diff options
author | Andrew J. Bennieston <andrew.bennieston@citrix.com> | 2014-06-04 10:30:46 +0100 |
---|---|---|
committer | David S. Miller <davem@davemloft.net> | 2014-06-04 14:48:17 -0700 |
commit | a2deb8b1e7a403fd63e08f87eb95217dd983e0d3 (patch) | |
tree | 48f3cced298b41ed38258f1b37fa1930438335de /include/xen | |
parent | 50ee60611bf0c7328e5cae438ea5c26590f3f747 (diff) | |
download | op-kernel-dev-a2deb8b1e7a403fd63e08f87eb95217dd983e0d3.zip op-kernel-dev-a2deb8b1e7a403fd63e08f87eb95217dd983e0d3.tar.gz |
xen-net{back, front}: Document multi-queue feature in netif.h
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@citrix.com>
Acked-by: Wei Liu <wei.liu2@citrix.com>
Acked-by: Ian Campbell <ian.campbell@citrix.com>
Reviewed-by: David Vrabel <david.vrabel@citrix.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Diffstat (limited to 'include/xen')
-rw-r--r-- | include/xen/interface/io/netif.h | 53 |
1 files changed, 53 insertions, 0 deletions
diff --git a/include/xen/interface/io/netif.h b/include/xen/interface/io/netif.h index c50061d..70054cc 100644 --- a/include/xen/interface/io/netif.h +++ b/include/xen/interface/io/netif.h @@ -51,6 +51,59 @@ */ /* + * Multiple transmit and receive queues: + * If supported, the backend will write the key "multi-queue-max-queues" to + * the directory for that vif, 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, which + * must be greater than zero, and no more than the value reported by the backend + * in "multi-queue-max-queues". + * + * Queues replicate the shared rings and event channels. + * "feature-split-event-channels" may optionally be used when using + * multiple queues, but is not mandatory. + * + * Each queue consists of one shared ring pair, i.e. there must be the same + * number of tx and rx rings. + * + * 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 (or event-channel-{tx,rx}) and {tx,rx}-ring-ref keys, + * instead writing those keys 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. For example, a frontend with two queues and split + * event channels must write the following set of queue-related keys: + * + * /local/domain/1/device/vif/0/multi-queue-num-queues = "2" + * /local/domain/1/device/vif/0/queue-0 = "" + * /local/domain/1/device/vif/0/queue-0/tx-ring-ref = "<ring-ref-tx0>" + * /local/domain/1/device/vif/0/queue-0/rx-ring-ref = "<ring-ref-rx0>" + * /local/domain/1/device/vif/0/queue-0/event-channel-tx = "<evtchn-tx0>" + * /local/domain/1/device/vif/0/queue-0/event-channel-rx = "<evtchn-rx0>" + * /local/domain/1/device/vif/0/queue-1 = "" + * /local/domain/1/device/vif/0/queue-1/tx-ring-ref = "<ring-ref-tx1>" + * /local/domain/1/device/vif/0/queue-1/rx-ring-ref = "<ring-ref-rx1" + * /local/domain/1/device/vif/0/queue-1/event-channel-tx = "<evtchn-tx1>" + * /local/domain/1/device/vif/0/queue-1/event-channel-rx = "<evtchn-rx1>" + * + * If there is any inconsistency in the XenStore data, the backend may + * choose not to connect any queues, instead treating the request as an + * error. This includes scenarios where more (or fewer) queues were + * requested than the frontend provided details for. + * + * 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. Guests must be + * prepared to receive packets on any queue they have requested be set up. + */ + +/* * "feature-no-csum-offload" should be used to turn IPv4 TCP/UDP checksum * offload off or on. If it is missing then the feature is assumed to be on. * "feature-ipv6-csum-offload" should be used to turn IPv6 TCP/UDP checksum |