]> git.proxmox.com Git - mirror_ubuntu-kernels.git/commit
mm/madvise: make MADV_POPULATE_(READ|WRITE) handle VM_FAULT_RETRY properly
authorDavid Hildenbrand <david@redhat.com>
Thu, 14 Mar 2024 16:12:59 +0000 (17:12 +0100)
committerAndrew Morton <akpm@linux-foundation.org>
Tue, 16 Apr 2024 22:39:48 +0000 (15:39 -0700)
commit631426ba1d45a8672b177ee85ad4cabe760dd131
treeefbba3e1052b75c03bc0c9fbd1463a6f3d5585a6
parent0bbac3facb5d6cc0171c45c9873a2dc96bea9680
mm/madvise: make MADV_POPULATE_(READ|WRITE) handle VM_FAULT_RETRY properly

Darrick reports that in some cases where pread() would fail with -EIO and
mmap()+access would generate a SIGBUS signal, MADV_POPULATE_READ /
MADV_POPULATE_WRITE will keep retrying forever and not fail with -EFAULT.

While the madvise() call can be interrupted by a signal, this is not the
desired behavior.  MADV_POPULATE_READ / MADV_POPULATE_WRITE should behave
like page faults in that case: fail and not retry forever.

A reproducer can be found at [1].

The reason is that __get_user_pages(), as called by
faultin_vma_page_range(), will not handle VM_FAULT_RETRY in a proper way:
it will simply return 0 when VM_FAULT_RETRY happened, making
madvise_populate()->faultin_vma_page_range() retry again and again, never
setting FOLL_TRIED->FAULT_FLAG_TRIED for __get_user_pages().

__get_user_pages_locked() does what we want, but duplicating that logic in
faultin_vma_page_range() feels wrong.

So let's use __get_user_pages_locked() instead, that will detect
VM_FAULT_RETRY and set FOLL_TRIED when retrying, making the fault handler
return VM_FAULT_SIGBUS (VM_FAULT_ERROR) at some point, propagating -EFAULT
from faultin_page() to __get_user_pages(), all the way to
madvise_populate().

But, there is an issue: __get_user_pages_locked() will end up re-taking
the MM lock and then __get_user_pages() will do another VMA lookup.  In
the meantime, the VMA layout could have changed and we'd fail with
different error codes than we'd want to.

As __get_user_pages() will currently do a new VMA lookup either way, let
it do the VMA handling in a different way, controlled by a new
FOLL_MADV_POPULATE flag, effectively moving these checks from
madvise_populate() + faultin_page_range() in there.

With this change, Darricks reproducer properly fails with -EFAULT, as
documented for MADV_POPULATE_READ / MADV_POPULATE_WRITE.

[1] https://lore.kernel.org/all/20240313171936.GN1927156@frogsfrogsfrogs/

Link: https://lkml.kernel.org/r/20240314161300.382526-1-david@redhat.com
Link: https://lkml.kernel.org/r/20240314161300.382526-2-david@redhat.com
Fixes: 4ca9b3859dac ("mm/madvise: introduce MADV_POPULATE_(READ|WRITE) to prefault page tables")
Signed-off-by: David Hildenbrand <david@redhat.com>
Reported-by: Darrick J. Wong <djwong@kernel.org>
Closes: https://lore.kernel.org/all/20240311223815.GW1927156@frogsfrogsfrogs/
Cc: Darrick J. Wong <djwong@kernel.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
mm/gup.c
mm/internal.h
mm/madvise.c