Re: [PATCH V5 4/8] io_uring: support SQE group

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

On Tue, Sep 10, 2024 at 09:31:45PM +0100, Pavel Begunkov wrote:
> On 9/10/24 16:04, Ming Lei wrote:
> > On Tue, Sep 10, 2024 at 02:12:53PM +0100, Pavel Begunkov wrote:
> > > On 9/7/24 10:36, Ming Lei wrote:
> > > ...
> > > > > > Wrt. ublk, group provides zero copy, and the ublk io(group) is generic
> > > > > > IO, sometime IO_LINK is really needed & helpful, such as in ublk-nbd,
> > > > > > send(tcp) requests need to be linked & zc. And we shouldn't limit IO_LINK
> > > > > > for generic io_uring IO.
> > > > > > 
> > > > > > > from nuances as such, which would be quite hard to track, the semantics
> > > > > > > of IOSQE_CQE_SKIP_SUCCESS is unclear.
> > > > > > 
> > > > > > IO group just follows every normal request.
> > > > > 
> > > > > It tries to mimic but groups don't and essentially can't do it the
> > > > > same way, at least in some aspects. E.g. IOSQE_CQE_SKIP_SUCCESS
> > > > > usually means that all following will be silenced. What if a
> > > > > member is CQE_SKIP, should it stop the leader from posting a CQE?
> > > > > And whatever the answer is, it'll be different from the link's
> > > > > behaviour.
> > > > 
> > > > Here it looks easier than link's:
> > > > 
> > > > - only leader's IOSQE_CQE_SKIP_SUCCESS follows linked request's rule
> > > > - all members just respects the flag for its own, and not related with
> > > > leader's
> > > > 
> > > > > 
> > > > > Regardless, let's forbid IOSQE_CQE_SKIP_SUCCESS and linked timeouts
> > > > > for groups, that can be discussed afterwards.
> > > > 
> > > > It should easy to forbid IOSQE_CQE_SKIP_SUCCESS which is per-sqe, will do
> > > > it in V6.
> > > > 
> > > > I am not sure if it is easy to disallow IORING_OP_LINK_TIMEOUT, which
> > > > covers all linked sqes, and group leader could be just one of them.
> > > > Can you share any idea about the implementation to forbid LINK_TIMEOUT
> > > > for sqe group?
> > > 
> > > diff --git a/io_uring/timeout.c b/io_uring/timeout.c
> > > index 671d6093bf36..83b5fd64b4e9 100644
> > > --- a/io_uring/timeout.c
> > > +++ b/io_uring/timeout.c
> > > @@ -542,6 +542,9 @@ static int __io_timeout_prep(struct io_kiocb *req,
> > >   	data->mode = io_translate_timeout_mode(flags);
> > >   	hrtimer_init(&data->timer, io_timeout_get_clock(data), data->mode);
> > > +	if (is_timeout_link && req->ctx->submit_state.group.head)
> > > +		return -EINVAL;
> > > +
> > >   	if (is_timeout_link) {
> > >   		struct io_submit_link *link = &req->ctx->submit_state.link;
> > > 
> > > This should do, they already look into the ctx's link list. Just move
> > > it into the "if (is_timeout_link)" block.
> > 
> > OK.
> > 
> > > 
> > > 
> > > > > > 1) fail in linked chain
> > > > > > - follows IO_LINK's behavior since io_fail_links() covers io group
> > > > > > 
> > > > > > 2) otherwise
> > > > > > - just respect IOSQE_CQE_SKIP_SUCCESS
> > > > > > 
> > > > > > > And also it doen't work with IORING_OP_LINK_TIMEOUT.
> > > > > > 
> > > > > > REQ_F_LINK_TIMEOUT can work on whole group(or group leader) only, and I
> > > > > > will document it in V6.
> > > > > 
> > > > > It would still be troublesome. When a linked timeout fires it searches
> > > > > for the request it's attached to and cancels it, however, group leaders
> > > > > that queued up their members are discoverable. But let's say you can find
> > > > > them in some way, then the only sensbile thing to do is cancel members,
> > > > > which should be doable by checking req->grp_leader, but might be easier
> > > > > to leave it to follow up patches.
> > > > 
> > > > We have changed sqe group to start queuing members after leader is
> > > > completed. link timeout will cancel leader with all its members via
> > > > leader->grp_link, this behavior should respect IORING_OP_LINK_TIMEOUT
> > > > completely.
> > > > 
> > > > Please see io_fail_links() and io_cancel_group_members().
> > > > 
> > > > > 
> > > > > 
> > > > > > > > +
> > > > > > > > +		lead->grp_refs += 1;
> > > > > > > > +		group->last->grp_link = req;
> > > > > > > > +		group->last = req;
> > > > > > > > +
> > > > > > > > +		if (req->flags & REQ_F_SQE_GROUP)
> > > > > > > > +			return NULL;
> > > > > > > > +
> > > > > > > > +		req->grp_link = NULL;
> > > > > > > > +		req->flags |= REQ_F_SQE_GROUP;
> > > > > > > > +		group->head = NULL;
> > > > > > > > +		if (lead->flags & REQ_F_FAIL) {
> > > > > > > > +			io_queue_sqe_fallback(lead);
> > > > > > > 
> > > > > > > Let's say the group was in the middle of a link, it'll
> > > > > > > complete that group and continue with assembling / executing
> > > > > > > the link when it should've failed it and honoured the
> > > > > > > request order.
> > > > > > 
> > > > > > OK, here we can simply remove the above two lines, and link submit
> > > > > > state can handle this failure in link chain.
> > > > > 
> > > > > If you just delete then nobody would check for REQ_F_FAIL and
> > > > > fail the request.
> > > > 
> > > > io_link_assembling() & io_link_sqe() checks for REQ_F_FAIL and call
> > > > io_queue_sqe_fallback() either if it is in link chain or
> > > > not.
> > > 
> > > The case we're talking about is failing a group, which is
> > > also in the middle of a link.
> > > 
> > > LINK_HEAD -> {GROUP_LEAD, GROUP_MEMBER}
> > > 
> > > Let's say GROUP_MEMBER fails and sets REQ_F_FAIL to the lead,
> > > then in v5 does:
> > > 
> > > if (lead->flags & REQ_F_FAIL) {
> > > 	io_queue_sqe_fallback(lead);
> > > 	return NULL;
> > > }
> > > 
> > > In which case it posts cqes for GROUP_LEAD and GROUP_MEMBER,
> > > and then try to execute LINK_HEAD (without failing it), which
> > > is wrong. So first we need:
> > > 
> > > if (state.linked_link.head)
> > > 	req_fail_link_node(state.linked_link.head);
> > 
> > For group leader, link advancing is always done via io_queue_next(), in
> > which io_disarm_next() is called for failing the whole remained link
> > if the current request is marked as FAIL.
> > 
> > > 
> > > And then we can't just remove io_queue_sqe_fallback(), because
> > > when a group is not linked there would be no io_link_sqe()
> > > to fail it. You can do:
> > 
> > If one request in group is marked as FAIL, io_link_assembling()
> > will return true, and io_link_sqe() will fail it.
> 
> Hmm, you're right, even though it's not a great way of doing it,
> i.e. pushing a req into io_link_sqe() even when linking has never
> been requested, but that's fine. I can drop a quick patch on
> top if it bothers me.

Yeah, it isn't very readable, but following the original logic.

Anyway, I will comment that non-linked request is covered by
the code block.


Thanks,
Ming