This is a note to let you know that I've just added the patch titled mm: fix page leak with multiple threads mapping the same page to the 5.15-stable tree which can be found at: http://www.kernel.org/git/?p=linux/kernel/git/stable/stable-queue.git;a=summary The filename of the patch is: mm-fix-page-leak-with-multiple-threads-mapping-the-same-page.patch and it can be found in the queue-5.15 subdirectory. If you, or anyone else, feels it should not be added to the stable tree, please let <stable@xxxxxxxxxxxxxxx> know about it. >From 3fe2895cfecd03ac74977f32102b966b6589f481 Mon Sep 17 00:00:00 2001 From: Josef Bacik <josef@xxxxxxxxxxxxxx> Date: Tue, 5 Jul 2022 16:00:36 -0400 Subject: mm: fix page leak with multiple threads mapping the same page From: Josef Bacik <josef@xxxxxxxxxxxxxx> commit 3fe2895cfecd03ac74977f32102b966b6589f481 upstream. 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. [josef@xxxxxxxxxxxxxx: v2] Link: https://lkml.kernel.org/r/e90c8f0dbae836632b669c2afc434006a00d4a67.1657721478.git.josef@xxxxxxxxxxxxxx Link: https://lkml.kernel.org/r/2b798acfd95c9ab9395fe85e8d5a835e2e10a920.1657051137.git.josef@xxxxxxxxxxxxxx Fixes: f9ce0be71d1f ("mm: Cleanup faultaround and finish_fault() codepaths") Signed-off-by: Josef Bacik <josef@xxxxxxxxxxxxxx> Signed-off-by: Rik van Riel <riel@xxxxxxxxxxx> Signed-off-by: Chris Mason <clm@xxxxxx> Acked-by: Kirill A. Shutemov <kirill.shutemov@xxxxxxxxxxxxxxx> Cc: Matthew Wilcox (Oracle) <willy@xxxxxxxxxxxxx> Cc: <stable@xxxxxxxxxxxxxxx> Signed-off-by: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> Signed-off-by: Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx> --- mm/memory.c | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-) --- a/mm/memory.c +++ b/mm/memory.c @@ -4071,9 +4071,12 @@ vm_fault_t finish_fault(struct vm_fault } } - /* See comment in handle_pte_fault() */ + /* + * See comment in handle_pte_fault() for how this scenario happens, we + * need to return NOPAGE so that we drop this page. + */ 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); Patches currently in stable-queue which might be from josef@xxxxxxxxxxxxxx are queue-5.15/mm-fix-page-leak-with-multiple-threads-mapping-the-same-page.patch