On Fri, May 29, 2020 at 9:22 PM Keith Busch <kbusch@xxxxxxxxxx> wrote: > > On Fri, May 29, 2020 at 11:39:46AM +0800, Ming Lei wrote: > > On Fri, May 29, 2020 at 4:19 AM Alan Adamson <alan.adamson@xxxxxxxxxx> wrote: > > That said NVMe's > > error handling becomes better after applying the patchs of '[PATCH > > 0/3] blk-mq/nvme: improve > > nvme-pci reset handler'. > > I think that's a bit debatable. Alan is synthesizing a truly broken > controller. The current code will abandon this controller after about 30 Not sure it can be thought as a truly broken controller. When waiting on nvme_wait_freeze() during reset, the controller has been in normal state. There is still chance to trigger timeout by any occasional event, just like any other timeout, which isn't special enough for us to have to kill the controller. > seconds. Your series will reset that broken controller indefinitely. > Which of those options is better? Removing controller is very horrible, because it becomes a brick basically, together with data loss. And we should retry enough before killing the controller. Mys series doesn't reset indefinitely since every request is just retried limited times(default is 5), at least chance should be provided to retry claimed times for IO requests. > > I think degrading to an admin-only mode at some point would be preferable. If the timeout event is occasional, this way gives up too early and doesn't retry claimed times, then peopele may complain for data loss. Thanks, Ming Lei
