On 2020/06/26 2:18, Kanchan Joshi wrote:
> [Revised as per feedback from Damien, Pavel, Jens, Christoph, Matias, Wilcox]
>
> This patchset enables zone-append using io-uring/linux-aio, on block IO path.
> Purpose is to provide zone-append consumption ability to applications which are
> using zoned-block-device directly.
>
> The application may specify RWF_ZONE_APPEND flag with write when it wants to
> send zone-append. RWF_* flags work with a certain subset of APIs e.g. uring,
> aio, and pwritev2. An error is reported if zone-append is requested using
> pwritev2. It is not in the scope of this patchset to support pwritev2 or any
> other sync write API for reasons described later.
>
> Zone-append completion result --->
> With zone-append, where write took place can only be known after completion.
> So apart from usual return value of write, additional mean is needed to obtain
> the actual written location.
>
> In aio, this is returned to application using res2 field of io_event -
>
> struct io_event {
> __u64 data; /* the data field from the iocb */
> __u64 obj; /* what iocb this event came from */
> __s64 res; /* result code for this event */
> __s64 res2; /* secondary result */
> };
>
> In io-uring, cqe->flags is repurposed for zone-append result.
>
> struct io_uring_cqe {
> __u64 user_data; /* sqe->data submission passed back */
> __s32 res; /* result code for this event */
> __u32 flags;
> };
>
> Since 32 bit flags is not sufficient, we choose to return zone-relative offset
> in sector/512b units. This can cover zone-size represented by chunk_sectors.
> Applications will have the trouble to combine this with zone start to know
> disk-relative offset. But if more bits are obtained by pulling from res field
> that too would compel application to interpret res field differently, and it
> seems more painstaking than the former option.
> To keep uniformity, even with aio, zone-relative offset is returned.
I am really not a fan of this, to say the least. The input is byte offset, the
output is 512B relative sector count... Arg... We really cannot do better than
that ?
At the very least, byte relative offset ? The main reason is that this is
_somewhat_ acceptable for raw block device accesses since the "sector"
abstraction has a clear meaning, but once we add iomap/zonefs async zone append
support, we really will want to have byte unit as the interface is regular
files, not block device file. We could argue that 512B sector unit is still
around even for files (e.g. block counts in file stat). Bu the different unit
for input and output of one operation is really ugly. This is not nice for the user.
>
> Append using io_uring fixed-buffer --->
> This is flagged as not-supported at the moment. Reason being, for fixed-buffer
> io-uring sends iov_iter of bvec type. But current append-infra in block-layer
> does not support such iov_iter.
>
> Block IO vs File IO --->
> For now, the user zone-append interface is supported only for zoned-block-device.
> Regular files/block-devices are not supported. Regular file-system (e.g. F2FS)
> will not need this anyway, because zone peculiarities are abstracted within FS.
> At this point, ZoneFS also likes to use append implicitly rather than explicitly.
> But if/when ZoneFS starts supporting explicit/on-demand zone-append, the check
> allowing-only-block-device should be changed.
Sure, but I think the interface is still a problem. I am not super happy about
the 512B sector unit. Zonefs will be the only file system that will be impacted
since other normal POSIX file system will not have zone append interface for
users. So this is a limited problem. Still, even for raw block device files
accesses, POSIX system calls use Byte unit everywhere. Let's try to use that.
For aio, it is easy since res2 is unsigned long long. For io_uring, as discussed
already, we can still 8 bits from the cqe res. All you need is to add a small
helper function in userspace iouring.h to simplify the work of the application
to get that result.
>
> Semantics --->
> Zone-append, by its nature, may perform write on a different location than what
> was specified. It does not fit into POSIX, and trying to fit may just undermine
> its benefit. It may be better to keep semantics as close to zone-append as
> possible i.e. specify zone-start location, and obtain the actual-write location
> post completion. Towards that goal, existing async APIs seem to fit fine.
> Async APIs (uring, linux aio) do not work on implicit write-pointer and demand
> explicit write offset (which is what we need for append). Neither write-pointer
What do you mean by "implicit write pointer" ? Are you referring to the behavior
of AIO write with a block device file open with O_APPEND ? The yes, it does not
work. But that is perfectly fine for regular files, that is for zonefs.
I would prefer that this paragraph simply state the semantic that is implemented
first. Then explain why the choice. But first, clarify how the API works, what
is allowed, what's not etc. That will also simplify reviewing the code as one
can then check the code against the goal.
> is taken as input, nor it is updated on completion. And there is a clear way to
> get zone-append result. Zone-aware applications while using these async APIs
> can be fine with, for the lack of better word, zone-append semantics itself.
>
> Sync APIs work with implicit write-pointer (at least few of those), and there is
> no way to obtain zone-append result, making it hard for user-space zone-append.
Sync API are executed under inode lock, at least for regular files. So there is
absolutely no problem to use zone append. zonefs does it already. The problem is
the lack of locking for block device file.
>
> Tests --->
> Using new interface in fio (uring and libaio engine) by extending zbd tests
> for zone-append: https://github.com/axboe/fio/pull/1026
>
> Changes since v1:
> - No new opcodes in uring or aio. Use RWF_ZONE_APPEND flag instead.
> - linux-aio changes vanish because of no new opcode
> - Fixed the overflow and other issues mentioned by Damien
> - Simplified uring support code, fixed the issues mentioned by Pavel
> - Added error checks
>
> Kanchan Joshi (1):
> fs,block: Introduce RWF_ZONE_APPEND and handling in direct IO path
>
> Selvakumar S (1):
> io_uring: add support for zone-append
>
> fs/block_dev.c | 28 ++++++++++++++++++++++++----
> fs/io_uring.c | 32 ++++++++++++++++++++++++++++++--
> include/linux/fs.h | 9 +++++++++
> include/uapi/linux/fs.h | 5 ++++-
> 4 files changed, 67 insertions(+), 7 deletions(-)
>
--
Damien Le Moal
Western Digital Research
