[PATCH 1/1] nvme-multipath: Skip nr_active increments in RETRY disposition

Keith Busch kbusch at kernel.org
Thu Sep 25 07:43:44 PDT 2025


On Wed, Sep 24, 2025 at 06:14:27PM -0700, Mohamed Khalfella wrote:
> On 2025-09-24 17:02:51 -0600, Keith Busch wrote:
> > On Wed, Sep 24, 2025 at 03:43:18PM -0700, Amit Chaudhary wrote:
> > >  static inline void nvme_start_request(struct request *rq)
> > >  {
> > > -	if (rq->cmd_flags & REQ_NVME_MPATH)
> > > +	if ((rq->cmd_flags & REQ_NVME_MPATH) && (!nvme_req(rq)->retries))
> > >  		nvme_mpath_start_request(rq);
> > >  	blk_mq_start_request(rq);
> > >  }
> > 
> > Using "retries" is bit indirect as a proxy for multipath active counts.
> > Could this be moved to the mpath start instead, directly using the flag
> > that accounts for the path? This also helps to keep track if the command
> > gets retried across a user toggling the policy to "qd".
> > 
> > ---
> > diff --git a/drivers/nvme/host/multipath.c b/drivers/nvme/host/multipath.c
> > index 3da980dc60d91..1c630967ddd40 100644
> > --- a/drivers/nvme/host/multipath.c
> > +++ b/drivers/nvme/host/multipath.c
> > @@ -182,7 +182,8 @@ void nvme_mpath_start_request(struct request *rq)
> >         struct nvme_ns *ns = rq->q->queuedata;
> >         struct gendisk *disk = ns->head->disk;
> > 
> > -       if (READ_ONCE(ns->head->subsys->iopolicy) == NVME_IOPOLICY_QD) {
> > +       if (READ_ONCE(ns->head->subsys->iopolicy) == NVME_IOPOLICY_QD &&
> > +           !(nvme_req(rq)->flags & NVME_MPATH_CNT_ACTIVE)) {
> >                 atomic_inc(&ns->ctrl->nr_active);
> >                 nvme_req(rq)->flags |= NVME_MPATH_CNT_ACTIVE;
> >         }
> > --
> 
> 193         nvme_req(rq)->flags |= NVME_MPATH_IO_STATS;
> 194         nvme_req(rq)->start_time = bdev_start_io_acct(disk->part0, req_op(rq),
> 195                                                       jiffies);
> 
> Doing it this way might messup with stats accounting because the two
> lines above will be executed on request retry. I do not think we need
> that, right?

Yeah, but we can use the other flag to know if it's already been
accounted:

--- a/drivers/nvme/host/multipath.c
+++ b/drivers/nvme/host/multipath.c
@@ -182,12 +182,14 @@ void nvme_mpath_start_request(struct request *rq)
        struct nvme_ns *ns = rq->q->queuedata;
        struct gendisk *disk = ns->head->disk;

-       if (READ_ONCE(ns->head->subsys->iopolicy) == NVME_IOPOLICY_QD) {
+       if (READ_ONCE(ns->head->subsys->iopolicy) == NVME_IOPOLICY_QD &&
+           !(nvme_req(rq)->flags & NVME_MPATH_CNT_ACTIVE)) {
                atomic_inc(&ns->ctrl->nr_active);
                nvme_req(rq)->flags |= NVME_MPATH_CNT_ACTIVE;
        }

-       if (!blk_queue_io_stat(disk->queue) || blk_rq_is_passthrough(rq))
+       if (!blk_queue_io_stat(disk->queue) || blk_rq_is_passthrough(rq) ||
+           nvme_req(rq)->flags & NVME_MPATH_IO_STATS)
                return;

        nvme_req(rq)->flags |= NVME_MPATH_IO_STATS;



More information about the Linux-nvme mailing list