Re: WARNING in account_page_dirtied

From: Steven Whitehouse
Date: Wed Apr 04 2018 - 08:57:07 EST


Hi,


On 04/04/18 13:36, Jan Kara wrote:
Hi,

On Wed 04-04-18 10:24:48, Steven Whitehouse wrote:
On 03/04/18 13:05, Jan Kara wrote:
Hello,

On Sun 01-04-18 10:01:02, syzbot wrote:
syzbot hit the following crash on upstream commit
10b84daddbec72c6b440216a69de9a9605127f7a (Sat Mar 31 17:59:00 2018 +0000)
Merge branch 'perf-urgent-for-linus' of
git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip
syzbot dashboard link:
https://syzkaller.appspot.com/bug?extid=b7772c65a1d88bfd8fca

C reproducer: https://syzkaller.appspot.com/x/repro.c?id=5705587757154304
syzkaller reproducer:
https://syzkaller.appspot.com/x/repro.syz?id=5644332530925568
Raw console output:
https://syzkaller.appspot.com/x/log.txt?id=5472755969425408
Kernel config:
https://syzkaller.appspot.com/x/.config?id=-2760467897697295172
compiler: gcc (GCC) 7.1.1 20170620

IMPORTANT: if you fix the bug, please add the following tag to the commit:
Reported-by: syzbot+b7772c65a1d88bfd8fca@xxxxxxxxxxxxxxxxxxxxxxxxx
It will help syzbot understand when the bug is fixed. See footer for
details.
If you forward the report, please keep this part and the footer.

gfs2: fsid=loop0.0: jid=0, already locked for use
gfs2: fsid=loop0.0: jid=0: Looking at journal...
gfs2: fsid=loop0.0: jid=0: Done
gfs2: fsid=loop0.0: first mount done, others may mount
gfs2: fsid=loop0.0: found 1 quota changes
WARNING: CPU: 0 PID: 4469 at ./include/linux/backing-dev.h:341 inode_to_wb
include/linux/backing-dev.h:338 [inline]
WARNING: CPU: 0 PID: 4469 at ./include/linux/backing-dev.h:341
account_page_dirtied+0x8f9/0xcb0 mm/page-writeback.c:2416
Kernel panic - not syncing: panic_on_warn set ...

CPU: 0 PID: 4469 Comm: syzkaller368843 Not tainted 4.16.0-rc7+ #9
Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS
Google 01/01/2011
Call Trace:
__dump_stack lib/dump_stack.c:17 [inline]
dump_stack+0x194/0x24d lib/dump_stack.c:53
panic+0x1e4/0x41c kernel/panic.c:183
__warn+0x1dc/0x200 kernel/panic.c:547
report_bug+0x1f4/0x2b0 lib/bug.c:186
fixup_bug.part.10+0x37/0x80 arch/x86/kernel/traps.c:178
fixup_bug arch/x86/kernel/traps.c:247 [inline]
do_error_trap+0x2d7/0x3e0 arch/x86/kernel/traps.c:296
do_invalid_op+0x1b/0x20 arch/x86/kernel/traps.c:315
invalid_op+0x1b/0x40 arch/x86/entry/entry_64.S:986
RIP: 0010:inode_to_wb include/linux/backing-dev.h:338 [inline]
RIP: 0010:account_page_dirtied+0x8f9/0xcb0 mm/page-writeback.c:2416
RSP: 0018:ffff8801d966e5c0 EFLAGS: 00010093
RAX: ffff8801acb7e600 RBX: 1ffff1003b2cdcba RCX: ffffffff818f47a9
RDX: 0000000000000000 RSI: ffff8801d3338148 RDI: 0000000000000082
RBP: ffff8801d966e698 R08: 1ffff1003b2cdc13 R09: 000000000000000c
R10: ffff8801d966e558 R11: 0000000000000002 R12: ffff8801c96f0368
R13: ffffea0006b12780 R14: ffff8801c96f01d8 R15: ffff8801c96f01d8
__set_page_dirty+0x100/0x4b0 fs/buffer.c:605
mark_buffer_dirty+0x454/0x5d0 fs/buffer.c:1126
Huh, I don't see how this could possibly happen. The warning is:

WARN_ON_ONCE(debug_locks &&
(!lockdep_is_held(&inode->i_lock) &&
!lockdep_is_held(&inode->i_mapping->tree_lock) &&
!lockdep_is_held(&inode->i_wb->list_lock)));

Now __set_page_dirty() which called account_page_dirtied() just did:

spin_lock_irqsave(&mapping->tree_lock, flags);

Now the fact is that account_page_dirtied() actually checks
mapping->host->i_mapping->tree_lock so if mapping->host->i_mapping doesn't
get us back to 'mapping', that would explain the warning. But then
something would have to be very wrong in the GFS2 land... Adding some GFS2
related CCs just in case they have some idea.
So I looked at this for some time trying to work out what is going on. I'm
sill not 100% sure now, but lets see if we can figure it out....

The stack trace shows a call path to the end of the journal flush code where
we are unpinning pages that have been through the journal. Assuming that
jdata is not in use (it is used for some internal files, even if it is not
selected by the user) then it is most likely that this applies to a metadata
page.

For recent gfs2, all the metadata pages are kept in an address space which
for inodes is in the relevant glock, and for resource groups is a single
address space kept for only that purpose in the super block. In both of
those cases the mapping->host points to the block device inode. Since the
inode's mapping->host reflects only the block device address space (unused
by gfs2) we would not expect it to point back to the relevant address space.

As far as I can tell this usage is ok, since it doesn't make much sense to
require lots of inodes to be hanging around uselessly just to keep metadata
pages in. That after all, is why the address space and inode are separate
structures in the first place since it is not a one to one relationship. So
I think that probably explains why this triggers, since the test is not
really a valid one in all cases,
The problem is we really do expect mapping->host->i_mapping == mapping as
we pass mapping and inode interchangebly in the mm code. The address_space
and inodes are separate structures because you can have many inodes
pointing to one address space (block devices). However it is not allowed
for several address_spaces to point to one inode! That way mm code may end
up using different address_spaces in different places although they should
be the same one as is the case in this assert... Probably you use these
address_spaces in a very limited way and so things seem to work but it is
really a pure coincidence. From a very quick look you seem to be using
these special address_spaces to track dirty metadata associated with an
inode? Anything else?

Honza

Yes, either an inode or a rgrp. However I'm fairly sure that we landed up doing that because we were told that inodes and address spaces were intended to be independent at some point in the past. They are used in a fairly limited way and mostly so that we can efficiently invalidate metadata belonging to a particular inode (or rgrp).

In the rgrp case we could just use the existing block dev inode's address space except that we'd have to make sure that we invalidated it on mount. The rgrps are easy because each one is a single extent only. For the inode metadata case, we did (a very long time ago) try tracking the metadata in a different way and it was not very efficient at all, so using a separate address space was the best solution we could find at the time.

We do not want to go back to having two struct inodes for each real inode since that took up a lot of memory in cases where there were lots of small files...

https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/fs/gfs2/glock.c?id=009d851837ab26cab18adda6169a813f70b0b21b


and now I remember that is also resolved an issue of a circular dependency between inodes used for the metadata address space and "proper" inodes too. When we introduced the change in the above patch, both inodes and glock were using the address spaces in the glock, however we further optimised the rgrps at a later date to share a single address space between them.

So while that doesn't solve the problem, it does, I hope, explain some of the background,

Steve.