Re: fuse: kernel panic while using splice (lru corruption)

[Miklos Szeredi <miklos@szeredi.hu>]

On Wed, Oct 20, 2021 at 05:26:59PM -0400, Frank Dinoff wrote:
> On Thu, Oct 7, 2021 at 8:54 AM Miklos Szeredi <miklos@szeredi.hu> wrote:
> >
> > Adding linux-mm/Johannes to Cc.
> >
> > On Wed, 6 Oct 2021 at 21:13, Frank Dinoff <fdinoff@google.com> wrote:
> > >
> > > I'm experiencing a kernel panic while using fuse related to SPLICE_F_MOVE.
> > >
> > > Some stack traces
> > >
> > > [   52.864466] CPU: 1 PID: 10619 Comm: cp Not tainted 5.15.0-upstream-DEV #7
> > > [   52.879137] Hardware name: Google Google Compute Engine/Google
> > > Compute Engine, BIOS Google 01/01/2011
> > > [   52.888490] RIP: 0010:__list_del_entry_valid+0x69/0x80
> > > [   52.893907] Code: 7f 12 84 31 c0 e8 2d 42 55 00 0f 0b 48 c7 c7 37
> > > 8e 10 84 31 c0 e8 1d 42 55 00 0f 0b 48 c7 c7 19 52 19 84 31 c0 e8 0d
> > > 42 55 00 <0f> 0b 48 c7 c7 8e e7 15 84 31 c0 e8 fd 41 55 00 0f 0b 00 00
> > > 00 cc
> > > [   52.980251] RSP: 0018:ffff8938ea093978 EFLAGS: 00010046
> > > [   52.994508] RAX: 0000000000000054 RBX: ffffd8d7c5914ec0 RCX: 45fe15c1d0642d00
> > > [   53.001774] RDX: ffff893939d230b8 RSI: ffff893939d17510 RDI: ffff893939d17510
> > > [   53.009039] RBP: ffff8938ea093978 R08: 0000000000000000 R09: ffffffff8492dbf0
> > > [   53.016312] R10: 00000000ffff7fff R11: 0000000000000000 R12: ffff8938ea093a98
> > > [   53.023575] R13: ffff8938ced23400 R14: 000000000000000d R15: ffff8938ced23400
> > > [   53.030843] FS:  00007f8bd4b1d740(0000) GS:ffff893939d00000(0000)
> > > knlGS:0000000000000000
> > > [   53.039064] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> > > [   53.044939] CR2: 000000000020c86e CR3: 0000000164572003 CR4: 00000000003706e0
> > > [   53.087945] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
> > > [   53.105377] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
> > > [   53.119206] Call Trace:
> > > [   53.121786]  release_pages+0x1d0/0x490
> > > [   53.125856]  __pagevec_release+0x4f/0x60
> > > [   53.129914]  invalidate_inode_pages2_range+0x5c4/0x600
> > > [   53.135186]  ? kmem_cache_free+0x7c/0x100
> > > [   53.139330]  invalidate_inode_pages2+0x17/0x20
> > > [   53.143907]  fuse_finish_open+0x75/0x150
> > > [   53.147976]  fuse_open_common+0x113/0x120
> > > [   53.152117]  fuse_open+0x10/0x20
> > > [   53.155487]  do_dentry_open+0x263/0x360
> > > [   53.167370]  vfs_open+0x2d/0x30
> > > [   53.173633]  path_openat+0xa0f/0xd90
> > > [   53.177353]  ? mntput+0x23/0x40
> > > [   53.180635]  ? path_put+0x1e/0x30
> > > [   53.184104]  do_filp_open+0xc7/0x170
> > > [   53.187933]  do_sys_openat2+0x91/0x170
> > > [   53.195012]  __x64_sys_openat+0x7e/0xa0
> > > [   53.198989]  do_syscall_64+0x44/0xa0
> > > [   53.202714]  ? exc_page_fault+0x71/0x160
> > > [   53.211294]  entry_SYSCALL_64_after_hwframe+0x44/0xae
> > > [   53.219256] RIP: 0033:0x7f8bd4c4bec2
> > > [   53.231375] Code: 8d 48 08 48 89 4d d8 8b 18 48 8b 05 90 8d 07 00
> > > 83 38 00 75 30 b8 01 01 00 00 41 89 da bf 9c ff ff ff 4c 89 f6 44 89
> > > fa 0f 05 <48> 89 c3 48 3d 00 f0 ff ff 77 42 89 d8 48 81 c4 c8 00 00 00
> > > 5b 41
> > > [   53.261026] RSP: 002b:00007ffd4ae55570 EFLAGS: 00000246 ORIG_RAX:
> > > 0000000000000101
> > > [   53.268738] RAX: ffffffffffffffda RBX: 0000000000000000 RCX: 00007f8bd4c4bec2
> > > [   53.284714] RDX: 0000000000000000 RSI: 00007ffd4ae57bbc RDI: 00000000ffffff9c
> > > [   53.305198] RBP: 00007ffd4ae55650 R08: 0000000000000000 R09: 00007ffd4ae55baf
> > > [   53.312469] R10: 0000000000000000 R11: 0000000000000246 R12: 0000000000000000
> > > [   53.319736] R13: 00000000000081a4 R14: 00007ffd4ae57bbc R15: 0000000000000000
> > > [   53.327003] Modules linked in: 9p 9pnet_virtio 9pnet vfat fat
> > > virtio_net net_failover failover
> > > [   53.335755] ---[ end trace 086000a6a6747ea3 ]---
> > >
> > > With CONFIG_DEBUG_VM it looks like we are trying to add a page to the lru that
> > > is already on the lru.
> >
> > Code in question is in fs/fuse/dev.c:fuse_try_move_page() after the
> > call to replace_page_cache_page().
> >
> > Looks like PIPE_BUF_FLAG_LRU isn't reliable., which means it's
> > completely useless.
> >
> > Johannes, any idea how to fix this?
> 
> Ping, any thoughts on how to fix this?

Turns out to be a bad interaction between anon pipe bufs and the way fuse clones
and steals pipe buffers.

Here's a patch that should fix it (survives your reproducer).

Thanks,
Miklos
---

From: Miklos Szeredi <mszeredi@redhat.com>
Subject: fuse: fix page stealing

It is possible to trigger a crash by splicing anon pipe bufs to the fuse
device.

The reason for this is that anon_pipe_buf_release() will reuse buf->page if
the refcount is 1, but that page might have already been stolen and its
flags modified (e.g. PG_lru added).

This happens in the unlikely case of fuse_dev_splice_write() getting around
to calling pipe_buf_release() after a page has been stolen, added to the
page cache and removed from the page cache.

Fix by calling pipe_buf_release() right after the page was inserted into
the page cache.  In this case the page has an elevated refcount so any
release function will know that the page isn't reusable.

Reported-by: Frank Dinoff <fdinoff@google.com>
Link: https://lore.kernel.org/r/CAAmZXrsGg2xsP1CK+cbuEMumtrqdvD-NKnWzhNcvn71RV3c1yw@mail.gmail.com/
Fixes: dd3bb14f44a6 ("fuse: support splice() writing to fuse device")
Cc: <stable@vger.kernel.org> # v2.6.35
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
---
 fs/fuse/dev.c |   14 ++++++++++++--
 1 file changed, 12 insertions(+), 2 deletions(-)

--- a/fs/fuse/dev.c
+++ b/fs/fuse/dev.c
@@ -847,6 +847,12 @@ static int fuse_try_move_page(struct fus
 
 	replace_page_cache_page(oldpage, newpage);
 
+	/*
+	 * Release while we have extra ref on stolen page.  Otherwise
+	 * anon_pipe_buf_release() might think the page can be reused.
+	 */
+	pipe_buf_release(cs->pipe, buf);
+
 	get_page(newpage);
 
 	if (!(buf->flags & PIPE_BUF_FLAG_LRU))
@@ -2031,8 +2037,12 @@ static ssize_t fuse_dev_splice_write(str
 
 	pipe_lock(pipe);
 out_free:
-	for (idx = 0; idx < nbuf; idx++)
-		pipe_buf_release(pipe, &bufs[idx]);
+	for (idx = 0; idx < nbuf; idx++) {
+		struct pipe_buffer *buf = &bufs[idx];
+
+		if (buf->ops)
+			pipe_buf_release(pipe, buf);
+	}
 	pipe_unlock(pipe);
 
 	kvfree(bufs);