Re: ata/scsi crash on halt

[Tejun Heo <htejun@gmail.com>]

Hello, James, Andrew.

James Bottomley wrote:
> On Sat, 2007-05-12 at 11:25 -0700, Andrew Morton wrote:
>> [  715.196000] sd 2:0:0:0: [sda] Synchronizing SCSI cache
>> [  715.196000] sd 2:0:0:0: [sda] Stopping disk
>> [  715.196000] ata3.00: DISK MIGHT NOT BE SPUN DOWN PROPERLY. UPDATE SHUTDOWN UTILITY
>> [  715.196000] ata3.00: For more info, visit http://linux-ata.org/shutdown.html
>> [  715.196000] BUG: scheduling while atomic: swapper/0x00000100/0
>> [  715.196000]  [<c02e282d>] __sched_text_start+0x55/0x504
>> [  715.196000]  [<c0239faa>] scsi_done+0x0/0x1d
>> [  715.196000]  [<c02e32fb>] schedule_timeout+0x78/0x95
>> [  715.196000]  [<c011ffc2>] process_timeout+0x0/0x5
>> [  715.196000]  [<c011ff87>] msleep+0x10/0x18
>> [  715.196000]  [<c024d455>] ata_scsi_start_stop_xlat+0xe3/0x11b
> 
> This is a bug in libata ... there's a ssleep in the command translation
> path.  The commit that did it was this one:
> 
> commit 920a4b1038e442700a1cfac77ea7e20bd615a2c3
> Author: Tejun Heo <htejun@gmail.com>
> Date:   Fri May 4 21:28:48 2007 +0200
> 
>     libata: implement libata.spindown_compat
> 
> I'm afraid the queuecommand path is not necessarily connected to a user
> process because of the way block scheduling works ... you can't call
> sleeping functions along it. 

Right, I missed that.

> Ordinarily, what you'd do if you want to delay calling ->done() is to
> schedule a timer to handle it ... unfortunately I don't see an easy way
> of doing that because of the way this particular layering works.

Yeah, thought about doing it with timer but it required modifications to
command issue hot path, so I tried to cheat there (apparently
unsuccessfully).  I'll think about something better.

Thanks.

-- 
tejun