Re: multipath-tools: add ANA support for NVMe device

From: Mike Snitzer
Date: Wed Nov 14 2018 - 14:26:26 EST


On Wed, Nov 14 2018 at 1:51pm -0500,
Hannes Reinecke <hare@xxxxxxx> wrote:

> On 11/14/18 6:47 PM, Mike Snitzer wrote:
> > On Wed, Nov 14 2018 at 2:49am -0500,
> > Hannes Reinecke <hare@xxxxxxx> wrote:
> >
> >> On 11/14/18 6:38 AM, Mike Snitzer wrote:
> >>> On Tue, Nov 13 2018 at 1:00pm -0500,
> >>> Mike Snitzer <snitzer@xxxxxxxxxx> wrote:
> >>>
> >>>> [1]: http://lists.infradead.org/pipermail/linux-nvme/2018-November/020765.html
> >>>> [2]: https://www.redhat.com/archives/dm-devel/2018-November/msg00072.html
> >>> ...
> >>>
> >>> I knew there had to be a pretty tight coupling between the NVMe driver's
> >>> native multipathing and ANA support... and that the simplicity of
> >>> Hannes' patch [1] was too good to be true.
> >>>
> >>> The real justification for not making Hannes' change is it'd effectively
> >>> be useless without first splitting out the ANA handling done during NVMe
> >>> request completion (NVME_SC_ANA_* cases in nvme_failover_req) that
> >>> triggers re-reading the ANA log page accordingly.
> >>>
> >>> So without the ability to drive the ANA workqueue to trigger
> >>> nvme_read_ana_log() from the nvme driver's completion path -- even if
> >>> nvme_core.multipath=N -- it really doesn't buy multipath-tools anything
> >>> to have the NVMe driver export the ana state via sysfs, because that ANA
> >>> state will never get updated.
> >>>
> >> Hmm. Indeed, I was more focussed on having the sysfs attributes
> >> displayed, so yes, indeed it needs some more work.
> > ...
> >>> Not holding my breath BUT:
> >>> if decoupling the reading of ANA state from native NVMe multipathing
> >>> specific work during nvme request completion were an acceptable
> >>> advancement I'd gladly do the work.
> >>>
> >> I'd be happy to work on that, given that we'll have to have 'real'
> >> ANA support for device-mapper anyway for SLE12 SP4 etc.
> >
> > I had a close enough look yesterday that I figured I'd just implement
> > what I reasoned through as one way forward, compile tested only (patch
> > relative to Jens' for-4.21/block):
> >
> > drivers/nvme/host/core.c | 14 +++++++---
> > drivers/nvme/host/multipath.c | 65 ++++++++++++++++++++++++++-----------------
> > drivers/nvme/host/nvme.h | 4 +++
> > 3 files changed, 54 insertions(+), 29 deletions(-)
> >
> > diff --git a/drivers/nvme/host/core.c b/drivers/nvme/host/core.c
> > index f172d63db2b5..05313ab5d91e 100644
> > --- a/drivers/nvme/host/core.c
> > +++ b/drivers/nvme/host/core.c
> > @@ -252,10 +252,16 @@ void nvme_complete_rq(struct request *req)
> > trace_nvme_complete_rq(req);
> >
> > if (unlikely(status != BLK_STS_OK && nvme_req_needs_retry(req))) {
> > - if ((req->cmd_flags & REQ_NVME_MPATH) &&
> > - blk_path_error(status)) {
> > - nvme_failover_req(req);
> > - return;
> > + if (blk_path_error(status)) {
> > + struct nvme_ns *ns = req->q->queuedata;
> > + u16 nvme_status = nvme_req(req)->status;
> > +
> > + if (req->cmd_flags & REQ_NVME_MPATH) {
> > + nvme_failover_req(req);
> > + nvme_update_ana(ns, nvme_status);
> > + return;
> > + }
> > + nvme_update_ana(ns, nvme_status);
> > }
> >
> > if (!blk_queue_dying(req->q)) {
> > diff --git a/drivers/nvme/host/multipath.c b/drivers/nvme/host/multipath.c
> > index 5e3cc8c59a39..f7fbc161dc8c 100644
> > --- a/drivers/nvme/host/multipath.c
> > +++ b/drivers/nvme/host/multipath.c
> > @@ -22,7 +22,7 @@ MODULE_PARM_DESC(multipath,
> >
> > inline bool nvme_ctrl_use_ana(struct nvme_ctrl *ctrl)
> > {
> > - return multipath && ctrl->subsys && (ctrl->subsys->cmic & (1 << 3));
> > + return ctrl->subsys && (ctrl->subsys->cmic & (1 << 3));
> > }
> >
> > /*
> > @@ -47,6 +47,17 @@ void nvme_set_disk_name(char *disk_name, struct nvme_ns *ns,
> > }
> > }
> >
> > +static bool nvme_ana_error(u16 status)
> > +{
> > + switch (status & 0x7ff) {
> > + case NVME_SC_ANA_TRANSITION:
> > + case NVME_SC_ANA_INACCESSIBLE:
> > + case NVME_SC_ANA_PERSISTENT_LOSS:
> > + return true;
> > + }
> > + return false;
> > +}
> > +
> > void nvme_failover_req(struct request *req)
> > {
> > struct nvme_ns *ns = req->q->queuedata;
> > @@ -58,10 +69,7 @@ void nvme_failover_req(struct request *req)
> > spin_unlock_irqrestore(&ns->head->requeue_lock, flags);
> > blk_mq_end_request(req, 0);
> >
> > - switch (status & 0x7ff) {
> > - case NVME_SC_ANA_TRANSITION:
> > - case NVME_SC_ANA_INACCESSIBLE:
> > - case NVME_SC_ANA_PERSISTENT_LOSS:
> > + if (nvme_ana_error(status)) {
> > /*
> > * If we got back an ANA error we know the controller is alive,
> > * but not ready to serve this namespaces. The spec suggests
> > @@ -69,31 +77,38 @@ void nvme_failover_req(struct request *req)
> > * that the admin and I/O queues are not serialized that is
> > * fundamentally racy. So instead just clear the current path,
> > * mark the the path as pending and kick of a re-read of the ANA
> > - * log page ASAP.
> > + * log page ASAP (see nvme_update_ana() below).
> > */
> > nvme_mpath_clear_current_path(ns);
> > - if (ns->ctrl->ana_log_buf) {
> > - set_bit(NVME_NS_ANA_PENDING, &ns->flags);
> > - queue_work(nvme_wq, &ns->ctrl->ana_work);
> > + } else {
> > + switch (status & 0x7ff) {
> > + case NVME_SC_HOST_PATH_ERROR:
> > + /*
> > + * Temporary transport disruption in talking to the
> > + * controller. Try to send on a new path.
> > + */
> > + nvme_mpath_clear_current_path(ns);
> > + break;
> > + default:
> > + /*
> > + * Reset the controller for any non-ANA error as we
> > + * don't know what caused the error.
> > + */
> > + nvme_reset_ctrl(ns->ctrl);
> > + break;
> > }
> > - break;
> > - case NVME_SC_HOST_PATH_ERROR:
> > - /*
> > - * Temporary transport disruption in talking to the controller.
> > - * Try to send on a new path.
> > - */
> > - nvme_mpath_clear_current_path(ns);
> > - break;
> > - default:
> > - /*
> > - * Reset the controller for any non-ANA error as we don't know
> > - * what caused the error.
> > - */
> > - nvme_reset_ctrl(ns->ctrl);
> > - break;
> > }
> Maybe move nvme_mpath_clear_current_path() out of the loop (it wouldn't
> matter if we clear the path if we reset the controller afterwards); that
> might even clean up the code even more.

Not completely sure what you're suggesting. But I was going for purely
functional equivalent of the existing upstream code -- just with
multipathing and ana split.

Could be that in future it wouldn't make sense to always
nvme_mpath_clear_current_path() for all cases?

>
> > +}
> >
> > - kblockd_schedule_work(&ns->head->requeue_work);
> > +void nvme_update_ana(struct nvme_ns *ns, u16 status)
> > +{
> > + if (nvme_ana_error(status) && ns->ctrl->ana_log_buf) {
> > + set_bit(NVME_NS_ANA_PENDING, &ns->flags);
> > + queue_work(nvme_wq, &ns->ctrl->ana_work);
> > + }
> > +
> > + if (multipath)
> > + kblockd_schedule_work(&ns->head->requeue_work);
> > }
>
> maybe use 'ns->head->disk' here; we only need to call this if we have a
> multipath disk.

Sure.

> Remaining bits are okay.

Thanks,
Mike