Message ID | 20210708081434.GV3840@techsingularity.net (mailing list archive) |
---|---|
State | New |
Headers | show |
Series | mm/page_alloc: Avoid page allocator recursion with pagesets.lock held | expand |
On Thu, Jul 08, 2021 at 09:14:34AM +0100, Mel Gorman wrote: > Syzbot is reporting potential deadlocks due to pagesets.lock when > PAGE_OWNER is enabled. One example from Desmond Cheong Zhi Xi is > as follows > > __alloc_pages_bulk() > local_lock_irqsave(&pagesets.lock, flags) <---- outer lock here > prep_new_page(): > post_alloc_hook(): > set_page_owner(): > __set_page_owner(): > save_stack(): > stack_depot_save(): > alloc_pages(): > alloc_page_interleave(): > __alloc_pages(): > get_page_from_freelist(): > rm_queue(): > rm_queue_pcplist(): > local_lock_irqsave(&pagesets.lock, flags); > *** DEADLOCK *** > > Zhang, Qiang also reported > > BUG: sleeping function called from invalid context at mm/page_alloc.c:5179 > in_atomic(): 0, irqs_disabled(): 1, non_block: 0, pid: 1, name: swapper/0 > ..... > __dump_stack lib/dump_stack.c:79 [inline] > dump_stack_lvl+0xcd/0x134 lib/dump_stack.c:96 > ___might_sleep.cold+0x1f1/0x237 kernel/sched/core.c:9153 > prepare_alloc_pages+0x3da/0x580 mm/page_alloc.c:5179 > __alloc_pages+0x12f/0x500 mm/page_alloc.c:5375 > alloc_page_interleave+0x1e/0x200 mm/mempolicy.c:2147 > alloc_pages+0x238/0x2a0 mm/mempolicy.c:2270 > stack_depot_save+0x39d/0x4e0 lib/stackdepot.c:303 > save_stack+0x15e/0x1e0 mm/page_owner.c:120 > __set_page_owner+0x50/0x290 mm/page_owner.c:181 > prep_new_page mm/page_alloc.c:2445 [inline] > __alloc_pages_bulk+0x8b9/0x1870 mm/page_alloc.c:5313 > alloc_pages_bulk_array_node include/linux/gfp.h:557 [inline] > vm_area_alloc_pages mm/vmalloc.c:2775 [inline] > __vmalloc_area_node mm/vmalloc.c:2845 [inline] > __vmalloc_node_range+0x39d/0x960 mm/vmalloc.c:2947 > __vmalloc_node mm/vmalloc.c:2996 [inline] > vzalloc+0x67/0x80 mm/vmalloc.c:3066 > > There are a number of ways it could be fixed. The page owner code could > be audited to strip GFP flags that allow sleeping but it'll impair the > functionality of PAGE_OWNER if allocations fail. The bulk allocator > could add a special case to release/reacquire the lock for prep_new_page > and lookup PCP after the lock is reacquired at the cost of performance. > The patches requiring prep could be tracked using the least significant > bit and looping through the array although it is more complicated for > the list interface. The options are relatively complex and the second > one still incurs a performance penalty when PAGE_OWNER is active so this > patch takes the simple approach -- disable bulk allocation of PAGE_OWNER is ^^^^ Minor nit: s/of/if > active. The caller will be forced to allocate one page at a time incurring > a performance penalty but PAGE_OWNER is already a performance penalty. > > Fixes: dbbee9d5cd83 ("mm/page_alloc: convert per-cpu list protection to local_lock") > Reported-by: Desmond Cheong Zhi Xi <desmondcheongzx@gmail.com> > Reported-by: "Zhang, Qiang" <Qiang.Zhang@windriver.com> > Reported-and-tested-by: syzbot+127fd7828d6eeb611703@syzkaller.appspotmail.com > Signed-off-by: Mel Gorman <mgorman@techsingularity.net> > --- > mm/page_alloc.c | 12 ++++++++++++ > 1 file changed, 12 insertions(+) > > diff --git a/mm/page_alloc.c b/mm/page_alloc.c > index 3b97e17806be..6ef86f338151 100644 > --- a/mm/page_alloc.c > +++ b/mm/page_alloc.c > @@ -5239,6 +5239,18 @@ unsigned long __alloc_pages_bulk(gfp_t gfp, int preferred_nid, > if (nr_pages - nr_populated == 1) > goto failed; > > +#ifdef CONFIG_PAGE_OWNER > + /* > + * PAGE_OWNER may recurse into the allocator to allocate space to > + * save the stack with pagesets.lock held. Releasing/reacquiring > + * removes much of the performance benefit of bulk allocation so > + * force the caller to allocate one page at a time as it'll have > + * similar performance to added complexity to the bulk allocator. > + */ > + if (static_branch_unlikely(&page_owner_inited)) > + goto failed; > +#endif > + > /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */ > gfp &= gfp_allowed_mask; > alloc_gfp = gfp; > Acked-by: Rafael Aquini <aquini@redhat.com>
On Thu, Jul 08, 2021 at 09:10:18AM -0400, Rafael Aquini wrote: > > There are a number of ways it could be fixed. The page owner code could > > be audited to strip GFP flags that allow sleeping but it'll impair the > > functionality of PAGE_OWNER if allocations fail. The bulk allocator > > could add a special case to release/reacquire the lock for prep_new_page > > and lookup PCP after the lock is reacquired at the cost of performance. > > The patches requiring prep could be tracked using the least significant If we're nitpicking changelogs, then s/patches/pages/ > > bit and looping through the array although it is more complicated for > > the list interface. The options are relatively complex and the second > > one still incurs a performance penalty when PAGE_OWNER is active so this > > patch takes the simple approach -- disable bulk allocation of PAGE_OWNER is > ^^^^ > Minor nit: s/of/if > > > active. The caller will be forced to allocate one page at a time incurring > > a performance penalty but PAGE_OWNER is already a performance penalty. The thought occurs that all pages allocated this way will have the same stack. Someone suitably motivated could: - Refcount the stacks managed by page_owner - Generate the stack in the caller of alloc_bulk - Pass the stack in to a new __set_page_owner() That's a lot of work; I'm not going to do it, and I don't think Mel should either. But if someone's looking for something to do ...
diff --git a/mm/page_alloc.c b/mm/page_alloc.c index 3b97e17806be..6ef86f338151 100644 --- a/mm/page_alloc.c +++ b/mm/page_alloc.c @@ -5239,6 +5239,18 @@ unsigned long __alloc_pages_bulk(gfp_t gfp, int preferred_nid, if (nr_pages - nr_populated == 1) goto failed; +#ifdef CONFIG_PAGE_OWNER + /* + * PAGE_OWNER may recurse into the allocator to allocate space to + * save the stack with pagesets.lock held. Releasing/reacquiring + * removes much of the performance benefit of bulk allocation so + * force the caller to allocate one page at a time as it'll have + * similar performance to added complexity to the bulk allocator. + */ + if (static_branch_unlikely(&page_owner_inited)) + goto failed; +#endif + /* May set ALLOC_NOFRAGMENT, fragmentation will return 1 page. */ gfp &= gfp_allowed_mask; alloc_gfp = gfp;
Syzbot is reporting potential deadlocks due to pagesets.lock when PAGE_OWNER is enabled. One example from Desmond Cheong Zhi Xi is as follows __alloc_pages_bulk() local_lock_irqsave(&pagesets.lock, flags) <---- outer lock here prep_new_page(): post_alloc_hook(): set_page_owner(): __set_page_owner(): save_stack(): stack_depot_save(): alloc_pages(): alloc_page_interleave(): __alloc_pages(): get_page_from_freelist(): rm_queue(): rm_queue_pcplist(): local_lock_irqsave(&pagesets.lock, flags); *** DEADLOCK *** Zhang, Qiang also reported BUG: sleeping function called from invalid context at mm/page_alloc.c:5179 in_atomic(): 0, irqs_disabled(): 1, non_block: 0, pid: 1, name: swapper/0 ..... __dump_stack lib/dump_stack.c:79 [inline] dump_stack_lvl+0xcd/0x134 lib/dump_stack.c:96 ___might_sleep.cold+0x1f1/0x237 kernel/sched/core.c:9153 prepare_alloc_pages+0x3da/0x580 mm/page_alloc.c:5179 __alloc_pages+0x12f/0x500 mm/page_alloc.c:5375 alloc_page_interleave+0x1e/0x200 mm/mempolicy.c:2147 alloc_pages+0x238/0x2a0 mm/mempolicy.c:2270 stack_depot_save+0x39d/0x4e0 lib/stackdepot.c:303 save_stack+0x15e/0x1e0 mm/page_owner.c:120 __set_page_owner+0x50/0x290 mm/page_owner.c:181 prep_new_page mm/page_alloc.c:2445 [inline] __alloc_pages_bulk+0x8b9/0x1870 mm/page_alloc.c:5313 alloc_pages_bulk_array_node include/linux/gfp.h:557 [inline] vm_area_alloc_pages mm/vmalloc.c:2775 [inline] __vmalloc_area_node mm/vmalloc.c:2845 [inline] __vmalloc_node_range+0x39d/0x960 mm/vmalloc.c:2947 __vmalloc_node mm/vmalloc.c:2996 [inline] vzalloc+0x67/0x80 mm/vmalloc.c:3066 There are a number of ways it could be fixed. The page owner code could be audited to strip GFP flags that allow sleeping but it'll impair the functionality of PAGE_OWNER if allocations fail. The bulk allocator could add a special case to release/reacquire the lock for prep_new_page and lookup PCP after the lock is reacquired at the cost of performance. The patches requiring prep could be tracked using the least significant bit and looping through the array although it is more complicated for the list interface. The options are relatively complex and the second one still incurs a performance penalty when PAGE_OWNER is active so this patch takes the simple approach -- disable bulk allocation of PAGE_OWNER is active. The caller will be forced to allocate one page at a time incurring a performance penalty but PAGE_OWNER is already a performance penalty. Fixes: dbbee9d5cd83 ("mm/page_alloc: convert per-cpu list protection to local_lock") Reported-by: Desmond Cheong Zhi Xi <desmondcheongzx@gmail.com> Reported-by: "Zhang, Qiang" <Qiang.Zhang@windriver.com> Reported-and-tested-by: syzbot+127fd7828d6eeb611703@syzkaller.appspotmail.com Signed-off-by: Mel Gorman <mgorman@techsingularity.net> --- mm/page_alloc.c | 12 ++++++++++++ 1 file changed, 12 insertions(+)