This is series mainly merges out-of-order mergeable and
non-mergeable Rx paths.
Doing so, we remove one path, and micro-benchmark does not
show any performance regression:
+---------+----------+----------+----------+----------+----------+
| V18.11 | v18.11 + Rx paths merge
|
+---------+----------+----------+----------+----------+----------+
| 1 queue | 2 queues | 3 queues | 1 queue | 2 queues | 3
queues |
+-----+---------+---------+----------+----------+----------+----------+----------+
| MRG | Rxonly | 16.22 | 32.47 | 48.81 | 16.31 | 32.62 | 48.88
|
| OFF | IO loop | 12.86 | 25.69 | 38.16 | 12.88 | 25.74 | 38.53
|
+-----+---------+---------+----------+----------+----------+----------+----------+
| MRG | Rxonly | 16.27 | 32.65 | 48.81 | 16.29 | 32.67 | 48.77
|
| ON | IO loop | 12.82 | 25.59 | 37.86 | 12.86 | 25.68 | 37.9
|
+-----+---------+---------+----------+----------+----------+----------+----------+
Note that to avoid vector path to be selected as this benchmark is done
without offloads, I had to do this small change to disable it in
virtio_dev_configure():
- hw->use_simple_rx = 1;
+ hw->use_simple_rx = 0;
The series also enables in-order path selection when mergeable buffers feature
is disabled. Offload and refill helpers are also changed to be inlined.
Maxime Coquelin (3):
net/virtio: inline refill and offload helpers
net/virtio: merge Rx mergeable and non-mergeable paths
net/virtio: add non-mergeable support to in-order path
drivers/net/virtio/virtio_ethdev.c | 16 +---
drivers/net/virtio/virtio_ethdev.h | 5 +-
drivers/net/virtio/virtio_rxtx.c | 133 ++++-------------------------
3 files changed, 22 insertions(+), 132 deletions(-)
--
2.17.2
