Re: [PATCH 3/3] Make scsi_free_queue() abort pending requests

[Mike Christie <michaelc@cs.wisc.edu>]

On 05/05/2012 01:07 AM, Bart Van Assche wrote:
> On 05/04/12 20:32, Mike Christie wrote:
> 
>> Oh not wait. I do not get the patch. After blk_cleanup_queue runs then
>> no IO should be running and no new IO can be queued can it?
>>
>>>>  	 */
>>>>  	blk_cleanup_queue(q);
>>>> +	blk_abort_queue(q);
>>>>  
>>>>  	if (sdev->is_visible) {
>>>>  		if (scsi_device_set_state(sdev, SDEV_CANCEL) != 0)
> 
> 
> After blk_cleanup_queue() finished no new requests will be queued to a
> SCSI LLD. However, that function doesn't wait for already queued
> requests to finish. I have verified with ib_srp LLD that the\

It does for me. It is supposed to isn't it? For BLOCK FS requests
blk_drain_queue will wait for the q->in_flight counters to go to zero.
They get incremented at scsi_request_fn-> blk_start_request ->
blk_dequeue_request time and decremented in scsi_end_request ->
blk_end_request ->  blk_end_bidi_request -> blk_finish_request ->
blk_put_request -> elv_completed_request. For BLOCK PC and other
requests there is the rq.count tracking.

I added some printk code in the blk_drain_queue loop, and if I force the
target to wait to complete a IO, then remove the device, blk_drain_queue
is looping until the IO is eventually completed.


> blk_abort_queue() call triggers the "SRP abort called" kernel log
> message generated by ib_srp when srp_abort() is called.

What type of IO is it?