| Message ID | 20200324110255.8385-1-bob.liu@oracle.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | dm zoned: extend the way of exposing zoned block device | expand |
On 3/24/20 12:02 PM, Bob Liu wrote: > Motivation: > dm-zoned exposes a zoned block device(ZBC) as a regular block device by storing > metadata and buffering random writes in its conventional zones. > This way is not flexible, there must be enough conventional zones and the > performance may be constrained. > > This patchset split the metadata from zoned device to an extra regular device, > with aim to increase the flexibility and potential performance. > For example, now we can store metadata in a faster device like persistent memory. > Also random writes can go to the regular devices in this version. > > Usage(will send user space patches later): >> dmzadm --format $zoned_dev --regular=$regu_dev --force >> echo "0 $size zoned $regu_dev $zoned_dev" | dmsetup create $dm-zoned-name > > v2: > * emulate regular device zone info > * support write both metadata and random writes to regular dev > > Bob Liu (3): > dm zoned: rename dev name to zoned_dev > dm zoned: introduce regular device to dm-zoned-target > dm zoned: add regular device info to metadata > > drivers/md/dm-zoned-metadata.c | 205 +++++++++++++++++++++++++++-------------- > drivers/md/dm-zoned-target.c | 205 +++++++++++++++++++++++------------------ > drivers/md/dm-zoned.h | 53 ++++++++++- > 3 files changed, 299 insertions(+), 164 deletions(-) > Well, surprise, surprise, both our patchsets are largely identical ... So how to proceed? I guess if you were using 'cdev' instead of 'regu_dm_dev' we should be having an overlap of about 90 percent. The main difference between our implementation is that I didn't move the metadata to the cache/regulard device, seeing that dmzadm will only write metadata onto the zoned device. I would rather keep it that way (ie storing metadata on the zoned device, too, if possible) as we would be keeping backwards compability with that. And we could always move metadata to the cache/regular device in a later patch; for doing it properly we'll need to update the metadata anyway as we need to introduce UUIDs to stitch those devices together. Remember, one my have more than one zoned device and regular device... Should I try to merge both patchsets and send them out as an RFC? Cheers, Hannes
On 3/24/20 7:52 PM, Hannes Reinecke wrote: > On 3/24/20 12:02 PM, Bob Liu wrote: >> Motivation: >> dm-zoned exposes a zoned block device(ZBC) as a regular block device by storing >> metadata and buffering random writes in its conventional zones. >> This way is not flexible, there must be enough conventional zones and the >> performance may be constrained. >> >> This patchset split the metadata from zoned device to an extra regular device, >> with aim to increase the flexibility and potential performance. >> For example, now we can store metadata in a faster device like persistent memory. >> Also random writes can go to the regular devices in this version. >> >> Usage(will send user space patches later): >>> dmzadm --format $zoned_dev --regular=$regu_dev --force >>> echo "0 $size zoned $regu_dev $zoned_dev" | dmsetup create $dm-zoned-name >> >> v2: >> * emulate regular device zone info >> * support write both metadata and random writes to regular dev >> >> Bob Liu (3): >> dm zoned: rename dev name to zoned_dev >> dm zoned: introduce regular device to dm-zoned-target >> dm zoned: add regular device info to metadata >> >> drivers/md/dm-zoned-metadata.c | 205 +++++++++++++++++++++++++++-------------- >> drivers/md/dm-zoned-target.c | 205 +++++++++++++++++++++++------------------ >> drivers/md/dm-zoned.h | 53 ++++++++++- >> 3 files changed, 299 insertions(+), 164 deletions(-) >> > Well, surprise, surprise, both our patchsets are largely identical ... > You may missed my first rfc. https://www.redhat.com/archives/dm-devel/2020-January/msg00024.html > So how to proceed? I guess if you were using 'cdev' instead of 'regu_dm_dev' we should be having an overlap of about 90 percent. > > The main difference between our implementation is that I didn't move the metadata to the cache/regulard device, seeing that dmzadm will only write metadata onto the zoned device. I also patched dmzadm(will sent out soon) a lot, now my implementation can compatible with original usage. It supports two different usage: - Original zoned device only: #: dmzadm --format $zoned_dev #: echo "0 $size zoned $zoned_dev" | dmsetup create $dm-zoned-name (All data in zoned device) - With regular device： dmzadm --format $zoned_dev --regular=$regu_dev echo "0 $size zoned $regu_dev $zoned_dev" | dmsetup create $dm-zoned-name (Metadata in regular device, other data spread in both regular device and zoned device.) I haven't thought about store metadata in zoned dev while other data in both cache/regular device and zoned dev. Actually in my first rfc I just plan to split metadata to a fast device, so as to get performance improvement. > I would rather keep it that way (ie storing metadata on the zoned device, too, if possible) as we would be keeping backwards compability with that. > And we could always move metadata to the cache/regular device in a later patch; for doing it properly we'll need to update the metadata anyway as we need to introduce UUIDs to stitch those devices together. Sure. > Remember, one my have more than one zoned device and regular device... > > Should I try to merge both patchsets and send them out as an RFC? > Fine to me, just please keep my signed-off-by. Thanks, Bob