We have an application with a lot of threads that use a shared mmap
backed by tmpfs mounted with -o huge=within_size. This application
started leaking loads of huge pages when we upgraded to a recent kernel.
Using the page ref tracepoints and a BPF program written by Tejun Heo we
were able to determine that these pages would have multiple refcounts
from the page fault path, but when it came to unmap time we wouldn't
drop the number of refs we had added from the faults.
I wrote a reproducer that mmap'ed a file backed by tmpfs with -o
huge=always, and then spawned 20 threads all looping faulting random
offsets in this map, while using madvise(MADV_DONTNEED) randomly for
huge page aligned ranges. This very quickly reproduced the problem.
The problem here is that we check for the case that we have multiple
threads faulting in a range that was previously unmapped. One thread
maps the PMD, the other thread loses the race and then returns 0.
However at this point we already have the page, and we are no longer
putting this page into the processes address space, and so we leak the
page. We actually did the correct thing prior to f9ce0be71d1f, however
it looks like Kirill copied what we do in the anonymous page case. In
the anonymous page case we don't yet have a page, so we don't have to
drop a reference on anything. Previously we did the correct thing for
file based faults by returning VM_FAULT_NOPAGE so we correctly drop the
reference on the page we faulted in.
Fix this by returning VM_FAULT_NOPAGE in the pmd_devmap_trans_unstable()
case, this makes us drop the ref on the page properly, and now my
reproducer no longer leaks the huge pages.
Fixes: f9ce0be71d1f ("mm: Cleanup faultaround and finish_fault() codepaths")
Cc: Kirill A. Shutemov <kirill.shutemov@xxxxxxxxxxxxxxx>
Cc: Matthew Wilcox (Oracle) <willy@xxxxxxxxxxxxx>
Signed-off-by: Josef Bacik <josef@xxxxxxxxxxxxxx>
Signed-off-by: Rik van Riel <riel@xxxxxxxxxxx>
Signed-off-by: Chris Mason <clm@xxxxxx>
---
mm/memory.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/mm/memory.c b/mm/memory.c
index 7a089145cad4..f10724d7dca3 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -4371,7 +4371,7 @@ vm_fault_t finish_fault(struct vm_fault *vmf)
/* See comment in handle_pte_fault() */
if (pmd_devmap_trans_unstable(vmf->pmd))
- return 0;
+ return VM_FAULT_NOPAGE;
vmf->pte = pte_offset_map_lock(vma->vm_mm, vmf->pmd,
vmf->address, &vmf->ptl);
--
2.36.1
