Re: [PATCH netdev 5/5] virtio-net, xsk: virtio-net support xsk zero copy tx

From: Michael S. Tsirkin
Date: Tue Jan 05 2021 - 08:23:28 EST


On Tue, Jan 05, 2021 at 05:11:43PM +0800, Xuan Zhuo wrote:
> Virtio net support xdp socket.
>
> We should open the module param "napi_tx" for using this feature.

what does this imply exactly?

> In fact, various virtio implementations have some problems:
> 1. The tx interrupt may be lost
> 2. The tx interrupt may have a relatively large delay
>
> This brings us to several questions:
>
> 1. Wakeup wakes up a tx interrupt or directly starts a napi on the
> current cpu, which will cause a delay in sending packets.
> 2. When the tx ring is full, the tx interrupt may be lost or delayed,
> resulting in untimely recovery.
>
> I choose to send part of the data directly during wakeup. If the sending
> has not been completed, I will start a napi to complete the subsequent
> sending work.
>
> Since the possible delay or loss of tx interrupt occurs when the tx ring
> is full, I added a timer to solve this problem.

A lost interrupt sounds like a bug somewhere.
Why isn't this device already broken then, even without zero copy?
Won't a full ring stall forever? And won't a significantly delayed
tx interrupt block userspace?

How about putting work arounds were in a separate patch for now,
so it's easier to figure out whether anything in the patch itself
is causing issues?

> The performance of udp sending based on virtio net + xsk is 6 times that
> of ordinary kernel udp send.
>
> * xsk_check_timeout: when the dev full or all xsk.hdr used, start timer
> to check the xsk.hdr is avail. the unit is us.
> * xsk_num_max: the xsk.hdr max num
> * xsk_num_percent: the max hdr num be the percent of the virtio ring
> size. The real xsk hdr num will the min of xsk_num_max and the percent
> of the num of virtio ring
> * xsk_budget: the budget for xsk run
>
> Signed-off-by: Xuan Zhuo <xuanzhuo@xxxxxxxxxxxxxxxxx>
> ---
> drivers/net/virtio_net.c | 437 ++++++++++++++++++++++++++++++++++++++++++++++-
> 1 file changed, 434 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
> index e744dce..76319e7 100644
> --- a/drivers/net/virtio_net.c
> +++ b/drivers/net/virtio_net.c
> @@ -22,10 +22,21 @@
> #include <net/route.h>
> #include <net/xdp.h>
> #include <net/net_failover.h>
> +#include <net/xdp_sock_drv.h>
>
> static int napi_weight = NAPI_POLL_WEIGHT;
> module_param(napi_weight, int, 0444);
>
> +static int xsk_check_timeout = 100;
> +static int xsk_num_max = 1024;
> +static int xsk_num_percent = 80;
> +static int xsk_budget = 128;
> +
> +module_param(xsk_check_timeout, int, 0644);
> +module_param(xsk_num_max, int, 0644);
> +module_param(xsk_num_percent, int, 0644);
> +module_param(xsk_budget, int, 0644);
> +
> static bool csum = true, gso = true, napi_tx = true;
> module_param(csum, bool, 0444);
> module_param(gso, bool, 0444);
> @@ -110,6 +121,9 @@ struct virtnet_xsk_hdr {
> u32 len;
> };
>
> +#define VIRTNET_STATE_XSK_WAKEUP BIT(0)
> +#define VIRTNET_STATE_XSK_TIMER BIT(1)
> +
> #define VIRTNET_SQ_STAT(m) offsetof(struct virtnet_sq_stats, m)
> #define VIRTNET_RQ_STAT(m) offsetof(struct virtnet_rq_stats, m)
>
> @@ -149,6 +163,32 @@ struct send_queue {
> struct virtnet_sq_stats stats;
>
> struct napi_struct napi;
> +
> + struct {
> + struct xsk_buff_pool __rcu *pool;
> + struct virtnet_xsk_hdr __rcu *hdr;
> +
> + unsigned long state;
> + u64 hdr_con;
> + u64 hdr_pro;
> + u64 hdr_n;
> + struct xdp_desc last_desc;
> + bool wait_slot;
> + /* tx interrupt issues
> + * 1. that may be lost
> + * 2. that too slow, 200/s or delay 10ms

I mean, we call virtqueue_enable_cb_delayed on each start_xmit.
The point is explicitly to reduce the # of tx interrupts,
is this the issue?

> + *
> + * timer for:
> + * 1. recycle the desc.(no check for performance, see below)
> + * 2. check the nic ring is avali. when nic ring is full
> + *
> + * Here, the regular check is performed for dev full. The
> + * application layer must ensure that the number of cq is
> + * sufficient, otherwise there may be insufficient cq in use.

Can't really parse this. cq as in control vq?

> + *
> + */
> + struct hrtimer timer;
> + } xsk;
> };
>
> /* Internal representation of a receive virtqueue */
> @@ -267,6 +307,8 @@ static void __free_old_xmit_ptr(struct send_queue *sq, bool in_napi,
> bool xsk_wakeup,
> unsigned int *_packets, unsigned int *_bytes);
> static void free_old_xmit_skbs(struct send_queue *sq, bool in_napi);
> +static int virtnet_xsk_run(struct send_queue *sq,
> + struct xsk_buff_pool *pool, int budget);
>
> static bool is_xdp_frame(void *ptr)
> {
> @@ -1439,6 +1481,40 @@ static int virtnet_receive(struct receive_queue *rq, int budget,
> return stats.packets;
> }
>
> +static void virt_xsk_complete(struct send_queue *sq, u32 num, bool xsk_wakeup)
> +{
> + struct xsk_buff_pool *pool;
> + int n;
> +
> + rcu_read_lock();
> +
> + WRITE_ONCE(sq->xsk.hdr_pro, sq->xsk.hdr_pro + num);

WRITE_ONCE without READ_ONCE anywhere looks a bit weird.
Add a comment explaining why this is right?

> +
> + pool = rcu_dereference(sq->xsk.pool);
> + if (!pool) {
> + if (sq->xsk.hdr_pro - sq->xsk.hdr_con == sq->xsk.hdr_n) {
> + kfree(sq->xsk.hdr);
> + rcu_assign_pointer(sq->xsk.hdr, NULL);
> + }
> + rcu_read_unlock();
> + return;
> + }
> +
> + xsk_tx_completed(pool, num);
> +
> + rcu_read_unlock();
> +
> + if (!xsk_wakeup || !sq->xsk.wait_slot)
> + return;
> +
> + n = sq->xsk.hdr_pro - sq->xsk.hdr_con;
> +
> + if (n > sq->xsk.hdr_n / 2) {
> + sq->xsk.wait_slot = false;
> + virtqueue_napi_schedule(&sq->napi, sq->vq);
> + }
> +}
> +
> static void __free_old_xmit_ptr(struct send_queue *sq, bool in_napi,
> bool xsk_wakeup,
> unsigned int *_packets, unsigned int *_bytes)
> @@ -1446,6 +1522,7 @@ static void __free_old_xmit_ptr(struct send_queue *sq, bool in_napi,
> unsigned int packets = 0;
> unsigned int bytes = 0;
> unsigned int len;
> + u64 xsknum = 0;
> struct virtnet_xdp_type *xtype;
> struct xdp_frame *frame;
> struct virtnet_xsk_hdr *xskhdr;
> @@ -1466,6 +1543,7 @@ static void __free_old_xmit_ptr(struct send_queue *sq, bool in_napi,
> if (xtype->type == XDP_TYPE_XSK) {
> xskhdr = (struct virtnet_xsk_hdr *)xtype;
> bytes += xskhdr->len;
> + xsknum += 1;
> } else {
> frame = xtype_got_ptr(xtype);
> xdp_return_frame(frame);
> @@ -1475,6 +1553,9 @@ static void __free_old_xmit_ptr(struct send_queue *sq, bool in_napi,
> packets++;
> }
>
> + if (xsknum)
> + virt_xsk_complete(sq, xsknum, xsk_wakeup);
> +
> *_packets = packets;
> *_bytes = bytes;
> }
> @@ -1595,6 +1676,8 @@ static int virtnet_poll_tx(struct napi_struct *napi, int budget)
> struct virtnet_info *vi = sq->vq->vdev->priv;
> unsigned int index = vq2txq(sq->vq);
> struct netdev_queue *txq;
> + struct xsk_buff_pool *pool;
> + int work = 0;
>
> if (unlikely(is_xdp_raw_buffer_queue(vi, index))) {
> /* We don't need to enable cb for XDP */
> @@ -1604,15 +1687,26 @@ static int virtnet_poll_tx(struct napi_struct *napi, int budget)
>
> txq = netdev_get_tx_queue(vi->dev, index);
> __netif_tx_lock(txq, raw_smp_processor_id());
> - free_old_xmit_skbs(sq, true);
> +
> + rcu_read_lock();
> + pool = rcu_dereference(sq->xsk.pool);
> + if (pool) {
> + work = virtnet_xsk_run(sq, pool, budget);
> + rcu_read_unlock();
> + } else {
> + rcu_read_unlock();
> + free_old_xmit_skbs(sq, true);
> + }
> +
> __netif_tx_unlock(txq);
>
> - virtqueue_napi_complete(napi, sq->vq, 0);
> + if (work < budget)
> + virtqueue_napi_complete(napi, sq->vq, 0);
>
> if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
> netif_tx_wake_queue(txq);
>
> - return 0;
> + return work;
> }
>
> static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
> @@ -2560,16 +2654,346 @@ static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
> return err;
> }
>
> +static enum hrtimer_restart virtnet_xsk_timeout(struct hrtimer *timer)
> +{
> + struct send_queue *sq;
> +
> + sq = container_of(timer, struct send_queue, xsk.timer);
> +
> + clear_bit(VIRTNET_STATE_XSK_TIMER, &sq->xsk.state);
> +
> + virtqueue_napi_schedule(&sq->napi, sq->vq);
> +
> + return HRTIMER_NORESTART;
> +}
> +
> +static int virtnet_xsk_pool_enable(struct net_device *dev,
> + struct xsk_buff_pool *pool,
> + u16 qid)
> +{
> + struct virtnet_info *vi = netdev_priv(dev);
> + struct send_queue *sq = &vi->sq[qid];
> + struct virtnet_xsk_hdr *hdr;
> + int n, ret = 0;
> +
> + if (qid >= dev->real_num_rx_queues || qid >= dev->real_num_tx_queues)
> + return -EINVAL;
> +
> + if (qid >= vi->curr_queue_pairs)
> + return -EINVAL;
> +
> + rcu_read_lock();
> +
> + ret = -EBUSY;
> + if (rcu_dereference(sq->xsk.pool))
> + goto end;
> +
> + /* check last xsk wait for hdr been free */
> + if (rcu_dereference(sq->xsk.hdr))
> + goto end;
> +
> + n = virtqueue_get_vring_size(sq->vq);
> + n = min(xsk_num_max, n * (xsk_num_percent % 100) / 100);
> +
> + ret = -ENOMEM;
> + hdr = kcalloc(n, sizeof(struct virtnet_xsk_hdr), GFP_ATOMIC);
> + if (!hdr)
> + goto end;
> +
> + memset(&sq->xsk, 0, sizeof(sq->xsk));
> +
> + sq->xsk.hdr_pro = n;
> + sq->xsk.hdr_n = n;
> +
> + hrtimer_init(&sq->xsk.timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL_PINNED);
> + sq->xsk.timer.function = virtnet_xsk_timeout;
> +
> + rcu_assign_pointer(sq->xsk.pool, pool);
> + rcu_assign_pointer(sq->xsk.hdr, hdr);
> +
> + ret = 0;
> +end:
> + rcu_read_unlock();
> +
> + return ret;
> +}
> +
> +static int virtnet_xsk_pool_disable(struct net_device *dev, u16 qid)
> +{
> + struct virtnet_info *vi = netdev_priv(dev);
> + struct send_queue *sq = &vi->sq[qid];
> +
> + if (qid >= dev->real_num_rx_queues || qid >= dev->real_num_tx_queues)
> + return -EINVAL;
> +
> + if (qid >= vi->curr_queue_pairs)
> + return -EINVAL;
> +
> + rcu_assign_pointer(sq->xsk.pool, NULL);
> +
> + hrtimer_cancel(&sq->xsk.timer);
> +
> + synchronize_rcu(); /* Sync with the XSK wakeup and with NAPI. */
> +
> + if (sq->xsk.hdr_pro - sq->xsk.hdr_con == sq->xsk.hdr_n) {
> + kfree(sq->xsk.hdr);
> + rcu_assign_pointer(sq->xsk.hdr, NULL);
> + synchronize_rcu();
> + }
> +
> + return 0;
> +}
> +
> static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp)
> {
> switch (xdp->command) {
> case XDP_SETUP_PROG:
> return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
> + case XDP_SETUP_XSK_POOL:
> + xdp->xsk.need_dma = false;
> + if (xdp->xsk.pool)
> + return virtnet_xsk_pool_enable(dev, xdp->xsk.pool,
> + xdp->xsk.queue_id);
> + else
> + return virtnet_xsk_pool_disable(dev, xdp->xsk.queue_id);
> default:
> return -EINVAL;
> }
> }
>
> +static int virtnet_xsk_xmit(struct send_queue *sq, struct xsk_buff_pool *pool,
> + struct xdp_desc *desc)
> +{
> + struct virtnet_info *vi = sq->vq->vdev->priv;
> + void *data, *ptr;
> + struct page *page;
> + struct virtnet_xsk_hdr *xskhdr;
> + u32 idx, offset, n, i, copy, copied;
> + u64 addr;
> + int err, m;
> +
> + addr = desc->addr;
> +
> + data = xsk_buff_raw_get_data(pool, addr);
> + offset = offset_in_page(data);
> +
> + /* one for hdr, one for the first page */
> + n = 2;
> + m = desc->len - (PAGE_SIZE - offset);
> + if (m > 0) {
> + n += m >> PAGE_SHIFT;
> + if (m & PAGE_MASK)
> + ++n;
> +
> + n = min_t(u32, n, ARRAY_SIZE(sq->sg));
> + }
> +
> + idx = sq->xsk.hdr_con % sq->xsk.hdr_n;
> + xskhdr = &sq->xsk.hdr[idx];
> +
> + /* xskhdr->hdr has been memset to zero, so not need to clear again */
> +
> + sg_init_table(sq->sg, n);
> + sg_set_buf(sq->sg, &xskhdr->hdr, vi->hdr_len);
> +
> + copied = 0;
> + for (i = 1; i < n; ++i) {
> + copy = min_t(int, desc->len - copied, PAGE_SIZE - offset);
> +
> + page = xsk_buff_raw_get_page(pool, addr + copied);
> +
> + sg_set_page(sq->sg + i, page, copy, offset);
> + copied += copy;
> + if (offset)
> + offset = 0;
> + }
> +
> + xskhdr->len = desc->len;
> + ptr = xdp_to_ptr(&xskhdr->type);
> +
> + err = virtqueue_add_outbuf(sq->vq, sq->sg, n, ptr, GFP_ATOMIC);
> + if (unlikely(err))
> + sq->xsk.last_desc = *desc;
> + else
> + sq->xsk.hdr_con++;
> +
> + return err;
> +}
> +
> +static bool virtnet_xsk_dev_is_full(struct send_queue *sq)
> +{
> + if (sq->vq->num_free < 2 + MAX_SKB_FRAGS)
> + return true;
> +
> + if (sq->xsk.hdr_con == sq->xsk.hdr_pro)
> + return true;
> +
> + return false;
> +}
> +
> +static int virtnet_xsk_xmit_zc(struct send_queue *sq,
> + struct xsk_buff_pool *pool, unsigned int budget)
> +{
> + struct xdp_desc desc;
> + int err, packet = 0;
> + int ret = -EAGAIN;
> +
> + if (sq->xsk.last_desc.addr) {
> + err = virtnet_xsk_xmit(sq, pool, &sq->xsk.last_desc);
> + if (unlikely(err))
> + return -EBUSY;
> +
> + ++packet;
> + sq->xsk.last_desc.addr = 0;
> + }
> +
> + while (budget-- > 0) {
> + if (virtnet_xsk_dev_is_full(sq)) {
> + ret = -EBUSY;
> + break;
> + }
> +
> + if (!xsk_tx_peek_desc(pool, &desc)) {
> + /* done */
> + ret = 0;
> + break;
> + }
> +
> + err = virtnet_xsk_xmit(sq, pool, &desc);
> + if (unlikely(err)) {
> + ret = -EBUSY;
> + break;
> + }
> +
> + ++packet;
> + }
> +
> + if (packet) {
> + xsk_tx_release(pool);
> +
> + if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
> + u64_stats_update_begin(&sq->stats.syncp);
> + sq->stats.kicks++;
> + u64_stats_update_end(&sq->stats.syncp);
> + }
> + }
> +
> + return ret;
> +}
> +
> +static int virtnet_xsk_run(struct send_queue *sq,
> + struct xsk_buff_pool *pool, int budget)
> +{
> + int err, ret = 0;
> + unsigned int _packets = 0;
> + unsigned int _bytes = 0;
> +
> + sq->xsk.wait_slot = false;
> +
> + if (test_and_clear_bit(VIRTNET_STATE_XSK_TIMER, &sq->xsk.state))
> + hrtimer_try_to_cancel(&sq->xsk.timer);
> +
> + __free_old_xmit_ptr(sq, true, false, &_packets, &_bytes);
> +
> + err = virtnet_xsk_xmit_zc(sq, pool, xsk_budget);
> + if (!err) {
> + struct xdp_desc desc;
> +
> + clear_bit(VIRTNET_STATE_XSK_WAKEUP, &sq->xsk.state);
> + xsk_set_tx_need_wakeup(pool);
> +
> + /* Race breaker. If new is coming after last xmit
> + * but before flag change
> + */


A bit more text explaining the rules for the two bits please.

> +
> + if (!xsk_tx_peek_desc(pool, &desc))
> + goto end;
> +
> + set_bit(VIRTNET_STATE_XSK_WAKEUP, &sq->xsk.state);
> + xsk_clear_tx_need_wakeup(pool);
> +
> + sq->xsk.last_desc = desc;
> + ret = budget;
> + goto end;
> + }
> +
> + xsk_clear_tx_need_wakeup(pool);
> +
> + if (err == -EAGAIN) {
> + ret = budget;
> + goto end;
> + }
> +
> + /* -EBUSY: wait tx ring avali.
> + * by tx interrupt or rx interrupt or start_xmit or timer


can't parse this either ...

> + */
> +
> + __free_old_xmit_ptr(sq, true, false, &_packets, &_bytes);
> +
> + if (!virtnet_xsk_dev_is_full(sq)) {
> + ret = budget;
> + goto end;
> + }
> +
> + sq->xsk.wait_slot = true;
> +
> + if (xsk_check_timeout) {
> + hrtimer_start(&sq->xsk.timer,
> + ns_to_ktime(xsk_check_timeout * 1000),
> + HRTIMER_MODE_REL_PINNED);
> +
> + set_bit(VIRTNET_STATE_XSK_TIMER, &sq->xsk.state);
> + }
> +
> + virtnet_sq_stop_check(sq, true);
> +
> +end:
> + return ret;
> +}
> +
> +static int virtnet_xsk_wakeup(struct net_device *dev, u32 qid, u32 flag)
> +{
> + struct virtnet_info *vi = netdev_priv(dev);
> + struct send_queue *sq;
> + struct xsk_buff_pool *pool;
> + struct netdev_queue *txq;
> + int work = 0;
> +
> + if (!netif_running(dev))
> + return -ENETDOWN;
> +
> + if (qid >= vi->curr_queue_pairs)
> + return -EINVAL;
> +
> + sq = &vi->sq[qid];
> +
> + rcu_read_lock();
> +
> + pool = rcu_dereference(sq->xsk.pool);
> + if (!pool)
> + goto end;
> +
> + if (test_and_set_bit(VIRTNET_STATE_XSK_WAKEUP, &sq->xsk.state))
> + goto end;
> +
> + txq = netdev_get_tx_queue(dev, qid);
> +
> + local_bh_disable();
> + __netif_tx_lock(txq, raw_smp_processor_id());
> +
> + work = virtnet_xsk_run(sq, pool, xsk_budget);
> +
> + __netif_tx_unlock(txq);
> + local_bh_enable();
> +
> + if (work == xsk_budget)
> + virtqueue_napi_schedule(&sq->napi, sq->vq);
> +
> +end:
> + rcu_read_unlock();
> + return 0;
> +}
> +
> static int virtnet_get_phys_port_name(struct net_device *dev, char *buf,
> size_t len)
> {
> @@ -2624,6 +3048,7 @@ static int virtnet_set_features(struct net_device *dev,
> .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
> .ndo_bpf = virtnet_xdp,
> .ndo_xdp_xmit = virtnet_xdp_xmit,
> + .ndo_xsk_wakeup = virtnet_xsk_wakeup,
> .ndo_features_check = passthru_features_check,
> .ndo_get_phys_port_name = virtnet_get_phys_port_name,
> .ndo_set_features = virtnet_set_features,
> @@ -2722,6 +3147,7 @@ static void free_receive_page_frags(struct virtnet_info *vi)
> static void free_unused_bufs(struct virtnet_info *vi)
> {
> void *buf;
> + u32 n;
> int i;
> struct send_queue *sq;
>
> @@ -2740,6 +3166,11 @@ static void free_unused_bufs(struct virtnet_info *vi)
> xdp_return_frame(xtype_got_ptr(xtype));
> }
> }
> +
> + n = sq->xsk.hdr_con + sq->xsk.hdr_n;
> + n -= sq->xsk.hdr_pro;
> + if (n)
> + virt_xsk_complete(sq, n, false);
> }
>
> for (i = 0; i < vi->max_queue_pairs; i++) {
> --
> 1.8.3.1