From: Ouyang Changchun <changchun.ouyang@intel.com>
To: dev@dpdk.org
Subject: [dpdk-dev] [RFC PATCH 16/17] virtio: Free mbuf's with threshold
Date: Mon, 8 Dec 2014 14:21:55 +0800 [thread overview]
Message-ID: <1418019716-4962-17-git-send-email-changchun.ouyang@intel.com> (raw)
In-Reply-To: <1418019716-4962-1-git-send-email-changchun.ouyang@intel.com>
This makes virtio driver work like ixgbe. Transmit buffers are
held until a transmit threshold is reached. The previous behavior
was to hold mbuf's until the ring entry was reused which caused
more memory usage than needed.
Signed-off-by: Changchun Ouyang <changchun.ouyang@intel.com>
Signed-off-by: Stephen Hemminger <stephen@networkplumber.org>
---
lib/librte_pmd_virtio/virtio_ethdev.c | 7 ++--
lib/librte_pmd_virtio/virtio_rxtx.c | 70 +++++++++++++++++++++++++++++------
lib/librte_pmd_virtio/virtqueue.h | 3 +-
3 files changed, 64 insertions(+), 16 deletions(-)
diff --git a/lib/librte_pmd_virtio/virtio_ethdev.c b/lib/librte_pmd_virtio/virtio_ethdev.c
index c5f21c1..1ec29e1 100644
--- a/lib/librte_pmd_virtio/virtio_ethdev.c
+++ b/lib/librte_pmd_virtio/virtio_ethdev.c
@@ -176,15 +176,16 @@ virtio_send_command(struct virtqueue *vq, struct virtio_pmd_ctrl *ctrl,
virtqueue_notify(vq);
- while (vq->vq_used_cons_idx == vq->vq_ring.used->idx)
+ rte_rmb();
+ while (vq->vq_used_cons_idx == vq->vq_ring.used->idx) {
+ rte_rmb();
usleep(100);
+ }
while (vq->vq_used_cons_idx != vq->vq_ring.used->idx) {
uint32_t idx, desc_idx, used_idx;
struct vring_used_elem *uep;
- virtio_rmb();
-
used_idx = (uint32_t)(vq->vq_used_cons_idx
& (vq->vq_nentries - 1));
uep = &vq->vq_ring.used->ring[used_idx];
diff --git a/lib/librte_pmd_virtio/virtio_rxtx.c b/lib/librte_pmd_virtio/virtio_rxtx.c
index b44f091..26c0a1d 100644
--- a/lib/librte_pmd_virtio/virtio_rxtx.c
+++ b/lib/librte_pmd_virtio/virtio_rxtx.c
@@ -129,9 +129,15 @@ virtqueue_dequeue_burst_rx(struct virtqueue *vq, struct rte_mbuf **rx_pkts,
return i;
}
+#ifndef DEFAULT_TX_FREE_THRESH
+#define DEFAULT_TX_FREE_THRESH 32
+#endif
+
+/* Cleanup from completed transmits. */
static void
-virtqueue_dequeue_pkt_tx(struct virtqueue *vq)
+virtio_xmit_cleanup(struct virtqueue *vq, uint16_t num)
{
+#if 0
struct vring_used_elem *uep;
uint16_t used_idx, desc_idx;
@@ -140,6 +146,25 @@ virtqueue_dequeue_pkt_tx(struct virtqueue *vq)
desc_idx = (uint16_t) uep->id;
vq->vq_used_cons_idx++;
vq_ring_free_chain(vq, desc_idx);
+#endif
+ uint16_t i, used_idx, desc_idx;
+ for (i = 0; i < num ; i++) {
+ struct vring_used_elem *uep;
+ struct vq_desc_extra *dxp;
+
+ used_idx = (uint16_t)(vq->vq_used_cons_idx & (vq->vq_nentries - 1));
+ uep = &vq->vq_ring.used->ring[used_idx];
+ dxp = &vq->vq_descx[used_idx];
+
+ desc_idx = (uint16_t) uep->id;
+ vq->vq_used_cons_idx++;
+ vq_ring_free_chain(vq, desc_idx);
+
+ if (dxp->cookie != NULL) {
+ rte_pktmbuf_free(dxp->cookie);
+ dxp->cookie = NULL;
+ }
+ }
}
@@ -203,8 +228,10 @@ virtqueue_enqueue_xmit(struct virtqueue *txvq, struct rte_mbuf *cookie)
idx = head_idx;
dxp = &txvq->vq_descx[idx];
+#if 0
if (dxp->cookie != NULL)
rte_pktmbuf_free(dxp->cookie);
+#endif
dxp->cookie = (void *)cookie;
dxp->ndescs = needed;
@@ -404,6 +431,7 @@ virtio_dev_tx_queue_setup(struct rte_eth_dev *dev,
{
uint8_t vtpci_queue_idx = 2 * queue_idx + VTNET_SQ_TQ_QUEUE_IDX;
struct virtqueue *vq;
+ uint16_t tx_free_thresh;
int ret;
PMD_INIT_FUNC_TRACE();
@@ -421,6 +449,21 @@ virtio_dev_tx_queue_setup(struct rte_eth_dev *dev,
return ret;
}
+ tx_free_thresh = tx_conf->tx_free_thresh;
+ if (tx_free_thresh == 0)
+ tx_free_thresh = RTE_MIN(vq->vq_nentries / 4, DEFAULT_TX_FREE_THRESH);
+
+ if (tx_free_thresh >= (vq->vq_nentries - 3)) {
+ RTE_LOG(ERR, PMD, "tx_free_thresh must be less than the "
+ "number of TX entries minus 3 (%u)."
+ " (tx_free_thresh=%u port=%u queue=%u)\n",
+ vq->vq_nentries - 3,
+ tx_free_thresh, dev->data->port_id, queue_idx);
+ return -EINVAL;
+ }
+
+ vq->vq_free_thresh = tx_free_thresh;
+
dev->data->tx_queues[queue_idx] = vq;
return 0;
}
@@ -688,11 +731,9 @@ virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
{
struct virtqueue *txvq = tx_queue;
struct rte_mbuf *txm;
- uint16_t nb_used, nb_tx, num;
+ uint16_t nb_used, nb_tx;
int error;
- nb_tx = 0;
-
if (unlikely(nb_pkts < 1))
return nb_pkts;
@@ -700,21 +741,26 @@ virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
nb_used = VIRTQUEUE_NUSED(txvq);
virtio_rmb();
+ if (likely(nb_used > txvq->vq_free_thresh))
+ virtio_xmit_cleanup(txvq, nb_used);
- num = (uint16_t)(likely(nb_used < VIRTIO_MBUF_BURST_SZ) ? nb_used : VIRTIO_MBUF_BURST_SZ);
+ nb_tx = 0;
while (nb_tx < nb_pkts) {
/* Need one more descriptor for virtio header. */
int need = tx_pkts[nb_tx]->nb_segs - txvq->vq_free_cnt + 1;
- int deq_cnt = RTE_MIN(need, (int)num);
- num -= (deq_cnt > 0) ? deq_cnt : 0;
- while (deq_cnt > 0) {
- virtqueue_dequeue_pkt_tx(txvq);
- deq_cnt--;
+ /*Positive value indicates it need free vring descriptors */
+ if (unlikely(need > 0)) {
+ nb_used = VIRTQUEUE_NUSED(txvq);
+ virtio_rmb();
+ need = RTE_MIN(need, (int)nb_used);
+
+ virtio_xmit_cleanup(txvq, need);
+ need = (int)tx_pkts[nb_tx]->nb_segs -
+ txvq->vq_free_cnt + 1;
}
- need = (int)tx_pkts[nb_tx]->nb_segs - txvq->vq_free_cnt + 1;
/*
* Zero or negative value indicates it has enough free
* descriptors to use for transmitting.
@@ -723,7 +769,7 @@ virtio_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
txm = tx_pkts[nb_tx];
/* Do VLAN tag insertion */
- if (txm->ol_flags & PKT_TX_VLAN_PKT) {
+ if (unlikely(txm->ol_flags & PKT_TX_VLAN_PKT)) {
error = rte_vlan_insert(&txm);
if (unlikely(error)) {
rte_pktmbuf_free(txm);
diff --git a/lib/librte_pmd_virtio/virtqueue.h b/lib/librte_pmd_virtio/virtqueue.h
index 4cb82f9..7672a12 100644
--- a/lib/librte_pmd_virtio/virtqueue.h
+++ b/lib/librte_pmd_virtio/virtqueue.h
@@ -164,6 +164,7 @@ struct virtqueue {
struct rte_mempool *mpool; /**< mempool for mbuf allocation */
uint16_t queue_id; /**< DPDK queue index. */
uint8_t port_id; /**< Device port identifier. */
+ uint16_t vq_queue_index; /**< PCI queue index */
void *vq_ring_virt_mem; /**< linear address of vring*/
unsigned int vq_ring_size;
@@ -172,7 +173,7 @@ struct virtqueue {
struct vring vq_ring; /**< vring keeping desc, used and avail */
uint16_t vq_free_cnt; /**< num of desc available */
uint16_t vq_nentries; /**< vring desc numbers */
- uint16_t vq_queue_index; /**< PCI queue index */
+ uint16_t vq_free_thresh; /**< free threshold */
/**
* Head of the free chain in the descriptor table. If
* there are no free descriptors, this will be set to
--
1.8.4.2
next prev parent reply other threads:[~2014-12-08 6:22 UTC|newest]
Thread overview: 30+ messages / expand[flat|nested] mbox.gz Atom feed top
2014-12-08 6:21 [dpdk-dev] [RFC PATCH 00/17] Single virtio implementation Ouyang Changchun
2014-12-08 6:21 ` [dpdk-dev] [RFC PATCH 01/17] virtio: Rearrange resource initialization Ouyang Changchun
2014-12-08 6:21 ` [dpdk-dev] [RFC PATCH 02/17] virtio: Use weaker barriers Ouyang Changchun
2014-12-08 6:21 ` [dpdk-dev] [RFC PATCH 03/17] virtio: Allow starting with link down Ouyang Changchun
2014-12-08 6:21 ` [dpdk-dev] [RFC PATCH 04/17] virtio: Add support for Link State interrupt Ouyang Changchun
2014-12-08 6:21 ` [dpdk-dev] [RFC PATCH 05/17] ether: Add soft vlan encap/decap functions Ouyang Changchun
2014-12-08 6:21 ` [dpdk-dev] [RFC PATCH 06/17] virtio: Use software vlan stripping Ouyang Changchun
2014-12-08 6:21 ` [dpdk-dev] [RFC PATCH 07/17] virtio: Remove unnecessary adapter structure Ouyang Changchun
2014-12-08 6:21 ` [dpdk-dev] [RFC PATCH 08/17] virtio: Remove redundant vq_alignment Ouyang Changchun
2014-12-08 6:21 ` [dpdk-dev] [RFC PATCH 09/17] virtio: Fix how states are handled during initialization Ouyang Changchun
2014-12-08 6:21 ` [dpdk-dev] [RFC PATCH 10/17] virtio: Make vtpci_get_status local Ouyang Changchun
2014-12-08 6:21 ` [dpdk-dev] [RFC PATCH 11/17] virtio: Check for packet headroom at compile time Ouyang Changchun
2014-12-08 6:21 ` [dpdk-dev] [RFC PATCH 12/17] virtio: Move allocation before initialization Ouyang Changchun
2014-12-08 6:21 ` [dpdk-dev] [RFC PATCH 13/17] virtio: Add support for vlan filtering Ouyang Changchun
2014-12-08 6:21 ` [dpdk-dev] [RFC PATCH 14/17] virtio: Add suport for multiple mac addresses Ouyang Changchun
2014-12-08 6:21 ` [dpdk-dev] [RFC PATCH 15/17] virtio: Add ability to set MAC address Ouyang Changchun
2014-12-08 6:21 ` Ouyang Changchun [this message]
2014-12-08 6:21 ` [dpdk-dev] [RFC PATCH 17/17] virtio: Use port IO to get PCI resource Ouyang Changchun
2014-12-08 9:30 ` [dpdk-dev] [RFC PATCH 00/17] Single virtio implementation Thomas Monjalon
2014-12-09 1:08 ` Ouyang, Changchun
2014-12-09 3:23 ` Qiu, Michael
2014-12-09 5:24 ` Stephen Hemminger
2014-12-09 5:41 ` Ouyang, Changchun
2014-12-09 6:11 ` Thomas Monjalon
2014-12-09 6:40 ` Ouyang, Changchun
2014-12-09 8:53 ` Thomas Monjalon
2014-12-09 9:46 ` Bruce Richardson
2014-12-09 14:08 ` Ouyang, Changchun
2014-12-09 16:03 ` Qiu, Michael
2014-12-10 0:29 ` Ouyang, Changchun
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=1418019716-4962-17-git-send-email-changchun.ouyang@intel.com \
--to=changchun.ouyang@intel.com \
--cc=dev@dpdk.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).