Re: IDE lockups with 2.5.28...

[Marcin Dalecki <dalecki@evision.ag>]

Petr Vandrovec wrote:

> Well, no. Both of these loop have completely different terminating conditions.
> You exit when IDE hardware is busy, while SCSI exits if hardware is busy,
> or when there is nothing to do. Fundamental difference.

Shit - you are right. We look until the next request sets IDE_BUSY as a 
side effect.... I just wanted to close the window between clear we clear
IDE_BUSY in ata_irq_handler just before recalling do_request to set it 
immediately on again.
Should be both of course.

>>Same allies to blk_stop_queue().
> 
> 
> So your request_fn is invoked for each of queues which had pending
> requests. Upper layer cannot expect that you are using two queues,
> but hardware really wants to use only one. Shared queue_lock is there
> for hardware which can start one request at a time (one set of
> registers...), but can have requests to the different devices
> in progress.

Yes theoretically yes. The problem is only that queue_lock doesn't as
advertized becouse the request_fn are *releasing* the spin lock at a 
point where the QUEUE_FLAG_STOP doesn't have any usefull value.


> P.S.: I did not saw IDE 105. Does it exist?

I think I did send it under a wrong topic. Please look for Re:
Linux-2.5.28.