Re: kernel BUG at fs/ext4/inode.c:2428!

[Jan Kara <jack@suse.cz>]

Hello,

On Mon 14-11-16 12:15:16, Nikolay Borisov wrote:
> So I hit the following BUG_ON on 3 separate servers: 
> 
> [1387898.597939] sh (14886): drop_caches: 3
> [1387945.259613] ------------[ cut here ]------------
> [1387945.259791] kernel BUG at fs/ext4/inode.c:2428!
> [1387945.259964] invalid opcode: 0000 [#1] SMP 
> [1387945.263921] CPU: 9 PID: 8987 Comm: kworker/u24:23 Tainted: P           O    4.4.26-clouder1 #3
> [1387945.264213] Hardware name: Supermicro X9DRD-iF/LF/X9DRD-iF, BIOS 3.2 01/16/2015
> [1387945.264512] Workqueue: writeback wb_workfn (flush-253:8)
> [1387945.264780] task: ffff880287ca5280 ti: ffff8800064dc000 task.ti: ffff8800064dc000
> [1387945.265073] RIP: 0010:[<ffffffff8122a3ac>]  [<ffffffff8122a3ac>] mpage_prepare_extent_to_map+0x29c/0x2d0
> [1387945.265426] RSP: 0018:ffff8800064df960  EFLAGS: 00010246
> [1387945.265596] RAX: 02fffc0000030039 RBX: ffff8800064dfad0 RCX: 0000000000000537
> [1387945.265881] RDX: 000000000000231b RSI: 0000000000000000 RDI: ffffffff81a052a6
> [1387945.266165] RBP: ffff8800064dfa28 R08: 0000000000000000 R09: 0000000000000000
> [1387945.266450] R10: 0000000000000001 R11: 0000000000000100 R12: ffff8800064df980
> [1387945.266734] R13: 0000000000003400 R14: ffffffffffffffff R15: ffffea000686fbc0
> [1387945.267024] FS:  0000000000000000(0000) GS:ffff88047fd20000(0000) knlGS:0000000000000000
> [1387945.267315] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
> [1387945.267487] CR2: ffffffffff600400 CR3: 00000004555ff000 CR4: 00000000000406e0
> [1387945.267765] Stack:
> [1387945.267925]  0000000000000000 ffff88015b3c2be0 ffff8800064df980 0000000000000538
> [1387945.268386]  0000000000000001 0000000000000000 ffffea000686fbc0 ffff88046bce1800
> [1387945.268850]  ffff8800064df9e8 ffffffff81274777 ffffffff02400040 0fd000086bce6800
> [1387945.269319] Call Trace:
> [1387945.269489]  [<ffffffff81274777>] ? jbd2__journal_start+0xe7/0x200
> [1387945.269663]  [<ffffffff8122e581>] ? ext4_writepages+0x3a1/0xcd0
> [1387945.269839]  [<ffffffff8125ba8d>] ? __ext4_journal_start_sb+0x6d/0x100
> [1387945.270013]  [<ffffffff8122e5b2>] ext4_writepages+0x3d2/0xcd0
> [1387945.270207]  [<ffffffffa07f878b>] ? leaf_space_used+0xcb/0x100 [btrfs]
> [1387945.270382]  [<ffffffff810823a1>] ? sched_clock_cpu+0xa1/0xb0
> [1387945.270556]  [<ffffffff8107beb4>] ? check_preempt_curr+0x54/0x90
> [1387945.270730]  [<ffffffff8107c07d>] ? ttwu_do_activate.constprop.92+0x5d/0x70
> [1387945.270905]  [<ffffffff8113735e>] do_writepages+0x1e/0x30
> [1387945.271076]  [<ffffffff811c6c85>] __writeback_single_inode+0x45/0x370
> [1387945.271250]  [<ffffffff811c74d2>] writeback_sb_inodes+0x252/0x570
> [1387945.271423]  [<ffffffff811c7879>] __writeback_inodes_wb+0x89/0xc0
> [1387945.271596]  [<ffffffff811c7bb8>] wb_writeback+0x268/0x300
> [1387945.271766]  [<ffffffff811c83e6>] wb_workfn+0x2d6/0x400
> [1387945.271938]  [<ffffffff81614ea8>] ? _raw_spin_lock_irqsave+0x18/0x50
> [1387945.272112]  [<ffffffff8106bf89>] process_one_work+0x159/0x450
> [1387945.272285]  [<ffffffff8106c639>] worker_thread+0x69/0x490
> [1387945.272456]  [<ffffffff8106c5d0>] ? rescuer_thread+0x350/0x350
> [1387945.272630]  [<ffffffff810717bf>] kthread+0xef/0x110
> [1387945.272803]  [<ffffffff810716d0>] ? kthread_park+0x60/0x60
> [1387945.272975]  [<ffffffff816156bf>] ret_from_fork+0x3f/0x70
> [1387945.273146]  [<ffffffff810716d0>] ? kthread_park+0x60/0x60
> [1387945.273322] Code: d2 e4 ff e8 67 6f 3e 00 48 8b 85 50 ff ff ff 49 39 c6 0f 83 15 fe ff ff 31 c0 eb a7 4c 89 ff e8 3b e8 ef ff e9 b8 fe ff ff 0f 0b <0f> 0b 48 8d bd 58 ff ff ff 89 85 48 ff ff ff e8 50 f8 f0 ff 8b 
> [1387945.276751] RIP  [<ffffffff8122a3ac>] mpage_prepare_extent_to_map+0x29c/0x2d0
> [1387945.277089]  RSP <ffff8800064df960>
> 
> So a user triggers drop_caches and ext4 crashes due to it trying to 
> write a page that isn't fs-owned. ffffffff8122a3ac is : head = page_buffers(page);
> which has this: BUG_ON(!PagePrivate(page)); 
> 
> page.flags is flags = 216168384067469369 (in binary: 0000001011111111111111000000000000000000000000110000000000111001) 
> so the 11th bit (PG_private) is not set, triggering the BUG_ON. The flags are 
> (PG_LOCKED|PG_UPTODATE|PG_DIRTY|MAPPEDTODISK|PG_RECLAIM). Do these flags seem 
> corrupt - uptodate and dirty being set at the same time? Maybe the page struct
> is being corrupted?
> 
> page.private is actually NULL. The page does have an associated address_space 
> mapping. I've validated this since address_space.host is the same as the 
> inode member of the passed mpd. 

Interesting. I didn't see this yet. What mount options does the filesystem
use? The file where this happened is a regular file I assume, right? What
is blocksize and page size for the filesystem?

The page flags actually look rather consistent. The only thing that is
strange is how a dirty page can be without buffers because whenever we
write to a page we create buffers for the page and then mark (some of
those) buffers dirty. ext4_releasepage() ends up calling
try_to_free_buffers() which does not remove dirty buffers. But apparently
the buffers attached to a page got somehow clean but the page remained
dirty, then ext4_releasepage() removed those buffers and then we crashed
when trying to write that page. If blocksize < pagesize, I could imagine
some subtle bugs in our code e.g. for pages straddling EOF...

								Honza
-- 
Jan Kara <jack@suse.com>
SUSE Labs, CR