[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Xen-devel] [PATCH] netif.h: Document xen-net{back, front} multi-queue feature

On 19/02/14 12:02, Ian Campbell wrote:
On Wed, 2014-02-19 at 11:47 +0000, Andrew Bennieston wrote:
On 19/02/14 11:06, Ian Campbell wrote:
On Mon, 2014-02-17 at 18:01 +0000, Andrew J. Bennieston wrote:
From: "Andrew J. Bennieston" <andrew.bennieston@xxxxxxxxxx>

Document the multi-queue feature in terms of XenStore keys to be written
by the backend and by the frontend.

Signed-off-by: Andrew J. Bennieston <andrew.bennieston@xxxxxxxxxx>
   xen/include/public/io/netif.h |   21 +++++++++++++++++++++
   1 file changed, 21 insertions(+)

diff --git a/xen/include/public/io/netif.h b/xen/include/public/io/netif.h
index d7fb771..90be2fc 100644
--- a/xen/include/public/io/netif.h
+++ b/xen/include/public/io/netif.h
@@ -69,6 +69,27 @@

+ * Multiple transmit and receive queues:
+ * If supported, the backend will write "multi-queue-max-queues" and set its
+ * value to the maximum supported number of queues.
+ * Frontends that are aware of this feature and wish to use it can write the
+ * key "multi-queue-num-queues", set to the number they wish to use.
+ *
+ * Queues replicate the shared rings and event channels, and
+ * "feature-split-event-channels" is required when using multiple queues.
+ *
+ * For frontends requesting just one queue, the usual event-channel and
+ * ring-ref keys are written as before, simplifying the backend processing
+ * to avoid distinguishing between a frontend that doesn't understand the
+ * multi-queue feature, and one that does, but requested only one queue.
+ *
+ * Frontends requesting two or more queues must not write the toplevel
+ * event-channel and ring-ref keys, instead writing them under sub-keys having
+ * the name "queue-N" where N is the integer ID of the queue for which those
+ * keys belong. Queues are indexed from zero.

If "feature-split-event-channels" is required then I think what should
be written is queue-N/event-channel-{tx,rx} and
queue-N/{tx,rx}-ring-ref, rather than queue-N/{event-channel,ring-ref}
as the final paragraph sort of implies?
I can change the wording to make this more clear.


(what a shame we have event-channel-DIR and DIR-ring-ref, oh well!)

Is it required to have the same number of RX and TX queues?
Strictly speaking, no. But the implementation assumes this to be the
case. Since the code already sets up one pair, simply looping over this
sufficient times to create N pairs was a fairly sane approach to this.
In practice, if you have an asymmetry between RX and TX queues, you will
end up hitting a bottleneck sooner in one direction than the other,
which seems impractical.

OK. So we should either mandate that there will always be the same
number, if that is going to be the case, or we should design the
xenstore layout to be extensible if not.

It sounds reasonable to me to mandate they be the same, but I don't
really know.

Are there any other properties/behaviours which should be documented,
e.g. relating to the selection of which queue to use for a given frame
(on either TX or RX)? If not and it is up to the relevant end to do what
it wants then I think it would be useful to say so.
There are no other requirements. The current implementation will
transmit anything it cannot hash sensibly on queue 0, but this is an
arbitrary choice (albeit a sensible one, since queue 0 should always
exist). I'll document this.

What is this undocumented/unnegotiated "hash" you speak of ;-)
Each end must decide how to split packets across queues. This decision is made by the transmitting guest, and the receiving guest will simply take packets on any queue and deal with them as before. This is why it is unnegotiated; the two ends could be using completely different mechanisms for this, or could always choose a single queue, with no adverse effect on anything except performance (i.e. if they do this badly, they will not benefit from the full value of multiple queues).

These properties should probably be documented, though. I'll add a paragraph to the effect of the one above. A complication arises if we wish to support Windows frontends in a logical manner, since Windows expects to be able to control the mapping of packets to RX queues. This means that for a Windows guest, we may need to write additional Xenstore keys to specify these parameters to the backend. However, that is not considered here because

a] the default behaviour of decoupling the TX and RX queue mappings is not affected by this; a Windows frontend would specifically request an
algorithm, while other frontends may remain silent about this.

b] Windows also wants the hash value (used to select the queue, via the Toeplitz hash) to be transmitted along with each packet. This would require either an extension to the ring protocol to pass this value in an extra slot, or some other mechanism - either of which would require separate negotiation anyway.

For these reasons, the default behaviour is to _not_ negotiate any hashing between the back- and frontends, leaving each free to do the most appropriate thing. I'd like to avoid writing unnecessary stuff into Xenstore, preferring to have a sensible default behaviour that can be overridden for the few specific cases that require something different.

As such, I think the docs should, for now, say something along the lines of:
  Mapping of packets to queues is considered to be a function of the
  transmitting system (backend or frontend) and is not negotiated
  between the two. Guests are free to transmit packets on any queue
  they choose, provided it has been set up correctly.


Xen-devel mailing list



Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.