Re: [PATCH] mm/page_alloc: avoid deadlocks for &pagesets.lock

[Mel Gorman]

On Wed, Jul 07, 2021 at 07:12:45PM +0800, Desmond Cheong Zhi Xi wrote:
> Syzbot reports a number of potential deadlocks for &pagesets.lock. It
> seems that this new lock is being used as both an inner and outer
> lock, which makes it prone to creating circular dependencies.
> 
> For example, one such call trace goes 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 ***
> 
> The common culprit for the lockdep splats seems to be the call to
> local_lock_irqsave(&pagesets.lock, flags) inside
> __alloc_pages_bulk(). &pagesets.lock becomes an outer lock if it's
> held during the call to prep_new_page().
> 
> As the local lock is used to protect the PCP structure, we adjust the
> locking in __alloc_pages_bulk so that only the necessary structures
> are protected.
> 
> Fixes: dbbee9d5cd83 ("mm/page_alloc: convert per-cpu list protection to local_lock")
> Reported-and-tested-by: syzbot+127fd7828d6eeb611703@xxxxxxxxxxxxxxxxxxxxxxxxx
> Signed-off-by: Desmond Cheong Zhi Xi <desmondcheongzx@xxxxxxxxx>

Hi Desmond,

Thanks for the report. Unfortunately, this patch incurs a performance
penalty for the bulk allocator even if PAGE_OWNER is disabled. Can you
try the following as an alternative please? It passed a build and boot
test but I didn't try triggering the actual bug.

--8<--
mm/page_alloc: Avoid page allocator recursion with pagesets.lock held

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.
Both 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@xxxxxxxxx>
Reported-by: "Zhang, Qiang" <Qiang.Zhang@xxxxxxxxxxxxx>
Reported-by: syzbot+127fd7828d6eeb611703@xxxxxxxxxxxxxxxxxxxxxxxxx
Signed-off-by: Mel Gorman <mgorman@xxxxxxxxxxxxxxxxxxx>
---
 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;