NFS BUG_ON in nfs_do_writepage

From: rercola
Date: Mon Apr 13 2009 - 01:53:10 EST

Hi world,
I've got a production server that's running as an NFSv4 client, along with a number of other machines.

All the other machines are perfectly happy, but this one is a bit of a bother. It's got a Core 2 Duo 6700, with a D975XBX2 motherboard and 4 GB of ECC RAM.

The problem is that, under heavy load, NFS will trip a BUG_ON in nfs_do_writepage, as follows:
------------[ cut here ]------------
kernel BUG at fs/nfs/write.c:252!
invalid opcode: 0000 [#1] SMP
last sysfs file: /sys/devices/virtual/block/dm-
0/range
CPU 0
Modules linked in: fuse autofs4 coretemp hwmon nfs lockd nfs_acl auth_rpcgss sunrpc ipv6 cpufreq_ondemand acpi_cpufreq freq_table kvm_intel kvm snd_hda_codec_idt snd_hda_intel snd_hda_codec snd_hwdep snd_seq_dummy snd_seq_oss snd_seq_midi_event snd_seq snd_seq_device snd_pcm_oss snd_mixer_oss snd_pcm snd_timer usb_storage snd cpia_usb e1000e soundcore cpia ppdev firewire_ohci snd_page_alloc firewire_core i2c_i801 videodev parport_pc pcspkr iTCO_wdt i2c_core v4l1_compat crc_itu_t parport iTCO_vendor_support v4l2_compat_ioctl32 i82975x_edac edac_core raid1
Pid: 309, comm: pdflush Not tainted 2.6.29.1 #1
RIP: 0010:[<ffffffffa0291a47>] [<ffffffffa0291a47>] nfs_do_writepage+0x106/0x1a2 [nfs]
RSP: 0018:ffff88012d805af0 EFLAGS: 00010282
RAX: 0000000000000001 RBX: ffffe20001f66878 RCX: 0000000000000015
RDX: 0000000000600020 RSI: 0000000000000000 RDI: ffff88000155789c
RBP: ffff88012d805b20 R08: ffff88012cd53460 R09: 0000000000000004
R10: ffff88009d421700 R11: ffffffffa02a98d0 R12: ffff88010253a300
R13: ffff88000155789c R14: ffffe20001f66878 R15: ffff88012d805c80
FS: 0000000000000000(0000) GS:ffffffff817df000(0000) knlGS:0000000000000000
CS: 0010 DS: 0018 ES: 0018 CR0: 000000008005003b
CR2: 00000000f7d2b000 CR3: 000000008708a000 CR4: 00000000000026e0
DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
DR3: 0000000000000000 DR6: 00000000ffff0ff0 DR7: 0000000000000400
Process pdflush (pid: 309, threadinfo ffff88012d804000, task ffff88012e4fdb80)
Stack:
ffff88012d805b20 ffffe20001f66878 ffffe20001f66878 0000000000000000
0000000000000001 0000000000000000 ffff88012d805b40 ffffffffa0291f5a
ffffe20001f66878 ffff88012d805e40 ffff88012d805c70 ffffffff810a9c1d
Call Trace:
[<ffffffffa0291f5a>] nfs_writepages_callback+0x14/0x25 [nfs]
[<ffffffff810a9c1d>] write_cache_pages+0x261/0x3a4
[<ffffffffa0291f46>] ? nfs_writepages_callback+0x0/0x25 [nfs]
[<ffffffffa0291f1c>] nfs_writepages+0xb5/0xdf [nfs]
[<ffffffffa02932bd>] ? nfs_flush_one+0x0/0xeb [nfs]
[<ffffffff81060f78>] ? bit_waitqueue+0x17/0xa4
[<ffffffff810a9db7>] do_writepages+0x2d/0x3d
[<ffffffff810f4a51>] __writeback_single_inode+0x1b2/0x347
[<ffffffff8100f7d4>] ? __switch_to+0xbe/0x3eb
[<ffffffff810f4ffb>] generic_sync_sb_inodes+0x24a/0x395
[<ffffffff810f5354>] writeback_inodes+0xa9/0x102
[<ffffffff810a9f26>] wb_kupdate+0xa8/0x11e
[<ffffffff810aac9d>] pdflush+0x173/0x236
[<ffffffff810a9e7e>] ? wb_kupdate+0x0/0x11e
[<ffffffff810aab2a>] ? pdflush+0x0/0x236
[<ffffffff810aab2a>] ? pdflush+0x0/0x236
[<ffffffff81060c9e>] kthread+0x4e/0x7b
[<ffffffff810126ca>] child_rip+0xa/0x20
[<ffffffff81011fe7>] ? restore_args+0x0/0x30
[<ffffffff81060c50>] ? kthread+0x0/0x7b
[<ffffffff810126c0>] ? child_rip+0x0/0x20
Code: 89 e7 e8 d5 cc ff ff 4c 89 e7 89 c3 e8 2a cd ff ff 85 db 74 a0 e9 83 00 00 00 41 f6 44 24 40 02 74 0d 4c 89 ef e8 e2 a5 d9 e0 90 <0f> 0b eb fe 4c 89 f7 e8 f5 7a e1 e0 85 c0 75 49 49 8b 46 18 ba
RIP [<ffffffffa0291a47>] nfs_do_writepage+0x106/0x1a2 [nfs]
RSP <ffff88012d805af0>
---[ end trace 6d60c9b253ebcf15 ]---

64bit kernel, 32bit userland. 2.6.29.1 vanilla, bug occurred as early as 2.6.28, bug still occurs with 2.6.30-rc1. I'm running bisect now, but there's a limit on how often I can reboot a production server, so I'll report back when I find it.

The unfortunate part, of course, is that when this bug occurs, the writepage never returns...meaning that the process in question is permanently locked in la-la-land (AKA state D). This renders this unfortunate bug a bit...inconvenient.

[No other clients, or the server, report anything interesting when this happens, AFAICS.]

- Rich
--
Ah say, son, you're about as sharp as a bowlin' ball.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/