Re: [PATCH v2 05/14] mm: Introduce mf_dax_kill_procs() for fsdax case

[Dan Williams]

Dan Williams wrote:
> Dan Williams wrote:
> > HORIGUCHI NAOYA(堀口　直也) wrote:
> > > On Wed, Aug 24, 2022 at 02:52:51PM -0700, Dan Williams wrote:
> > > > Shiyang Ruan wrote:
> > > > > This new function is a variant of mf_generic_kill_procs that accepts a
> > > > > file, offset pair instead of a struct to support multiple files sharing
> > > > > a DAX mapping.  It is intended to be called by the file systems as part
> > > > > of the memory_failure handler after the file system performed a reverse
> > > > > mapping from the storage address to the file and file offset.
> > > > > 
> > > > > Signed-off-by: Shiyang Ruan <ruansy.fnst@xxxxxxxxxxx>
> > > > > Reviewed-by: Dan Williams <dan.j.williams@xxxxxxxxx>
> > > > > Reviewed-by: Christoph Hellwig <hch@xxxxxx>
> > > > > Reviewed-by: Darrick J. Wong <djwong@xxxxxxxxxx>
> > > > > Reviewed-by: Miaohe Lin <linmiaohe@xxxxxxxxxx>
> > > > > ---
> > > > >  include/linux/mm.h  |  2 +
> > > > >  mm/memory-failure.c | 96 ++++++++++++++++++++++++++++++++++++++++-----
> > > > >  2 files changed, 88 insertions(+), 10 deletions(-)
> > > > 
> > > > Unfortunately my test suite was only running the "non-destructive" set
> > > > of 'ndctl' tests which skipped some of the complex memory-failure cases.
> > > > Upon fixing that, bisect flags this commit as the source of the following
> > > > crash regression:
> > > 
> > > Thank you for testing/reporting.
> > > 
> > > > 
> > > >  kernel BUG at mm/memory-failure.c:310!
> > > >  invalid opcode: 0000 [#1] PREEMPT SMP PTI
> > > >  CPU: 26 PID: 1252 Comm: dax-pmd Tainted: G           OE     5.19.0-rc4+ #58
> > > >  Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 0.0.0 02/06/2015
> > > >  RIP: 0010:add_to_kill+0x304/0x400
> > > > [..]
> > > >  Call Trace:
> > > >   <TASK>
> > > >   collect_procs.part.0+0x2c8/0x470
> > > >   memory_failure+0x979/0xf30
> > > >   do_madvise.part.0.cold+0x9c/0xd3
> > > >   ? lock_is_held_type+0xe3/0x140
> > > >   ? find_held_lock+0x2b/0x80
> > > >   ? lock_release+0x145/0x2f0
> > > >   ? lock_is_held_type+0xe3/0x140
> > > >   ? syscall_enter_from_user_mode+0x20/0x70
> > > >   __x64_sys_madvise+0x56/0x70
> > > >   do_syscall_64+0x3a/0x80
> > > >   entry_SYSCALL_64_after_hwframe+0x46/0xb0
> > > 
> > > This stacktrace shows that VM_BUG_ON_VMA() in dev_pagemap_mapping_shift()
> > > was triggered.  I think that BUG_ON is too harsh here because address ==
> > > -EFAULT means that there's no mapping for the address.  The subsequent
> > > code considers "tk->size_shift == 0" as "no mapping" cases, so
> > > dev_pagemap_mapping_shift() can return 0 in such a case?
> > > 
> > > Could the following diff work for the issue?
> > 
> > This passes the "dax-ext4.sh" and "dax-xfs.sh" tests from the ndctl
> > suite.

So that diff works to avoid the BUG_ON, but it does not work to handle
the error case. I think the problem comes from:

    vma->vm_file->f_mapping != folio->mapping

...where page_folio(page)->mapping is likely not setup correctly for DAX
pages. This goes back to the broken nature of DAX page reference
counting which I am fixing now, but this folio association also needs to
be fixed up.