[PATCH] NVMe: Re-introduce nvmeq->q_suspended

[Jens Axboe]

On 12/09/2014 03:08 PM, Sam Bradshaw (sbradshaw) wrote:
>
>
>> -----Original Message-----
>> From: Jens Axboe [mailto:axboe at kernel.dk]
>> Sent: Friday, November 21, 2014 5:12 PM
>> To: Keith Busch; Sam Bradshaw (sbradshaw)
>> Cc: linux-nvme at lists.infradead.org; m at bjorling.me
>> Subject: Re: [PATCH] NVMe: Re-introduce nvmeq->q_suspended
>>
>> On 11/21/2014 06:07 PM, Keith Busch wrote:
>>> On Fri, 21 Nov 2014, Sam Bradshaw wrote:
>>>> At some point in the blk-mq transition, q_suspended was removed.
>>>> This patch adds it back to address two failure modes, both
>>>> illustrated in the attached snippet.  First, it protects the ioctl
>>>> path from inadvertently dereferencing invalid nvme_queue data
>>>> structures during device shutdown/reset.  Second, it protects
>> against
>>>> a double
>>>> free_irq() if the shutdown/reset/resume sequence doesn't recover a
>>>> flaky controller.
>>>
>>> I could have sworn there was a better way to suspend at the
>>> request_queue level with blk-mq so we don't need this in the driver.
>>> The blk_start/stop_queue is close, but doesn't work in the ioctl path.
>>> Maybe I'm thinking blk-mq freeze, but that's not exported for drivers.
>>
>> The freeze stuff should work. I've got a prelim patch that allows
>> attach/detach of the namespace/block parts, the first prep patch is
>> exporting the freeze bits.
>
> Would it be possible either to incorporate my patch or your patch exporting freeze bits so I could build similar functionality on top of it?  I'm not able to test any interesting hot plug cases with the current state of the nvme driver in for-3.19/drivers.

We can definitely export the freeze bits. If you have a tested patch 
with that approach, by all means, lets get it reviewed and incorporated.


-- 
Jens Axboe