On Fri, May 4, 2018 at 4:28 PM, jianchao.wang <jianchao.w.w...@oracle.com> wrote: > Hi ming > > On 05/04/2018 04:02 PM, Ming Lei wrote: >>> nvme_error_handler should invoke nvme_reset_ctrl instead of introducing >>> another interface. >>> Then it is more convenient to ensure that there will be only one resetting >>> instance running. >>> >> But as you mentioned above, reset_work has to be splitted into two >> contexts for handling IO timeout during wait_freeze in reset_work, >> so single instance of nvme_reset_ctrl() may not work well. > > I mean the EH kthread and the reset_work which both could reset the ctrl > instead of > the pre and post rest context. > > Honestly, I suspect a bit that whether it is worthy to try to recover from > [1]. > The Eh kthread solution could make things easier, but the codes for recovery > from [1] has > made code really complicated. It is more difficult to unify the nvme-pci, > rdma and fc.
Another choice may be nested EH, which should be easier to implement: - run the whole recovery procedures(shutdown & reset) in one single context - and start a new context to handle new timeout during last recovery in the same way The two approaches is just like sync IO vs AIO. Thanks, Ming Lei
