Re: BUG: using smp_processor_id() in preemptible [00000000] code: icedove-bin/5449

[Divyesh Shah]

On Tue, Jun 1, 2010 at 12:53 AM, Jens Axboe <axboe@xxxxxxxxx> wrote:
> On Tue, Jun 01 2010, Ingo Molnar wrote:
>>
>> * Jens Axboe <axboe@xxxxxxxxx> wrote:
>>
>> > On Tue, Jun 01 2010, Ingo Molnar wrote:
>> > >
>> > > * Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>> > >
>> > > > On Mon, 2010-05-24 at 05:03 +0200, Piotr Hosowicz wrote:
>> > > > > [  720.313607] BUG: using smp_processor_id() in preemptible [00000000] code: icedove-bin/5449
>> > > > > [  720.313612] caller is native_sched_clock+0x3c/0x68
>> > > > > [  720.313616] Pid: 5449, comm: icedove-bin Tainted: P            2.6.34-20100524-0407 #1
>> > > > > [  720.313618] Call Trace:
>> > > > > [  720.313624]  [<ffffffff811a533b>] debug_smp_processor_id+0xc7/0xe0
>> > > > > [  720.313629]  [<ffffffff81009b87>] native_sched_clock+0x3c/0x68
>> > > > > [  720.313634]  [<ffffffff81009a4d>] sched_clock+0x9/0xd
>> > > > > [  720.313637]  [<ffffffff811823ec>] blk_rq_init+0x92/0x9d
>> > > > > [  720.313641]  [<ffffffff81184227>] get_request+0x1bf/0x2c7
>> > > > > [  720.313646]  [<ffffffff8118435c>] get_request_wait+0x2d/0x19d
>> > > >
>> > > > This comes from wreckage in the blk tree..
>> > > >
>> > > > ---
>> > > > commit 9195291e5f05e01d67f9a09c756b8aca8f009089
>> > > > Author: Divyesh Shah <dpshah@xxxxxxxxxx>
>> > > > Date:   Thu Apr 1 15:01:41 2010 -0700
>> > > >
>> > > >     blkio: Increment the blkio cgroup stats for real now
>> > >
>> > > Jens, this regression is still in .35-rc1 and triggers in about 25% of all
>> > > -tip boot tests.
>> > >
>> > > The above commit is using sched_clock() in an unsafe way - please fix it or
>> > > revert it.
>> > >
>> > > The local_clock() changes PeterZ is working on are still WIP, it's not sure
>> > > we'll have it before .36.
>> >
>> > OK, I guess we'll have to solve this differently for .35 - I'll cook up
>> > something simple, if need be revert the change.
>>
>> I suspect you can put get_cpu/put_cpu around it and use cpu_clock(). The
>> cross-CPU effects will still be there and there might be weird stats.
>
> It'll shut it up at least, which is the primary concern at this point.

Jens,
     Thanks for the temporary fix (adding preempt_enable/disable()
calls). I was able to repro the issue and and have a patch that
replaces use of sched_clock() in block layer w/
ktime_to_ns(ktime_get()). We will lose resolution when running w/o
highres timers but should be better than unbounded drift between cpus.
I'll send the patch next.

-Divyesh

>
> --
> Jens Axboe
>
>
--
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/