1 /* A network driver using virtio.
3 * Copyright 2007 Rusty Russell <rusty@rustcorp.com.au> IBM Corporation
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation; either version 2 of the License, or
8 * (at your option) any later version.
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, see <http://www.gnu.org/licenses/>.
19 #include <linux/netdevice.h>
20 #include <linux/etherdevice.h>
21 #include <linux/ethtool.h>
22 #include <linux/module.h>
23 #include <linux/virtio.h>
24 #include <linux/virtio_net.h>
25 #include <linux/bpf.h>
26 #include <linux/scatterlist.h>
27 #include <linux/if_vlan.h>
28 #include <linux/slab.h>
29 #include <linux/cpu.h>
30 #include <linux/average.h>
31 #include <net/busy_poll.h>
33 static int napi_weight = NAPI_POLL_WEIGHT;
34 module_param(napi_weight, int, 0444);
36 static bool csum = true, gso = true;
37 module_param(csum, bool, 0444);
38 module_param(gso, bool, 0444);
40 /* FIXME: MTU in config. */
41 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
42 #define GOOD_COPY_LEN 128
44 /* RX packet size EWMA. The average packet size is used to determine the packet
45 * buffer size when refilling RX rings. As the entire RX ring may be refilled
46 * at once, the weight is chosen so that the EWMA will be insensitive to short-
47 * term, transient changes in packet size.
49 DECLARE_EWMA(pkt_len, 1, 64)
51 /* Minimum alignment for mergeable packet buffers. */
52 #define MERGEABLE_BUFFER_ALIGN max(L1_CACHE_BYTES, 256)
54 #define VIRTNET_DRIVER_VERSION "1.0.0"
56 struct virtnet_stats {
57 struct u64_stats_sync tx_syncp;
58 struct u64_stats_sync rx_syncp;
66 /* Internal representation of a send virtqueue */
68 /* Virtqueue associated with this send _queue */
71 /* TX: fragments + linear part + virtio header */
72 struct scatterlist sg[MAX_SKB_FRAGS + 2];
74 /* Name of the send queue: output.$index */
78 /* Internal representation of a receive virtqueue */
79 struct receive_queue {
80 /* Virtqueue associated with this receive_queue */
83 struct napi_struct napi;
85 struct bpf_prog __rcu *xdp_prog;
87 /* Chain pages by the private ptr. */
90 /* Average packet length for mergeable receive buffers. */
91 struct ewma_pkt_len mrg_avg_pkt_len;
93 /* Page frag for packet buffer allocation. */
94 struct page_frag alloc_frag;
96 /* RX: fragments + linear part + virtio header */
97 struct scatterlist sg[MAX_SKB_FRAGS + 2];
99 /* Name of this receive queue: input.$index */
103 struct virtnet_info {
104 struct virtio_device *vdev;
105 struct virtqueue *cvq;
106 struct net_device *dev;
107 struct send_queue *sq;
108 struct receive_queue *rq;
111 /* Max # of queue pairs supported by the device */
114 /* # of queue pairs currently used by the driver */
115 u16 curr_queue_pairs;
117 /* # of XDP queue pairs currently used by the driver */
120 /* I like... big packets and I cannot lie! */
123 /* Host will merge rx buffers for big packets (shake it! shake it!) */
124 bool mergeable_rx_bufs;
126 /* Has control virtqueue */
129 /* Host can handle any s/g split between our header and packet data */
132 /* Packet virtio header size */
135 /* Active statistics */
136 struct virtnet_stats __percpu *stats;
138 /* Work struct for refilling if we run low on memory. */
139 struct delayed_work refill;
141 /* Work struct for config space updates */
142 struct work_struct config_work;
144 /* Does the affinity hint is set for virtqueues? */
145 bool affinity_hint_set;
147 /* CPU hotplug instances for online & dead */
148 struct hlist_node node;
149 struct hlist_node node_dead;
151 /* Control VQ buffers: protected by the rtnl lock */
152 struct virtio_net_ctrl_hdr ctrl_hdr;
153 virtio_net_ctrl_ack ctrl_status;
154 struct virtio_net_ctrl_mq ctrl_mq;
159 /* Ethtool settings */
164 struct padded_vnet_hdr {
165 struct virtio_net_hdr_mrg_rxbuf hdr;
167 * hdr is in a separate sg buffer, and data sg buffer shares same page
168 * with this header sg. This padding makes next sg 16 byte aligned
174 /* Converting between virtqueue no. and kernel tx/rx queue no.
175 * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
177 static int vq2txq(struct virtqueue *vq)
179 return (vq->index - 1) / 2;
182 static int txq2vq(int txq)
187 static int vq2rxq(struct virtqueue *vq)
189 return vq->index / 2;
192 static int rxq2vq(int rxq)
197 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
199 return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
203 * private is used to chain pages for big packets, put the whole
204 * most recent used list in the beginning for reuse
206 static void give_pages(struct receive_queue *rq, struct page *page)
210 /* Find end of list, sew whole thing into vi->rq.pages. */
211 for (end = page; end->private; end = (struct page *)end->private);
212 end->private = (unsigned long)rq->pages;
216 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
218 struct page *p = rq->pages;
221 rq->pages = (struct page *)p->private;
222 /* clear private here, it is used to chain pages */
225 p = alloc_page(gfp_mask);
229 static void skb_xmit_done(struct virtqueue *vq)
231 struct virtnet_info *vi = vq->vdev->priv;
233 /* Suppress further interrupts. */
234 virtqueue_disable_cb(vq);
236 /* We were probably waiting for more output buffers. */
237 netif_wake_subqueue(vi->dev, vq2txq(vq));
240 static unsigned int mergeable_ctx_to_buf_truesize(unsigned long mrg_ctx)
242 unsigned int truesize = mrg_ctx & (MERGEABLE_BUFFER_ALIGN - 1);
243 return (truesize + 1) * MERGEABLE_BUFFER_ALIGN;
246 static void *mergeable_ctx_to_buf_address(unsigned long mrg_ctx)
248 return (void *)(mrg_ctx & -MERGEABLE_BUFFER_ALIGN);
252 static unsigned long mergeable_buf_to_ctx(void *buf, unsigned int truesize)
254 unsigned int size = truesize / MERGEABLE_BUFFER_ALIGN;
255 return (unsigned long)buf | (size - 1);
258 /* Called from bottom half context */
259 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
260 struct receive_queue *rq,
261 struct page *page, unsigned int offset,
262 unsigned int len, unsigned int truesize)
265 struct virtio_net_hdr_mrg_rxbuf *hdr;
266 unsigned int copy, hdr_len, hdr_padded_len;
269 p = page_address(page) + offset;
271 /* copy small packet so we can reuse these pages for small data */
272 skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
276 hdr = skb_vnet_hdr(skb);
278 hdr_len = vi->hdr_len;
279 if (vi->mergeable_rx_bufs)
280 hdr_padded_len = sizeof *hdr;
282 hdr_padded_len = sizeof(struct padded_vnet_hdr);
284 memcpy(hdr, p, hdr_len);
287 offset += hdr_padded_len;
291 if (copy > skb_tailroom(skb))
292 copy = skb_tailroom(skb);
293 memcpy(skb_put(skb, copy), p, copy);
298 if (vi->mergeable_rx_bufs) {
300 skb_add_rx_frag(skb, 0, page, offset, len, truesize);
307 * Verify that we can indeed put this data into a skb.
308 * This is here to handle cases when the device erroneously
309 * tries to receive more than is possible. This is usually
310 * the case of a broken device.
312 if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
313 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
317 BUG_ON(offset >= PAGE_SIZE);
319 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
320 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
321 frag_size, truesize);
323 page = (struct page *)page->private;
328 give_pages(rq, page);
333 static u32 do_xdp_prog(struct virtnet_info *vi,
334 struct bpf_prog *xdp_prog,
335 struct page *page, int offset, int len)
342 buf = page_address(page) + offset;
344 if (vi->mergeable_rx_bufs)
345 hdr_padded_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
347 hdr_padded_len = sizeof(struct padded_vnet_hdr);
349 xdp.data = buf + hdr_padded_len;
350 xdp.data_end = xdp.data + (len - vi->hdr_len);
352 act = bpf_prog_run_xdp(xdp_prog, &xdp);
357 bpf_warn_invalid_xdp_action(act);
365 static struct sk_buff *receive_small(struct virtnet_info *vi, void *buf, unsigned int len)
367 struct sk_buff * skb = buf;
375 static struct sk_buff *receive_big(struct net_device *dev,
376 struct virtnet_info *vi,
377 struct receive_queue *rq,
381 struct bpf_prog *xdp_prog;
382 struct page *page = buf;
386 xdp_prog = rcu_dereference(rq->xdp_prog);
388 struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
391 if (unlikely(hdr->hdr.gso_type || hdr->hdr.flags))
393 act = do_xdp_prog(vi, xdp_prog, page, 0, len);
399 skb = page_to_skb(vi, rq, page, 0, len, PAGE_SIZE);
408 dev->stats.rx_dropped++;
409 give_pages(rq, page);
413 static struct sk_buff *receive_mergeable(struct net_device *dev,
414 struct virtnet_info *vi,
415 struct receive_queue *rq,
419 void *buf = mergeable_ctx_to_buf_address(ctx);
420 struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
421 u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
422 struct page *page = virt_to_head_page(buf);
423 int offset = buf - page_address(page);
424 struct sk_buff *head_skb, *curr_skb;
425 struct bpf_prog *xdp_prog;
426 unsigned int truesize;
429 xdp_prog = rcu_dereference(rq->xdp_prog);
433 /* No known backend devices should send packets with
434 * more than a single buffer when XDP conditions are
435 * met. However it is not strictly illegal so the case
436 * is handled as an exception and a warning is thrown.
438 if (unlikely(num_buf > 1)) {
439 bpf_warn_invalid_xdp_buffer();
443 /* Transient failure which in theory could occur if
444 * in-flight packets from before XDP was enabled reach
445 * the receive path after XDP is loaded. In practice I
446 * was not able to create this condition.
448 if (unlikely(hdr->hdr.gso_type || hdr->hdr.flags))
451 act = do_xdp_prog(vi, xdp_prog, page, offset, len);
457 truesize = max(len, mergeable_ctx_to_buf_truesize(ctx));
458 head_skb = page_to_skb(vi, rq, page, offset, len, truesize);
461 if (unlikely(!curr_skb))
466 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &len);
467 if (unlikely(!ctx)) {
468 pr_debug("%s: rx error: %d buffers out of %d missing\n",
470 virtio16_to_cpu(vi->vdev,
472 dev->stats.rx_length_errors++;
476 buf = mergeable_ctx_to_buf_address(ctx);
477 page = virt_to_head_page(buf);
479 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
480 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
481 struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
485 if (curr_skb == head_skb)
486 skb_shinfo(curr_skb)->frag_list = nskb;
488 curr_skb->next = nskb;
490 head_skb->truesize += nskb->truesize;
493 truesize = max(len, mergeable_ctx_to_buf_truesize(ctx));
494 if (curr_skb != head_skb) {
495 head_skb->data_len += len;
496 head_skb->len += len;
497 head_skb->truesize += truesize;
499 offset = buf - page_address(page);
500 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
502 skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
505 skb_add_rx_frag(curr_skb, num_skb_frags, page,
506 offset, len, truesize);
510 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
518 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &len);
519 if (unlikely(!ctx)) {
520 pr_debug("%s: rx error: %d buffers missing\n",
522 dev->stats.rx_length_errors++;
525 page = virt_to_head_page(mergeable_ctx_to_buf_address(ctx));
529 dev->stats.rx_dropped++;
530 dev_kfree_skb(head_skb);
534 static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
535 void *buf, unsigned int len)
537 struct net_device *dev = vi->dev;
538 struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
540 struct virtio_net_hdr_mrg_rxbuf *hdr;
542 if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
543 pr_debug("%s: short packet %i\n", dev->name, len);
544 dev->stats.rx_length_errors++;
545 if (vi->mergeable_rx_bufs) {
546 unsigned long ctx = (unsigned long)buf;
547 void *base = mergeable_ctx_to_buf_address(ctx);
548 put_page(virt_to_head_page(base));
549 } else if (vi->big_packets) {
557 if (vi->mergeable_rx_bufs)
558 skb = receive_mergeable(dev, vi, rq, (unsigned long)buf, len);
559 else if (vi->big_packets)
560 skb = receive_big(dev, vi, rq, buf, len);
562 skb = receive_small(vi, buf, len);
567 hdr = skb_vnet_hdr(skb);
569 u64_stats_update_begin(&stats->rx_syncp);
570 stats->rx_bytes += skb->len;
572 u64_stats_update_end(&stats->rx_syncp);
574 if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
575 skb->ip_summed = CHECKSUM_UNNECESSARY;
577 if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
578 virtio_is_little_endian(vi->vdev))) {
579 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
580 dev->name, hdr->hdr.gso_type,
585 skb->protocol = eth_type_trans(skb, dev);
586 pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
587 ntohs(skb->protocol), skb->len, skb->pkt_type);
589 napi_gro_receive(&rq->napi, skb);
593 dev->stats.rx_frame_errors++;
597 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
601 struct virtio_net_hdr_mrg_rxbuf *hdr;
604 skb = __netdev_alloc_skb_ip_align(vi->dev, GOOD_PACKET_LEN, gfp);
608 skb_put(skb, GOOD_PACKET_LEN);
610 hdr = skb_vnet_hdr(skb);
611 sg_init_table(rq->sg, 2);
612 sg_set_buf(rq->sg, hdr, vi->hdr_len);
613 skb_to_sgvec(skb, rq->sg + 1, 0, skb->len);
615 err = virtqueue_add_inbuf(rq->vq, rq->sg, 2, skb, gfp);
622 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
625 struct page *first, *list = NULL;
629 sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
631 /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
632 for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
633 first = get_a_page(rq, gfp);
636 give_pages(rq, list);
639 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
641 /* chain new page in list head to match sg */
642 first->private = (unsigned long)list;
646 first = get_a_page(rq, gfp);
648 give_pages(rq, list);
651 p = page_address(first);
653 /* rq->sg[0], rq->sg[1] share the same page */
654 /* a separated rq->sg[0] for header - required in case !any_header_sg */
655 sg_set_buf(&rq->sg[0], p, vi->hdr_len);
657 /* rq->sg[1] for data packet, from offset */
658 offset = sizeof(struct padded_vnet_hdr);
659 sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
661 /* chain first in list head */
662 first->private = (unsigned long)list;
663 err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
666 give_pages(rq, first);
671 static unsigned int get_mergeable_buf_len(struct ewma_pkt_len *avg_pkt_len)
673 const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
676 len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
677 GOOD_PACKET_LEN, PAGE_SIZE - hdr_len);
678 return ALIGN(len, MERGEABLE_BUFFER_ALIGN);
681 static int add_recvbuf_mergeable(struct receive_queue *rq, gfp_t gfp)
683 struct page_frag *alloc_frag = &rq->alloc_frag;
687 unsigned int len, hole;
689 len = get_mergeable_buf_len(&rq->mrg_avg_pkt_len);
690 if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
693 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
694 ctx = mergeable_buf_to_ctx(buf, len);
695 get_page(alloc_frag->page);
696 alloc_frag->offset += len;
697 hole = alloc_frag->size - alloc_frag->offset;
699 /* To avoid internal fragmentation, if there is very likely not
700 * enough space for another buffer, add the remaining space to
701 * the current buffer. This extra space is not included in
702 * the truesize stored in ctx.
705 alloc_frag->offset += hole;
708 sg_init_one(rq->sg, buf, len);
709 err = virtqueue_add_inbuf(rq->vq, rq->sg, 1, (void *)ctx, gfp);
711 put_page(virt_to_head_page(buf));
717 * Returns false if we couldn't fill entirely (OOM).
719 * Normally run in the receive path, but can also be run from ndo_open
720 * before we're receiving packets, or from refill_work which is
721 * careful to disable receiving (using napi_disable).
723 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
731 if (vi->mergeable_rx_bufs)
732 err = add_recvbuf_mergeable(rq, gfp);
733 else if (vi->big_packets)
734 err = add_recvbuf_big(vi, rq, gfp);
736 err = add_recvbuf_small(vi, rq, gfp);
738 oom = err == -ENOMEM;
741 } while (rq->vq->num_free);
742 virtqueue_kick(rq->vq);
746 static void skb_recv_done(struct virtqueue *rvq)
748 struct virtnet_info *vi = rvq->vdev->priv;
749 struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
751 /* Schedule NAPI, Suppress further interrupts if successful. */
752 if (napi_schedule_prep(&rq->napi)) {
753 virtqueue_disable_cb(rvq);
754 __napi_schedule(&rq->napi);
758 static void virtnet_napi_enable(struct receive_queue *rq)
760 napi_enable(&rq->napi);
762 /* If all buffers were filled by other side before we napi_enabled, we
763 * won't get another interrupt, so process any outstanding packets
764 * now. virtnet_poll wants re-enable the queue, so we disable here.
765 * We synchronize against interrupts via NAPI_STATE_SCHED */
766 if (napi_schedule_prep(&rq->napi)) {
767 virtqueue_disable_cb(rq->vq);
769 __napi_schedule(&rq->napi);
774 static void refill_work(struct work_struct *work)
776 struct virtnet_info *vi =
777 container_of(work, struct virtnet_info, refill.work);
781 for (i = 0; i < vi->curr_queue_pairs; i++) {
782 struct receive_queue *rq = &vi->rq[i];
784 napi_disable(&rq->napi);
785 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
786 virtnet_napi_enable(rq);
788 /* In theory, this can happen: if we don't get any buffers in
789 * we will *never* try to fill again.
792 schedule_delayed_work(&vi->refill, HZ/2);
796 static int virtnet_receive(struct receive_queue *rq, int budget)
798 struct virtnet_info *vi = rq->vq->vdev->priv;
799 unsigned int len, received = 0;
802 while (received < budget &&
803 (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
804 receive_buf(vi, rq, buf, len);
808 if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) {
809 if (!try_fill_recv(vi, rq, GFP_ATOMIC))
810 schedule_delayed_work(&vi->refill, 0);
816 static int virtnet_poll(struct napi_struct *napi, int budget)
818 struct receive_queue *rq =
819 container_of(napi, struct receive_queue, napi);
820 unsigned int r, received;
822 received = virtnet_receive(rq, budget);
824 /* Out of packets? */
825 if (received < budget) {
826 r = virtqueue_enable_cb_prepare(rq->vq);
827 napi_complete_done(napi, received);
828 if (unlikely(virtqueue_poll(rq->vq, r)) &&
829 napi_schedule_prep(napi)) {
830 virtqueue_disable_cb(rq->vq);
831 __napi_schedule(napi);
838 #ifdef CONFIG_NET_RX_BUSY_POLL
839 /* must be called with local_bh_disable()d */
840 static int virtnet_busy_poll(struct napi_struct *napi)
842 struct receive_queue *rq =
843 container_of(napi, struct receive_queue, napi);
844 struct virtnet_info *vi = rq->vq->vdev->priv;
845 int r, received = 0, budget = 4;
847 if (!(vi->status & VIRTIO_NET_S_LINK_UP))
848 return LL_FLUSH_FAILED;
850 if (!napi_schedule_prep(napi))
851 return LL_FLUSH_BUSY;
853 virtqueue_disable_cb(rq->vq);
856 received += virtnet_receive(rq, budget);
858 r = virtqueue_enable_cb_prepare(rq->vq);
859 clear_bit(NAPI_STATE_SCHED, &napi->state);
860 if (unlikely(virtqueue_poll(rq->vq, r)) &&
861 napi_schedule_prep(napi)) {
862 virtqueue_disable_cb(rq->vq);
863 if (received < budget) {
867 __napi_schedule(napi);
873 #endif /* CONFIG_NET_RX_BUSY_POLL */
875 static int virtnet_open(struct net_device *dev)
877 struct virtnet_info *vi = netdev_priv(dev);
880 for (i = 0; i < vi->max_queue_pairs; i++) {
881 if (i < vi->curr_queue_pairs)
882 /* Make sure we have some buffers: if oom use wq. */
883 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
884 schedule_delayed_work(&vi->refill, 0);
885 virtnet_napi_enable(&vi->rq[i]);
891 static void free_old_xmit_skbs(struct send_queue *sq)
895 struct virtnet_info *vi = sq->vq->vdev->priv;
896 struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
898 while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
899 pr_debug("Sent skb %p\n", skb);
901 u64_stats_update_begin(&stats->tx_syncp);
902 stats->tx_bytes += skb->len;
904 u64_stats_update_end(&stats->tx_syncp);
906 dev_kfree_skb_any(skb);
910 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
912 struct virtio_net_hdr_mrg_rxbuf *hdr;
913 const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
914 struct virtnet_info *vi = sq->vq->vdev->priv;
916 unsigned hdr_len = vi->hdr_len;
919 pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
921 can_push = vi->any_header_sg &&
922 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
923 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
924 /* Even if we can, don't push here yet as this would skew
925 * csum_start offset below. */
927 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
929 hdr = skb_vnet_hdr(skb);
931 if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
932 virtio_is_little_endian(vi->vdev)))
935 if (vi->mergeable_rx_bufs)
936 hdr->num_buffers = 0;
938 sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
940 __skb_push(skb, hdr_len);
941 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
942 /* Pull header back to avoid skew in tx bytes calculations. */
943 __skb_pull(skb, hdr_len);
945 sg_set_buf(sq->sg, hdr, hdr_len);
946 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
948 return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
951 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
953 struct virtnet_info *vi = netdev_priv(dev);
954 int qnum = skb_get_queue_mapping(skb);
955 struct send_queue *sq = &vi->sq[qnum];
957 struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
958 bool kick = !skb->xmit_more;
960 /* Free up any pending old buffers before queueing new ones. */
961 free_old_xmit_skbs(sq);
963 /* timestamp packet in software */
964 skb_tx_timestamp(skb);
966 /* Try to transmit */
967 err = xmit_skb(sq, skb);
969 /* This should not happen! */
971 dev->stats.tx_fifo_errors++;
974 "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
975 dev->stats.tx_dropped++;
976 dev_kfree_skb_any(skb);
980 /* Don't wait up for transmitted skbs to be freed. */
984 /* If running out of space, stop queue to avoid getting packets that we
985 * are then unable to transmit.
986 * An alternative would be to force queuing layer to requeue the skb by
987 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
988 * returned in a normal path of operation: it means that driver is not
989 * maintaining the TX queue stop/start state properly, and causes
990 * the stack to do a non-trivial amount of useless work.
991 * Since most packets only take 1 or 2 ring slots, stopping the queue
992 * early means 16 slots are typically wasted.
994 if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
995 netif_stop_subqueue(dev, qnum);
996 if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
997 /* More just got used, free them then recheck. */
998 free_old_xmit_skbs(sq);
999 if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
1000 netif_start_subqueue(dev, qnum);
1001 virtqueue_disable_cb(sq->vq);
1006 if (kick || netif_xmit_stopped(txq))
1007 virtqueue_kick(sq->vq);
1009 return NETDEV_TX_OK;
1013 * Send command via the control virtqueue and check status. Commands
1014 * supported by the hypervisor, as indicated by feature bits, should
1015 * never fail unless improperly formatted.
1017 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
1018 struct scatterlist *out)
1020 struct scatterlist *sgs[4], hdr, stat;
1021 unsigned out_num = 0, tmp;
1023 /* Caller should know better */
1024 BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
1026 vi->ctrl_status = ~0;
1027 vi->ctrl_hdr.class = class;
1028 vi->ctrl_hdr.cmd = cmd;
1030 sg_init_one(&hdr, &vi->ctrl_hdr, sizeof(vi->ctrl_hdr));
1031 sgs[out_num++] = &hdr;
1034 sgs[out_num++] = out;
1036 /* Add return status. */
1037 sg_init_one(&stat, &vi->ctrl_status, sizeof(vi->ctrl_status));
1038 sgs[out_num] = &stat;
1040 BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1041 virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1043 if (unlikely(!virtqueue_kick(vi->cvq)))
1044 return vi->ctrl_status == VIRTIO_NET_OK;
1046 /* Spin for a response, the kick causes an ioport write, trapping
1047 * into the hypervisor, so the request should be handled immediately.
1049 while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1050 !virtqueue_is_broken(vi->cvq))
1053 return vi->ctrl_status == VIRTIO_NET_OK;
1056 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1058 struct virtnet_info *vi = netdev_priv(dev);
1059 struct virtio_device *vdev = vi->vdev;
1061 struct sockaddr *addr;
1062 struct scatterlist sg;
1064 addr = kmalloc(sizeof(*addr), GFP_KERNEL);
1067 memcpy(addr, p, sizeof(*addr));
1069 ret = eth_prepare_mac_addr_change(dev, addr);
1073 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1074 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1075 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1076 VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1077 dev_warn(&vdev->dev,
1078 "Failed to set mac address by vq command.\n");
1082 } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1083 !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1086 /* Naturally, this has an atomicity problem. */
1087 for (i = 0; i < dev->addr_len; i++)
1088 virtio_cwrite8(vdev,
1089 offsetof(struct virtio_net_config, mac) +
1090 i, addr->sa_data[i]);
1093 eth_commit_mac_addr_change(dev, p);
1101 static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
1102 struct rtnl_link_stats64 *tot)
1104 struct virtnet_info *vi = netdev_priv(dev);
1108 for_each_possible_cpu(cpu) {
1109 struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
1110 u64 tpackets, tbytes, rpackets, rbytes;
1113 start = u64_stats_fetch_begin_irq(&stats->tx_syncp);
1114 tpackets = stats->tx_packets;
1115 tbytes = stats->tx_bytes;
1116 } while (u64_stats_fetch_retry_irq(&stats->tx_syncp, start));
1119 start = u64_stats_fetch_begin_irq(&stats->rx_syncp);
1120 rpackets = stats->rx_packets;
1121 rbytes = stats->rx_bytes;
1122 } while (u64_stats_fetch_retry_irq(&stats->rx_syncp, start));
1124 tot->rx_packets += rpackets;
1125 tot->tx_packets += tpackets;
1126 tot->rx_bytes += rbytes;
1127 tot->tx_bytes += tbytes;
1130 tot->tx_dropped = dev->stats.tx_dropped;
1131 tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1132 tot->rx_dropped = dev->stats.rx_dropped;
1133 tot->rx_length_errors = dev->stats.rx_length_errors;
1134 tot->rx_frame_errors = dev->stats.rx_frame_errors;
1139 #ifdef CONFIG_NET_POLL_CONTROLLER
1140 static void virtnet_netpoll(struct net_device *dev)
1142 struct virtnet_info *vi = netdev_priv(dev);
1145 for (i = 0; i < vi->curr_queue_pairs; i++)
1146 napi_schedule(&vi->rq[i].napi);
1150 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1153 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1154 VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1155 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1159 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1161 struct scatterlist sg;
1162 struct net_device *dev = vi->dev;
1164 if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1167 vi->ctrl_mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1168 sg_init_one(&sg, &vi->ctrl_mq, sizeof(vi->ctrl_mq));
1170 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1171 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1172 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1176 vi->curr_queue_pairs = queue_pairs;
1177 /* virtnet_open() will refill when device is going to up. */
1178 if (dev->flags & IFF_UP)
1179 schedule_delayed_work(&vi->refill, 0);
1185 static int virtnet_close(struct net_device *dev)
1187 struct virtnet_info *vi = netdev_priv(dev);
1190 /* Make sure refill_work doesn't re-enable napi! */
1191 cancel_delayed_work_sync(&vi->refill);
1193 for (i = 0; i < vi->max_queue_pairs; i++)
1194 napi_disable(&vi->rq[i].napi);
1199 static void virtnet_set_rx_mode(struct net_device *dev)
1201 struct virtnet_info *vi = netdev_priv(dev);
1202 struct scatterlist sg[2];
1203 struct virtio_net_ctrl_mac *mac_data;
1204 struct netdev_hw_addr *ha;
1210 /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1211 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1214 vi->ctrl_promisc = ((dev->flags & IFF_PROMISC) != 0);
1215 vi->ctrl_allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1217 sg_init_one(sg, &vi->ctrl_promisc, sizeof(vi->ctrl_promisc));
1219 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1220 VIRTIO_NET_CTRL_RX_PROMISC, sg))
1221 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1222 vi->ctrl_promisc ? "en" : "dis");
1224 sg_init_one(sg, &vi->ctrl_allmulti, sizeof(vi->ctrl_allmulti));
1226 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1227 VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1228 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1229 vi->ctrl_allmulti ? "en" : "dis");
1231 uc_count = netdev_uc_count(dev);
1232 mc_count = netdev_mc_count(dev);
1233 /* MAC filter - use one buffer for both lists */
1234 buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1235 (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1240 sg_init_table(sg, 2);
1242 /* Store the unicast list and count in the front of the buffer */
1243 mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1245 netdev_for_each_uc_addr(ha, dev)
1246 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1248 sg_set_buf(&sg[0], mac_data,
1249 sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1251 /* multicast list and count fill the end */
1252 mac_data = (void *)&mac_data->macs[uc_count][0];
1254 mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1256 netdev_for_each_mc_addr(ha, dev)
1257 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1259 sg_set_buf(&sg[1], mac_data,
1260 sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1262 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1263 VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1264 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1269 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1270 __be16 proto, u16 vid)
1272 struct virtnet_info *vi = netdev_priv(dev);
1273 struct scatterlist sg;
1276 sg_init_one(&sg, &vi->ctrl_vid, sizeof(vi->ctrl_vid));
1278 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1279 VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1280 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1284 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1285 __be16 proto, u16 vid)
1287 struct virtnet_info *vi = netdev_priv(dev);
1288 struct scatterlist sg;
1291 sg_init_one(&sg, &vi->ctrl_vid, sizeof(vi->ctrl_vid));
1293 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1294 VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1295 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1299 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1303 if (vi->affinity_hint_set) {
1304 for (i = 0; i < vi->max_queue_pairs; i++) {
1305 virtqueue_set_affinity(vi->rq[i].vq, -1);
1306 virtqueue_set_affinity(vi->sq[i].vq, -1);
1309 vi->affinity_hint_set = false;
1313 static void virtnet_set_affinity(struct virtnet_info *vi)
1318 /* In multiqueue mode, when the number of cpu is equal to the number of
1319 * queue pairs, we let the queue pairs to be private to one cpu by
1320 * setting the affinity hint to eliminate the contention.
1322 if (vi->curr_queue_pairs == 1 ||
1323 vi->max_queue_pairs != num_online_cpus()) {
1324 virtnet_clean_affinity(vi, -1);
1329 for_each_online_cpu(cpu) {
1330 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1331 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1332 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
1336 vi->affinity_hint_set = true;
1339 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
1341 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1343 virtnet_set_affinity(vi);
1347 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
1349 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1351 virtnet_set_affinity(vi);
1355 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
1357 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1360 virtnet_clean_affinity(vi, cpu);
1364 static enum cpuhp_state virtionet_online;
1366 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
1370 ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
1373 ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1377 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1381 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
1383 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1384 cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1388 static void virtnet_get_ringparam(struct net_device *dev,
1389 struct ethtool_ringparam *ring)
1391 struct virtnet_info *vi = netdev_priv(dev);
1393 ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1394 ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1395 ring->rx_pending = ring->rx_max_pending;
1396 ring->tx_pending = ring->tx_max_pending;
1400 static void virtnet_get_drvinfo(struct net_device *dev,
1401 struct ethtool_drvinfo *info)
1403 struct virtnet_info *vi = netdev_priv(dev);
1404 struct virtio_device *vdev = vi->vdev;
1406 strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1407 strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1408 strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1412 /* TODO: Eliminate OOO packets during switching */
1413 static int virtnet_set_channels(struct net_device *dev,
1414 struct ethtool_channels *channels)
1416 struct virtnet_info *vi = netdev_priv(dev);
1417 u16 queue_pairs = channels->combined_count;
1420 /* We don't support separate rx/tx channels.
1421 * We don't allow setting 'other' channels.
1423 if (channels->rx_count || channels->tx_count || channels->other_count)
1426 if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
1429 /* For now we don't support modifying channels while XDP is loaded
1430 * also when XDP is loaded all RX queues have XDP programs so we only
1431 * need to check a single RX queue.
1433 if (vi->rq[0].xdp_prog)
1437 err = virtnet_set_queues(vi, queue_pairs);
1439 netif_set_real_num_tx_queues(dev, queue_pairs);
1440 netif_set_real_num_rx_queues(dev, queue_pairs);
1442 virtnet_set_affinity(vi);
1449 static void virtnet_get_channels(struct net_device *dev,
1450 struct ethtool_channels *channels)
1452 struct virtnet_info *vi = netdev_priv(dev);
1454 channels->combined_count = vi->curr_queue_pairs;
1455 channels->max_combined = vi->max_queue_pairs;
1456 channels->max_other = 0;
1457 channels->rx_count = 0;
1458 channels->tx_count = 0;
1459 channels->other_count = 0;
1462 /* Check if the user is trying to change anything besides speed/duplex */
1463 static bool virtnet_validate_ethtool_cmd(const struct ethtool_cmd *cmd)
1465 struct ethtool_cmd diff1 = *cmd;
1466 struct ethtool_cmd diff2 = {};
1468 /* cmd is always set so we need to clear it, validate the port type
1469 * and also without autonegotiation we can ignore advertising
1471 ethtool_cmd_speed_set(&diff1, 0);
1472 diff2.port = PORT_OTHER;
1473 diff1.advertising = 0;
1477 return !memcmp(&diff1, &diff2, sizeof(diff1));
1480 static int virtnet_set_settings(struct net_device *dev, struct ethtool_cmd *cmd)
1482 struct virtnet_info *vi = netdev_priv(dev);
1485 speed = ethtool_cmd_speed(cmd);
1486 /* don't allow custom speed and duplex */
1487 if (!ethtool_validate_speed(speed) ||
1488 !ethtool_validate_duplex(cmd->duplex) ||
1489 !virtnet_validate_ethtool_cmd(cmd))
1492 vi->duplex = cmd->duplex;
1497 static int virtnet_get_settings(struct net_device *dev, struct ethtool_cmd *cmd)
1499 struct virtnet_info *vi = netdev_priv(dev);
1501 ethtool_cmd_speed_set(cmd, vi->speed);
1502 cmd->duplex = vi->duplex;
1503 cmd->port = PORT_OTHER;
1508 static void virtnet_init_settings(struct net_device *dev)
1510 struct virtnet_info *vi = netdev_priv(dev);
1512 vi->speed = SPEED_UNKNOWN;
1513 vi->duplex = DUPLEX_UNKNOWN;
1516 static const struct ethtool_ops virtnet_ethtool_ops = {
1517 .get_drvinfo = virtnet_get_drvinfo,
1518 .get_link = ethtool_op_get_link,
1519 .get_ringparam = virtnet_get_ringparam,
1520 .set_channels = virtnet_set_channels,
1521 .get_channels = virtnet_get_channels,
1522 .get_ts_info = ethtool_op_get_ts_info,
1523 .get_settings = virtnet_get_settings,
1524 .set_settings = virtnet_set_settings,
1527 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog)
1529 unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr);
1530 struct virtnet_info *vi = netdev_priv(dev);
1531 struct bpf_prog *old_prog;
1532 u16 xdp_qp = 0, curr_qp;
1535 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1536 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6)) {
1537 netdev_warn(dev, "can't set XDP while host is implementing LRO, disable LRO first\n");
1541 if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
1542 netdev_warn(dev, "XDP expects header/data in single page, any_header_sg required\n");
1546 if (dev->mtu > max_sz) {
1547 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz);
1551 curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
1553 xdp_qp = nr_cpu_ids;
1555 /* XDP requires extra queues for XDP_TX */
1556 if (curr_qp + xdp_qp > vi->max_queue_pairs) {
1557 netdev_warn(dev, "request %i queues but max is %i\n",
1558 curr_qp + xdp_qp, vi->max_queue_pairs);
1562 err = virtnet_set_queues(vi, curr_qp + xdp_qp);
1564 dev_warn(&dev->dev, "XDP Device queue allocation failure.\n");
1569 prog = bpf_prog_add(prog, vi->max_queue_pairs - 1);
1571 virtnet_set_queues(vi, curr_qp);
1572 return PTR_ERR(prog);
1576 vi->xdp_queue_pairs = xdp_qp;
1577 netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
1579 for (i = 0; i < vi->max_queue_pairs; i++) {
1580 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
1581 rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
1583 bpf_prog_put(old_prog);
1589 static bool virtnet_xdp_query(struct net_device *dev)
1591 struct virtnet_info *vi = netdev_priv(dev);
1594 for (i = 0; i < vi->max_queue_pairs; i++) {
1595 if (vi->rq[i].xdp_prog)
1601 static int virtnet_xdp(struct net_device *dev, struct netdev_xdp *xdp)
1603 switch (xdp->command) {
1604 case XDP_SETUP_PROG:
1605 return virtnet_xdp_set(dev, xdp->prog);
1606 case XDP_QUERY_PROG:
1607 xdp->prog_attached = virtnet_xdp_query(dev);
1614 static const struct net_device_ops virtnet_netdev = {
1615 .ndo_open = virtnet_open,
1616 .ndo_stop = virtnet_close,
1617 .ndo_start_xmit = start_xmit,
1618 .ndo_validate_addr = eth_validate_addr,
1619 .ndo_set_mac_address = virtnet_set_mac_address,
1620 .ndo_set_rx_mode = virtnet_set_rx_mode,
1621 .ndo_get_stats64 = virtnet_stats,
1622 .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
1623 .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
1624 #ifdef CONFIG_NET_POLL_CONTROLLER
1625 .ndo_poll_controller = virtnet_netpoll,
1627 #ifdef CONFIG_NET_RX_BUSY_POLL
1628 .ndo_busy_poll = virtnet_busy_poll,
1630 .ndo_xdp = virtnet_xdp,
1633 static void virtnet_config_changed_work(struct work_struct *work)
1635 struct virtnet_info *vi =
1636 container_of(work, struct virtnet_info, config_work);
1639 if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
1640 struct virtio_net_config, status, &v) < 0)
1643 if (v & VIRTIO_NET_S_ANNOUNCE) {
1644 netdev_notify_peers(vi->dev);
1645 virtnet_ack_link_announce(vi);
1648 /* Ignore unknown (future) status bits */
1649 v &= VIRTIO_NET_S_LINK_UP;
1651 if (vi->status == v)
1656 if (vi->status & VIRTIO_NET_S_LINK_UP) {
1657 netif_carrier_on(vi->dev);
1658 netif_tx_wake_all_queues(vi->dev);
1660 netif_carrier_off(vi->dev);
1661 netif_tx_stop_all_queues(vi->dev);
1665 static void virtnet_config_changed(struct virtio_device *vdev)
1667 struct virtnet_info *vi = vdev->priv;
1669 schedule_work(&vi->config_work);
1672 static void virtnet_free_queues(struct virtnet_info *vi)
1676 for (i = 0; i < vi->max_queue_pairs; i++) {
1677 napi_hash_del(&vi->rq[i].napi);
1678 netif_napi_del(&vi->rq[i].napi);
1681 /* We called napi_hash_del() before netif_napi_del(),
1682 * we need to respect an RCU grace period before freeing vi->rq
1690 static void free_receive_bufs(struct virtnet_info *vi)
1692 struct bpf_prog *old_prog;
1696 for (i = 0; i < vi->max_queue_pairs; i++) {
1697 while (vi->rq[i].pages)
1698 __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
1700 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
1701 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
1703 bpf_prog_put(old_prog);
1708 static void free_receive_page_frags(struct virtnet_info *vi)
1711 for (i = 0; i < vi->max_queue_pairs; i++)
1712 if (vi->rq[i].alloc_frag.page)
1713 put_page(vi->rq[i].alloc_frag.page);
1716 static void free_unused_bufs(struct virtnet_info *vi)
1721 for (i = 0; i < vi->max_queue_pairs; i++) {
1722 struct virtqueue *vq = vi->sq[i].vq;
1723 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
1727 for (i = 0; i < vi->max_queue_pairs; i++) {
1728 struct virtqueue *vq = vi->rq[i].vq;
1730 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1731 if (vi->mergeable_rx_bufs) {
1732 unsigned long ctx = (unsigned long)buf;
1733 void *base = mergeable_ctx_to_buf_address(ctx);
1734 put_page(virt_to_head_page(base));
1735 } else if (vi->big_packets) {
1736 give_pages(&vi->rq[i], buf);
1744 static void virtnet_del_vqs(struct virtnet_info *vi)
1746 struct virtio_device *vdev = vi->vdev;
1748 virtnet_clean_affinity(vi, -1);
1750 vdev->config->del_vqs(vdev);
1752 virtnet_free_queues(vi);
1755 static int virtnet_find_vqs(struct virtnet_info *vi)
1757 vq_callback_t **callbacks;
1758 struct virtqueue **vqs;
1763 /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
1764 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
1765 * possible control vq.
1767 total_vqs = vi->max_queue_pairs * 2 +
1768 virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
1770 /* Allocate space for find_vqs parameters */
1771 vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
1774 callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
1777 names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
1781 /* Parameters for control virtqueue, if any */
1783 callbacks[total_vqs - 1] = NULL;
1784 names[total_vqs - 1] = "control";
1787 /* Allocate/initialize parameters for send/receive virtqueues */
1788 for (i = 0; i < vi->max_queue_pairs; i++) {
1789 callbacks[rxq2vq(i)] = skb_recv_done;
1790 callbacks[txq2vq(i)] = skb_xmit_done;
1791 sprintf(vi->rq[i].name, "input.%d", i);
1792 sprintf(vi->sq[i].name, "output.%d", i);
1793 names[rxq2vq(i)] = vi->rq[i].name;
1794 names[txq2vq(i)] = vi->sq[i].name;
1797 ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
1803 vi->cvq = vqs[total_vqs - 1];
1804 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
1805 vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
1808 for (i = 0; i < vi->max_queue_pairs; i++) {
1809 vi->rq[i].vq = vqs[rxq2vq(i)];
1810 vi->sq[i].vq = vqs[txq2vq(i)];
1829 static int virtnet_alloc_queues(struct virtnet_info *vi)
1833 vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
1836 vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
1840 INIT_DELAYED_WORK(&vi->refill, refill_work);
1841 for (i = 0; i < vi->max_queue_pairs; i++) {
1842 vi->rq[i].pages = NULL;
1843 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
1846 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
1847 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
1848 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
1859 static int init_vqs(struct virtnet_info *vi)
1863 /* Allocate send & receive queues */
1864 ret = virtnet_alloc_queues(vi);
1868 ret = virtnet_find_vqs(vi);
1873 virtnet_set_affinity(vi);
1879 virtnet_free_queues(vi);
1885 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
1886 struct rx_queue_attribute *attribute, char *buf)
1888 struct virtnet_info *vi = netdev_priv(queue->dev);
1889 unsigned int queue_index = get_netdev_rx_queue_index(queue);
1890 struct ewma_pkt_len *avg;
1892 BUG_ON(queue_index >= vi->max_queue_pairs);
1893 avg = &vi->rq[queue_index].mrg_avg_pkt_len;
1894 return sprintf(buf, "%u\n", get_mergeable_buf_len(avg));
1897 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
1898 __ATTR_RO(mergeable_rx_buffer_size);
1900 static struct attribute *virtio_net_mrg_rx_attrs[] = {
1901 &mergeable_rx_buffer_size_attribute.attr,
1905 static const struct attribute_group virtio_net_mrg_rx_group = {
1906 .name = "virtio_net",
1907 .attrs = virtio_net_mrg_rx_attrs
1911 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
1913 const char *fname, const char *dname)
1915 if (!virtio_has_feature(vdev, fbit))
1918 dev_err(&vdev->dev, "device advertises feature %s but not %s",
1924 #define VIRTNET_FAIL_ON(vdev, fbit, dbit) \
1925 virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
1927 static bool virtnet_validate_features(struct virtio_device *vdev)
1929 if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
1930 (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
1931 "VIRTIO_NET_F_CTRL_VQ") ||
1932 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
1933 "VIRTIO_NET_F_CTRL_VQ") ||
1934 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
1935 "VIRTIO_NET_F_CTRL_VQ") ||
1936 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
1937 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
1938 "VIRTIO_NET_F_CTRL_VQ"))) {
1945 #define MIN_MTU ETH_MIN_MTU
1946 #define MAX_MTU ETH_MAX_MTU
1948 static int virtnet_probe(struct virtio_device *vdev)
1951 struct net_device *dev;
1952 struct virtnet_info *vi;
1953 u16 max_queue_pairs;
1956 if (!vdev->config->get) {
1957 dev_err(&vdev->dev, "%s failure: config access disabled\n",
1962 if (!virtnet_validate_features(vdev))
1965 /* Find if host supports multiqueue virtio_net device */
1966 err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
1967 struct virtio_net_config,
1968 max_virtqueue_pairs, &max_queue_pairs);
1970 /* We need at least 2 queue's */
1971 if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
1972 max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
1973 !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1974 max_queue_pairs = 1;
1976 /* Allocate ourselves a network device with room for our info */
1977 dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
1981 /* Set up network device as normal. */
1982 dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
1983 dev->netdev_ops = &virtnet_netdev;
1984 dev->features = NETIF_F_HIGHDMA;
1986 dev->ethtool_ops = &virtnet_ethtool_ops;
1987 SET_NETDEV_DEV(dev, &vdev->dev);
1989 /* Do we support "hardware" checksums? */
1990 if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
1991 /* This opens up the world of extra features. */
1992 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
1994 dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
1996 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
1997 dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
1998 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
2000 /* Individual feature bits: what can host handle? */
2001 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
2002 dev->hw_features |= NETIF_F_TSO;
2003 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
2004 dev->hw_features |= NETIF_F_TSO6;
2005 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
2006 dev->hw_features |= NETIF_F_TSO_ECN;
2007 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
2008 dev->hw_features |= NETIF_F_UFO;
2010 dev->features |= NETIF_F_GSO_ROBUST;
2013 dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
2014 /* (!csum && gso) case will be fixed by register_netdev() */
2016 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
2017 dev->features |= NETIF_F_RXCSUM;
2019 dev->vlan_features = dev->features;
2021 /* MTU range: 68 - 65535 */
2022 dev->min_mtu = MIN_MTU;
2023 dev->max_mtu = MAX_MTU;
2025 /* Configuration may specify what MAC to use. Otherwise random. */
2026 if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
2027 virtio_cread_bytes(vdev,
2028 offsetof(struct virtio_net_config, mac),
2029 dev->dev_addr, dev->addr_len);
2031 eth_hw_addr_random(dev);
2033 /* Set up our device-specific information */
2034 vi = netdev_priv(dev);
2038 vi->stats = alloc_percpu(struct virtnet_stats);
2040 if (vi->stats == NULL)
2043 for_each_possible_cpu(i) {
2044 struct virtnet_stats *virtnet_stats;
2045 virtnet_stats = per_cpu_ptr(vi->stats, i);
2046 u64_stats_init(&virtnet_stats->tx_syncp);
2047 u64_stats_init(&virtnet_stats->rx_syncp);
2050 INIT_WORK(&vi->config_work, virtnet_config_changed_work);
2052 /* If we can receive ANY GSO packets, we must allocate large ones. */
2053 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2054 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2055 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
2056 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
2057 vi->big_packets = true;
2059 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
2060 vi->mergeable_rx_bufs = true;
2062 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
2063 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2064 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2066 vi->hdr_len = sizeof(struct virtio_net_hdr);
2068 if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
2069 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2070 vi->any_header_sg = true;
2072 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2075 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2076 mtu = virtio_cread16(vdev,
2077 offsetof(struct virtio_net_config,
2079 if (mtu < dev->min_mtu) {
2080 __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
2087 if (vi->any_header_sg)
2088 dev->needed_headroom = vi->hdr_len;
2090 /* Enable multiqueue by default */
2091 if (num_online_cpus() >= max_queue_pairs)
2092 vi->curr_queue_pairs = max_queue_pairs;
2094 vi->curr_queue_pairs = num_online_cpus();
2095 vi->max_queue_pairs = max_queue_pairs;
2097 /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
2103 if (vi->mergeable_rx_bufs)
2104 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
2106 netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
2107 netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
2109 virtnet_init_settings(dev);
2111 err = register_netdev(dev);
2113 pr_debug("virtio_net: registering device failed\n");
2117 virtio_device_ready(vdev);
2119 err = virtnet_cpu_notif_add(vi);
2121 pr_debug("virtio_net: registering cpu notifier failed\n");
2122 goto free_unregister_netdev;
2126 virtnet_set_queues(vi, vi->curr_queue_pairs);
2129 /* Assume link up if device can't report link status,
2130 otherwise get link status from config. */
2131 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
2132 netif_carrier_off(dev);
2133 schedule_work(&vi->config_work);
2135 vi->status = VIRTIO_NET_S_LINK_UP;
2136 netif_carrier_on(dev);
2139 pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
2140 dev->name, max_queue_pairs);
2144 free_unregister_netdev:
2145 vi->vdev->config->reset(vdev);
2147 unregister_netdev(dev);
2149 cancel_delayed_work_sync(&vi->refill);
2150 free_receive_page_frags(vi);
2151 virtnet_del_vqs(vi);
2153 free_percpu(vi->stats);
2159 static void remove_vq_common(struct virtnet_info *vi)
2161 vi->vdev->config->reset(vi->vdev);
2163 /* Free unused buffers in both send and recv, if any. */
2164 free_unused_bufs(vi);
2166 free_receive_bufs(vi);
2168 free_receive_page_frags(vi);
2170 virtnet_del_vqs(vi);
2173 static void virtnet_remove(struct virtio_device *vdev)
2175 struct virtnet_info *vi = vdev->priv;
2177 virtnet_cpu_notif_remove(vi);
2179 /* Make sure no work handler is accessing the device. */
2180 flush_work(&vi->config_work);
2182 unregister_netdev(vi->dev);
2184 remove_vq_common(vi);
2186 free_percpu(vi->stats);
2187 free_netdev(vi->dev);
2190 #ifdef CONFIG_PM_SLEEP
2191 static int virtnet_freeze(struct virtio_device *vdev)
2193 struct virtnet_info *vi = vdev->priv;
2196 virtnet_cpu_notif_remove(vi);
2198 /* Make sure no work handler is accessing the device */
2199 flush_work(&vi->config_work);
2201 netif_device_detach(vi->dev);
2202 cancel_delayed_work_sync(&vi->refill);
2204 if (netif_running(vi->dev)) {
2205 for (i = 0; i < vi->max_queue_pairs; i++)
2206 napi_disable(&vi->rq[i].napi);
2209 remove_vq_common(vi);
2214 static int virtnet_restore(struct virtio_device *vdev)
2216 struct virtnet_info *vi = vdev->priv;
2223 virtio_device_ready(vdev);
2225 if (netif_running(vi->dev)) {
2226 for (i = 0; i < vi->curr_queue_pairs; i++)
2227 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
2228 schedule_delayed_work(&vi->refill, 0);
2230 for (i = 0; i < vi->max_queue_pairs; i++)
2231 virtnet_napi_enable(&vi->rq[i]);
2234 netif_device_attach(vi->dev);
2237 virtnet_set_queues(vi, vi->curr_queue_pairs);
2240 err = virtnet_cpu_notif_add(vi);
2248 static struct virtio_device_id id_table[] = {
2249 { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
2253 #define VIRTNET_FEATURES \
2254 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
2256 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
2257 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
2258 VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
2259 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
2260 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
2261 VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
2262 VIRTIO_NET_F_CTRL_MAC_ADDR, \
2265 static unsigned int features[] = {
2269 static unsigned int features_legacy[] = {
2272 VIRTIO_F_ANY_LAYOUT,
2275 static struct virtio_driver virtio_net_driver = {
2276 .feature_table = features,
2277 .feature_table_size = ARRAY_SIZE(features),
2278 .feature_table_legacy = features_legacy,
2279 .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
2280 .driver.name = KBUILD_MODNAME,
2281 .driver.owner = THIS_MODULE,
2282 .id_table = id_table,
2283 .probe = virtnet_probe,
2284 .remove = virtnet_remove,
2285 .config_changed = virtnet_config_changed,
2286 #ifdef CONFIG_PM_SLEEP
2287 .freeze = virtnet_freeze,
2288 .restore = virtnet_restore,
2292 static __init int virtio_net_driver_init(void)
2296 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "AP_VIRT_NET_ONLINE",
2298 virtnet_cpu_down_prep);
2301 virtionet_online = ret;
2302 ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "VIRT_NET_DEAD",
2303 NULL, virtnet_cpu_dead);
2307 ret = register_virtio_driver(&virtio_net_driver);
2312 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
2314 cpuhp_remove_multi_state(virtionet_online);
2318 module_init(virtio_net_driver_init);
2320 static __exit void virtio_net_driver_exit(void)
2322 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
2323 cpuhp_remove_multi_state(virtionet_online);
2324 unregister_virtio_driver(&virtio_net_driver);
2326 module_exit(virtio_net_driver_exit);
2328 MODULE_DEVICE_TABLE(virtio, id_table);
2329 MODULE_DESCRIPTION("Virtio network driver");
2330 MODULE_LICENSE("GPL");