Hi Christoph Many thanks for your kindly response.
On 01/04/2018 06:35 PM, Christoph Hellwig wrote: > On Wed, Jan 03, 2018 at 06:31:44AM +0800, Jianchao Wang wrote: >> NVME_CTRL_RESETTING used to indicate the range of nvme initializing >> strictly in fd634f41(nvme: merge probe_work and reset_work), but it >> is not now. The NVME_CTRL_RESETTING is set before queue the >> reset_work, there could be a big gap before the reset work handles >> the outstanding requests. So when the NVME_CTRL_RESETTING is set, >> nvme_timeout will not only meet the admin requests from the >> initializing procedure, but also the IO and admin requests from >> previous work before nvme_dev_disable is invoked. >> >> To fix it, introduce a flag NVME_DEV_FLAG_INITIALIZING to mark the >> range of initializing. When this flag is not set, handle the expried >> requests as nvme_cancel_request. Otherwise, the requests should be >> from the initializing procedure. Handle them as before. Because the >> nvme_reset_work will see the error and disable the dev itself, so >> discard the nvme_dev_disable here. > > Instead of a parallel set of states we'll need to split > NVME_CTRL_RESET into NVME_CTRL_RESET_SCHEDULED and NVME_CTRL_RESETTING. > > And if my memory doesn't fail me we were already considering that a while > ago. > Yes, it is indeed more reasonable to split current NVME_CTRL_RESETTING into two states, but the nvme_dev_disable() in nvme_reset_work() should be the boundary. After that, all the in-flight requests are requeued and request queue is quiesced, the nvme driver is clear. So the new state maybe something like NEW_CTRL_RESET_PREPARE.:) Thanks Jianchao