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.
So what is it you're trying to solve?
Cheers,
Hannes
--
Dr. Hannes Reinecke Kernel Storage Architect
hare@xxxxxxx +49 911 74053 688
SUSE Software Solutions Germany GmbH, 90409 Nürnberg
GF: F. Imendörffer, HRB 36809 (AG Nürnberg)
