From: Paolo Bonzini <pbonzini@redhat.com>
To: Stefan Hajnoczi <stefanha@redhat.com>
Cc: Kevin Wolf <kwolf@redhat.com>, Fam Zheng <fam@euphon.net>,
qemu-devel <qemu-devel@nongnu.org>,
qemu-block@nongnu.org, Max Reitz <mreitz@redhat.com>
Subject: Re: [PATCH] util/async: make bh_aio_poll() O(1)
Date: Wed, 19 Feb 2020 20:05:12 +0100 [thread overview]
Message-ID: <CABgObfYY+tr3xuoBFF3Q3zn2khrZCSfGK-P63NYJ_phS1vCp5w@mail.gmail.com> (raw)
In-Reply-To: <20200219165425.GB1089598@stefanha-x1.localdomain>
[-- Attachment #1: Type: text/plain, Size: 3995 bytes --]
Il mer 19 feb 2020, 18:58 Stefan Hajnoczi <stefanha@redhat.com> ha scritto:
> On Wed, Feb 19, 2020 at 12:09:48PM +0100, Paolo Bonzini wrote:
> > Really a great idea, though I have some remarks on the implementation
> below.
> >
> > On 19/02/20 11:00, Stefan Hajnoczi wrote:
> > > + * Each aio_bh_poll() call carves off a slice of the BH list. This
> way newly
> > > + * scheduled BHs are not processed until the next aio_bh_poll()
> call. This
> > > + * concept extends to nested aio_bh_poll() calls because slices are
> chained
> > > + * together.
> >
> > This is the tricky part so I would expand a bit on why it's needed:
> >
> > /*
> > * Each aio_bh_poll() call carves off a slice of the BH list, so that
> > * newly scheduled BHs are not processed until the next aio_bh_poll()
> > * call. All active aio_bh_poll() calls chained their slices together
> > * in a list, so that nested aio_bh_poll() calls process all scheduled
> > * bottom halves.
> > */
>
> Thanks, will fix in v2.
>
> > > +struct BHListSlice {
> > > + QEMUBH *first_bh;
> > > + BHListSlice *next;
> > > +};
> > > +
> >
> > Using QLIST and QSLIST removes the need to create your own lists, since
> > you can use QSLIST_MOVE_ATOMIC and QSLIST_INSERT_HEAD_ATOMIC. For
> example:
> >
> > struct BHListSlice {
> > QSLIST_HEAD(, QEMUBH) first_bh;
> > QLIST_ENTRY(BHListSlice) next;
> > };
> >
> > ...
> >
> > QSLIST_HEAD(, QEMUBH) active_bh;
> > QLIST_HEAD(, BHListSlice) bh_list;
>
> I thought about this but chose the explicit tail pointer approach
> because it lets aio_compute_timeout() and aio_ctx_check() iterate over
> both the active BH list and slices in a single for loop :). But
> thinking about it more, maybe it can still be done by replacing
> active_bh with a permanently present first BHListSlice element.
>
Probably not so easy since the idea was to empty the slices list entirely
via the FIFO order.
But since you are rewriting everything anyway, can you add a flag for
whether there are any non-idle bottom halves anywhere in the list? It need
not be computed perfectly, because any non-idle bh will cause the idle
bottom halves to be triggered as well; you can just set in qemu_bh_schedule
and clear it at the end of aio_bh_poll.
Then if there is any active bh or slice you consult the flag and use it to
return the timeout, which will be either 0 or 10ms depending on the flag.
That's truly O(1). (More precisely, this patch goes from O(#created-bh) to
O(#scheduled-bh), which of course is optimal for aio_bh_poll but not for
aio_compute_timeout; making timeout computation O(1) can lower latency a
bit by decreasing the constant factor).
Paolo
> >
> > Related to this, in the aio_bh_poll() loop:
> >
> > for (s = ctx->bh_list.next; s; s = s->next) {
> > }
> >
> > You can actually do the removal inside the loop. This is slightly more
> > efficient since you can remove slices early from the nested
> > aio_bh_poll(). Not that it's relevant for performance, but I think the
> > FIFO order for slices is also more intuitive than LIFO.
> >
> > Putting this idea together with the QLIST one, you would get:
> >
> > /*
> > * If a bottom half causes a recursive call, this slice will be
> > * removed by the nested aio_bh_poll().
> > */
> > QSLIST_MOVE_ATOMIC(&slice.first_bh, ctx->active_bh);
> > QLIST_INSERT_TAIL(&ctx->bh_list, slice);
> > while ((s = QLIST_FIRST(&ctx->bh_list)) {
> > while ((bh = aio_bh_dequeue(&s, &flags))) {
> > }
> > QLIST_REMOVE_HEAD(s, next);
> > }
>
> Cool, reusing "queue.h" is nice.
>
> >
> > > /* Multiple occurrences of aio_bh_poll cannot be called concurrently.
> > > * The count in ctx->list_lock is incremented before the call, and is
> > > * not affected by the call.
> >
> > The second sentence is now stale.
>
> Thanks, will fix in v2.
>
> Stefan
>
[-- Attachment #2: Type: text/html, Size: 5256 bytes --]
next prev parent reply other threads:[~2020-02-19 19:06 UTC|newest]
Thread overview: 5+ messages / expand[flat|nested] mbox.gz Atom feed top
2020-02-19 10:00 [PATCH] util/async: make bh_aio_poll() O(1) Stefan Hajnoczi
2020-02-19 11:09 ` Paolo Bonzini
2020-02-19 16:54 ` Stefan Hajnoczi
2020-02-19 19:05 ` Paolo Bonzini [this message]
2020-02-21 13:47 ` Stefan Hajnoczi
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=CABgObfYY+tr3xuoBFF3Q3zn2khrZCSfGK-P63NYJ_phS1vCp5w@mail.gmail.com \
--to=pbonzini@redhat.com \
--cc=fam@euphon.net \
--cc=kwolf@redhat.com \
--cc=mreitz@redhat.com \
--cc=qemu-block@nongnu.org \
--cc=qemu-devel@nongnu.org \
--cc=stefanha@redhat.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).