RE: [PATCH] hv: do not lose pending heartbeat vmbus packets

From: Long Li
Date: Fri Sep 30 2016 - 17:04:48 EST




> -----Original Message-----
> From: Vitaly Kuznetsov [mailto:vkuznets@xxxxxxxxxx]
> Sent: Thursday, September 29, 2016 2:22 AM
> To: KY Srinivasan <kys@xxxxxxxxxxxxx>; Long Li <longli@xxxxxxxxxxxxx>
> Cc: Haiyang Zhang <haiyangz@xxxxxxxxxxxxx>;
> devel@xxxxxxxxxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx
> Subject: Re: [PATCH] hv: do not lose pending heartbeat vmbus packets
>
> Long Li <longli@xxxxxxxxxxxxxxxxxxxxxx> writes:
>
> > From: Long Li <longli@xxxxxxxxxxxxx>
> >
> > The host keeps sending heartbeat packets independent of guest
> responding to them. In some situations, there might be multiple heartbeat
> packets pending in the ring buffer. Don't lose them, read them all.
> >
> > Signed-off-by: Long Li <longli@xxxxxxxxxxxxx>
>
> Long, K. Y.,
>
> it seems this patch didn't make it to char-misc tree and it looks like an
> important fix. A couple of nitpicks below,
>
> > ---
> > drivers/hv/hv_util.c | 10 +++++++---
> > 1 file changed, 7 insertions(+), 3 deletions(-)
> >
> > diff --git a/drivers/hv/hv_util.c b/drivers/hv/hv_util.c index
> > d5acaa2..9dc6372 100644
> > --- a/drivers/hv/hv_util.c
> > +++ b/drivers/hv/hv_util.c
> > @@ -283,10 +283,14 @@ static void heartbeat_onchannelcallback(void
> *context)
> > u8 *hbeat_txf_buf = util_heartbeat.recv_buffer;
> > struct icmsg_negotiate *negop = NULL;
> >
> > - vmbus_recvpacket(channel, hbeat_txf_buf,
> > - PAGE_SIZE, &recvlen, &requestid);
> > + while (1) {
> > +
> > + vmbus_recvpacket(channel, hbeat_txf_buf,
> > + PAGE_SIZE, &recvlen, &requestid);
>
> We should check vmbus_recvpacket() return value as well. E.g.
> hv_ringbuffer_read() may return -EAGAIN in case we didn't receive the
> whole packet (and we do this check in other drivers, see
> storvsc_on_channel_callback() for example).

I agree with you, we should check for -EAGAIN. This should also be done in storvsc_on_channel_callback.

I think the chance of hv_ringbuffer_read() returning -EAGAIN is almost zero. Because read_index and write_index are updated after the whole packet is written to the ring buffer, and protected by memory barriers. So getting a partial read is impossible, unless the host is doing something wrong.

Checking for recvlen is safe, because it's always set to 0 at the beginning of hv_ringbuffer_read().

Anyway, we should check for -EAGAIN for all hyperv drivers on read. I think this is a separate issue on how we deal with a buggy host. Will send another set of patches .

>
> > +
> > + if (!recvlen)
>
> so this should be 'if (ret || !recvlen)'
>
> > + break;
> >
> > - if (recvlen > 0) {
> > icmsghdrp = (struct icmsg_hdr *)&hbeat_txf_buf[
> > sizeof(struct vmbuspipe_hdr)];
>
> --
> Vitaly