Re: block completion races

[Jens Axboe]

On Tue, Jan 15 2002, Andre Hedrick wrote:
> 
> We have a more interesting problem!
> This does not show up in 2.5.1 which is patch base for 2.5.3-pre1.
> It does show up in 2.5.2-pre10 but I have not walked the patch through
> 2.5.2preX series.
> 
> We have a very bad queue race that is PIO specific but really the whole
> darn driver before the patch was applied.  ACB only tighten the driver's
> alignment to the NCITS standards.  One should note the direct access via
> the ioctl does not lock the driver.  Only coming down from BLOCK will this
> event occur.
> 
> Repeatable test "hdparm -d0 -t /dev/hdx"
> 
> If you apply the acb-io patch to 2.5.1 this does not happen.
> 
> In the introduction of BIO, there were no "q->queue_lock" applied to
> protecting the queue.
> 
>         /*
>          * Is meant to protect the queue in the future instead of
>          * io_request_lock
>          */
>         spinlock_t              queue_lock;
> 
> Well we pulled "io_request_lock" but did we forgot to insert or add
> q->queue_lock spinlocks?
> 
> It is going to be a LONG LONG NIGHT :-(

ide_lock == q->queue_lock

ide-probe.c:ide_init_queue():	blk_init_queue(q, do_ide_request, &ide_lock);

-- 
Jens Axboe