Re: [PATCH V3 5/9] io_uring: support SQE group

[Ming Lei <ming.lei@redhat.com>]

On Sun, Jun 16, 2024 at 07:14:37PM +0100, Pavel Begunkov wrote:
> On 6/11/24 14:32, Ming Lei wrote:
> > On Mon, Jun 10, 2024 at 02:55:22AM +0100, Pavel Begunkov wrote:
> > > On 5/21/24 03:58, Ming Lei wrote:
> > > > On Sat, May 11, 2024 at 08:12:08AM +0800, Ming Lei wrote:
> > > > > SQE group is defined as one chain of SQEs starting with the first SQE that
> > > > > has IOSQE_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.
> > > > > 
> > > > > Not like linked SQEs, each sqe is issued after the previous one is completed.
> > > > > All SQEs in one group are submitted in parallel, so there isn't any dependency
> > > > > among SQEs in one group.
> > > > > 
> > > > > The 1st SQE is group leader, and the other SQEs are group member. The whole
> > > > > group share single IOSQE_IO_LINK and IOSQE_IO_DRAIN from group leader, and
> > > > > the two flags are ignored for group members.
> > > > > 
> > > > > When the group is in one link chain, this group isn't submitted until the
> > > > > previous SQE or group is completed. And the following SQE or group can't
> > > > > be started if this group isn't completed. Failure from any group member will
> > > > > fail the group leader, then the link chain can be terminated.
> > > > > 
> > > > > When IOSQE_IO_DRAIN is set for group leader, all requests in this group and
> > > > > previous requests submitted are drained. Given IOSQE_IO_DRAIN can be set for
> > > > > group leader only, we respect IO_DRAIN by always completing group leader as
> > > > > the last one in the group.
> > > > > 
> > > > > Working together with IOSQE_IO_LINK, SQE group provides flexible way to
> > > > > support N:M dependency, such as:
> > > > > 
> > > > > - group A is chained with group B together
> > > > > - group A has N SQEs
> > > > > - group B has M SQEs
> > > > > 
> > > > > then M SQEs in group B depend on N SQEs in group A.
> > > > > 
> > > > > N:M dependency can support some interesting use cases in efficient way:
> > > > > 
> > > > > 1) read from multiple files, then write the read data into single file
> > > > > 
> > > > > 2) read from single file, and write the read data into multiple files
> > > > > 
> > > > > 3) write same data into multiple files, and read data from multiple files and
> > > > > compare if correct data is written
> > > > > 
> > > > > Also IOSQE_SQE_GROUP takes the last bit in sqe->flags, but we still can
> > > > > extend sqe->flags with one uring context flag, such as use __pad3 for
> > > > > non-uring_cmd OPs and part of uring_cmd_flags for uring_cmd OP.
> > > > > 
> > > > > Suggested-by: Kevin Wolf <kwolf@redhat.com>
> > > > > Signed-off-by: Ming Lei <ming.lei@redhat.com>
> > > > 
> > > > BTW, I wrote one link-grp-cp.c liburing/example which is based on sqe group,
> > > > and keep QD not changed, just re-organize IOs in the following ways:
> > > > 
> > > > - each group have 4 READ IOs, linked by one single write IO for writing
> > > >     the read data in sqe group to destination file
> > > 
> > > IIUC it's comparing 1 large write request with 4 small, and
> > 
> > It is actually reasonable from storage device viewpoint, concurrent
> > small READs are often fast than single big READ, but concurrent small
> > writes are usually slower.
> 
> It is, but that doesn't make the comparison apple to apple.
> Even what I described, even though it's better (same number
> of syscalls but better parallelism as you don't block next
> batch of reads by writes), you can argues it's not a
> completely fair comparison either since needs different number
> of buffers, etc.
> 
> > > it's not exactly anything close to fair. And you can do same
> > > in userspace (without links). And having control in userspace
> > 
> > No, you can't do it with single syscall.
> 
> That's called you _can_ do it. And syscalls is not everything,

For ublk, syscall does mean something, because each ublk IO is
handled by io_uring, if more syscalls are introduced for each ublk IO,
performance definitely degrades a lot because IOPS can be million level.
Now syscall PTI overhead does make difference, please see:

https://lwn.net/Articles/752587/

> context switching turned to be a bigger problem, and to execute
> links it does exactly that.

If that is true, IO_LINK shouldn't have been needed, cause you can model
dependency via io_uring syscall, unfortunately it isn't true. IO_LINK not
only simplifies application programming, but also avoids extra syscall.

If you compare io_uring-cp.c(282 LOC) with link-cp.c(193 LOC) in
liburing/examples, you can see io_uring-cp.c is more complicated. Adding
one extra syscall(wait point) makes application hard to write, especially in
modern async/.await programming environment.


Thanks,
Ming