Thread (18 messages) flat view 18 messages, 3 authors, 2014-08-08

Re: [PATCH net-next 0/2] xen-netback: Changes around carrier handling

From: Zoltan Kiss <hidden>
Date: 2014-08-07 16:49:42
Also in: lkml

On 06/08/14 00:07, David Miller wrote:
From: Zoltan Kiss <redacted>
Date: Mon, 4 Aug 2014 16:20:56 +0100
quoted
This series starts using carrier off as a way to purge packets when the guest is
not able (or willing) to receive them. It is a much faster way to get rid of
packets waiting for an overwhelmed guest.
The first patch changes current netback code where it relies currently on
netif_carrier_ok.
The second turns off the carrier if the guest times out on a queue, and only
turn it on again if that queue (or queues) resurrects.

Signed-off-by: Zoltan Kiss <redacted>
Signed-off-by: David Vrabel <redacted>
Applied, but I have some reservations:

1) This is starting to bleed what is normally qdisc type policy into the
    driver.

2) There are other drivers that could run into this kind of situation and
    have similar concerns, therefore we should make sure we have a consistent
    approach that such entities use to handle this problem.

Part of the problem is that netif_carrier_off() only partially mimicks
the situation.  It expresses the "transmitter is down so packets
aren't going onto the wire" part, which keeps the watchdog from
spitting out log messages ever time it fires.  But it doesn't deal
with packet freeing policy meanwhile, which I guess is the part that
this patch series is largely trying to address.
David Vrabel pointed out an important question in a reply to the 
previous version of this series: this patch deschedule NAPI if the 
carrier goes down. The backend doesn't receive packets from the guest. 
DavidVr and others said we shouldn't do this, the guest should be able 
to transmit even if it's not able/willing to receive. Other drivers 
doesn't deschedule NAPI at carrier off as well, however the "carrier 
off" information comes from the hardware, not from an untrusted guest 
who is not posting buffers on the receive ring.
I don't have any good argument why I did it the current way, other than 
a hunch that it feels more natural.
David, do you have an opinion on that?

Zoli
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help