[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
RE: [PATCH V2 0/4] hw/block/nvme: Implement ZNS finish-zone ZDC AEN
From: |
Clay Mayers |
Subject: |
RE: [PATCH V2 0/4] hw/block/nvme: Implement ZNS finish-zone ZDC AEN |
Date: |
Wed, 23 Nov 2022 01:03:13 +0000 |
> From: Klaus Jensen <its@irrelevant.dk>
> Sent: Wednesday, November 9, 2022 12:40 AM
>
> On Oct 21 16:10, clay.mayers@kioxia.com wrote:
> > From: Clay Mayers <clay.mayers@kioxia.com>
> >
> > ZNS controllers have the option to limit the time a zone can remain in
> > the active state. It begins with a background process in the controller
> > setting the finish-zone-recommended FZR attribute for a zone. As part of
> > setting this attribute, the zone's id is added to the namespace's
> > zone-descriptor-changed (ZDC) log page. If enabled, items added to the
> > ZDC log page generate a ZDC "asynchronous event notification" AEN.
> Optionally,
> > the control can induce a "zone excursion" forcing the zone into the finished
> > state that also generates a ZDC event.
> >
> > Zone enabled applications need to properly handle ZDC events. In a real
> device,
> > the timeout is many hours making testing an application difficult.
> > Implemented is the generation of FZR ZDC events to speed up O/S and
> application
> > testing.
> >
> > Added to the zoned NVMe command set is an optional, per-namespace timer
> > (zoned.finish_time) to set the FZR attr for long-lived active zones; A per
> > namespace ZDC log page; AEN results to including CQE.DW1 (the NSID of the
> ZDC
> > AEN) and generating a ZDC AEN if it's been enabled. Zone excursions are not
> > modeled.
> >
> > See section 5.5 of the NVMe Zoned Namespace Command Set Specification
> v1.1
> > for more details.
> >
> > Changes since v1
> > - Fixed offset length checking in zdc log page
> > - Moved zdc_event_queued to the patch 4
> > - Unwatched zdc events in nvme_exit()
> >
> > Clay Mayers (4):
> > hw/block/nvme: add ZONE_FINISH_RECOMMENDED functionality
> > hw/block/nvme: add zone descriptor changed log page
> > hw/block/nvme: supply dw1 for aen result
> > hw/block/nvme: add zone descriptor changed AEN
> >
> > docs/system/devices/nvme.rst | 5 +
> > hw/nvme/ctrl.c | 174 +++++++++++++++++++++++++++++++++--
> > hw/nvme/ns.c | 15 +++
> > hw/nvme/nvme.h | 37 +++++++-
> > hw/nvme/trace-events | 3 +-
> > include/block/nvme.h | 14 ++-
> > 6 files changed, 233 insertions(+), 15 deletions(-)
> >
> > --
> > 2.27.0
> >
>
> Nice work Clay!
Thanks! I think you had implemented excursions at one point and I'm
not sure why it didn't make it in. I originally left out excursions because
all I needed was the AEN to test my Linux patch. I have a V3 in the works
with excursions so I can test what happens to libzbd applications. I can't
tell if there is real interest in this but I've tested rocksDB using zenfs and
it has difficulties with zone excursions. What I don't know is if excursions
are a corner case or not - in my world, they are not.
>
> Series looks good to me,
>
> Reviewed-by: Klaus Jensen <k.jensen@samsung.com>