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/scatterlist.h>
26 #include <linux/if_vlan.h>
27 #include <linux/slab.h>
28 #include <linux/cpu.h>
29 #include <linux/average.h>
30 #include <net/busy_poll.h>
32 static int napi_weight = NAPI_POLL_WEIGHT;
33 module_param(napi_weight, int, 0444);
35 static bool csum = true, gso = true;
36 module_param(csum, bool, 0444);
37 module_param(gso, bool, 0444);
39 /* FIXME: MTU in config. */
40 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
41 #define GOOD_COPY_LEN 128
43 /* Weight used for the RX packet size EWMA. The average packet size is used to
44 * determine the packet buffer size when refilling RX rings. As the entire RX
45 * ring may be refilled at once, the weight is chosen so that the EWMA will be
46 * insensitive to short-term, transient changes in packet size.
48 #define RECEIVE_AVG_WEIGHT 64
50 /* Minimum alignment for mergeable packet buffers. */
51 #define MERGEABLE_BUFFER_ALIGN max(L1_CACHE_BYTES, 256)
53 #define VIRTNET_DRIVER_VERSION "1.0.0"
55 struct virtnet_stats {
56 struct u64_stats_sync tx_syncp;
57 struct u64_stats_sync rx_syncp;
65 /* Internal representation of a send virtqueue */
67 /* Virtqueue associated with this send _queue */
70 /* TX: fragments + linear part + virtio header */
71 struct scatterlist sg[MAX_SKB_FRAGS + 2];
73 /* Name of the send queue: output.$index */
77 /* Internal representation of a receive virtqueue */
78 struct receive_queue {
79 /* Virtqueue associated with this receive_queue */
82 struct napi_struct napi;
84 /* Chain pages by the private ptr. */
87 /* Average packet length for mergeable receive buffers. */
88 struct ewma mrg_avg_pkt_len;
90 /* Page frag for packet buffer allocation. */
91 struct page_frag alloc_frag;
93 /* RX: fragments + linear part + virtio header */
94 struct scatterlist sg[MAX_SKB_FRAGS + 2];
96 /* Name of this receive queue: input.$index */
100 struct virtnet_info {
101 struct virtio_device *vdev;
102 struct virtqueue *cvq;
103 struct net_device *dev;
104 struct send_queue *sq;
105 struct receive_queue *rq;
108 /* Max # of queue pairs supported by the device */
111 /* # of queue pairs currently used by the driver */
112 u16 curr_queue_pairs;
114 /* I like... big packets and I cannot lie! */
117 /* Host will merge rx buffers for big packets (shake it! shake it!) */
118 bool mergeable_rx_bufs;
120 /* Has control virtqueue */
123 /* Host can handle any s/g split between our header and packet data */
126 /* Packet virtio header size */
129 /* Active statistics */
130 struct virtnet_stats __percpu *stats;
132 /* Work struct for refilling if we run low on memory. */
133 struct delayed_work refill;
135 /* Work struct for config space updates */
136 struct work_struct config_work;
138 /* Does the affinity hint is set for virtqueues? */
139 bool affinity_hint_set;
141 /* CPU hot plug notifier */
142 struct notifier_block nb;
145 struct padded_vnet_hdr {
146 struct virtio_net_hdr_mrg_rxbuf hdr;
148 * hdr is in a separate sg buffer, and data sg buffer shares same page
149 * with this header sg. This padding makes next sg 16 byte aligned
155 /* Converting between virtqueue no. and kernel tx/rx queue no.
156 * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
158 static int vq2txq(struct virtqueue *vq)
160 return (vq->index - 1) / 2;
163 static int txq2vq(int txq)
168 static int vq2rxq(struct virtqueue *vq)
170 return vq->index / 2;
173 static int rxq2vq(int rxq)
178 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
180 return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
184 * private is used to chain pages for big packets, put the whole
185 * most recent used list in the beginning for reuse
187 static void give_pages(struct receive_queue *rq, struct page *page)
191 /* Find end of list, sew whole thing into vi->rq.pages. */
192 for (end = page; end->private; end = (struct page *)end->private);
193 end->private = (unsigned long)rq->pages;
197 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
199 struct page *p = rq->pages;
202 rq->pages = (struct page *)p->private;
203 /* clear private here, it is used to chain pages */
206 p = alloc_page(gfp_mask);
210 static void skb_xmit_done(struct virtqueue *vq)
212 struct virtnet_info *vi = vq->vdev->priv;
214 /* Suppress further interrupts. */
215 virtqueue_disable_cb(vq);
217 /* We were probably waiting for more output buffers. */
218 netif_wake_subqueue(vi->dev, vq2txq(vq));
221 static unsigned int mergeable_ctx_to_buf_truesize(unsigned long mrg_ctx)
223 unsigned int truesize = mrg_ctx & (MERGEABLE_BUFFER_ALIGN - 1);
224 return (truesize + 1) * MERGEABLE_BUFFER_ALIGN;
227 static void *mergeable_ctx_to_buf_address(unsigned long mrg_ctx)
229 return (void *)(mrg_ctx & -MERGEABLE_BUFFER_ALIGN);
233 static unsigned long mergeable_buf_to_ctx(void *buf, unsigned int truesize)
235 unsigned int size = truesize / MERGEABLE_BUFFER_ALIGN;
236 return (unsigned long)buf | (size - 1);
239 /* Called from bottom half context */
240 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
241 struct receive_queue *rq,
242 struct page *page, unsigned int offset,
243 unsigned int len, unsigned int truesize)
246 struct virtio_net_hdr_mrg_rxbuf *hdr;
247 unsigned int copy, hdr_len, hdr_padded_len;
250 p = page_address(page) + offset;
252 /* copy small packet so we can reuse these pages for small data */
253 skb = netdev_alloc_skb_ip_align(vi->dev, GOOD_COPY_LEN);
257 hdr = skb_vnet_hdr(skb);
259 hdr_len = vi->hdr_len;
260 if (vi->mergeable_rx_bufs)
261 hdr_padded_len = sizeof *hdr;
263 hdr_padded_len = sizeof(struct padded_vnet_hdr);
265 memcpy(hdr, p, hdr_len);
268 offset += hdr_padded_len;
272 if (copy > skb_tailroom(skb))
273 copy = skb_tailroom(skb);
274 memcpy(skb_put(skb, copy), p, copy);
279 if (vi->mergeable_rx_bufs) {
281 skb_add_rx_frag(skb, 0, page, offset, len, truesize);
288 * Verify that we can indeed put this data into a skb.
289 * This is here to handle cases when the device erroneously
290 * tries to receive more than is possible. This is usually
291 * the case of a broken device.
293 if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
294 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
298 BUG_ON(offset >= PAGE_SIZE);
300 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
301 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
302 frag_size, truesize);
304 page = (struct page *)page->private;
309 give_pages(rq, page);
314 static struct sk_buff *receive_small(struct virtnet_info *vi, void *buf, unsigned int len)
316 struct sk_buff * skb = buf;
324 static struct sk_buff *receive_big(struct net_device *dev,
325 struct virtnet_info *vi,
326 struct receive_queue *rq,
330 struct page *page = buf;
331 struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len, PAGE_SIZE);
339 dev->stats.rx_dropped++;
340 give_pages(rq, page);
344 static struct sk_buff *receive_mergeable(struct net_device *dev,
345 struct virtnet_info *vi,
346 struct receive_queue *rq,
350 void *buf = mergeable_ctx_to_buf_address(ctx);
351 struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
352 u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
353 struct page *page = virt_to_head_page(buf);
354 int offset = buf - page_address(page);
355 unsigned int truesize = max(len, mergeable_ctx_to_buf_truesize(ctx));
357 struct sk_buff *head_skb = page_to_skb(vi, rq, page, offset, len,
359 struct sk_buff *curr_skb = head_skb;
361 if (unlikely(!curr_skb))
366 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &len);
367 if (unlikely(!ctx)) {
368 pr_debug("%s: rx error: %d buffers out of %d missing\n",
370 virtio16_to_cpu(vi->vdev,
372 dev->stats.rx_length_errors++;
376 buf = mergeable_ctx_to_buf_address(ctx);
377 page = virt_to_head_page(buf);
379 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
380 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
381 struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
385 if (curr_skb == head_skb)
386 skb_shinfo(curr_skb)->frag_list = nskb;
388 curr_skb->next = nskb;
390 head_skb->truesize += nskb->truesize;
393 truesize = max(len, mergeable_ctx_to_buf_truesize(ctx));
394 if (curr_skb != head_skb) {
395 head_skb->data_len += len;
396 head_skb->len += len;
397 head_skb->truesize += truesize;
399 offset = buf - page_address(page);
400 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
402 skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
405 skb_add_rx_frag(curr_skb, num_skb_frags, page,
406 offset, len, truesize);
410 ewma_add(&rq->mrg_avg_pkt_len, head_skb->len);
416 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &len);
417 if (unlikely(!ctx)) {
418 pr_debug("%s: rx error: %d buffers missing\n",
420 dev->stats.rx_length_errors++;
423 page = virt_to_head_page(mergeable_ctx_to_buf_address(ctx));
427 dev->stats.rx_dropped++;
428 dev_kfree_skb(head_skb);
432 static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
433 void *buf, unsigned int len)
435 struct net_device *dev = vi->dev;
436 struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
438 struct virtio_net_hdr_mrg_rxbuf *hdr;
440 if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
441 pr_debug("%s: short packet %i\n", dev->name, len);
442 dev->stats.rx_length_errors++;
443 if (vi->mergeable_rx_bufs) {
444 unsigned long ctx = (unsigned long)buf;
445 void *base = mergeable_ctx_to_buf_address(ctx);
446 put_page(virt_to_head_page(base));
447 } else if (vi->big_packets) {
455 if (vi->mergeable_rx_bufs)
456 skb = receive_mergeable(dev, vi, rq, (unsigned long)buf, len);
457 else if (vi->big_packets)
458 skb = receive_big(dev, vi, rq, buf, len);
460 skb = receive_small(vi, buf, len);
465 hdr = skb_vnet_hdr(skb);
467 u64_stats_update_begin(&stats->rx_syncp);
468 stats->rx_bytes += skb->len;
470 u64_stats_update_end(&stats->rx_syncp);
472 if (hdr->hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
473 pr_debug("Needs csum!\n");
474 if (!skb_partial_csum_set(skb,
475 virtio16_to_cpu(vi->vdev, hdr->hdr.csum_start),
476 virtio16_to_cpu(vi->vdev, hdr->hdr.csum_offset)))
478 } else if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID) {
479 skb->ip_summed = CHECKSUM_UNNECESSARY;
482 skb->protocol = eth_type_trans(skb, dev);
483 pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
484 ntohs(skb->protocol), skb->len, skb->pkt_type);
486 if (hdr->hdr.gso_type != VIRTIO_NET_HDR_GSO_NONE) {
488 switch (hdr->hdr.gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
489 case VIRTIO_NET_HDR_GSO_TCPV4:
490 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
492 case VIRTIO_NET_HDR_GSO_UDP:
499 "host using disabled UFO feature; please fix it\n");
501 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
504 case VIRTIO_NET_HDR_GSO_TCPV6:
505 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
508 net_warn_ratelimited("%s: bad gso type %u.\n",
509 dev->name, hdr->hdr.gso_type);
513 if (hdr->hdr.gso_type & VIRTIO_NET_HDR_GSO_ECN)
514 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
516 skb_shinfo(skb)->gso_size = virtio16_to_cpu(vi->vdev,
518 if (skb_shinfo(skb)->gso_size == 0) {
519 net_warn_ratelimited("%s: zero gso size.\n", dev->name);
523 /* Header must be checked, and gso_segs computed. */
524 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
525 skb_shinfo(skb)->gso_segs = 0;
528 skb_mark_napi_id(skb, &rq->napi);
530 netif_receive_skb(skb);
534 dev->stats.rx_frame_errors++;
538 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
542 struct virtio_net_hdr_mrg_rxbuf *hdr;
545 skb = __netdev_alloc_skb_ip_align(vi->dev, GOOD_PACKET_LEN, gfp);
549 skb_put(skb, GOOD_PACKET_LEN);
551 hdr = skb_vnet_hdr(skb);
552 sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
553 sg_set_buf(rq->sg, hdr, vi->hdr_len);
554 skb_to_sgvec(skb, rq->sg + 1, 0, skb->len);
556 err = virtqueue_add_inbuf(rq->vq, rq->sg, 2, skb, gfp);
563 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
566 struct page *first, *list = NULL;
570 sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
572 /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
573 for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
574 first = get_a_page(rq, gfp);
577 give_pages(rq, list);
580 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
582 /* chain new page in list head to match sg */
583 first->private = (unsigned long)list;
587 first = get_a_page(rq, gfp);
589 give_pages(rq, list);
592 p = page_address(first);
594 /* rq->sg[0], rq->sg[1] share the same page */
595 /* a separated rq->sg[0] for header - required in case !any_header_sg */
596 sg_set_buf(&rq->sg[0], p, vi->hdr_len);
598 /* rq->sg[1] for data packet, from offset */
599 offset = sizeof(struct padded_vnet_hdr);
600 sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
602 /* chain first in list head */
603 first->private = (unsigned long)list;
604 err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
607 give_pages(rq, first);
612 static unsigned int get_mergeable_buf_len(struct ewma *avg_pkt_len)
614 const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
617 len = hdr_len + clamp_t(unsigned int, ewma_read(avg_pkt_len),
618 GOOD_PACKET_LEN, PAGE_SIZE - hdr_len);
619 return ALIGN(len, MERGEABLE_BUFFER_ALIGN);
622 static int add_recvbuf_mergeable(struct receive_queue *rq, gfp_t gfp)
624 struct page_frag *alloc_frag = &rq->alloc_frag;
628 unsigned int len, hole;
630 len = get_mergeable_buf_len(&rq->mrg_avg_pkt_len);
631 if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
634 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
635 ctx = mergeable_buf_to_ctx(buf, len);
636 get_page(alloc_frag->page);
637 alloc_frag->offset += len;
638 hole = alloc_frag->size - alloc_frag->offset;
640 /* To avoid internal fragmentation, if there is very likely not
641 * enough space for another buffer, add the remaining space to
642 * the current buffer. This extra space is not included in
643 * the truesize stored in ctx.
646 alloc_frag->offset += hole;
649 sg_init_one(rq->sg, buf, len);
650 err = virtqueue_add_inbuf(rq->vq, rq->sg, 1, (void *)ctx, gfp);
652 put_page(virt_to_head_page(buf));
658 * Returns false if we couldn't fill entirely (OOM).
660 * Normally run in the receive path, but can also be run from ndo_open
661 * before we're receiving packets, or from refill_work which is
662 * careful to disable receiving (using napi_disable).
664 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
672 if (vi->mergeable_rx_bufs)
673 err = add_recvbuf_mergeable(rq, gfp);
674 else if (vi->big_packets)
675 err = add_recvbuf_big(vi, rq, gfp);
677 err = add_recvbuf_small(vi, rq, gfp);
679 oom = err == -ENOMEM;
682 } while (rq->vq->num_free);
683 virtqueue_kick(rq->vq);
687 static void skb_recv_done(struct virtqueue *rvq)
689 struct virtnet_info *vi = rvq->vdev->priv;
690 struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
692 /* Schedule NAPI, Suppress further interrupts if successful. */
693 if (napi_schedule_prep(&rq->napi)) {
694 virtqueue_disable_cb(rvq);
695 __napi_schedule(&rq->napi);
699 static void virtnet_napi_enable(struct receive_queue *rq)
701 napi_enable(&rq->napi);
703 /* If all buffers were filled by other side before we napi_enabled, we
704 * won't get another interrupt, so process any outstanding packets
705 * now. virtnet_poll wants re-enable the queue, so we disable here.
706 * We synchronize against interrupts via NAPI_STATE_SCHED */
707 if (napi_schedule_prep(&rq->napi)) {
708 virtqueue_disable_cb(rq->vq);
710 __napi_schedule(&rq->napi);
715 static void refill_work(struct work_struct *work)
717 struct virtnet_info *vi =
718 container_of(work, struct virtnet_info, refill.work);
722 for (i = 0; i < vi->curr_queue_pairs; i++) {
723 struct receive_queue *rq = &vi->rq[i];
725 napi_disable(&rq->napi);
726 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
727 virtnet_napi_enable(rq);
729 /* In theory, this can happen: if we don't get any buffers in
730 * we will *never* try to fill again.
733 schedule_delayed_work(&vi->refill, HZ/2);
737 static int virtnet_receive(struct receive_queue *rq, int budget)
739 struct virtnet_info *vi = rq->vq->vdev->priv;
740 unsigned int len, received = 0;
743 while (received < budget &&
744 (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
745 receive_buf(vi, rq, buf, len);
749 if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) {
750 if (!try_fill_recv(vi, rq, GFP_ATOMIC))
751 schedule_delayed_work(&vi->refill, 0);
757 static int virtnet_poll(struct napi_struct *napi, int budget)
759 struct receive_queue *rq =
760 container_of(napi, struct receive_queue, napi);
761 unsigned int r, received = 0;
763 received += virtnet_receive(rq, budget - received);
765 /* Out of packets? */
766 if (received < budget) {
767 r = virtqueue_enable_cb_prepare(rq->vq);
769 if (unlikely(virtqueue_poll(rq->vq, r)) &&
770 napi_schedule_prep(napi)) {
771 virtqueue_disable_cb(rq->vq);
772 __napi_schedule(napi);
779 #ifdef CONFIG_NET_RX_BUSY_POLL
780 /* must be called with local_bh_disable()d */
781 static int virtnet_busy_poll(struct napi_struct *napi)
783 struct receive_queue *rq =
784 container_of(napi, struct receive_queue, napi);
785 struct virtnet_info *vi = rq->vq->vdev->priv;
786 int r, received = 0, budget = 4;
788 if (!(vi->status & VIRTIO_NET_S_LINK_UP))
789 return LL_FLUSH_FAILED;
791 if (!napi_schedule_prep(napi))
792 return LL_FLUSH_BUSY;
794 virtqueue_disable_cb(rq->vq);
797 received += virtnet_receive(rq, budget);
799 r = virtqueue_enable_cb_prepare(rq->vq);
800 clear_bit(NAPI_STATE_SCHED, &napi->state);
801 if (unlikely(virtqueue_poll(rq->vq, r)) &&
802 napi_schedule_prep(napi)) {
803 virtqueue_disable_cb(rq->vq);
804 if (received < budget) {
808 __napi_schedule(napi);
814 #endif /* CONFIG_NET_RX_BUSY_POLL */
816 static int virtnet_open(struct net_device *dev)
818 struct virtnet_info *vi = netdev_priv(dev);
821 for (i = 0; i < vi->max_queue_pairs; i++) {
822 if (i < vi->curr_queue_pairs)
823 /* Make sure we have some buffers: if oom use wq. */
824 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
825 schedule_delayed_work(&vi->refill, 0);
826 virtnet_napi_enable(&vi->rq[i]);
832 static void free_old_xmit_skbs(struct send_queue *sq)
836 struct virtnet_info *vi = sq->vq->vdev->priv;
837 struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
839 while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
840 pr_debug("Sent skb %p\n", skb);
842 u64_stats_update_begin(&stats->tx_syncp);
843 stats->tx_bytes += skb->len;
845 u64_stats_update_end(&stats->tx_syncp);
847 dev_kfree_skb_any(skb);
851 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
853 struct virtio_net_hdr_mrg_rxbuf *hdr;
854 const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
855 struct virtnet_info *vi = sq->vq->vdev->priv;
857 unsigned hdr_len = vi->hdr_len;
860 pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
862 can_push = vi->any_header_sg &&
863 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
864 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
865 /* Even if we can, don't push here yet as this would skew
866 * csum_start offset below. */
868 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
870 hdr = skb_vnet_hdr(skb);
872 if (skb->ip_summed == CHECKSUM_PARTIAL) {
873 hdr->hdr.flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
874 hdr->hdr.csum_start = cpu_to_virtio16(vi->vdev,
875 skb_checksum_start_offset(skb));
876 hdr->hdr.csum_offset = cpu_to_virtio16(vi->vdev,
880 hdr->hdr.csum_offset = hdr->hdr.csum_start = 0;
883 if (skb_is_gso(skb)) {
884 hdr->hdr.hdr_len = cpu_to_virtio16(vi->vdev, skb_headlen(skb));
885 hdr->hdr.gso_size = cpu_to_virtio16(vi->vdev,
886 skb_shinfo(skb)->gso_size);
887 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
888 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
889 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
890 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
893 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCP_ECN)
894 hdr->hdr.gso_type |= VIRTIO_NET_HDR_GSO_ECN;
896 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE;
897 hdr->hdr.gso_size = hdr->hdr.hdr_len = 0;
900 if (vi->mergeable_rx_bufs)
901 hdr->num_buffers = 0;
903 sg_init_table(sq->sg, MAX_SKB_FRAGS + 2);
905 __skb_push(skb, hdr_len);
906 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
907 /* Pull header back to avoid skew in tx bytes calculations. */
908 __skb_pull(skb, hdr_len);
910 sg_set_buf(sq->sg, hdr, hdr_len);
911 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
913 return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
916 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
918 struct virtnet_info *vi = netdev_priv(dev);
919 int qnum = skb_get_queue_mapping(skb);
920 struct send_queue *sq = &vi->sq[qnum];
922 struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
923 bool kick = !skb->xmit_more;
925 /* Free up any pending old buffers before queueing new ones. */
926 free_old_xmit_skbs(sq);
928 /* Try to transmit */
929 err = xmit_skb(sq, skb);
931 /* This should not happen! */
933 dev->stats.tx_fifo_errors++;
936 "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
937 dev->stats.tx_dropped++;
938 dev_kfree_skb_any(skb);
942 /* Don't wait up for transmitted skbs to be freed. */
946 /* Apparently nice girls don't return TX_BUSY; stop the queue
947 * before it gets out of hand. Naturally, this wastes entries. */
948 if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
949 netif_stop_subqueue(dev, qnum);
950 if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
951 /* More just got used, free them then recheck. */
952 free_old_xmit_skbs(sq);
953 if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
954 netif_start_subqueue(dev, qnum);
955 virtqueue_disable_cb(sq->vq);
960 if (kick || netif_xmit_stopped(txq))
961 virtqueue_kick(sq->vq);
967 * Send command via the control virtqueue and check status. Commands
968 * supported by the hypervisor, as indicated by feature bits, should
969 * never fail unless improperly formatted.
971 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
972 struct scatterlist *out)
974 struct scatterlist *sgs[4], hdr, stat;
975 struct virtio_net_ctrl_hdr ctrl;
976 virtio_net_ctrl_ack status = ~0;
977 unsigned out_num = 0, tmp;
979 /* Caller should know better */
980 BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
985 sg_init_one(&hdr, &ctrl, sizeof(ctrl));
986 sgs[out_num++] = &hdr;
989 sgs[out_num++] = out;
991 /* Add return status. */
992 sg_init_one(&stat, &status, sizeof(status));
993 sgs[out_num] = &stat;
995 BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
996 virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
998 if (unlikely(!virtqueue_kick(vi->cvq)))
999 return status == VIRTIO_NET_OK;
1001 /* Spin for a response, the kick causes an ioport write, trapping
1002 * into the hypervisor, so the request should be handled immediately.
1004 while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1005 !virtqueue_is_broken(vi->cvq))
1008 return status == VIRTIO_NET_OK;
1011 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1013 struct virtnet_info *vi = netdev_priv(dev);
1014 struct virtio_device *vdev = vi->vdev;
1016 struct sockaddr *addr = p;
1017 struct scatterlist sg;
1019 ret = eth_prepare_mac_addr_change(dev, p);
1023 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1024 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1025 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1026 VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1027 dev_warn(&vdev->dev,
1028 "Failed to set mac address by vq command.\n");
1031 } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1032 !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1035 /* Naturally, this has an atomicity problem. */
1036 for (i = 0; i < dev->addr_len; i++)
1037 virtio_cwrite8(vdev,
1038 offsetof(struct virtio_net_config, mac) +
1039 i, addr->sa_data[i]);
1042 eth_commit_mac_addr_change(dev, p);
1047 static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
1048 struct rtnl_link_stats64 *tot)
1050 struct virtnet_info *vi = netdev_priv(dev);
1054 for_each_possible_cpu(cpu) {
1055 struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
1056 u64 tpackets, tbytes, rpackets, rbytes;
1059 start = u64_stats_fetch_begin_irq(&stats->tx_syncp);
1060 tpackets = stats->tx_packets;
1061 tbytes = stats->tx_bytes;
1062 } while (u64_stats_fetch_retry_irq(&stats->tx_syncp, start));
1065 start = u64_stats_fetch_begin_irq(&stats->rx_syncp);
1066 rpackets = stats->rx_packets;
1067 rbytes = stats->rx_bytes;
1068 } while (u64_stats_fetch_retry_irq(&stats->rx_syncp, start));
1070 tot->rx_packets += rpackets;
1071 tot->tx_packets += tpackets;
1072 tot->rx_bytes += rbytes;
1073 tot->tx_bytes += tbytes;
1076 tot->tx_dropped = dev->stats.tx_dropped;
1077 tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1078 tot->rx_dropped = dev->stats.rx_dropped;
1079 tot->rx_length_errors = dev->stats.rx_length_errors;
1080 tot->rx_frame_errors = dev->stats.rx_frame_errors;
1085 #ifdef CONFIG_NET_POLL_CONTROLLER
1086 static void virtnet_netpoll(struct net_device *dev)
1088 struct virtnet_info *vi = netdev_priv(dev);
1091 for (i = 0; i < vi->curr_queue_pairs; i++)
1092 napi_schedule(&vi->rq[i].napi);
1096 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1099 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1100 VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1101 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1105 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1107 struct scatterlist sg;
1108 struct virtio_net_ctrl_mq s;
1109 struct net_device *dev = vi->dev;
1111 if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1114 s.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1115 sg_init_one(&sg, &s, sizeof(s));
1117 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1118 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1119 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1123 vi->curr_queue_pairs = queue_pairs;
1124 /* virtnet_open() will refill when device is going to up. */
1125 if (dev->flags & IFF_UP)
1126 schedule_delayed_work(&vi->refill, 0);
1132 static int virtnet_close(struct net_device *dev)
1134 struct virtnet_info *vi = netdev_priv(dev);
1137 /* Make sure refill_work doesn't re-enable napi! */
1138 cancel_delayed_work_sync(&vi->refill);
1140 for (i = 0; i < vi->max_queue_pairs; i++)
1141 napi_disable(&vi->rq[i].napi);
1146 static void virtnet_set_rx_mode(struct net_device *dev)
1148 struct virtnet_info *vi = netdev_priv(dev);
1149 struct scatterlist sg[2];
1150 u8 promisc, allmulti;
1151 struct virtio_net_ctrl_mac *mac_data;
1152 struct netdev_hw_addr *ha;
1158 /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1159 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1162 promisc = ((dev->flags & IFF_PROMISC) != 0);
1163 allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1165 sg_init_one(sg, &promisc, sizeof(promisc));
1167 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1168 VIRTIO_NET_CTRL_RX_PROMISC, sg))
1169 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1170 promisc ? "en" : "dis");
1172 sg_init_one(sg, &allmulti, sizeof(allmulti));
1174 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1175 VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1176 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1177 allmulti ? "en" : "dis");
1179 uc_count = netdev_uc_count(dev);
1180 mc_count = netdev_mc_count(dev);
1181 /* MAC filter - use one buffer for both lists */
1182 buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1183 (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1188 sg_init_table(sg, 2);
1190 /* Store the unicast list and count in the front of the buffer */
1191 mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1193 netdev_for_each_uc_addr(ha, dev)
1194 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1196 sg_set_buf(&sg[0], mac_data,
1197 sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1199 /* multicast list and count fill the end */
1200 mac_data = (void *)&mac_data->macs[uc_count][0];
1202 mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1204 netdev_for_each_mc_addr(ha, dev)
1205 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1207 sg_set_buf(&sg[1], mac_data,
1208 sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1210 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1211 VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1212 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1217 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1218 __be16 proto, u16 vid)
1220 struct virtnet_info *vi = netdev_priv(dev);
1221 struct scatterlist sg;
1223 sg_init_one(&sg, &vid, sizeof(vid));
1225 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1226 VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1227 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1231 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1232 __be16 proto, u16 vid)
1234 struct virtnet_info *vi = netdev_priv(dev);
1235 struct scatterlist sg;
1237 sg_init_one(&sg, &vid, sizeof(vid));
1239 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1240 VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1241 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1245 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1249 if (vi->affinity_hint_set) {
1250 for (i = 0; i < vi->max_queue_pairs; i++) {
1251 virtqueue_set_affinity(vi->rq[i].vq, -1);
1252 virtqueue_set_affinity(vi->sq[i].vq, -1);
1255 vi->affinity_hint_set = false;
1259 static void virtnet_set_affinity(struct virtnet_info *vi)
1264 /* In multiqueue mode, when the number of cpu is equal to the number of
1265 * queue pairs, we let the queue pairs to be private to one cpu by
1266 * setting the affinity hint to eliminate the contention.
1268 if (vi->curr_queue_pairs == 1 ||
1269 vi->max_queue_pairs != num_online_cpus()) {
1270 virtnet_clean_affinity(vi, -1);
1275 for_each_online_cpu(cpu) {
1276 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1277 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1278 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
1282 vi->affinity_hint_set = true;
1285 static int virtnet_cpu_callback(struct notifier_block *nfb,
1286 unsigned long action, void *hcpu)
1288 struct virtnet_info *vi = container_of(nfb, struct virtnet_info, nb);
1290 switch(action & ~CPU_TASKS_FROZEN) {
1292 case CPU_DOWN_FAILED:
1294 virtnet_set_affinity(vi);
1296 case CPU_DOWN_PREPARE:
1297 virtnet_clean_affinity(vi, (long)hcpu);
1306 static void virtnet_get_ringparam(struct net_device *dev,
1307 struct ethtool_ringparam *ring)
1309 struct virtnet_info *vi = netdev_priv(dev);
1311 ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1312 ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1313 ring->rx_pending = ring->rx_max_pending;
1314 ring->tx_pending = ring->tx_max_pending;
1318 static void virtnet_get_drvinfo(struct net_device *dev,
1319 struct ethtool_drvinfo *info)
1321 struct virtnet_info *vi = netdev_priv(dev);
1322 struct virtio_device *vdev = vi->vdev;
1324 strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1325 strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1326 strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1330 /* TODO: Eliminate OOO packets during switching */
1331 static int virtnet_set_channels(struct net_device *dev,
1332 struct ethtool_channels *channels)
1334 struct virtnet_info *vi = netdev_priv(dev);
1335 u16 queue_pairs = channels->combined_count;
1338 /* We don't support separate rx/tx channels.
1339 * We don't allow setting 'other' channels.
1341 if (channels->rx_count || channels->tx_count || channels->other_count)
1344 if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
1348 err = virtnet_set_queues(vi, queue_pairs);
1350 netif_set_real_num_tx_queues(dev, queue_pairs);
1351 netif_set_real_num_rx_queues(dev, queue_pairs);
1353 virtnet_set_affinity(vi);
1360 static void virtnet_get_channels(struct net_device *dev,
1361 struct ethtool_channels *channels)
1363 struct virtnet_info *vi = netdev_priv(dev);
1365 channels->combined_count = vi->curr_queue_pairs;
1366 channels->max_combined = vi->max_queue_pairs;
1367 channels->max_other = 0;
1368 channels->rx_count = 0;
1369 channels->tx_count = 0;
1370 channels->other_count = 0;
1373 static const struct ethtool_ops virtnet_ethtool_ops = {
1374 .get_drvinfo = virtnet_get_drvinfo,
1375 .get_link = ethtool_op_get_link,
1376 .get_ringparam = virtnet_get_ringparam,
1377 .set_channels = virtnet_set_channels,
1378 .get_channels = virtnet_get_channels,
1382 #define MAX_MTU 65535
1384 static int virtnet_change_mtu(struct net_device *dev, int new_mtu)
1386 if (new_mtu < MIN_MTU || new_mtu > MAX_MTU)
1392 static const struct net_device_ops virtnet_netdev = {
1393 .ndo_open = virtnet_open,
1394 .ndo_stop = virtnet_close,
1395 .ndo_start_xmit = start_xmit,
1396 .ndo_validate_addr = eth_validate_addr,
1397 .ndo_set_mac_address = virtnet_set_mac_address,
1398 .ndo_set_rx_mode = virtnet_set_rx_mode,
1399 .ndo_change_mtu = virtnet_change_mtu,
1400 .ndo_get_stats64 = virtnet_stats,
1401 .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
1402 .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
1403 #ifdef CONFIG_NET_POLL_CONTROLLER
1404 .ndo_poll_controller = virtnet_netpoll,
1406 #ifdef CONFIG_NET_RX_BUSY_POLL
1407 .ndo_busy_poll = virtnet_busy_poll,
1411 static void virtnet_config_changed_work(struct work_struct *work)
1413 struct virtnet_info *vi =
1414 container_of(work, struct virtnet_info, config_work);
1417 if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
1418 struct virtio_net_config, status, &v) < 0)
1421 if (v & VIRTIO_NET_S_ANNOUNCE) {
1422 netdev_notify_peers(vi->dev);
1423 virtnet_ack_link_announce(vi);
1426 /* Ignore unknown (future) status bits */
1427 v &= VIRTIO_NET_S_LINK_UP;
1429 if (vi->status == v)
1434 if (vi->status & VIRTIO_NET_S_LINK_UP) {
1435 netif_carrier_on(vi->dev);
1436 netif_tx_wake_all_queues(vi->dev);
1438 netif_carrier_off(vi->dev);
1439 netif_tx_stop_all_queues(vi->dev);
1443 static void virtnet_config_changed(struct virtio_device *vdev)
1445 struct virtnet_info *vi = vdev->priv;
1447 schedule_work(&vi->config_work);
1450 static void virtnet_free_queues(struct virtnet_info *vi)
1454 for (i = 0; i < vi->max_queue_pairs; i++)
1455 netif_napi_del(&vi->rq[i].napi);
1461 static void free_receive_bufs(struct virtnet_info *vi)
1465 for (i = 0; i < vi->max_queue_pairs; i++) {
1466 while (vi->rq[i].pages)
1467 __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
1471 static void free_receive_page_frags(struct virtnet_info *vi)
1474 for (i = 0; i < vi->max_queue_pairs; i++)
1475 if (vi->rq[i].alloc_frag.page)
1476 put_page(vi->rq[i].alloc_frag.page);
1479 static void free_unused_bufs(struct virtnet_info *vi)
1484 for (i = 0; i < vi->max_queue_pairs; i++) {
1485 struct virtqueue *vq = vi->sq[i].vq;
1486 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
1490 for (i = 0; i < vi->max_queue_pairs; i++) {
1491 struct virtqueue *vq = vi->rq[i].vq;
1493 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1494 if (vi->mergeable_rx_bufs) {
1495 unsigned long ctx = (unsigned long)buf;
1496 void *base = mergeable_ctx_to_buf_address(ctx);
1497 put_page(virt_to_head_page(base));
1498 } else if (vi->big_packets) {
1499 give_pages(&vi->rq[i], buf);
1507 static void virtnet_del_vqs(struct virtnet_info *vi)
1509 struct virtio_device *vdev = vi->vdev;
1511 virtnet_clean_affinity(vi, -1);
1513 vdev->config->del_vqs(vdev);
1515 virtnet_free_queues(vi);
1518 static int virtnet_find_vqs(struct virtnet_info *vi)
1520 vq_callback_t **callbacks;
1521 struct virtqueue **vqs;
1526 /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
1527 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
1528 * possible control vq.
1530 total_vqs = vi->max_queue_pairs * 2 +
1531 virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
1533 /* Allocate space for find_vqs parameters */
1534 vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
1537 callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
1540 names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
1544 /* Parameters for control virtqueue, if any */
1546 callbacks[total_vqs - 1] = NULL;
1547 names[total_vqs - 1] = "control";
1550 /* Allocate/initialize parameters for send/receive virtqueues */
1551 for (i = 0; i < vi->max_queue_pairs; i++) {
1552 callbacks[rxq2vq(i)] = skb_recv_done;
1553 callbacks[txq2vq(i)] = skb_xmit_done;
1554 sprintf(vi->rq[i].name, "input.%d", i);
1555 sprintf(vi->sq[i].name, "output.%d", i);
1556 names[rxq2vq(i)] = vi->rq[i].name;
1557 names[txq2vq(i)] = vi->sq[i].name;
1560 ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
1566 vi->cvq = vqs[total_vqs - 1];
1567 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
1568 vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
1571 for (i = 0; i < vi->max_queue_pairs; i++) {
1572 vi->rq[i].vq = vqs[rxq2vq(i)];
1573 vi->sq[i].vq = vqs[txq2vq(i)];
1592 static int virtnet_alloc_queues(struct virtnet_info *vi)
1596 vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
1599 vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
1603 INIT_DELAYED_WORK(&vi->refill, refill_work);
1604 for (i = 0; i < vi->max_queue_pairs; i++) {
1605 vi->rq[i].pages = NULL;
1606 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
1608 napi_hash_add(&vi->rq[i].napi);
1610 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
1611 ewma_init(&vi->rq[i].mrg_avg_pkt_len, 1, RECEIVE_AVG_WEIGHT);
1612 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
1623 static int init_vqs(struct virtnet_info *vi)
1627 /* Allocate send & receive queues */
1628 ret = virtnet_alloc_queues(vi);
1632 ret = virtnet_find_vqs(vi);
1637 virtnet_set_affinity(vi);
1643 virtnet_free_queues(vi);
1649 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
1650 struct rx_queue_attribute *attribute, char *buf)
1652 struct virtnet_info *vi = netdev_priv(queue->dev);
1653 unsigned int queue_index = get_netdev_rx_queue_index(queue);
1656 BUG_ON(queue_index >= vi->max_queue_pairs);
1657 avg = &vi->rq[queue_index].mrg_avg_pkt_len;
1658 return sprintf(buf, "%u\n", get_mergeable_buf_len(avg));
1661 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
1662 __ATTR_RO(mergeable_rx_buffer_size);
1664 static struct attribute *virtio_net_mrg_rx_attrs[] = {
1665 &mergeable_rx_buffer_size_attribute.attr,
1669 static const struct attribute_group virtio_net_mrg_rx_group = {
1670 .name = "virtio_net",
1671 .attrs = virtio_net_mrg_rx_attrs
1675 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
1677 const char *fname, const char *dname)
1679 if (!virtio_has_feature(vdev, fbit))
1682 dev_err(&vdev->dev, "device advertises feature %s but not %s",
1688 #define VIRTNET_FAIL_ON(vdev, fbit, dbit) \
1689 virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
1691 static bool virtnet_validate_features(struct virtio_device *vdev)
1693 if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
1694 (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
1695 "VIRTIO_NET_F_CTRL_VQ") ||
1696 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
1697 "VIRTIO_NET_F_CTRL_VQ") ||
1698 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
1699 "VIRTIO_NET_F_CTRL_VQ") ||
1700 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
1701 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
1702 "VIRTIO_NET_F_CTRL_VQ"))) {
1709 static int virtnet_probe(struct virtio_device *vdev)
1712 struct net_device *dev;
1713 struct virtnet_info *vi;
1714 u16 max_queue_pairs;
1716 if (!virtnet_validate_features(vdev))
1719 /* Find if host supports multiqueue virtio_net device */
1720 err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
1721 struct virtio_net_config,
1722 max_virtqueue_pairs, &max_queue_pairs);
1724 /* We need at least 2 queue's */
1725 if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
1726 max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
1727 !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1728 max_queue_pairs = 1;
1730 /* Allocate ourselves a network device with room for our info */
1731 dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
1735 /* Set up network device as normal. */
1736 dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
1737 dev->netdev_ops = &virtnet_netdev;
1738 dev->features = NETIF_F_HIGHDMA;
1740 dev->ethtool_ops = &virtnet_ethtool_ops;
1741 SET_NETDEV_DEV(dev, &vdev->dev);
1743 /* Do we support "hardware" checksums? */
1744 if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
1745 /* This opens up the world of extra features. */
1746 dev->hw_features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1748 dev->features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1750 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
1751 dev->hw_features |= NETIF_F_TSO
1752 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
1754 /* Individual feature bits: what can host handle? */
1755 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
1756 dev->hw_features |= NETIF_F_TSO;
1757 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
1758 dev->hw_features |= NETIF_F_TSO6;
1759 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
1760 dev->hw_features |= NETIF_F_TSO_ECN;
1763 dev->features |= dev->hw_features & NETIF_F_ALL_TSO;
1764 /* (!csum && gso) case will be fixed by register_netdev() */
1766 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
1767 dev->features |= NETIF_F_RXCSUM;
1769 dev->vlan_features = dev->features;
1771 /* Configuration may specify what MAC to use. Otherwise random. */
1772 if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
1773 virtio_cread_bytes(vdev,
1774 offsetof(struct virtio_net_config, mac),
1775 dev->dev_addr, dev->addr_len);
1777 eth_hw_addr_random(dev);
1779 /* Set up our device-specific information */
1780 vi = netdev_priv(dev);
1784 vi->stats = alloc_percpu(struct virtnet_stats);
1786 if (vi->stats == NULL)
1789 for_each_possible_cpu(i) {
1790 struct virtnet_stats *virtnet_stats;
1791 virtnet_stats = per_cpu_ptr(vi->stats, i);
1792 u64_stats_init(&virtnet_stats->tx_syncp);
1793 u64_stats_init(&virtnet_stats->rx_syncp);
1796 INIT_WORK(&vi->config_work, virtnet_config_changed_work);
1798 /* If we can receive ANY GSO packets, we must allocate large ones. */
1799 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1800 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1801 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN))
1802 vi->big_packets = true;
1804 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
1805 vi->mergeable_rx_bufs = true;
1807 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
1808 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
1809 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1811 vi->hdr_len = sizeof(struct virtio_net_hdr);
1813 if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT))
1814 vi->any_header_sg = true;
1816 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1819 if (vi->any_header_sg)
1820 dev->needed_headroom = vi->hdr_len;
1822 /* Use single tx/rx queue pair as default */
1823 vi->curr_queue_pairs = 1;
1824 vi->max_queue_pairs = max_queue_pairs;
1826 /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
1832 if (vi->mergeable_rx_bufs)
1833 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
1835 netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
1836 netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
1838 err = register_netdev(dev);
1840 pr_debug("virtio_net: registering device failed\n");
1844 virtio_device_ready(vdev);
1846 /* Last of all, set up some receive buffers. */
1847 for (i = 0; i < vi->curr_queue_pairs; i++) {
1848 try_fill_recv(vi, &vi->rq[i], GFP_KERNEL);
1850 /* If we didn't even get one input buffer, we're useless. */
1851 if (vi->rq[i].vq->num_free ==
1852 virtqueue_get_vring_size(vi->rq[i].vq)) {
1853 free_unused_bufs(vi);
1855 goto free_recv_bufs;
1859 vi->nb.notifier_call = &virtnet_cpu_callback;
1860 err = register_hotcpu_notifier(&vi->nb);
1862 pr_debug("virtio_net: registering cpu notifier failed\n");
1863 goto free_recv_bufs;
1866 /* Assume link up if device can't report link status,
1867 otherwise get link status from config. */
1868 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
1869 netif_carrier_off(dev);
1870 schedule_work(&vi->config_work);
1872 vi->status = VIRTIO_NET_S_LINK_UP;
1873 netif_carrier_on(dev);
1876 pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
1877 dev->name, max_queue_pairs);
1882 vi->vdev->config->reset(vdev);
1884 free_receive_bufs(vi);
1885 unregister_netdev(dev);
1887 cancel_delayed_work_sync(&vi->refill);
1888 free_receive_page_frags(vi);
1889 virtnet_del_vqs(vi);
1891 free_percpu(vi->stats);
1897 static void remove_vq_common(struct virtnet_info *vi)
1899 vi->vdev->config->reset(vi->vdev);
1901 /* Free unused buffers in both send and recv, if any. */
1902 free_unused_bufs(vi);
1904 free_receive_bufs(vi);
1906 free_receive_page_frags(vi);
1908 virtnet_del_vqs(vi);
1911 static void virtnet_remove(struct virtio_device *vdev)
1913 struct virtnet_info *vi = vdev->priv;
1915 unregister_hotcpu_notifier(&vi->nb);
1917 /* Make sure no work handler is accessing the device. */
1918 flush_work(&vi->config_work);
1920 unregister_netdev(vi->dev);
1922 remove_vq_common(vi);
1924 free_percpu(vi->stats);
1925 free_netdev(vi->dev);
1928 #ifdef CONFIG_PM_SLEEP
1929 static int virtnet_freeze(struct virtio_device *vdev)
1931 struct virtnet_info *vi = vdev->priv;
1934 unregister_hotcpu_notifier(&vi->nb);
1936 /* Make sure no work handler is accessing the device */
1937 flush_work(&vi->config_work);
1939 netif_device_detach(vi->dev);
1940 cancel_delayed_work_sync(&vi->refill);
1942 if (netif_running(vi->dev)) {
1943 for (i = 0; i < vi->max_queue_pairs; i++) {
1944 napi_disable(&vi->rq[i].napi);
1945 napi_hash_del(&vi->rq[i].napi);
1946 netif_napi_del(&vi->rq[i].napi);
1950 remove_vq_common(vi);
1955 static int virtnet_restore(struct virtio_device *vdev)
1957 struct virtnet_info *vi = vdev->priv;
1964 virtio_device_ready(vdev);
1966 if (netif_running(vi->dev)) {
1967 for (i = 0; i < vi->curr_queue_pairs; i++)
1968 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1969 schedule_delayed_work(&vi->refill, 0);
1971 for (i = 0; i < vi->max_queue_pairs; i++)
1972 virtnet_napi_enable(&vi->rq[i]);
1975 netif_device_attach(vi->dev);
1978 virtnet_set_queues(vi, vi->curr_queue_pairs);
1981 err = register_hotcpu_notifier(&vi->nb);
1989 static struct virtio_device_id id_table[] = {
1990 { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
1994 static unsigned int features[] = {
1995 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM,
1996 VIRTIO_NET_F_GSO, VIRTIO_NET_F_MAC,
1997 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_TSO6,
1998 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6,
1999 VIRTIO_NET_F_GUEST_ECN,
2000 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ,
2001 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN,
2002 VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ,
2003 VIRTIO_NET_F_CTRL_MAC_ADDR,
2004 VIRTIO_F_ANY_LAYOUT,
2007 static struct virtio_driver virtio_net_driver = {
2008 .feature_table = features,
2009 .feature_table_size = ARRAY_SIZE(features),
2010 .driver.name = KBUILD_MODNAME,
2011 .driver.owner = THIS_MODULE,
2012 .id_table = id_table,
2013 .probe = virtnet_probe,
2014 .remove = virtnet_remove,
2015 .config_changed = virtnet_config_changed,
2016 #ifdef CONFIG_PM_SLEEP
2017 .freeze = virtnet_freeze,
2018 .restore = virtnet_restore,
2022 module_virtio_driver(virtio_net_driver);
2024 MODULE_DEVICE_TABLE(virtio, id_table);
2025 MODULE_DESCRIPTION("Virtio network driver");
2026 MODULE_LICENSE("GPL");