Re: [PATCH] [fuse] alloc_page nofs avoid deadlock

From: Miklos Szeredi
Date: Wed Aug 18 2021 - 05:25:04 EST


On Tue, 13 Jul 2021 at 04:42, Ed Tsai <ed.tsai@xxxxxxxxxxxx> wrote:
>
> On Tue, 2021-06-08 at 17:30 +0200, Miklos Szeredi wrote:
> > On Thu, 3 Jun 2021 at 14:52, chenguanyou <chenguanyou9338@xxxxxxxxx>
> > wrote:
> > >
> > > ABA deadlock
> > >
> > > PID: 17172 TASK: ffffffc0c162c000 CPU: 6 COMMAND: "Thread-21"
> > > 0 [ffffff802d16b400] __switch_to at ffffff8008086a4c
> > > 1 [ffffff802d16b470] __schedule at ffffff80091ffe58
> > > 2 [ffffff802d16b4d0] schedule at ffffff8009200348
> > > 3 [ffffff802d16b4f0] bit_wait at ffffff8009201098
> > > 4 [ffffff802d16b510] __wait_on_bit at ffffff8009200a34
> > > 5 [ffffff802d16b5b0] inode_wait_for_writeback at ffffff800830e1e8
> > > 6 [ffffff802d16b5e0] evict at ffffff80082fb15c
> > > 7 [ffffff802d16b620] iput at ffffff80082f9270
> > > 8 [ffffff802d16b680] dentry_unlink_inode at ffffff80082f4c90
> > > 9 [ffffff802d16b6a0] __dentry_kill at ffffff80082f1710
> > > 10 [ffffff802d16b6d0] shrink_dentry_list at ffffff80082f1c34
> > > 11 [ffffff802d16b750] prune_dcache_sb at ffffff80082f18a8
> > > 12 [ffffff802d16b770] super_cache_scan at ffffff80082d55ac
> > > 13 [ffffff802d16b860] shrink_slab at ffffff8008266170
> > > 14 [ffffff802d16b900] shrink_node at ffffff800826b420
> > > 15 [ffffff802d16b980] do_try_to_free_pages at ffffff8008268460
> > > 16 [ffffff802d16ba60] try_to_free_pages at ffffff80082680d0
> > > 17 [ffffff802d16bbe0] __alloc_pages_nodemask at ffffff8008256514
> > > 18 [ffffff802d16bc60] fuse_copy_fill at ffffff8008438268
> > > 19 [ffffff802d16bd00] fuse_dev_do_read at ffffff8008437654
> > > 20 [ffffff802d16bdc0] fuse_dev_splice_read at ffffff8008436f40
> > > 21 [ffffff802d16be60] sys_splice at ffffff8008315d18
> > > 22 [ffffff802d16bff0] __sys_trace at ffffff8008084014
> > >
> > > PID: 9652 TASK: ffffffc0c9ce0000 CPU: 4 COMMAND: "kworker/u16:8"
> > > 0 [ffffff802e793650] __switch_to at ffffff8008086a4c
> > > 1 [ffffff802e7936c0] __schedule at ffffff80091ffe58
> > > 2 [ffffff802e793720] schedule at ffffff8009200348
> > > 3 [ffffff802e793770] __fuse_request_send at ffffff8008435760
> > > 4 [ffffff802e7937b0] fuse_simple_request at ffffff8008435b14
> > > 5 [ffffff802e793930] fuse_flush_times at ffffff800843a7a0
> > > 6 [ffffff802e793950] fuse_write_inode at ffffff800843e4dc
> > > 7 [ffffff802e793980] __writeback_single_inode at ffffff8008312740
> > > 8 [ffffff802e793aa0] writeback_sb_inodes at ffffff80083117e4
> > > 9 [ffffff802e793b00] __writeback_inodes_wb at ffffff8008311d98
> > > 10 [ffffff802e793c00] wb_writeback at ffffff8008310cfc
> > > 11 [ffffff802e793d00] wb_workfn at ffffff800830e4a8
> > > 12 [ffffff802e793d90] process_one_work at ffffff80080e4fac
> > > 13 [ffffff802e793e00] worker_thread at ffffff80080e5670
> > > 14 [ffffff802e793e60] kthread at ffffff80080eb650
> >
> > The issue is real.
> >
> > The fix, however, is not the right one. The fundamental problem is
> > that fuse_write_inode() blocks on a request to userspace.
> >
> > This is the same issue that fuse_writepage/fuse_writepages face. In
> > that case the solution was to copy the page contents to a temporary
> > buffer and return immediately as if the writeback already completed.
> >
> > Something similar needs to be done here: send the FUSE_SETATTR
> > request
> > asynchronously and return immediately from fuse_write_inode(). The
> > tricky part is to make sure that multiple time updates for the same
> > inode aren't mixed up...
> >
> > Thanks,
> > Miklos
>
> Dear Szeredi,
>
> Writeback thread calls fuse_write_inode() and wait for user Daemon to
> complete this write inode request. The user daemon will alloc_page()
> after taking this request, and a deadlock could happen when we try to
> shrink dentry list under memory pressure.
>
> We (Mediatek) glad to work on this issue for mainline and also LTS. So
> another problem is that we should not change the protocol or feature
> for stable kernel.
>
> Use GFP_NOFS | __GFP_HIGHMEM can really avoid this by skip the dentry
> shirnker. It works but degrade the alloc_page success rate. In a more
> fundamental way, we could cache the contents and return immediately.
> But how to ensure the request will be done successfully, e.g., always
> retry if it fails from daemon.

Key is where the the dirty metadata is flushed. To prevent deadlock
it must not be flushed from memory reclaim, so must make sure that it
is flushed on close(2) and munmap(2) and not dirtied after that.

I'm working on this currently and hope to get it ready for the next
merge window.

Thanks,
Miklos