On Tue, Mar 19, 2019 at 09:54:03AM +0100, Jens Freimann wrote:
> On Tue, Mar 19, 2019 at 02:43:06PM +0800, Tiwei Bie wrote:
> > Cache the AVAIL, USED and WRITE bits to avoid calculating
> > them as much as possible. Note that, the WRITE bit isn't
> > cached for control queue.
> >
> > Signed-off-by: Tiwei Bie <tiwei....@intel.com>
> > ---
> > drivers/net/virtio/virtio_ethdev.c | 35 ++++++++++++++----------------
> > drivers/net/virtio/virtio_rxtx.c | 31 ++++++++++----------------
> > drivers/net/virtio/virtqueue.h | 8 +++----
> > 3 files changed, 32 insertions(+), 42 deletions(-)
> >
> > diff --git a/drivers/net/virtio/virtio_ethdev.c
> > b/drivers/net/virtio/virtio_ethdev.c
> > index ff16fb63e..9060b6b33 100644
> > --- a/drivers/net/virtio/virtio_ethdev.c
> > +++ b/drivers/net/virtio/virtio_ethdev.c
> > @@ -149,7 +149,7 @@ virtio_send_command_packed(struct virtnet_ctl *cvq,
> > int head;
> > struct vring_packed_desc *desc = vq->ring_packed.desc_packed;
> > struct virtio_pmd_ctrl *result;
> > - bool avail_wrap_counter;
> > + uint16_t flags;
> > int sum = 0;
> > int nb_descs = 0;
> > int k;
> > @@ -161,14 +161,15 @@ virtio_send_command_packed(struct virtnet_ctl *cvq,
> > * One RX packet for ACK.
> > */
> > head = vq->vq_avail_idx;
> > - avail_wrap_counter = vq->avail_wrap_counter;
> > + flags = vq->cached_flags;
> > desc[head].addr = cvq->virtio_net_hdr_mem;
> > desc[head].len = sizeof(struct virtio_net_ctrl_hdr);
> > vq->vq_free_cnt--;
> > nb_descs++;
> > if (++vq->vq_avail_idx >= vq->vq_nentries) {
> > vq->vq_avail_idx -= vq->vq_nentries;
> > - vq->avail_wrap_counter ^= 1;
> > + vq->cached_flags ^=
> > + VRING_DESC_F_AVAIL(1) | VRING_DESC_F_USED(1);
>
> Maybe name it avail_used_flags instead of cached flags. Also we could
> use a constant value.
It also contains the WRITE bit (not just AVAIL and USED bits)
for Rx path. That's why I didn't name it as avail_used_flags.
>
> > }
> >
> > for (k = 0; k < pkt_num; k++) {
> > @@ -177,34 +178,31 @@ virtio_send_command_packed(struct virtnet_ctl *cvq,
> > + sizeof(ctrl->status) + sizeof(uint8_t) * sum;
> > desc[vq->vq_avail_idx].len = dlen[k];
> > desc[vq->vq_avail_idx].flags = VRING_DESC_F_NEXT |
> > - VRING_DESC_F_AVAIL(vq->avail_wrap_counter) |
> > - VRING_DESC_F_USED(!vq->avail_wrap_counter);
> > + vq->cached_flags;
> > sum += dlen[k];
> > vq->vq_free_cnt--;
> > nb_descs++;
> > if (++vq->vq_avail_idx >= vq->vq_nentries) {
> > vq->vq_avail_idx -= vq->vq_nentries;
> > - vq->avail_wrap_counter ^= 1;
> > + vq->cached_flags ^=
> > + VRING_DESC_F_AVAIL(1) | VRING_DESC_F_USED(1);
> > }
> > }
> >
> > desc[vq->vq_avail_idx].addr = cvq->virtio_net_hdr_mem
> > + sizeof(struct virtio_net_ctrl_hdr);
> > desc[vq->vq_avail_idx].len = sizeof(ctrl->status);
> > - desc[vq->vq_avail_idx].flags = VRING_DESC_F_WRITE |
> > - VRING_DESC_F_AVAIL(vq->avail_wrap_counter) |
> > - VRING_DESC_F_USED(!vq->avail_wrap_counter);
> > + desc[vq->vq_avail_idx].flags = VRING_DESC_F_WRITE | vq->cached_flags;
> > vq->vq_free_cnt--;
> > nb_descs++;
> > if (++vq->vq_avail_idx >= vq->vq_nentries) {
> > vq->vq_avail_idx -= vq->vq_nentries;
> > - vq->avail_wrap_counter ^= 1;
> > + vq->cached_flags ^=
> > + VRING_DESC_F_AVAIL(1) | VRING_DESC_F_USED(1);
> > }
> >
> > virtio_wmb(vq->hw->weak_barriers);
> > - desc[head].flags = VRING_DESC_F_NEXT |
> > - VRING_DESC_F_AVAIL(avail_wrap_counter) |
> > - VRING_DESC_F_USED(!avail_wrap_counter);
> > + desc[head].flags = VRING_DESC_F_NEXT | flags;
> >
> > virtio_wmb(vq->hw->weak_barriers);
> > virtqueue_notify(vq);
> > @@ -226,12 +224,12 @@ virtio_send_command_packed(struct virtnet_ctl *cvq,
> > PMD_INIT_LOG(DEBUG, "vq->vq_free_cnt=%d\n"
> > "vq->vq_avail_idx=%d\n"
> > "vq->vq_used_cons_idx=%d\n"
> > - "vq->avail_wrap_counter=%d\n"
> > + "vq->cached_flags=0x%x\n"
> > "vq->used_wrap_counter=%d\n",
> > vq->vq_free_cnt,
> > vq->vq_avail_idx,
> > vq->vq_used_cons_idx,
> > - vq->avail_wrap_counter,
> > + vq->cached_flags,
> > vq->used_wrap_counter);
> >
> > result = cvq->virtio_net_hdr_mz->addr;
> > @@ -491,11 +489,10 @@ virtio_init_queue(struct rte_eth_dev *dev, uint16_t
> > vtpci_queue_idx)
> > vq->vq_nentries = vq_size;
> > vq->event_flags_shadow = 0;
> > if (vtpci_packed_queue(hw)) {
> > - vq->avail_wrap_counter = 1;
> > vq->used_wrap_counter = 1;
> > - vq->avail_used_flags =
> > - VRING_DESC_F_AVAIL(vq->avail_wrap_counter) |
> > - VRING_DESC_F_USED(!vq->avail_wrap_counter);
> > + vq->cached_flags = VRING_DESC_F_AVAIL(1);
> > + if (queue_type == VTNET_RQ)
> > + vq->cached_flags |= VRING_DESC_F_WRITE;
> > }
> >
> > /*
> > diff --git a/drivers/net/virtio/virtio_rxtx.c
> > b/drivers/net/virtio/virtio_rxtx.c
> > index 771d3c3f6..3c354baef 100644
> > --- a/drivers/net/virtio/virtio_rxtx.c
> > +++ b/drivers/net/virtio/virtio_rxtx.c
> > @@ -431,7 +431,7 @@ virtqueue_enqueue_recv_refill_packed(struct virtqueue
> > *vq,
> > struct rte_mbuf **cookie, uint16_t num)
> > {
> > struct vring_packed_desc *start_dp = vq->ring_packed.desc_packed;
> > - uint16_t flags = VRING_DESC_F_WRITE | vq->avail_used_flags;
> > + uint16_t flags = vq->cached_flags;
> > struct virtio_hw *hw = vq->hw;
> > struct vq_desc_extra *dxp;
> > uint16_t idx;
> > @@ -460,11 +460,9 @@ virtqueue_enqueue_recv_refill_packed(struct virtqueue
> > *vq,
> > start_dp[idx].flags = flags;
> > if (++vq->vq_avail_idx >= vq->vq_nentries) {
> > vq->vq_avail_idx -= vq->vq_nentries;
> > - vq->avail_wrap_counter ^= 1;
> > - vq->avail_used_flags =
> > - VRING_DESC_F_AVAIL(vq->avail_wrap_counter) |
> > - VRING_DESC_F_USED(!vq->avail_wrap_counter);
> > - flags = VRING_DESC_F_WRITE | vq->avail_used_flags;
> > + vq->cached_flags ^=
> > + VRING_DESC_F_AVAIL(1) | VRING_DESC_F_USED(1);
> > + flags = vq->cached_flags;
>
> same here. it's not really cached, it's pre-calculated. And here we
> could also use a pre-calculated constand/define.
For pre-calculated constant/define, do you mean VRING_DESC_F_AVAIL(1)
and VRING_DESC_F_USED(1)? There is still little code in virtio-user
using them without constantly passing 1. I planned to fully get rid
of them in a separate patch later (but I can do it in this series if
anyone wants).
>
> Otherwise looks good! Did you see any performance improvements?
Yeah, I saw slightly better performance in a quick test.
Thanks,
Tiwei
>
>
> regards,
> Jens
