On 4/23/24 14:08, Kevin Wolf wrote:
Am 22.04.2024 um 20:27 hat Jens Axboe geschrieben:
On 4/7/24 7:03 PM, Ming Lei wrote:
SQE group is defined as one chain of SQEs starting with the first sqe that
has IOSQE_EXT_SQE_GROUP set, and ending with the first subsequent sqe that
doesn't have it set, and it is similar with chain of linked sqes.
The 1st SQE is group leader, and the other SQEs are group member. The group
leader is always freed after all members are completed. Group members
aren't submitted until the group leader is completed, and there isn't any
dependency among group members, and IOSQE_IO_LINK can't be set for group
members, same with IOSQE_IO_DRAIN.
Typically the group leader provides or makes resource, and the other members
consume the resource, such as scenario of multiple backup, the 1st SQE is to
read data from source file into fixed buffer, the other SQEs write data from
the same buffer into other destination files. SQE group provides very
efficient way to complete this task: 1) fs write SQEs and fs read SQE can be
submitted in single syscall, no need to submit fs read SQE first, and wait
until read SQE is completed, 2) no need to link all write SQEs together, then
write SQEs can be submitted to files concurrently. Meantime application is
simplified a lot in this way.
Another use case is to for supporting generic device zero copy:
- the lead SQE is for providing device buffer, which is owned by device or
kernel, can't be cross userspace, otherwise easy to cause leak for devil
application or panic
- member SQEs reads or writes concurrently against the buffer provided by lead
SQE
In concept, this looks very similar to "sqe bundles" that I played with
in the past:
https://git.kernel.dk/cgit/linux/log/?h=io_uring-bundle
Didn't look too closely yet at the implementation, but in spirit it's
about the same in that the first entry is processed first, and there's
no ordering implied between the test of the members of the bundle /
group.
When I first read this patch, I wondered if it wouldn't make sense to
allow linking a group with subsequent requests, e.g. first having a few
requests that run in parallel and once all of them have completed
continue with the next linked one sequentially.
For SQE bundles, you reused the LINK flag, which doesn't easily allow
this. Ming's patch uses a new flag for groups, so the interface would be
more obvious, you simply set the LINK flag on the last member of the
group (or on the leader, doesn't really matter). Of course, this doesn't
mean it has to be implemented now, but there is a clear way forward if
it's wanted.
Putting zc aside, links, graphs, groups, it all sounds interesting in
concept but let's not fool anyone, all the different ordering
relationships between requests proved to be a bad idea.
I can complaint for long, error handling is miserable, user handling
resubmitting a part of a link is horrible, the concept of errors is
hard coded (time to appreciate "beautifulness" of IOSQE_IO_HARDLINK
and the MSG_WAITALL workaround). The handling and workarounds are
leaking into generic paths, e.g. we can't init files when it's the most
convenient. For cancellation we're walking links, which need more care
than just looking at a request (is cancellation by user_data of a
"linked" to a group request even supported?). The list goes on
And what does it achieve? The infra has matured since early days,
it saves user-kernel transitions at best but not context switching
overhead, and not even that if you do wait(1) and happen to catch
middle CQEs. And it disables LAZY_WAKE, so CQ side batching with
timers and what not is effectively useless with links.
So, please, please! instead of trying to invent a new uber scheme
of request linking, which surely wouldn't step on same problems
over and over again, and would definitely be destined to overshadow
all previous attempts and finally conquer the world, let's rather
focus on minimasing the damage from this patchset's zero copy if
it's going to be taken.
Piggy backing bits on top of links should be just fine. May help
to save space in io_kiocb by unionising with links. And half
of the overhead (including completely destroying all inlining
in the submission patch) can be mitigated by folding it into
REQ_F_LINK handling and generally borowing the code structure
for it in the submission path.
The part that looks a bit arbitrary in Ming's patch is that the group
leader is always completed before the rest starts. It makes perfect
sense in the context that this series is really after (enabling zero
copy for ublk), but it doesn't really allow the case you mention in the
SQE bundle commit message, running everything in parallel and getting a
single CQE for the whole group.
I suppose you could hack around the sequential nature of the first
request by using an extra NOP as the group leader - which isn't any
worse than having an IORING_OP_BUNDLE really, just looks a bit odd - but
the group completion would still be missing. (Of course, removing the
sequential first operation would mean that ublk wouldn't have the buffer
ready any more when the other requests try to use it, so that would
defeat the purpose of the series...)
I wonder if we can still combine both approaches and create some
generally useful infrastructure and not something where it's visible
that it was designed mostly for ublk's special case and other use cases
just happened to be enabled as a side effect.
--
Pavel Begunkov
