Re: [PATCH RFC] nfsd: avoid recursive locking through fsnotify

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Mon 21-03-22 13:56:47, Amir Goldstein wrote:
> On Mon, Mar 21, 2022 at 1:23 PM Jan Kara <jack@xxxxxxx> wrote:
> >
> > On Sat 19-03-22 11:36:13, Amir Goldstein wrote:
> > > On Sat, Mar 19, 2022 at 9:02 AM Trond Myklebust <trondmy@xxxxxxxxxxxxxxx> wrote:
> > > >
> > > > On Fri, 2022-03-18 at 17:16 -0700, Khazhismel Kumykov wrote:
> > > > > fsnotify_add_inode_mark may allocate with GFP_KERNEL, which may
> > > > > result
> > > > > in recursing back into nfsd, resulting in deadlock. See below stack.
> > > > >
> > > > > nfsd            D    0 1591536      2 0x80004080
> > > > > Call Trace:
> > > > >  __schedule+0x497/0x630
> > > > >  schedule+0x67/0x90
> > > > >  schedule_preempt_disabled+0xe/0x10
> > > > >  __mutex_lock+0x347/0x4b0
> > > > >  fsnotify_destroy_mark+0x22/0xa0
> > > > >  nfsd_file_free+0x79/0xd0 [nfsd]
> > > > >  nfsd_file_put_noref+0x7c/0x90 [nfsd]
> > > > >  nfsd_file_lru_dispose+0x6d/0xa0 [nfsd]
> > > > >  nfsd_file_lru_scan+0x57/0x80 [nfsd]
> > > > >  do_shrink_slab+0x1f2/0x330
> > > > >  shrink_slab+0x244/0x2f0
> > > > >  shrink_node+0xd7/0x490
> > > > >  do_try_to_free_pages+0x12f/0x3b0
> > > > >  try_to_free_pages+0x43f/0x540
> > > > >  __alloc_pages_slowpath+0x6ab/0x11c0
> > > > >  __alloc_pages_nodemask+0x274/0x2c0
> > > > >  alloc_slab_page+0x32/0x2e0
> > > > >  new_slab+0xa6/0x8b0
> > > > >  ___slab_alloc+0x34b/0x520
> > > > >  kmem_cache_alloc+0x1c4/0x250
> > > > >  fsnotify_add_mark_locked+0x18d/0x4c0
> > > > >  fsnotify_add_mark+0x48/0x70
> > > > >  nfsd_file_acquire+0x570/0x6f0 [nfsd]
> > > > >  nfsd_read+0xa7/0x1c0 [nfsd]
> > > > >  nfsd3_proc_read+0xc1/0x110 [nfsd]
> > > > >  nfsd_dispatch+0xf7/0x240 [nfsd]
> > > > >  svc_process_common+0x2f4/0x610 [sunrpc]
> > > > >  svc_process+0xf9/0x110 [sunrpc]
> > > > >  nfsd+0x10e/0x180 [nfsd]
> > > > >  kthread+0x130/0x140
> > > > >  ret_from_fork+0x35/0x40
> > > > >
> > > > > Signed-off-by: Khazhismel Kumykov <khazhy@xxxxxxxxxx>
> > > > > ---
> > > > >  fs/nfsd/filecache.c | 4 ++++
> > > > >  1 file changed, 4 insertions(+)
> > > > >
> > > > > Marking this RFC since I haven't actually had a chance to test this,
> > > > > we
> > > > > we're seeing this deadlock for some customers.
> > > > >
> > > > > diff --git a/fs/nfsd/filecache.c b/fs/nfsd/filecache.c
> > > > > index fdf89fcf1a0c..a14760f9b486 100644
> > > > > --- a/fs/nfsd/filecache.c
> > > > > +++ b/fs/nfsd/filecache.c
> > > > > @@ -121,6 +121,7 @@ nfsd_file_mark_find_or_create(struct nfsd_file
> > > > > *nf)
> > > > >         struct fsnotify_mark    *mark;
> > > > >         struct nfsd_file_mark   *nfm = NULL, *new;
> > > > >         struct inode *inode = nf->nf_inode;
> > > > > +       unsigned int pflags;
> > > > >
> > > > >         do {
> > > > >                 mutex_lock(&nfsd_file_fsnotify_group->mark_mutex);
> > > > > @@ -149,7 +150,10 @@ nfsd_file_mark_find_or_create(struct nfsd_file
> > > > > *nf)
> > > > >                 new->nfm_mark.mask = FS_ATTRIB|FS_DELETE_SELF;
> > > > >                 refcount_set(&new->nfm_ref, 1);
> > > > >
> > > > > +               /* fsnotify allocates, avoid recursion back into nfsd
> > > > > */
> > > > > +               pflags = memalloc_nofs_save();
> > > > >                 err = fsnotify_add_inode_mark(&new->nfm_mark, inode,
> > > > > 0);
> > > > > +               memalloc_nofs_restore(pflags);
> > > > >
> > > > >                 /*
> > > > >                  * If the add was successful, then return the object.
> > > >
> > > > Isn't that stack trace showing a slab direct reclaim, and not a
> > > > filesystem writeback situation?
> > > >
> > > > Does memalloc_nofs_save()/restore() really fix this problem? It seems
> > > > to me that it cannot, particularly since knfsd is not a filesystem, and
> > > > so does not ever handle writeback of dirty pages.
> > > >
> > >
> > > Maybe NOFS throttles direct reclaims to the point that the problem is
> > > harder to hit?
> > >
> > > This report came in at good timing for me.
> > >
> > > It demonstrates an issue I did not predict for "volatile"' fanotify marks [1].
> > > As far as I can tell, nfsd filecache is currently the only fsnotify backend that
> > > frees fsnotify marks in memory shrinker. "volatile" fanotify marks would also
> > > be evictable in that way, so they would expose fanotify to this deadlock.
> > >
> > > For the short term, maybe nfsd filecache can avoid the problem by checking
> > > mutex_is_locked(&nfsd_file_fsnotify_group->mark_mutex) and abort the
> > > shrinker. I wonder if there is a place for a helper mutex_is_locked_by_me()?
> > >
> > > Jan,
> > >
> > > A relatively simple fix would be to allocate fsnotify_mark_connector in
> > > fsnotify_add_mark() and free it, if a connector already exists for the object.
> > > I don't think there is a good reason to optimize away this allocation
> > > for the case of a non-first group to set a mark on an object?
> >
> > Indeed, nasty. Volatile marks will add group->mark_mutex into a set of
> > locks grabbed during inode slab reclaim. So any allocation under
> > group->mark_mutex has to be GFP_NOFS now. This is not just about connector
> > allocations but also mark allocations for fanotify. Moving allocations from
> > under mark_mutex is also possible solution but passing preallocated memory
> > around is kind of ugly as well.
> 
> Yes, kind of, here is how it looks:
> https://github.com/amir73il/linux/commit/643bb6b9f664f70f68ea0393a06338673c4966b3
> https://github.com/amir73il/linux/commit/66f27fc99e46b12f1078e8e2915793040ce50ee7

Yup, not an act of beauty but bearable in the worst case :).

> > So the cleanest solution I currently see is
> > to come up with helpers like "fsnotify_lock_group() &
> > fsnotify_unlock_group()" which will lock/unlock mark_mutex and also do
> > memalloc_nofs_save / restore magic.
> >
> 
> Sounds good. Won't this cause a regression - more failures to setup new mark
> under memory pressure?

Well, yes, the chances of hitting ENOMEM under heavy memory pressure are
higher. But I don't think that much memory is consumed by connectors or
marks that the reduced chances for direct reclaim would really
substantially matter for the system as a whole.

> Should we maintain a flag in the group FSNOTIFY_GROUP_SHRINKABLE?
> and set NOFS state only in that case, so at least we don't cause regression
> for existing applications?

So that's a possibility I've left in my sleeve ;). We could do it but then
we'd also have to tell lockdep that there are two kinds of mark_mutex locks
so that it does not complain about possible reclaim deadlocks. Doable but
at this point I didn't consider it worth it unless someone comes with a bug
report from a real user scenario.

								Honza
-- 
Jan Kara <jack@xxxxxxxx>
SUSE Labs, CR



[Index of Archives]     [Linux Filesystem Development]     [Linux USB Development]     [Linux Media Development]     [Video for Linux]     [Linux NILFS]     [Linux Audio Users]     [Yosemite Info]     [Linux SCSI]

  Powered by Linux