[PATCH] NVMe: Force cancel commands on hot-removal

[Keith Busch]

On Thu, 10 Sep 2015, Mohana Goli wrote:
> I called nvme_freeze_queues before calling nvme_ns_remove in nvme_dev_remove,which seems to be solved  the problem.

Great! I thought that would do it, but I've some tests that seem to
get 1 IO lost somewhere in the block layer (thinking maybe it's in the
ctx->rq_list, still trying to confirm). That failure doesn't happen very
often though.


> On Thu, Sep 10, 2015 at 4:14 PM, Mohana Goli <mohana.goli at seagate.com> wrote:
>       Yes,i think we should call "nvme_freeze_queues" before deleting the gendisk.Right now that is no op since we are calling after deleting the name space.
> 
> On Wed, Sep 9, 2015 at 10:00 PM, Keith Busch <keith.busch at intel.com> wrote:
>       On Wed, 9 Sep 2015, Mohana Goli wrote:
>             I ran another test after creating filesystem on the nvme
>             devices.Unfortunately i have seen the IO Timeouts again.
>             I used the following steps:
> 
>
>       I guess we need stop the h/w queues first, then. I thought killing it
>       would have sufficed, but inflight commands will make it to the driver.
> 
> 
> 
> 
>