Hello,
A filesystem stress test script racer.sh caused a problem in
2.6.26-rc8-mm1. I am not sure it is caused by some changes to memory
management or handling dentry in proc_flush_task_mnt().
Is this problem already known and fixed?
- 2.6.26-rc8-mm1, no extra modules, no outer patch.
- racer.sh was introduced by Andreas Dilger in
http://marc.info/?l=linux-kernel&m=120518906530609&w=2
- While I changed the script a little, I don't think it affected the
result.
- If .config is necessary, I will post it later.
Junjiro Okajima
BUG: unable to handle kernel paging request at 6b6b6b6b
IP: [<c01949ad>] __d_lookup+0x64/0x1f2
*pde = 00000000
Oops: 0000 [#1] PREEMPT SMP
last sysfs file: /sys/devices/virtual/net/lo/operstate
Modules linked in: nfs rtc ide_cd_mod cdrom evdev
Pid: 29643, comm: mv Not tainted (2.6.26-rc8-mm1jrousD #36)
EIP: 0060:[<c01949ad>] EFLAGS: 00010292 CPU: 0
EIP is at __d_lookup+0x64/0x1f2
EAX: 00000027 EBX: 6b6b6b6b ECX: c04672f8 EDX: 00000000
ESI: 00000006 EDI: c3401920 EBP: c1407dbc ESP: c1407d68
DS: 007b ES: 007b FS: 00d8 GS: 0033 SS: 0068
Process mv (pid: 29643, ti=c1406000 task=c2de30b0 task.ti=c1406000)
Stack: c0420081 c2de32a4 000073cb 00000001 c340199c 00000006 c2f96006 6b6b6b6b
c3401920 c360f6f8 00000000 c1407da8 c1407e2c c3810f68 c1407e40 00000006
cbb0569d c2f96006 c2f9600c c1407e2c c1407f04 c1407de4 c018a5c1 c1407e40
Call Trace:
[<c018a5c1>] ? do_lookup+0x1e/0x1a0
[<c021842a>] ? security_inode_permission+0x3a/0x43
[<c018c366>] ? __link_path_walk+0x766/0xe29
[<c02369e4>] ? debug_smp_processor_id+0x28/0xcc
[<c013e3bf>] ? trace_hardirqs_off+0xb/0xd
[<c036bf49>] ? add_preempt_count+0x4a/0x8c
[<c036beb2>] ? sub_preempt_count+0x62/0xaf
[<c018cd8a>] ? do_path_lookup+0x12e/0x210
[<c018ca62>] ? path_walk+0x39/0x75
[<c018cccd>] ? do_path_lookup+0x71/0x210
[<c018da5c>] ? __path_lookup_intent_open+0x44/0x99
[<c018db35>] ? path_lookup_open+0x21/0x27
[<c018ddc5>] ? do_filp_open+0x90/0x7c5
[<c02369e4>] ? debug_smp_processor_id+0x28/0xcc
[<c036bf49>] ? add_preempt_count+0x4a/0x8c
[<c013e426>] ? put_lock_stats+0x27/0x38
[<c018106f>] ? get_unused_fd_flags+0xc7/0xe0
[<c036beb2>] ? sub_preempt_count+0x62/0xaf
[<c0369798>] ? _spin_unlock+0x41/0x4d
[<c01810e6>] ? do_sys_open+0x5e/0xf3
[<c0232a84>] ? trace_hardirqs_on_thunk+0xc/0x10
[<c01811c7>] ? sys_open+0x23/0x2b
[<c0103009>] ? sysenter_past_esp+0x6a/0xa5
=======================
Code: 0d 38 99 4a c0 d3 e8 31 c3 23 1d 34 99 4a c0 01 db 01 db 03 1d 3c 99 4a c0 e8 1e 3a fc ff 8b 03 89 c3 85 c0 75 37 e9 20 01 00 00 <8b> 03 0f 18 00 90 8d 43 d4 89 45 d0 8d 83 d4 94 94 94 3d ff 00
EIP: [<c01949ad>] __d_lookup+0x64/0x1f2 SS:ESP 0068:c1407d68
On Wednesday 30 July 2008 16:40, [email protected] wrote:
> Hello,
>
> A filesystem stress test script racer.sh caused a problem in
> 2.6.26-rc8-mm1. I am not sure it is caused by some changes to memory
> management or handling dentry in proc_flush_task_mnt().
> Is this problem already known and fixed?
This looks a bit like the PREEMPT_RCU problem that was fixed just
before 2.6.26. Do you have PREEMPT_RCU=y and HOTPLUG_CPU=n in your
config?
Nick Piggin:
> This looks a bit like the PREEMPT_RCU problem that was fixed just
> before 2.6.26. Do you have PREEMPT_RCU=y and HOTPLUG_CPU=n in your
> config?
Exactly.
So next -mm release will not have this problem. Anyway I will check it
again.
Thanks.
$ gunzip -c /proc/config.gz | egrep '(_RCU|HOTPLUG)'
CONFIG_HOTPLUG=y
# CONFIG_CLASSIC_RCU is not set
CONFIG_PREEMPT_RCU=y
# CONFIG_RCU_TRACE is not set
# CONFIG_HOTPLUG_CPU is not set
# CONFIG_HOTPLUG_PCI is not set
Junjiro Okajima