Re: [PATCH v3 0/2] block,scsi: fixup blk_get_request dead queue scenarios

[Jens Axboe <axboe@kernel.dk>]

On 08/26/2014 03:27 PM, Jeff Moyer wrote:
> Jens Axboe <axboe@kernel.dk> writes:
> 
>> On 08/26/2014 11:24 AM, Jeff Moyer wrote:
>>> Joe Lawrence <joe.lawrence@stratus.com> writes:
>>>
>>>> v2->v3: rebase to 3.16-rc2, consider return values from the
>>>>         blk_mq_alloc_request leg of the blk_get_request callchain
>>>>         (noted by Jeff), noted in the second patch changelog.
>>>>
>>>> 	blk_mq_queue_enter may return 0 or errno, which
>>>>         blk_mq_alloc_request can propogate out via ERR_PTR.
>>>> 	__blk_mq_alloc_request doesn't include any blk_queue_dying
>>>> 	checks, so I'm assuming that its failures can be attributed
>>>>         to -EWOULDBLOCK under !GFP_WAIT conditions.
>>>>
>>>> v1->v2: incorporate Jeff's feedback in bsg_map_hdr() and Reviewed-by
>>>>         tags.
>>>>
>>>> Joe Lawrence (2):
>>>>   block,scsi: verify return pointer from blk_get_request
>>>>   block,scsi: fixup blk_get_request dead queue scenarios
>>>
>>> Jens,
>>>
>>> Did this patch set fall through the cracks again?
>>
>> Falling through the cracks implies that I meant to apply it and did not,
>> which was not the case. 
> 
> Sorry, I was mislead by our earlier conversation on this (mail inline
> below).

I changed my mind, it didn't feel fully baked to me.

>> But I think we're at the point now where I'm finally comfortable with
>> applying it. So, Joe, could you ensure that it applies to 3.17-rc2,
>> then I will roll it in to the updates for 3.18.
> 
> Joe, you will have one hunk to modify for sure, in scsi_ioctl.c.  A
> previous patch added a check for null, but ended up returning the wrong
> value (ENOMEM instead of ENODEV).

I have applied the first one, will look over the second one and hand
apply it. Seems the NULL return was completely removed, so we _should_
be ok on the IS_ERR() conversion, though that sort of thing always
worries me a little bit. A NULL return can quickly show up again, and
then they would all fail.

-- 
Jens Axboe