From 1c00f9368628dde7337defd3699025e3611a816f Mon Sep 17 00:00:00 2001 From: Dave Chinner Date: Tue, 30 Apr 2024 15:28:23 +1000 Subject: mm: lift gfp_kmemleak_mask() to gfp.h Patch series "mm: fix nested allocation context filtering". This patchset is the followup to the comment I made earlier today: https://lore.kernel.org/linux-xfs/ZjAyIWUzDipofHFJ@dread.disaster.area/ Tl;dr: Memory allocations that are done inside the public memory allocation API need to obey the reclaim recursion constraints placed on the allocation by the original caller, including the "don't track recursion for this allocation" case defined by __GFP_NOLOCKDEP. These nested allocations are generally in debug code that is tracking something about the allocation (kmemleak, KASAN, etc) and so are allocating private kernel objects that only that debug system will use. Neither the page-owner code nor the stack depot code get this right. They also also clear GFP_ZONEMASK as a separate operation, which is completely redundant because the constraint filter applied immediately after guarantees that GFP_ZONEMASK bits are cleared. kmemleak gets this filtering right. It preserves the allocation constraints for deadlock prevention and clears all other context flags whilst also ensuring that the nested allocation will fail quickly, silently and without depleting emergency kernel reserves if there is no memory available. This can be made much more robust, immune to whack-a-mole games and the code greatly simplified by lifting gfp_kmemleak_mask() to include/linux/gfp.h and using that everywhere. Also document it so that there is no excuse for not knowing about it when writing new debug code that nests allocations. Tested with lockdep, KASAN + page_owner=on and kmemleak=on over multiple fstests runs with XFS. This patch (of 3): Any "internal" nested allocation done from within an allocation context needs to obey the high level allocation gfp_mask constraints. This is necessary for debug code like KASAN, kmemleak, lockdep, etc that allocate memory for saving stack traces and other information during memory allocation. If they don't obey things like __GFP_NOLOCKDEP or __GFP_NOWARN, they produce false positive failure detections. kmemleak gets this right by using gfp_kmemleak_mask() to pass through the relevant context flags to the nested allocation to ensure that the allocation follows the constraints of the caller context. KASAN recently was foudn to be missing __GFP_NOLOCKDEP due to stack depot allocations, and even more recently the page owner tracking code was also found to be missing __GFP_NOLOCKDEP support. We also don't wan't want KASAN or lockdep to drive the system into OOM kill territory by exhausting emergency reserves. This is something that kmemleak also gets right by adding (__GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN) to the allocation mask. Hence it is clear that we need to define a common nested allocation filter mask for these sorts of third party nested allocations used in debug code. So to start this process, lift gfp_kmemleak_mask() to gfp.h and rename it to gfp_nested_mask(), and convert the kmemleak callers to use it. Link: https://lkml.kernel.org/r/20240430054604.4169568-1-david@fromorbit.com Link: https://lkml.kernel.org/r/20240430054604.4169568-2-david@fromorbit.com Signed-off-by: Dave Chinner Reviewed-by: Marco Elver Reviewed-by: Christoph Hellwig Reviewed-by: Vlastimil Babka Reviewed-by: Oscar Salvador Cc: Andrey Konovalov Signed-off-by: Andrew Morton --- include/linux/gfp.h | 25 +++++++++++++++++++++++++ mm/kmemleak.c | 12 ++++-------- 2 files changed, 29 insertions(+), 8 deletions(-) diff --git a/include/linux/gfp.h b/include/linux/gfp.h index 450c2cbcf04b..7f9691d375f0 100644 --- a/include/linux/gfp.h +++ b/include/linux/gfp.h @@ -156,6 +156,31 @@ static inline int gfp_zonelist(gfp_t flags) return ZONELIST_FALLBACK; } +/* + * gfp flag masking for nested internal allocations. + * + * For code that needs to do allocations inside the public allocation API (e.g. + * memory allocation tracking code) the allocations need to obey the caller + * allocation context constrains to prevent allocation context mismatches (e.g. + * GFP_KERNEL allocations in GFP_NOFS contexts) from potential deadlock + * situations. + * + * It is also assumed that these nested allocations are for internal kernel + * object storage purposes only and are not going to be used for DMA, etc. Hence + * we strip out all the zone information and leave just the context information + * intact. + * + * Further, internal allocations must fail before the higher level allocation + * can fail, so we must make them fail faster and fail silently. We also don't + * want them to deplete emergency reserves. Hence nested allocations must be + * prepared for these allocations to fail. + */ +static inline gfp_t gfp_nested_mask(gfp_t flags) +{ + return ((flags & (GFP_KERNEL | GFP_ATOMIC | __GFP_NOLOCKDEP)) | + (__GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN)); +} + /* * We get the zone list from the current node and the gfp_mask. * This zone list contains a maximum of MAX_NUMNODES*MAX_NR_ZONES zones. diff --git a/mm/kmemleak.c b/mm/kmemleak.c index fdcf01f62202..d5b6fba44fc9 100644 --- a/mm/kmemleak.c +++ b/mm/kmemleak.c @@ -114,12 +114,6 @@ #define BYTES_PER_POINTER sizeof(void *) -/* GFP bitmask for kmemleak internal allocations */ -#define gfp_kmemleak_mask(gfp) (((gfp) & (GFP_KERNEL | GFP_ATOMIC | \ - __GFP_NOLOCKDEP)) | \ - __GFP_NORETRY | __GFP_NOMEMALLOC | \ - __GFP_NOWARN) - /* scanning area inside a memory block */ struct kmemleak_scan_area { struct hlist_node node; @@ -463,7 +457,8 @@ static struct kmemleak_object *mem_pool_alloc(gfp_t gfp) /* try the slab allocator first */ if (object_cache) { - object = kmem_cache_alloc_noprof(object_cache, gfp_kmemleak_mask(gfp)); + object = kmem_cache_alloc_noprof(object_cache, + gfp_nested_mask(gfp)); if (object) return object; } @@ -947,7 +942,8 @@ static void add_scan_area(unsigned long ptr, size_t size, gfp_t gfp) untagged_objp = (unsigned long)kasan_reset_tag((void *)object->pointer); if (scan_area_cache) - area = kmem_cache_alloc_noprof(scan_area_cache, gfp_kmemleak_mask(gfp)); + area = kmem_cache_alloc_noprof(scan_area_cache, + gfp_nested_mask(gfp)); raw_spin_lock_irqsave(&object->lock, flags); if (!area) { -- cgit v1.2.3 From 70c435ca8dcb64e3d7983a30a14484aa163bb2d2 Mon Sep 17 00:00:00 2001 From: Dave Chinner Date: Tue, 30 Apr 2024 15:28:24 +1000 Subject: stackdepot: use gfp_nested_mask() instead of open coded masking The stackdepot code is used by KASAN and lockdep for recoding stack traces. Both of these track allocation context information, and so their internal allocations must obey the caller allocation contexts to avoid generating their own false positive warnings that have nothing to do with the code they are instrumenting/tracking. We also don't want recording stack traces to deplete emergency memory reserves - debug code is useless if it creates new issues that can't be replicated when the debug code is disabled. Switch the stackdepot allocation masking to use gfp_nested_mask() to address these issues. gfp_nested_mask() also strips GFP_ZONEMASK naturally, so that greatly simplifies this code. Link: https://lkml.kernel.org/r/20240430054604.4169568-3-david@fromorbit.com Signed-off-by: Dave Chinner Reviewed-by: Marco Elver Reviewed-by: Christoph Hellwig Reviewed-by: Vlastimil Babka Reviewed-by: Oscar Salvador Cc: Andrey Konovalov Signed-off-by: Andrew Morton --- lib/stackdepot.c | 11 ++--------- 1 file changed, 2 insertions(+), 9 deletions(-) diff --git a/lib/stackdepot.c b/lib/stackdepot.c index cd8f23455285..5ed34cc963fc 100644 --- a/lib/stackdepot.c +++ b/lib/stackdepot.c @@ -624,15 +624,8 @@ depot_stack_handle_t stack_depot_save_flags(unsigned long *entries, * we won't be able to do that under the lock. */ if (unlikely(can_alloc && !READ_ONCE(new_pool))) { - /* - * Zero out zone modifiers, as we don't have specific zone - * requirements. Keep the flags related to allocation in atomic - * contexts, I/O, nolockdep. - */ - alloc_flags &= ~GFP_ZONEMASK; - alloc_flags &= (GFP_ATOMIC | GFP_KERNEL | __GFP_NOLOCKDEP); - alloc_flags |= __GFP_NOWARN; - page = alloc_pages(alloc_flags, DEPOT_POOL_ORDER); + page = alloc_pages(gfp_nested_mask(alloc_flags), + DEPOT_POOL_ORDER); if (page) prealloc = page_address(page); } -- cgit v1.2.3 From 99b80ac45f7ec351c2d1c9fbfec702213dcae566 Mon Sep 17 00:00:00 2001 From: Dave Chinner Date: Tue, 30 Apr 2024 15:28:25 +1000 Subject: mm/page-owner: use gfp_nested_mask() instead of open coded masking The page-owner tracking code records stack traces during page allocation. To do this, it must do a memory allocation for the stack information from inside an existing memory allocation context. This internal allocation must obey the high level caller allocation constraints to avoid generating false positive warnings that have nothing to do with the code they are instrumenting/tracking (e.g. through lockdep reclaim state tracking) We also don't want recording stack traces to deplete emergency memory reserves - debug code is useless if it creates new issues that can't be replicated when the debug code is disabled. Switch the stack tracking allocation masking to use gfp_nested_mask() to address these issues. gfp_nested_mask() naturally strips GFP_ZONEMASK, too, which greatly simplifies this code. Link: https://lkml.kernel.org/r/20240430054604.4169568-4-david@fromorbit.com Signed-off-by: Dave Chinner Reviewed-by: Christoph Hellwig Reviewed-by: Vlastimil Babka Reviewed-by: Oscar Salvador Cc: Andrey Konovalov Cc: Marco Elver Signed-off-by: Andrew Morton --- mm/page_owner.c | 7 +------ 1 file changed, 1 insertion(+), 6 deletions(-) diff --git a/mm/page_owner.c b/mm/page_owner.c index 0a448d76e0af..2d6360eaccbb 100644 --- a/mm/page_owner.c +++ b/mm/page_owner.c @@ -168,13 +168,8 @@ static void add_stack_record_to_list(struct stack_record *stack_record, unsigned long flags; struct stack *stack; - /* Filter gfp_mask the same way stackdepot does, for consistency */ - gfp_mask &= ~GFP_ZONEMASK; - gfp_mask &= (GFP_ATOMIC | GFP_KERNEL | __GFP_NOLOCKDEP); - gfp_mask |= __GFP_NOWARN; - set_current_in_page_owner(); - stack = kmalloc(sizeof(*stack), gfp_mask); + stack = kmalloc(sizeof(*stack), gfp_nested_mask(gfp_mask)); if (!stack) { unset_current_in_page_owner(); return; -- cgit v1.2.3