On Fri, Apr 16 2021 at 11:20am -0400, Hannes Reinecke <hare@xxxxxxx> wrote: > On 4/16/21 4:53 PM, Mike Snitzer wrote: > > On Fri, Apr 16 2021 at 10:01am -0400, > > Hannes Reinecke <hare@xxxxxxx> wrote: > > > >> On 4/16/21 1:15 AM, Mike Snitzer wrote: > >>> From: Chao Leng <lengchao@xxxxxxxxxx> > >>> > >>> REQ_FAILFAST_TRANSPORT was designed for SCSI, because the SCSI protocol > >>> does not define the local retry mechanism. SCSI implements a fuzzy > >>> local retry mechanism, so REQ_FAILFAST_TRANSPORT is needed to allow > >>> higher-level multipathing software to perform failover/retry. > >>> > >>> NVMe is different with SCSI about this. It defines a local retry > >>> mechanism and path error codes, so NVMe should retry local for non > >>> path error. If path related error, whether to retry and how to retry > >>> is still determined by higher-level multipathing's failover. > >>> > >>> Unlike SCSI, NVMe shouldn't prevent retry if REQ_FAILFAST_TRANSPORT > >>> because NVMe's local retry is needed -- as is NVMe specific logic to > >>> categorize whether an error is path related. > >>> > >>> In this way, the mechanism of NVMe multipath or other multipath are > >>> now equivalent. The mechanism is: non path related error will be > >>> retried locally, path related error is handled by multipath. > >>> > >>> Signed-off-by: Chao Leng <lengchao@xxxxxxxxxx> > >>> [snitzer: edited header for grammar and clarity, also added code comment] > >>> Signed-off-by: Mike Snitzer <snitzer@xxxxxxxxxx> > >>> --- > >>> drivers/nvme/host/core.c | 9 ++++++++- > >>> 1 file changed, 8 insertions(+), 1 deletion(-) > >>> > >>> diff --git a/drivers/nvme/host/core.c b/drivers/nvme/host/core.c > >>> index 540d6fd8ffef..4134cf3c7e48 100644 > >>> --- a/drivers/nvme/host/core.c > >>> +++ b/drivers/nvme/host/core.c > >>> @@ -306,7 +306,14 @@ static inline enum nvme_disposition nvme_decide_disposition(struct request *req) > >>> if (likely(nvme_req(req)->status == 0)) > >>> return COMPLETE; > >>> > >>> - if (blk_noretry_request(req) || > >>> + /* > >>> + * REQ_FAILFAST_TRANSPORT is set by upper layer software that > >>> + * handles multipathing. Unlike SCSI, NVMe's error handling was > >>> + * specifically designed to handle local retry for non-path errors. > >>> + * As such, allow NVMe's local retry mechanism to be used for > >>> + * requests marked with REQ_FAILFAST_TRANSPORT. > >>> + */ > >>> + if ((req->cmd_flags & (REQ_FAILFAST_DEV | REQ_FAILFAST_DRIVER)) || > >>> (nvme_req(req)->status & NVME_SC_DNR) || > >>> nvme_req(req)->retries >= nvme_max_retries) > >>> return COMPLETE; > >>> > >> Huh? > >> > >> #define blk_noretry_request(rq) \ > >> ((rq)->cmd_flags & (REQ_FAILFAST_DEV|REQ_FAILFAST_TRANSPORT| \ > >> REQ_FAILFAST_DRIVER)) > >> > >> making the only _actual_ change in your patch _not_ evaluating the > >> REQ_FAILFAST_DRIVER, which incidentally is only used by the NVMe core. > > > > No, not sure how you got there. I'd have thought the 5 references to > > "REQ_FAILFAST_TRANSPORT" would've been sufficient ;) > > > > Ah. Misread stuff. You're excluding the REQ_FAILFAST_TRANSPORT here. > But then it's _actually_ similar to the next patch (which I've also > commented). > > Wouldn't it be better to fold them into one patch and discuss things > together; especially as my comment to the next one might actually > achieve the same thing? 2 discrete things. This patch enables local retry. Patch 3 allows proper failover via upper layer multipathing. And as I replied, your suggestion about using DNR doesn't achieve the same thing (said as much in reply to the patch 3 thread). Mike
