_clear_page_tags (page_address (page)); }arch/arm64/mm/copypage.c +0
-9@@ -23,15 +23,6 @@ void copy_highpage(struct page *to, struct page *from)if (system_supports_mte () &&test_bit (PG_mte_tagged, &from->flags)) {set_bit (PG_mte_tagged, &to->flags); -page_kasan_tag_reset (to); -/* -* We need smp_wmb() in between setting the flags and clearing the -* tags because if another thread reads page->flags and builds a -* tagged address out of it, there is an actual dependency to the -* memory access, but on the current thread we do not guarantee that -* the new page->flags are visible before the tags were updated. -*/ -smp_wmb ();mte_copy_page_tags (kto, kfrom); } }arch/arm64/mm/mteswap.c +0
-9@@ -53,15 +53,6 @@ bool mte_restore_tags(swp_entry_t entry, struct page *page)if (!tags)return false; -page_kasan_tag_reset (page); -/* -* We need smp_wmb() in between setting the flags and clearing the -* tags because if another thread reads page->flags and builds a -* tagged address out of it, there is an actual dependency to the -* memory access, but on the current thread we do not guarantee that -* the new page->flags are visible before the tags were updated. -*/ -smp_wmb ();mte_restore_page_tags (page_address (page), tags);return true;
Signature · 2 lines
Regards, Vincenzo
Quoted 11 lines · Catalin Marinas
That's a second attempt on fixing the race race between setting the allocation (in-memory) tags in a page and the corresponding logical tag in
page->flags. Initial version here:https://lore.kernel.org/r/[email protected]
This new series does not introduce any new GFP flags but instead always skips unpoisoning of the user pages (we already skip the poisoning on free). Any unpoisoned page will have the
page->flagstag reset.[...]
Applied to arm64 (for-next/mte), thanks!
[1/4] mm: kasan: Ensure the tags are visible before the tag in page->flags
https://git.kernel.org/arm64/c/ed0a6d1d973e [2/4] mm: kasan: Skip unpoisoning of user pages https://git.kernel.org/arm64/c/70c248aca9e7 [3/4] mm: kasan: Skip page unpoisoning only if __GFP_SKIP_KASAN_UNPOISON https://git.kernel.org/arm64/c/6d05141a3930 [4/4] arm64: kasan: Revert "arm64: mte: reset the page tag in page->flags" https://git.kernel.org/arm64/c/20794545c146
Cheers,
Signature · 5 lines
Will https://fixes.arm64.dev https://next.arm64.dev https://will.arm64.dev
Quoted 39 lines · Catalin Marinas
Hi,
That's a second attempt on fixing the race race between setting the allocation (in-memory) tags in a page and the corresponding logical tag in
page->flags. Initial version here:https://lore.kernel.org/r/[email protected]
This new series does not introduce any new GFP flags but instead always skips unpoisoning of the user pages (we already skip the poisoning on free). Any unpoisoned page will have the
page->flagstag reset.For the background:
On a system with MTE and
KASAN_HW_TAGSenabled, when a page is allocatedkasan_unpoison_pages()sets a random tag and saves it inpage->flagsso thatpage_to_virt()re-creates the correct tagged pointer. We need to ensure that the in-memory tags are visible before setting thepage->flags:P0 (
__kasan_unpoison_range): P1 (access viavirt_to_page):
Wtags=x Rflags=x| | | DMB | address dependency V V Wflags=x Rtags=xThe first patch changes the order of page unpoisoning with the tag storing in
page->flags.page_kasan_tag_set()has the right barriers throughtry_cmpxchg().If a page is mapped in user-space with
PROT_MTE, the architecture code will set the allocation tag to 0 and a subsequentpage_to_virt()dereference will fault. We currently try to fix this by resetting the tag inpage->flagsso that it is0xff(match-all, not faulting). However, setting the tags and flags can race with another CPU reading the flags (page_to_virt()) and barriers can't help, e.g.:P0 (
mte_sync_page_tags): P1 (memcpy fromvirt_to_page):
Quoted 8 lines
Rflags!=0xffWflags=0xff
DMB (doesn't help)
Wtags=0
Rtags=0 // faultSince clearing the flags in the arch code doesn't work, to do this at page allocation time when
__GFP_SKIP_KASAN_UNPOISONis passed.
I've picked this up, thanks.
An alternative solution might be to use a seqlock (if you can find somewhere to put it) so that virt_to_page() spins briefly while the tags and flags are being updated.
Will
Quoted 8 lines · Catalin Marinas
Hi,
That's a second attempt on fixing the race race between setting the allocation (in-memory) tags in a page and the corresponding logical tag
inpage->flags. Initial version here:
https://lore.kernel.org/r/[email protected]
Quoted 55 lines
This new series does not introduce any new GFP flags but instead always
skips unpoisoning of the user pages (we already skip the poisoning on free). Any unpoisoned page will have thepage->flagstag reset.For the background:
On a system with MTE and
KASAN_HW_TAGSenabled, when a page is allocatedkasan_unpoison_pages()sets a random tag and saves it inpage->flagsso
thatpage_to_virt()re-creates the correct tagged pointer. We need to ensure that the in-memory tags are visible before setting thepage->flags:P0 (
__kasan_unpoison_range): P1 (access viavirt_to_page):
Wtags=x Rflags=x| | | DMB | address dependency V V Wflags=x Rtags=xThe first patch changes the order of page unpoisoning with the tag storing in
page->flags.page_kasan_tag_set()has the right barriers throughtry_cmpxchg().If a page is mapped in user-space with
PROT_MTE, the architecture code
will set the allocation tag to 0 and a subsequentpage_to_virt()dereference will fault. We currently try to fix this by resetting the tag inpage->flagsso that it is0xff(match-all, not faulting). However, setting the tags and flags can race with another CPU reading the flags (page_to_virt()) and barriers can't help, e.g.:P0 (
mte_sync_page_tags): P1 (memcpy fromvirt_to_page):Rflags!=0xffWflags=0xff
DMB (doesn't help)
Wtags=0
Rtags=0 // faultSince clearing the flags in the arch code doesn't work, to do this at page allocation time when
__GFP_SKIP_KASAN_UNPOISONis passed.Thanks.
Catalin Marinas (4):
mm: kasan: Ensure the tags are visible before the tag in page-Quoted 1 line
flags
mm: kasan: Skip unpoisoning of user pages
mm: kasan: Skip page unpoisoning only if__GFP_SKIP_KASAN_UNPOISON
arm64: kasan: Revert "arm64: mte: reset the page tag in page-Quoted 1 line
flags"
Quoted 10 lines
arch/arm64/kernel/hibernate.c | 5 ----- arch/arm64/kernel/mte.c | 9 --------- arch/arm64/mm/copypage.c | 9 --------- arch/arm64/mm/fault.c | 1 - arch/arm64/mm/mteswap.c | 9 --------- include/linux/gfp.h | 2 +- mm/kasan/common.c | 3 ++- mm/page_alloc.c | 19 ++++++++++--------- 8 files changed, 13 insertions(+), 44 deletions(-)
Hi kasan maintainers,
We hit the following issue on the android-6.1 devices with MTE and HW tag kasan enabled.
I observe that the anon flag doesn't have skip_kasan_poison and skip_kasan_unpoison flag and kasantag is weird.
AFAIK, kasantag of anon flag needs to be 0x0.
[ 71.953938] [T1403598] FramePolicy: [name:report&]========================================================= ========= [ 71.955305] [T1403598] FramePolicy: [name:report&]BUG: KASAN: invalid-access in copy_page+0x10/0xd0 [ 71.956476] [T1403598] FramePolicy: [name:report&]Read at addr f0ffff81332a8000 by task FramePolicy/3598 [ 71.957673] [T1403598] FramePolicy: [name:report_hw_tags&]Pointer tag: [f0], memory tag: [ff] [ 71.958746] [T1403598] FramePolicy: [name:report&] [ 71.959354] [T1403598] FramePolicy: CPU: 4 PID: 3598 Comm: FramePolicy Tainted: G S W OE 6.1.0-mainline-android14-0- ga8a53f83b9e4 #1 [ 71.960978] [T1403598] FramePolicy: Hardware name: MT6985(ENG) (DT) [ 71.961767] [T1403598] FramePolicy: Call trace: [ 71.962338] [T1403598] FramePolicy: dump_backtrace+0x108/0x158 [ 71.963097] [T1403598] FramePolicy: show_stack+0x20/0x48 [ 71.963782] [T1403598] FramePolicy: dump_stack_lvl+0x6c/0x88 [ 71.964512] [T1403598] FramePolicy: print_report+0x2cc/0xa64 [ 71.965263] [T1403598] FramePolicy: kasan_report+0xb8/0x138 [ 71.965986] [T1403598] FramePolicy: __do_kernel_fault+0xd4/0x248 [ 71.966782] [T1403598] FramePolicy: do_bad_area+0x38/0xe8 [ 71.967484] [T1403598] FramePolicy: do_tag_check_fault+0x24/0x38 [ 71.968261] [T1403598] FramePolicy: do_mem_abort+0x48/0xb0 [ 71.968973] [T1403598] FramePolicy: el1_abort+0x44/0x68 [ 71.969646] [T1403598] FramePolicy: el1h_64_sync_handler+0x68/0xb8 [ 71.970440] [T1403598] FramePolicy: el1h_64_sync+0x68/0x6c [ 71.971146] [T1403598] FramePolicy: copy_page+0x10/0xd0 [ 71.971824] [T1403598] FramePolicy: copy_user_highpage+0x20/0x40 [ 71.972603] [T1403598] FramePolicy: wp_page_copy+0xd0/0x9f8 [ 71.973344] [T1403598] FramePolicy: do_wp_page+0x374/0x3b0 [ 71.974056] [T1403598] FramePolicy: handle_mm_fault+0x3ec/0x119c [ 71.974833] [T1403598] FramePolicy: do_page_fault+0x344/0x4ac [ 71.975583] [T1403598] FramePolicy: do_mem_abort+0x48/0xb0 [ 71.976294] [T1403598] FramePolicy: el0_da+0x4c/0xe0 [ 71.976934] [T1403598] FramePolicy: el0t_64_sync_handler+0xd4/0xfc [ 71.977725] [T1403598] FramePolicy: el0t_64_sync+0x1a0/0x1a4 [ 71.978451] [T1403598] FramePolicy: [name:report&] [ 71.979057] [T1403598] FramePolicy: [name:report&]The buggy address belongs to the physical page: [ 71.980173] [T1403598] FramePolicy: [name:debug&]page:fffffffe04ccaa00 refcount:14 mapcount:13 mapping:0000000000000000 index:0x7884c74 pfn:0x1732a8 [ 71.981849] [T1403598] FramePolicy: [name:debug&]memcg:faffff80c0241000 [ 71.982680] [T1403598] FramePolicy: [name:debug&]anon flags: 0x43c000000048003e(referenced|uptodate|dirty|lru|active|swapbacked|arch _2|zone=1|kasantag=0xf) [ 71.984446] [T1403598] FramePolicy: raw: 43c000000048003e fffffffe04b99648 fffffffe04cca308 f2ffff8103390831 [ 71.985684] [T1403598] FramePolicy: raw: 0000000007884c74 0000000000000000 0000000e0000000c faffff80c0241000 [ 71.986919] [T1403598] FramePolicy: [name:debug&]page dumped because: kasan: bad access detected [ 71.988022] [T1403598] FramePolicy: [name:report&] [ 71.988624] [T1403598] FramePolicy: [name:report&]Memory state around the buggy address: [ 71.989641] [T1403598] FramePolicy: ffffff81332a7e00: fe fe fe fe fe fe fe fe fe fe fe fe fe fe fe fe [ 71.990811] [T1403598] FramePolicy: ffffff81332a7f00: fe fe fe fe fe fe fe fe fe fe fe fe fe fe fe fe [ 71.991982] [T1403598] FramePolicy: >ffffff81332a8000: ff ff ff ff f0 f0 fc fc fc fc fc fc fc f0 f0 f3 [ 71.993149] [T1403598] FramePolicy: [name:report&] ^ [ 71.993972] [T1403598] FramePolicy: ffffff81332a8100: f3 f3 f3 f3 f3 f3 f0 f0 f8 f8 f8 f8 f8 f8 f8 f0 [ 71.995141] [T1403598] FramePolicy: ffffff81332a8200: f0 fb fb fb fb fb fb fb f0 f0 fe fe fe fe fe fe [ 71.996332] [T1403598] FramePolicy:
[name:report&]=========================================================
=========
Originally, I suspect that some userspace pages have been migrated so the page->flags will be lost and page->flags is re-generated by alloc_pages().
I try the following diff, but it didn't help.
mm/migrate.c +1 -0
@@ -51,6 +51,7 @@#include <linux/random.h> #include <linux/sched/sysctl.h> #include <linux/memory-tiers.h> +#include <linux/kasan.h> #include <asm/tlbflush.h>
@@ -611,6 +612,14 @@ void folio_migrate_flags(struct folio *newfolio,
struct folio *folio)
if (!folio_test_hugetlb(folio))
mem_cgroup_migrate(folio, newfolio);
+
+#ifdef CONFIG_KASAN_HW_TAGS
+ if (kasan_hw_tags_enabled()) {
+ if (folio_test_skip_kasan_poison(folio))
+ folio_set_skip_kasan_poison(newfolio);
+ page_kasan_tag_set(&newfolio->page,page_kasan_tag(&folio->page));
+ }
+#endif
}
EXPORT_SYMBOL(folio_migrate_flags);
After I revert this patchset (4 patches), this issue disappear.
Quoted 1 line
Andrey Konovalov
reviewerQuoted 157 lines · Kuan-Ying Lee (李冠穎)
Quoted 8 lines · Catalin Marinas
Hi,
That's a second attempt on fixing the race race between setting the allocation (in-memory) tags in a page and the corresponding logical tag
inpage->flags. Initial version here:https://lore.kernel.org/r/[email protected]
Quoted 55 lines
This new series does not introduce any new GFP flags but instead always
skips unpoisoning of the user pages (we already skip the poisoning on free). Any unpoisoned page will have thepage->flagstag reset.For the background:
On a system with MTE and
KASAN_HW_TAGSenabled, when a page is allocatedkasan_unpoison_pages()sets a random tag and saves it inpage->flagsso
thatpage_to_virt()re-creates the correct tagged pointer. We need to ensure that the in-memory tags are visible before setting thepage->flags:P0 (
__kasan_unpoison_range): P1 (access viavirt_to_page):
Wtags=x Rflags=x| | | DMB | address dependency V V Wflags=x Rtags=xThe first patch changes the order of page unpoisoning with the tag storing in
page->flags.page_kasan_tag_set()has the right barriers throughtry_cmpxchg().If a page is mapped in user-space with
PROT_MTE, the architecture code
will set the allocation tag to 0 and a subsequentpage_to_virt()dereference will fault. We currently try to fix this by resetting the tag inpage->flagsso that it is0xff(match-all, not faulting). However, setting the tags and flags can race with another CPU reading the flags (page_to_virt()) and barriers can't help, e.g.:P0 (
mte_sync_page_tags): P1 (memcpy fromvirt_to_page):Rflags!=0xffWflags=0xff
DMB (doesn't help)
Wtags=0
Rtags=0 // faultSince clearing the flags in the arch code doesn't work, to do this at page allocation time when
__GFP_SKIP_KASAN_UNPOISONis passed.Thanks.
Catalin Marinas (4):
mm: kasan: Ensure the tags are visible before the tag in page-Quoted 1 line
flags
mm: kasan: Skip unpoisoning of user pages
mm: kasan: Skip page unpoisoning only if__GFP_SKIP_KASAN_UNPOISON
arm64: kasan: Revert "arm64: mte: reset the page tag in page-Quoted 1 line
flags"
Quoted 10 lines
arch/arm64/kernel/hibernate.c | 5 ----- arch/arm64/kernel/mte.c | 9 --------- arch/arm64/mm/copypage.c | 9 --------- arch/arm64/mm/fault.c | 1 - arch/arm64/mm/mteswap.c | 9 --------- include/linux/gfp.h | 2 +- mm/kasan/common.c | 3 ++- mm/page_alloc.c | 19 ++++++++++--------- 8 files changed, 13 insertions(+), 44 deletions(-)Hi kasan maintainers,
We hit the following issue on the android-6.1 devices with MTE and HW tag kasan enabled.
I observe that the anon flag doesn't have
skip_kasan_poisonandskip_kasan_unpoisonflag and kasantag is weird.AFAIK, kasantag of anon flag needs to be
0x0.[ 71.953938] [T1403598] FramePolicy: [name:report&]========================================================= ========= [ 71.955305] [T1403598] FramePolicy: [name:report&]BUG: KASAN: invalid-access in copy_page+0x10/0xd0 [ 71.956476] [T1403598] FramePolicy: [name:report&]Read at addr f0ffff81332a8000 by task FramePolicy/3598 [ 71.957673] [T1403598] FramePolicy: [name:report_hw_tags&]Pointer tag: [f0], memory tag: [ff] [ 71.958746] [T1403598] FramePolicy: [name:report&] [ 71.959354] [T1403598] FramePolicy: CPU: 4 PID: 3598 Comm: FramePolicy Tainted: G S W OE 6.1.0-mainline-android14-0- ga8a53f83b9e4 #1 [ 71.960978] [T1403598] FramePolicy: Hardware name: MT6985(ENG) (DT) [ 71.961767] [T1403598] FramePolicy: Call trace: [ 71.962338] [T1403598] FramePolicy: dump_backtrace+0x108/0x158 [ 71.963097] [T1403598] FramePolicy: show_stack+0x20/0x48 [ 71.963782] [T1403598] FramePolicy: dump_stack_lvl+0x6c/0x88 [ 71.964512] [T1403598] FramePolicy: print_report+0x2cc/0xa64 [ 71.965263] [T1403598] FramePolicy: kasan_report+0xb8/0x138 [ 71.965986] [T1403598] FramePolicy: __do_kernel_fault+0xd4/0x248 [ 71.966782] [T1403598] FramePolicy: do_bad_area+0x38/0xe8 [ 71.967484] [T1403598] FramePolicy: do_tag_check_fault+0x24/0x38 [ 71.968261] [T1403598] FramePolicy: do_mem_abort+0x48/0xb0 [ 71.968973] [T1403598] FramePolicy: el1_abort+0x44/0x68 [ 71.969646] [T1403598] FramePolicy: el1h_64_sync_handler+0x68/0xb8 [ 71.970440] [T1403598] FramePolicy: el1h_64_sync+0x68/0x6c [ 71.971146] [T1403598] FramePolicy: copy_page+0x10/0xd0 [ 71.971824] [T1403598] FramePolicy: copy_user_highpage+0x20/0x40 [ 71.972603] [T1403598] FramePolicy: wp_page_copy+0xd0/0x9f8 [ 71.973344] [T1403598] FramePolicy: do_wp_page+0x374/0x3b0 [ 71.974056] [T1403598] FramePolicy: handle_mm_fault+0x3ec/0x119c [ 71.974833] [T1403598] FramePolicy: do_page_fault+0x344/0x4ac [ 71.975583] [T1403598] FramePolicy: do_mem_abort+0x48/0xb0 [ 71.976294] [T1403598] FramePolicy: el0_da+0x4c/0xe0 [ 71.976934] [T1403598] FramePolicy: el0t_64_sync_handler+0xd4/0xfc [ 71.977725] [T1403598] FramePolicy: el0t_64_sync+0x1a0/0x1a4 [ 71.978451] [T1403598] FramePolicy: [name:report&] [ 71.979057] [T1403598] FramePolicy: [name:report&]The buggy address belongs to the physical page: [ 71.980173] [T1403598] FramePolicy: [name:debug&]page:fffffffe04ccaa00 refcount:14 mapcount:13 mapping:0000000000000000 index:0x7884c74 pfn:0x1732a8 [ 71.981849] [T1403598] FramePolicy: [name:debug&]memcg:faffff80c0241000 [ 71.982680] [T1403598] FramePolicy: [name:debug&]anon flags: 0x43c000000048003e(referenced|uptodate|dirty|lru|active|swapbacked|arch _2|zone=1|kasantag=0xf) [ 71.984446] [T1403598] FramePolicy: raw: 43c000000048003e fffffffe04b99648 fffffffe04cca308 f2ffff8103390831 [ 71.985684] [T1403598] FramePolicy: raw: 0000000007884c74 0000000000000000 0000000e0000000c faffff80c0241000 [ 71.986919] [T1403598] FramePolicy: [name:debug&]page dumped because: kasan: bad access detected [ 71.988022] [T1403598] FramePolicy: [name:report&] [ 71.988624] [T1403598] FramePolicy: [name:report&]Memory state around the buggy address: [ 71.989641] [T1403598] FramePolicy: ffffff81332a7e00: fe fe fe fe fe fe fe fe fe fe fe fe fe fe fe fe [ 71.990811] [T1403598] FramePolicy: ffffff81332a7f00: fe fe fe fe fe fe fe fe fe fe fe fe fe fe fe fe [ 71.991982] [T1403598] FramePolicy: >ffffff81332a8000: ff ff ff ff f0 f0 fc fc fc fc fc fc fc f0 f0 f3 [ 71.993149] [T1403598] FramePolicy: [name:report&] ^ [ 71.993972] [T1403598] FramePolicy: ffffff81332a8100: f3 f3 f3 f3 f3 f3 f0 f0 f8 f8 f8 f8 f8 f8 f8 f0 [ 71.995141] [T1403598] FramePolicy: ffffff81332a8200: f0 fb fb fb fb fb fb fb f0 f0 fe fe fe fe fe fe [ 71.996332] [T1403598] FramePolicy:
Quoted 6 lines
[name:report&]=========================================================
=========Originally, I suspect that some userspace pages have been migrated so the
page->flagswill be lost andpage->flagsis re-generated byalloc_pages().
Hi Kuan-Ying,
There recently was a similar crash due to incorrectly implemented sampling.
Do you have the following patch in your tree?
https://android.googlesource.com/kernel/common/+/9f7f5a25f335e6e1484695da9180281a728db7e2
If not, please sync your 6.1 tree with the Android common kernel. Hopefully this will fix the issue.
Thanks!
Quoted 6 lines · Andrey Konovalov
Quoted 4 lines · Kuan-Ying Lee (李冠穎)
Quoted 2 lines · Catalin Marinas
Hi,
Quoted 10 lines
Quoted 10 lines
Quoted 8 lines
That's a second attempt on fixing the race race between setting the
allocation (in-memory) tags in a page and the corresponding logical
tag
inpage->flags. Initial version here:
https://lore.kernel.org/r/[email protected]
Quoted 183 lines
Quoted 174 lines
Quoted 65 lines
This new series does not introduce any new GFP flags but instead always
skips unpoisoning of the user pages (we already skip the poisoning on
free). Any unpoisoned page will have thepage->flagstag reset.For the background:
On a system with MTE and
KASAN_HW_TAGSenabled, when a page is allocatedkasan_unpoison_pages()sets a random tag and saves it in page-Quoted 1 line
flags
so
thatpage_to_virt()re-creates the correct tagged pointer. We need to
ensure that the in-memory tags are visible before setting thepage->flags:P0 (
__kasan_unpoison_range): P1 (access viavirt_to_page):
Wtags=x Rflags=x| | | DMB | address dependency V V Wflags=x Rtags=xThe first patch changes the order of page unpoisoning with the tag
storing inpage->flags.page_kasan_tag_set()has the right barriers
throughtry_cmpxchg().If a page is mapped in user-space with
PROT_MTE, the architecture code
will set the allocation tag to 0 and a subsequentpage_to_virt()dereference will fault. We currently try to fix this by resetting the
tag inpage->flagsso that it is0xff(match-all, not faulting). However, setting the tags and flags can race with another CPU reading
the flags (page_to_virt()) and barriers can't help, e.g.:P0 (
mte_sync_page_tags): P1 (memcpy fromvirt_to_page):Rflags!=0xffWflags=0xff
DMB (doesn't help)
Wtags=0
Rtags=0 // faultSince clearing the flags in the arch code doesn't work, to do this at
page allocation time when__GFP_SKIP_KASAN_UNPOISONis passed.Thanks.
Catalin Marinas (4):
mm: kasan: Ensure the tags are visible before the tag in page-Quoted 1 line
flags
mm: kasan: Skip unpoisoning of user pages
mm: kasan: Skip page unpoisoning only if__GFP_SKIP_KASAN_UNPOISON
arm64: kasan: Revert "arm64: mte: reset the page tag in page-Quoted 1 line
flags"
Quoted 10 lines
arch/arm64/kernel/hibernate.c | 5 ----- arch/arm64/kernel/mte.c | 9 --------- arch/arm64/mm/copypage.c | 9 --------- arch/arm64/mm/fault.c | 1 - arch/arm64/mm/mteswap.c | 9 --------- include/linux/gfp.h | 2 +- mm/kasan/common.c | 3 ++- mm/page_alloc.c | 19 ++++++++++--------- 8 files changed, 13 insertions(+), 44 deletions(-)Hi kasan maintainers,
We hit the following issue on the android-6.1 devices with MTE and HW
tag kasan enabled.I observe that the anon flag doesn't have
skip_kasan_poisonandskip_kasan_unpoisonflag and kasantag is weird.AFAIK, kasantag of anon flag needs to be
0x0.[ 71.953938] [T1403598] FramePolicy: [name:report&]===================================================== ==== ========= [ 71.955305] [T1403598] FramePolicy: [name:report&]BUG: KASAN: invalid-access in copy_page+0x10/0xd0 [ 71.956476] [T1403598] FramePolicy: [name:report&]Read at addr f0ffff81332a8000 by task FramePolicy/3598 [ 71.957673] [T1403598] FramePolicy: [name:report_hw_tags&]Pointer tag: [f0], memory tag: [ff] [ 71.958746] [T1403598] FramePolicy: [name:report&] [ 71.959354] [T1403598] FramePolicy: CPU: 4 PID: 3598 Comm: FramePolicy Tainted: G S W OE 6.1.0-mainline-android14- 0- ga8a53f83b9e4 #1 [ 71.960978] [T1403598] FramePolicy: Hardware name: MT6985(ENG) (DT) [ 71.961767] [T1403598] FramePolicy: Call trace: [ 71.962338] [T1403598] FramePolicy: dump_backtrace+0x108/0x158 [ 71.963097] [T1403598] FramePolicy: show_stack+0x20/0x48 [ 71.963782] [T1403598] FramePolicy: dump_stack_lvl+0x6c/0x88 [ 71.964512] [T1403598] FramePolicy: print_report+0x2cc/0xa64 [ 71.965263] [T1403598] FramePolicy: kasan_report+0xb8/0x138 [ 71.965986] [T1403598] FramePolicy: __do_kernel_fault+0xd4/0x248 [ 71.966782] [T1403598] FramePolicy: do_bad_area+0x38/0xe8 [ 71.967484] [T1403598] FramePolicy: do_tag_check_fault+0x24/0x38 [ 71.968261] [T1403598] FramePolicy: do_mem_abort+0x48/0xb0 [ 71.968973] [T1403598] FramePolicy: el1_abort+0x44/0x68 [ 71.969646] [T1403598] FramePolicy: el1h_64_sync_handler+0x68/0xb8 [ 71.970440] [T1403598] FramePolicy: el1h_64_sync+0x68/0x6c [ 71.971146] [T1403598] FramePolicy: copy_page+0x10/0xd0 [ 71.971824] [T1403598] FramePolicy: copy_user_highpage+0x20/0x40 [ 71.972603] [T1403598] FramePolicy: wp_page_copy+0xd0/0x9f8 [ 71.973344] [T1403598] FramePolicy: do_wp_page+0x374/0x3b0 [ 71.974056] [T1403598] FramePolicy: handle_mm_fault+0x3ec/0x119c [ 71.974833] [T1403598] FramePolicy: do_page_fault+0x344/0x4ac [ 71.975583] [T1403598] FramePolicy: do_mem_abort+0x48/0xb0 [ 71.976294] [T1403598] FramePolicy: el0_da+0x4c/0xe0 [ 71.976934] [T1403598] FramePolicy: el0t_64_sync_handler+0xd4/0xfc [ 71.977725] [T1403598] FramePolicy: el0t_64_sync+0x1a0/0x1a4 [ 71.978451] [T1403598] FramePolicy: [name:report&] [ 71.979057] [T1403598] FramePolicy: [name:report&]The buggy address belongs to the physical page: [ 71.980173] [T1403598] FramePolicy: [name:debug&]page:fffffffe04ccaa00 refcount:14 mapcount:13 mapping:0000000000000000 index:0x7884c74 pfn:0x1732a8 [ 71.981849] [T1403598] FramePolicy: [name:debug&]memcg:faffff80c0241000 [ 71.982680] [T1403598] FramePolicy: [name:debug&]anon flags: 0x43c000000048003e(referenced|uptodate|dirty|lru|active|swapbacked| arch _2|zone=1|kasantag=0xf) [ 71.984446] [T1403598] FramePolicy: raw: 43c000000048003e fffffffe04b99648 fffffffe04cca308 f2ffff8103390831 [ 71.985684] [T1403598] FramePolicy: raw: 0000000007884c74 0000000000000000 0000000e0000000c faffff80c0241000 [ 71.986919] [T1403598] FramePolicy: [name:debug&]page dumped because: kasan: bad access detected [ 71.988022] [T1403598] FramePolicy: [name:report&] [ 71.988624] [T1403598] FramePolicy: [name:report&]Memory state around the buggy address: [ 71.989641] [T1403598] FramePolicy: ffffff81332a7e00: fe fe fe fe fe fe fe fe fe fe fe fe fe fe fe fe [ 71.990811] [T1403598] FramePolicy: ffffff81332a7f00: fe fe fe fe fe fe fe fe fe fe fe fe fe fe fe fe [ 71.991982] [T1403598] FramePolicy: >ffffff81332a8000: ff ff ff ff f0 f0 fc fc fc fc fc fc fc f0 f0 f3 [ 71.993149] [T1403598] FramePolicy: [name:report&] ^ [ 71.993972] [T1403598] FramePolicy: ffffff81332a8100: f3 f3 f3 f3 f3 f3 f0 f0 f8 f8 f8 f8 f8 f8 f8 f0 [ 71.995141] [T1403598] FramePolicy: ffffff81332a8200: f0 fb fb fb fb fb fb fb f0 f0 fe fe fe fe fe fe [ 71.996332] [T1403598] FramePolicy:Quoted 8 lines
[name:report&]=====================================================
====
=========Originally, I suspect that some userspace pages have been migrated so
thepage->flagswill be lost andpage->flagsis re-generated byalloc_pages().
Quoted 8 lines
Hi Kuan-Ying,
There recently was a similar crash due to incorrectly implemented sampling.
Do you have the following patch in your tree?
Quoted 6 lines
If not, please sync your 6.1 tree with the Android common kernel. Hopefully this will fix the issue.
Thanks!
Hi Andrey,
Thanks for your advice.
I saw this patch is to fix ("kasan: allow sampling page_alloc allocations for HW_TAGS").
But our 6.1 tree doesn't have following two commits now. ("FROMGIT: kasan: allow sampling page_alloc allocations for HW_TAGS") (FROMLIST: kasan: reset page tags properly with sampling)
Andrey Konovalov
reviewerQuoted 19 lines · Kuan-Ying Lee (李冠穎)
Quoted 9 lines
Hi Kuan-Ying,
There recently was a similar crash due to incorrectly implemented sampling.
Do you have the following patch in your tree?
Quoted 6 lines
If not, please sync your 6.1 tree with the Android common kernel. Hopefully this will fix the issue.
Thanks!
Hi Andrey,
Quoted 8 lines
Thanks for your advice.
I saw this patch is to fix ("kasan: allow sampling
page_allocallocations forHW_TAGS").But our 6.1 tree doesn't have following two commits now. ("FROMGIT: kasan: allow sampling
page_allocallocations forHW_TAGS") (FROMLIST: kasan: reset page tags properly with sampling)
Hi Kuan-Ying,
Just to clarify: these two patches were applied twice: once here on Jan 13:
https://android.googlesource.com/kernel/common/+/a2a9e34d164e90fc08d35fd097a164b9101d72ef
https://android.googlesource.com/kernel/common/+/435e2a6a6c8ba8d0eb55f9aaade53e7a3957322b
but then reverted here on Jan 20:
https://android.googlesource.com/kernel/common/+/5503dbe454478fe54b9cac3fc52d4477f52efdc9
https://android.googlesource.com/kernel/common/+/4573a3cf7e18735a477845426238d46d96426bb6
And then once again via the link I sent before together with a fix on Jan 25.
It might be that you still have to former two patches in your tree if you synced it before the revert.
However, if this is not the case:
Which 6.1 commit is your tree based on?
Do you have any private MTE-related changes in the kernel? Do you have userspace MTE enabled?
Thanks!
Quoted 5 lines · Andrey Konovalov
Quoted 3 lines · Kuan-Ying Lee (李冠穎)
Quoted 3 lines
Hi Kuan-Ying,
Quoted 8 lines
Quoted 8 lines
Quoted 6 lines
There recently was a similar crash due to incorrectly implemented sampling.
Do you have the following patch in your tree?
Quoted 14 lines
Quoted 14 lines
Quoted 6 lines
If not, please sync your 6.1 tree with the Android common kernel. Hopefully this will fix the issue.
Thanks!
Hi Andrey,
Thanks for your advice.
I saw this patch is to fix ("kasan: allow sampling
page_allocallocations forHW_TAGS").
Quoted 7 lines
Quoted 4 lines
But our 6.1 tree doesn't have following two commits now. ("FROMGIT: kasan: allow sampling
page_allocallocations forHW_TAGS")
(FROMLIST: kasan: reset page tags properly with sampling)Hi Kuan-Ying,
Hi Andrey,
I'll stand in for Kuan-Ying as he's out of office. Thanks for your help!
Quoted 4 lines
Just to clarify: these two patches were applied twice: once here on Jan 13:
Quoted 2 lines
Quoted 2 lines
Our codebase does not contain these two patches.
Quoted 3 lines
but then reverted here on Jan 20:
Quoted 2 lines
Quoted 11 lines
And then once again via the link I sent before together with a fix on Jan 25.
It might be that you still have to former two patches in your tree if you synced it before the revert.
However, if this is not the case:
Which 6.1 commit is your tree based on?
https://android.googlesource.com/kernel/common/+/53b3a7721b7aec74d8fa2ee55c2480044cc7c1b8
(53b3a77 Merge 6.1.1 into android14-6.1) is the latest commit in our tree.
Quoted 1 line
Do you have any private MTE-related changes in the kernel?
No, all the MTE-related code is the same as Android Common Kernel.
Quoted 1 line
Do you have userspace MTE enabled?
Yes, we have enabled MTE for both EL1 and EL0.
Quoted 2 lines
Thanks!
Quoted 78 lines · Qun-wei Lin (林群崴)
Quoted 5 lines · Andrey Konovalov
Quoted 3 lines · Kuan-Ying Lee (李冠穎)
Quoted 3 lines
Hi Kuan-Ying,
Quoted 8 lines
Quoted 8 lines
Quoted 6 lines
There recently was a similar crash due to incorrectly implemented sampling.
Do you have the following patch in your tree?
Quoted 14 lines
Quoted 14 lines
Quoted 6 lines
If not, please sync your 6.1 tree with the Android common kernel. Hopefully this will fix the issue.
Thanks!
Hi Andrey,
Thanks for your advice.
I saw this patch is to fix ("kasan: allow sampling
page_allocallocations forHW_TAGS").Quoted 7 lines
Quoted 4 lines
But our 6.1 tree doesn't have following two commits now. ("FROMGIT: kasan: allow sampling
page_allocallocations forHW_TAGS")
(FROMLIST: kasan: reset page tags properly with sampling)Hi Kuan-Ying,
Hi Andrey,
I'll stand in for Kuan-Ying as he's out of office. Thanks for your help!Quoted 4 lines
Just to clarify: these two patches were applied twice: once here on Jan 13:
Quoted 2 lines
Quoted 2 lines
Our codebase does not contain these two patches.
Quoted 3 lines
but then reverted here on Jan 20:
Quoted 2 lines
Quoted 11 lines
And then once again via the link I sent before together with a fix on Jan 25.
It might be that you still have to former two patches in your tree if you synced it before the revert.
However, if this is not the case:
Which 6.1 commit is your tree based on?
https://android.googlesource.com/kernel/common/+/53b3a7721b7aec74d8fa2ee55c2480044cc7c1b8
(53b3a77 Merge 6.1.1 into android14-6.1) is the latest commit in our tree.
Quoted 7 lines
Quoted 1 line
Do you have any private MTE-related changes in the kernel?
No, all the MTE-related code is the same as Android Common Kernel.
Quoted 1 line
Do you have userspace MTE enabled?
Yes, we have enabled MTE for both EL1 and EL0.
Hi Qun-wei,
Thanks for the information. We encountered a similar issue internally with the Android 5.15 common kernel. We tracked it down to an issue with page migration, where the source page was a userspace page with MTE tags, and the target page was allocated using KASAN (i.e. having a non-zero KASAN tag). This caused tag check faults when the page was subsequently accessed by the kernel as a result of the mismatching tags from userspace. Given the number of different ways that page migration target pages can be allocated, the simplest fix that we could think of was to synchronize the KASAN tag in copy_highpage().
Can you try the patch below and let us know whether it fixes the issue?
arch/arm64/mm/copypage.c +2 -0
@@ -23,6 +23,8 @@ void copy_highpage(struct page *to, struct page *from)if (system_supports_mte () &&test_bit (PG_mte_tagged, &from->flags)) {set_bit (PG_mte_tagged, &to->flags); +if (kasan_hw_tags_enabled ()) +page_kasan_tag_set (to,page_kasan_tag (from));mte_copy_page_tags (kto, kfrom); } }
Catalin, please let us know what you think of the patch above. It effectively partially undoes commit 20794545c146 ("arm64: kasan: Revert "arm64: mte: reset the page tag in page->flags""), but this seems okay to me because the mentioned race condition shouldn't affect "new" pages such as those being used as migration targets. The smp_wmb() that was there before doesn't seem necessary for the same reason.
If the patch is okay, we should apply it to the 6.1 stable kernel. The problem appears to be "fixed" in the mainline kernel because of a bad merge conflict resolution on my part; when I rebased commit e059853d14ca ("arm64: mte: Fix/clarify the PG_mte_tagged semantics") past commit 20794545c146, it looks like I accidentally brought back the page_kasan_tag_reset() line removed in the latter. But we should align the mainline kernel with whatever we decide to do on 6.1.
Peter
Catalin Marinas
maintainerHi Peter,
Quoted 17 lines · Peter Collingbourne
Thanks for the information. We encountered a similar issue internally with the Android 5.15 common kernel. We tracked it down to an issue with page migration, where the source page was a userspace page with MTE tags, and the target page was allocated using KASAN (i.e. having a non-zero KASAN tag). This caused tag check faults when the page was subsequently accessed by the kernel as a result of the mismatching tags from userspace. Given the number of different ways that page migration target pages can be allocated, the simplest fix that we could think of was to synchronize the KASAN tag in
copy_highpage().Can you try the patch below and let us know whether it fixes the issue?
arch/arm64/mm/copypage.c +0
-0
arch/arm64/mm/copypage.c void copy_highpage(struct page *to, struct page *from) · line 28
Quoted 6 lines
arch/arm64/mm/copypage.c +2
-0@@ -23,6 +23,8 @@ void copy_highpage(struct page *to, struct page *from)if (system_supports_mte () &&test_bit (PG_mte_tagged, &from->flags)) {set_bit (PG_mte_tagged, &to->flags); +if (kasan_hw_tags_enabled ()) +page_kasan_tag_set (to,page_kasan_tag (from));mte_copy_page_tags (kto, kfrom);
Why not just page_kasan_tag_reset(to)? If PG_mte_tagged is set on the 'from' page, the tags are random anyway and page_kasan_tag(from) should already be 0xff. It makes more sense to do the same for the 'to' page rather than copying the tag from the 'from' page. IOW, we are copying user-controlled tags into a page, the kernel should have a match-all tag in page->flags.
Quoted 7 lines
Catalin, please let us know what you think of the patch above. It effectively partially undoes commit
20794545c146("arm64: kasan: Revert "arm64: mte: reset the page tag inpage->flags""), but this seems okay to me because the mentioned race condition shouldn't affect "new" pages such as those being used as migration targets. Thesmp_wmb()that was there before doesn't seem necessary for the same reason.
Quoted 7 lines
If the patch is okay, we should apply it to the 6.1 stable kernel. The problem appears to be "fixed" in the mainline kernel because of a bad merge conflict resolution on my part; when I rebased commit
e059853d14ca("arm64: mte: Fix/clarify thePG_mte_taggedsemantics") past commit20794545c146, it looks like I accidentally brought back thepage_kasan_tag_reset()line removed in the latter. But we should align the mainline kernel with whatever we decide to do on 6.1.
Happy accident ;). When I reverted such calls in commit 20794545c146, my assumption was that we always get a page that went through post_alloc_hook() and the tags were reset. But it seems that's not always the case (and probably wasteful anyway if we have to zero the tags and data on a page we know we are going to override via copy_highpage() anyway). The barrier doesn't help, so we shouldn't add it back.
So, I'm fine with a stable fix but I wonder whether we should backport the whole "Fix/clarify the PG_mte_tagged semantics" series instead.
Signature · 1 line
Catalin
Hi Catalin,
Quoted 22 lines · Catalin Marinas
Hi Peter,
Quoted 18 lines · Peter Collingbourne
Thanks for the information. We encountered a similar issue internally with the Android 5.15 common kernel. We tracked it down to an issue with page migration, where the source page was a userspace page with MTE tags, and the target page was allocated using KASAN (i.e. having a non-zero KASAN tag). This caused tag check faults when the page was subsequently accessed by the kernel as a result of the mismatching tags from userspace. Given the number of different ways that page migration target pages can be allocated, the simplest fix that we could think of was to synchronize the KASAN tag in
copy_highpage().Can you try the patch below and let us know whether it fixes the issue?
arch/arm64/mm/copypage.c +0
-0@@ -23,6 +23,8 @@ void copy_highpage(struct page *to, struct page *from)
Quoted 12 lines
arch/arm64/mm/copypage.c void copy_highpage(struct page *to, struct page *from) · line 28
Quoted 5 lines
arch/arm64/mm/copypage.c +2
-0if (system_supports_mte () &&test_bit (PG_mte_tagged, &from->flags)) {set_bit (PG_mte_tagged, &to->flags); +if (kasan_hw_tags_enabled ()) +page_kasan_tag_set (to,page_kasan_tag (from));mte_copy_page_tags (kto, kfrom);Why not just
page_kasan_tag_reset(to)? IfPG_mte_taggedis set on the 'from' page, the tags are random anyway andpage_kasan_tag(from)should already be0xff. It makes more sense to do the same for the 'to' page rather than copying the tag from the 'from' page. IOW, we are copying user-controlled tags into a page, the kernel should have a match-all tag inpage->flags.
That would also work, but I was thinking that if copy_highpage() were being used to copy a KASAN page we should keep the original tag in order to maintain tag checks for page accesses.
Quoted 23 lines
Quoted 7 lines
Catalin, please let us know what you think of the patch above. It effectively partially undoes commit
20794545c146("arm64: kasan: Revert "arm64: mte: reset the page tag inpage->flags""), but this seems okay to me because the mentioned race condition shouldn't affect "new" pages such as those being used as migration targets. Thesmp_wmb()that was there before doesn't seem necessary for the same reason.Quoted 7 lines
If the patch is okay, we should apply it to the 6.1 stable kernel. The problem appears to be "fixed" in the mainline kernel because of a bad merge conflict resolution on my part; when I rebased commit
e059853d14ca("arm64: mte: Fix/clarify thePG_mte_taggedsemantics") past commit20794545c146, it looks like I accidentally brought back thepage_kasan_tag_reset()line removed in the latter. But we should align the mainline kernel with whatever we decide to do on 6.1.Happy accident ;). When I reverted such calls in commit
20794545c146, my assumption was that we always get a page that went throughpost_alloc_hook()and the tags were reset. But it seems that's not always the case (and probably wasteful anyway if we have to zero the tags and data on a page we know we are going to override viacopy_highpage()anyway). The barrier doesn't help, so we shouldn't add it back.
Quoted 2 lines
So, I'm fine with a stable fix but I wonder whether we should backport the whole "Fix/clarify the
PG_mte_taggedsemantics" series instead.
That seems fine to me (or as well as the above patch if we decide to copy the tag).
Peter
Catalin Marinas
maintainerQuoted 34 lines · Peter Collingbourne
Quoted 19 lines · Catalin Marinas
Quoted 18 lines · Peter Collingbourne
Thanks for the information. We encountered a similar issue internally with the Android 5.15 common kernel. We tracked it down to an issue with page migration, where the source page was a userspace page with MTE tags, and the target page was allocated using KASAN (i.e. having a non-zero KASAN tag). This caused tag check faults when the page was subsequently accessed by the kernel as a result of the mismatching tags from userspace. Given the number of different ways that page migration target pages can be allocated, the simplest fix that we could think of was to synchronize the KASAN tag in
copy_highpage().Can you try the patch below and let us know whether it fixes the issue?
arch/arm64/mm/copypage.c +0
-0@@ -23,6 +23,8 @@ void copy_highpage(struct page *to, struct page *from)Quoted 12 lines
arch/arm64/mm/copypage.c void copy_highpage(struct page *to, struct page *from) · line 28
Quoted 5 lines
arch/arm64/mm/copypage.c +2
-0if (system_supports_mte () &&test_bit (PG_mte_tagged, &from->flags)) {set_bit (PG_mte_tagged, &to->flags); +if (kasan_hw_tags_enabled ()) +page_kasan_tag_set (to,page_kasan_tag (from));mte_copy_page_tags (kto, kfrom);Why not just
page_kasan_tag_reset(to)? IfPG_mte_taggedis set on the 'from' page, the tags are random anyway andpage_kasan_tag(from)should already be0xff. It makes more sense to do the same for the 'to' page rather than copying the tag from the 'from' page. IOW, we are copying user-controlled tags into a page, the kernel should have a match-all tag inpage->flags.
Quoted 3 lines
That would also work, but I was thinking that if
copy_highpage()were being used to copy a KASAN page we should keep the original tag in order to maintain tag checks for page accesses.
If PG_mte_tagged is set on the source, it means that the tags are no longer trusted and we should reset to match-all. Otherwise if copy_highpage() is called on a page that was never mapped as PROT_MTE in user space, PG_mte_tagged would not be set on the source and no tags copied. In such case, we should keep the original KASAN tag in the destination. Unless I misunderstood what you meant.
Signature · 1 line
Catalin
Quoted 40 lines · Catalin Marinas
Quoted 34 lines · Peter Collingbourne
Quoted 19 lines · Catalin Marinas
Quoted 18 lines · Peter Collingbourne
Thanks for the information. We encountered a similar issue internally with the Android 5.15 common kernel. We tracked it down to an issue with page migration, where the source page was a userspace page with MTE tags, and the target page was allocated using KASAN (i.e. having a non-zero KASAN tag). This caused tag check faults when the page was subsequently accessed by the kernel as a result of the mismatching tags from userspace. Given the number of different ways that page migration target pages can be allocated, the simplest fix that we could think of was to synchronize the KASAN tag in
copy_highpage().Can you try the patch below and let us know whether it fixes the issue?
arch/arm64/mm/copypage.c +0
-0@@ -23,6 +23,8 @@ void copy_highpage(struct page *to, struct page *from)Quoted 12 lines
arch/arm64/mm/copypage.c void copy_highpage(struct page *to, struct page *from) · line 28
Quoted 5 lines
arch/arm64/mm/copypage.c +2
-0if (system_supports_mte () &&test_bit (PG_mte_tagged, &from->flags)) {set_bit (PG_mte_tagged, &to->flags); +if (kasan_hw_tags_enabled ()) +page_kasan_tag_set (to,page_kasan_tag (from));mte_copy_page_tags (kto, kfrom);Why not just
page_kasan_tag_reset(to)? IfPG_mte_taggedis set on the 'from' page, the tags are random anyway andpage_kasan_tag(from)should already be0xff. It makes more sense to do the same for the 'to' page rather than copying the tag from the 'from' page. IOW, we are copying user-controlled tags into a page, the kernel should have a match-all tag inpage->flags.Quoted 3 lines
That would also work, but I was thinking that if
copy_highpage()were being used to copy a KASAN page we should keep the original tag in order to maintain tag checks for page accesses.
Quoted 6 lines
If
PG_mte_taggedis set on the source, it means that the tags are no longer trusted and we should reset to match-all. Otherwise ifcopy_highpage()is called on a page that was never mapped asPROT_MTEin user space,PG_mte_taggedwould not be set on the source and no tags copied. In such case, we should keep the original KASAN tag in the destination. Unless I misunderstood what you meant.
I was thinking that it might be possible for PG_mte_tagged to be set on a page with KASAN tags. But as far as I can tell, this isn't actually possible (or not intended to be possible, at least). So I agree, we can just call page_kasan_tag_reset() here.
Patch sent (with stable backport instructions) here:
https://lore.kernel.org/all/[email protected]/
Peter
Quoted 11 lines · Peter Collingbourne
Quoted 10 lines · Qun-wei Lin (林群崴)
Quoted 9 lines · Andrey Konovalov
Quoted 7 lines · Kuan-Ying Lee (李冠穎)
Quoted 7 lines
Hi Kuan-Ying,
There recently was a similar crash due to incorrectly implemented
sampling.
Quoted 7 lines
Quoted 7 lines
Quoted 5 lines
Quoted 5 lines
Quoted 3 lines
Do you have the following patch in your tree?
Quoted 27 lines
Quoted 27 lines
Quoted 15 lines
Quoted 15 lines
Quoted 7 lines
If not, please sync your 6.1 tree with the Android common kernel.
Hopefully this will fix the issue.Thanks!
Hi Andrey,
Thanks for your advice.
I saw this patch is to fix ("kasan: allow sampling
page_allocallocations forHW_TAGS").Quoted 7 lines
Quoted 4 lines
But our 6.1 tree doesn't have following two commits now. ("FROMGIT: kasan: allow sampling
page_allocallocations forHW_TAGS")
(FROMLIST: kasan: reset page tags properly with sampling)Hi Kuan-Ying,
Hi Andrey,
I'll stand in for Kuan-Ying as he's out of office. Thanks for your help!
Quoted 7 lines
Quoted 7 lines
Quoted 5 lines
Just to clarify: these two patches were applied twice: once here on
Jan 13:
Quoted 5 lines
Quoted 5 lines
Quoted 3 lines
Quoted 11 lines
Quoted 11 lines
Quoted 3 lines
Our codebase does not contain these two patches.
Quoted 3 lines
but then reverted here on Jan 20:
Quoted 5 lines
Quoted 5 lines
Quoted 3 lines
Quoted 11 lines
Quoted 11 lines
Quoted 11 lines
And then once again via the link I sent before together with a fix on
Jan 25.It might be that you still have to former two patches in your tree if
you synced it before the revert.
Quoted 6 lines
Quoted 6 lines
Quoted 3 lines
However, if this is not the case:
Which 6.1 commit is your tree based on?
Quoted 44 lines
Quoted 5 lines
(53b3a77 Merge 6.1.1 into android14-6.1) is the latest commit in our
tree.Quoted 7 lines
Quoted 1 line
Do you have any private MTE-related changes in the kernel?
No, all the MTE-related code is the same as Android Common Kernel.
Quoted 1 line
Do you have userspace MTE enabled?
Yes, we have enabled MTE for both EL1 and EL0.
Hi Qun-wei,
Thanks for the information. We encountered a similar issue internally with the Android 5.15 common kernel. We tracked it down to an issue with page migration, where the source page was a userspace page with MTE tags, and the target page was allocated using KASAN (i.e. having a non-zero KASAN tag). This caused tag check faults when the page was subsequently accessed by the kernel as a result of the mismatching tags
from userspace. Given the number of different ways that page migration
target pages can be allocated, the simplest fix that we could think of
was to synchronize the KASAN tag incopy_highpage().Can you try the patch below and let us know whether it fixes the issue?
arch/arm64/mm/copypage.c +2
-0@@ -23,6 +23,8 @@ void copy_highpage(struct page *to, struct page *from)if (system_supports_mte () &&test_bit (PG_mte_tagged, &from- >flags)) {set_bit (PG_mte_tagged, &to->flags); +if (kasan_hw_tags_enabled ()) +page_kasan_tag_set (to,page_kasan_tag (from));mte_copy_page_tags (kto, kfrom);
Quoted 3 lines
}
}
Thank you so much, this patch has solved the problem.
Quoted 22 lines
Catalin, please let us know what you think of the patch above. It effectively partially undoes commit
20794545c146("arm64: kasan:
Revert
"arm64: mte: reset the page tag inpage->flags""), but this seems okay
to me because the mentioned race condition shouldn't affect "new" pages
such as those being used as migration targets. Thesmp_wmb()that was there before doesn't seem necessary for the same reason.If the patch is okay, we should apply it to the 6.1 stable kernel. The
problem appears to be "fixed" in the mainline kernel because of a bad merge conflict resolution on my part; when I rebased commite059853d14ca("arm64: mte: Fix/clarify thePG_mte_taggedsemantics") past commit20794545c146, it looks like I accidentally brought back thepage_kasan_tag_reset()line removed in the latter. But we should align
the mainline kernel with whatever we decide to do on 6.1.Peter