All of
 help / color / mirror / Atom feed
From: Samuel Thibault <>
To: Juergen Gross <>
Subject: Re: [PATCH v3] xenbus: support large messages
Date: Wed, 15 Sep 2021 16:30:48 +0200	[thread overview]
Message-ID: <20210915143048.lluaekufonze6pic@begin> (raw)
In-Reply-To: <>


Juergen Gross, le mer. 15 sept. 2021 16:08:15 +0200, a ecrit:
> +    while (off != len)
> +    {
> +        wait_event(xb_waitq, xenstore_buf->rsp_prod != xenstore_buf->rsp_cons);
> +
> +        prod = xenstore_buf->rsp_prod;
> +        cons = xenstore_buf->rsp_cons;
> +        DEBUG("Rsp_cons %d, rsp_prod %d.\n", cons, prod);
> +        size = min(len - off, prod - cons);
> +
> +        rmb();   /* Make sure data read from ring is ordered with rsp_prod. */
> +        memcpy_from_ring(xenstore_buf->rsp, buf + off,
> +                         MASK_XENSTORE_IDX(cons), size);
> +        off += size;
> +        xenstore_buf->rsp_cons += size;
> +        wmb();
> +        if (xenstore_buf->rsp_prod - cons >= XENSTORE_RING_SIZE)
> +            notify_remote_via_evtchn(xenbus_evtchn);
> +    }

Reading it again, I'm still not convinced we covered all cases. There is
at least one thing: Linux does

	memcpy(data, src, avail);
	/* Other side must not see free space until we've copied out */
	intf->rsp_cons += avail;

which makes sense to me: we don't want the compiler or anything to
reorder the write to rsp_cons respectively to the memcpy. So I believe
we also need a full barrier before 

	xenstore_buf->rsp_cons += size;

in mini-os.

Then there is

	xenstore_buf->rsp_cons += size;
	if (xenstore_buf->rsp_prod - cons >= XENSTORE_RING_SIZE)

The Linux code does

	intf->rsp_cons += avail;
	/* Implies mb(): other side will see the updated consumer. */
	if (intf->rsp_prod - cons >= XENSTORE_RING_SIZE)

(Note that the "Implies mb()" comment is about the notify_remote_via_evtchn call)

I believe the Linux code itself is missing a full barrier here: before
reading intf->rsp_prod, we want to make sure that our update of rsp_cons
is seen by the producer. Otherwise it may happen that the producer
doesn't see the rsp_cons and updates its rsp_prod (filling the ring
and going to sleep), and the consumer does not see the rsp_prod update
either, and thus we miss a notification and the producer stays stuck.

Extremely rare scenario etc. but I believe we do want a full barrier
between rsp_cons += and the if, both in mini-os and Linux.


      reply	other threads:[~2021-09-15 14:31 UTC|newest]

Thread overview: 2+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2021-09-15 14:08 [PATCH v3] xenbus: support large messages Juergen Gross
2021-09-15 14:30 ` Samuel Thibault [this message]

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:

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20210915143048.lluaekufonze6pic@begin \ \ \ \ \ \

* 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 an external index of several public inboxes,
see mirroring instructions on how to clone and mirror
all data and code used by this external index.