Re: [PATCH v4 6/6] io_uring: add support for zone-append

[Kanchan Joshi]

On Fri, Sep 25, 2020 at 8:22 AM Damien Le Moal <Damien.LeMoal@xxxxxxx> wrote:
>
> On 2020/09/25 2:20, Kanchan Joshi wrote:
> > On Tue, Sep 8, 2020 at 8:48 PM hch@xxxxxxxxxxxxx <hch@xxxxxxxxxxxxx> wrote:
> >>
> >> On Mon, Sep 07, 2020 at 12:31:42PM +0530, Kanchan Joshi wrote:
> >>> But there are use-cases which benefit from supporting zone-append on
> >>> raw block-dev path.
> >>> Certain user-space log-structured/cow FS/DB will use the device that
> >>> way. Aerospike is one example.
> >>> Pass-through is synchronous, and we lose the ability to use io-uring.
> >>
> >> So use zonefs, which is designed exactly for that use case.
> >
> > Not specific to zone-append, but in general it may not be good to lock
> > new features/interfaces to ZoneFS alone, given that direct-block
> > interface has its own merits.
> > Mapping one file to a one zone is good for some use-cases, but
> > limiting for others.
> > Some user-space FS/DBs would be more efficient (less meta, indirection)
> > with the freedom to decide file-to-zone mapping/placement.
>
> There is no metadata in zonefs. One file == one zone and the mapping between
> zonefs files and zones is static, determined at mount time simply using report
> zones. Zonefs files cannot be renamed nor deleted in anyway. Choosing a zonefs
> file *is* the same as choosing a zone. Zonfes is *not* a POSIX file system doing
> dynamic block allocation to files. The backing storage of files in zonefs is
> static and fixed to the zone they represent. The difference between zonefs vs
> raw zoned block device is the API that has to be used by the application, that
> is, file descriptor representing the entire disk for raw disk vs file descriptor
> representing one zone in zonefs. Note that the later has *a lot* of advantages
> over the former: enables O_APPEND use, protects against bugs with user write
> offsets mistakes, adds consistency of cached data against zone resets, and more.
>
> > - Rocksdb and those LSM style DBs would map SSTable to zone, but
> > SSTable file may be two small (initially) and may become too large
> > (after compaction) for a zone.
>
> You are contradicting yourself here. If a SSTable is mapped to a zone, then its
> size cannot exceed the zone capacity, regardless of the interface used to access
> the zones. And except for L0 tables which can be smaller (and are in memory
> anyway), all levels tables have the same maximum size, which for zoned drives
> must be the zone capacity. In any case, solving any problem in this area does
> not depend in any way on zonefs vs raw disk interface. The implementation will
> differ depending on the chosen interface, but what needs to be done to map
> SSTables to zones is the same in both cases.
>
> > - The internal parallelism of a single zone is a design-choice, and
> > depends on the drive. Writing multiple zones parallely (striped/raid
> > way) can give better performance than writing on one. In that case one
> > would want to file that seamlessly combines multiple-zones in a
> > striped fashion.
>
> Then write a FS for that... Or have a library do it in user space. For the
> library case, the implementation will differ for zonefs vs raw disk due to the
> different API (regular file vs block devicer file), but the principles to follow
> for stripping zones into a single storage object remain the same.

ZoneFS is better when it is about dealing at single-zone granularity,
and direct-block seems better when it is about grouping zones (in
various ways including striping). The latter case (i.e. grouping
zones) requires more involved mapping, and I agree that it can be left
to application (for both ZoneFS and raw-block backends).
But when an application tries that on ZoneFS, apart from mapping there
would be additional cost of indirection/fd-management (due to
file-on-files).
And if new features (zone-append for now) are available only on
ZoneFS, it forces application to use something that maynot be most
optimal for its need.

Coming to the original problem of plumbing append - I think divergence
started because RWF_APPEND did not have any meaning for block device.
Did I miss any other reason?
How about write-anywhere semantics (RWF_RELAXED_WRITE or
RWF_ANONYMOUS_WRITE flag) on block-dev.
Zone-append works a lot like write-anywhere on block-dev (or on any
other file that combines multiple-zones, in non-sequential fashion).

> > Also it seems difficult (compared to block dev) to fit simple-copy TP
> > in ZoneFS. The new
> > command needs: one NVMe drive, list of source LBAs and one destination
> > LBA. In ZoneFS, we would deal with N+1 file-descriptors (N source zone
> > file, and one destination zone file) for that. While with block
> > interface, we do not need  more than one file-descriptor representing
> > the entire device. With more zone-files, we face open/close overhead too.
>
> Are you expecting simple-copy to allow requests that are not zone aligned ? I do
> not think that will ever happen. Otherwise, the gotcha cases for it would be far
> too numerous. Simple-copy is essentially an optimized regular write command.
> Similarly to that command, it will not allow copies over zone boundaries and
> will need the destination LBA to be aligned to the destination zone WP. I have
> not checked the TP though and given the NVMe NDA, I will stop the discussion here.

TP is ratified, if that is the problem you are referring to.

> filesend() could be used as the interface for simple-copy. Implementing that in
> zonefs would not be that hard. What is your plan for simple-copy interface for
> raw block device ? An  ioctl ? filesend() too ? As as with any other user level
> API, we should not be restricted to a particular device type if we can avoid it,
> so in-kernel emulation of the feature is needed for devices that do not have
> simple-copy or scsi extended copy. filesend() seems to me like the best choice
> since all of that is already implemented there.

At this moment, ioctl as sync and io-uring for async. sendfile() and
copy_file_range() takes two fds....with that we can represent copy
from one source zone to another zone.
But it does not fit to represent larger copy (from N source zones to
one destination zone).
Not sure if I am clear, perhaps sending RFC would be better for
discussion on simple-copy.

> As for the open()/close() overhead for zonefs, may be some use cases may suffer
> from it, but my tests with LevelDB+zonefs did not show any significant
> difference. zonefs open()/close() operations are way faster than for a regular
> file system since there is no metadata and all inodes always exist in-memory.
> And zonefs() now supports MAR/MOR limits for O_WRONLY open(). That can simplify
> things for the user.
>
>
> --
> Damien Le Moal
> Western Digital Research



-- 
Joshi