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 /* enable config space updates */
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 /* Lock for config space updates */
139 struct mutex config_lock;
141 /* Does the affinity hint is set for virtqueues? */
142 bool affinity_hint_set;
144 /* CPU hot plug notifier */
145 struct notifier_block nb;
148 struct skb_vnet_hdr {
150 struct virtio_net_hdr hdr;
151 struct virtio_net_hdr_mrg_rxbuf mhdr;
155 struct padded_vnet_hdr {
156 struct virtio_net_hdr hdr;
158 * virtio_net_hdr should be in a separated sg buffer because of a
159 * QEMU bug, and data sg buffer shares same page with this header sg.
160 * This padding makes next sg 16 byte aligned after virtio_net_hdr.
165 /* Converting between virtqueue no. and kernel tx/rx queue no.
166 * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
168 static int vq2txq(struct virtqueue *vq)
170 return (vq->index - 1) / 2;
173 static int txq2vq(int txq)
178 static int vq2rxq(struct virtqueue *vq)
180 return vq->index / 2;
183 static int rxq2vq(int rxq)
188 static inline struct skb_vnet_hdr *skb_vnet_hdr(struct sk_buff *skb)
190 return (struct skb_vnet_hdr *)skb->cb;
194 * private is used to chain pages for big packets, put the whole
195 * most recent used list in the beginning for reuse
197 static void give_pages(struct receive_queue *rq, struct page *page)
201 /* Find end of list, sew whole thing into vi->rq.pages. */
202 for (end = page; end->private; end = (struct page *)end->private);
203 end->private = (unsigned long)rq->pages;
207 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
209 struct page *p = rq->pages;
212 rq->pages = (struct page *)p->private;
213 /* clear private here, it is used to chain pages */
216 p = alloc_page(gfp_mask);
220 static void skb_xmit_done(struct virtqueue *vq)
222 struct virtnet_info *vi = vq->vdev->priv;
224 /* Suppress further interrupts. */
225 virtqueue_disable_cb(vq);
227 /* We were probably waiting for more output buffers. */
228 netif_wake_subqueue(vi->dev, vq2txq(vq));
231 static unsigned int mergeable_ctx_to_buf_truesize(unsigned long mrg_ctx)
233 unsigned int truesize = mrg_ctx & (MERGEABLE_BUFFER_ALIGN - 1);
234 return (truesize + 1) * MERGEABLE_BUFFER_ALIGN;
237 static void *mergeable_ctx_to_buf_address(unsigned long mrg_ctx)
239 return (void *)(mrg_ctx & -MERGEABLE_BUFFER_ALIGN);
243 static unsigned long mergeable_buf_to_ctx(void *buf, unsigned int truesize)
245 unsigned int size = truesize / MERGEABLE_BUFFER_ALIGN;
246 return (unsigned long)buf | (size - 1);
249 /* Called from bottom half context */
250 static struct sk_buff *page_to_skb(struct receive_queue *rq,
251 struct page *page, unsigned int offset,
252 unsigned int len, unsigned int truesize)
254 struct virtnet_info *vi = rq->vq->vdev->priv;
256 struct skb_vnet_hdr *hdr;
257 unsigned int copy, hdr_len, hdr_padded_len;
260 p = page_address(page) + offset;
262 /* copy small packet so we can reuse these pages for small data */
263 skb = netdev_alloc_skb_ip_align(vi->dev, GOOD_COPY_LEN);
267 hdr = skb_vnet_hdr(skb);
269 if (vi->mergeable_rx_bufs) {
270 hdr_len = sizeof hdr->mhdr;
271 hdr_padded_len = sizeof hdr->mhdr;
273 hdr_len = sizeof hdr->hdr;
274 hdr_padded_len = sizeof(struct padded_vnet_hdr);
277 memcpy(hdr, p, hdr_len);
280 offset += hdr_padded_len;
284 if (copy > skb_tailroom(skb))
285 copy = skb_tailroom(skb);
286 memcpy(skb_put(skb, copy), p, copy);
291 if (vi->mergeable_rx_bufs) {
293 skb_add_rx_frag(skb, 0, page, offset, len, truesize);
300 * Verify that we can indeed put this data into a skb.
301 * This is here to handle cases when the device erroneously
302 * tries to receive more than is possible. This is usually
303 * the case of a broken device.
305 if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
306 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
310 BUG_ON(offset >= PAGE_SIZE);
312 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
313 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
314 frag_size, truesize);
316 page = (struct page *)page->private;
321 give_pages(rq, page);
326 static struct sk_buff *receive_small(void *buf, unsigned int len)
328 struct sk_buff * skb = buf;
330 len -= sizeof(struct virtio_net_hdr);
336 static struct sk_buff *receive_big(struct net_device *dev,
337 struct receive_queue *rq,
341 struct page *page = buf;
342 struct sk_buff *skb = page_to_skb(rq, page, 0, len, PAGE_SIZE);
350 dev->stats.rx_dropped++;
351 give_pages(rq, page);
355 static struct sk_buff *receive_mergeable(struct net_device *dev,
356 struct receive_queue *rq,
360 void *buf = mergeable_ctx_to_buf_address(ctx);
361 struct skb_vnet_hdr *hdr = buf;
362 int num_buf = hdr->mhdr.num_buffers;
363 struct page *page = virt_to_head_page(buf);
364 int offset = buf - page_address(page);
365 unsigned int truesize = max(len, mergeable_ctx_to_buf_truesize(ctx));
367 struct sk_buff *head_skb = page_to_skb(rq, page, offset, len, truesize);
368 struct sk_buff *curr_skb = head_skb;
370 if (unlikely(!curr_skb))
375 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &len);
376 if (unlikely(!ctx)) {
377 pr_debug("%s: rx error: %d buffers out of %d missing\n",
378 dev->name, num_buf, hdr->mhdr.num_buffers);
379 dev->stats.rx_length_errors++;
383 buf = mergeable_ctx_to_buf_address(ctx);
384 page = virt_to_head_page(buf);
386 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
387 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
388 struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
392 if (curr_skb == head_skb)
393 skb_shinfo(curr_skb)->frag_list = nskb;
395 curr_skb->next = nskb;
397 head_skb->truesize += nskb->truesize;
400 truesize = max(len, mergeable_ctx_to_buf_truesize(ctx));
401 if (curr_skb != head_skb) {
402 head_skb->data_len += len;
403 head_skb->len += len;
404 head_skb->truesize += truesize;
406 offset = buf - page_address(page);
407 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
409 skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
412 skb_add_rx_frag(curr_skb, num_skb_frags, page,
413 offset, len, truesize);
417 ewma_add(&rq->mrg_avg_pkt_len, head_skb->len);
423 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &len);
424 if (unlikely(!ctx)) {
425 pr_debug("%s: rx error: %d buffers missing\n",
427 dev->stats.rx_length_errors++;
430 page = virt_to_head_page(mergeable_ctx_to_buf_address(ctx));
434 dev->stats.rx_dropped++;
435 dev_kfree_skb(head_skb);
439 static void receive_buf(struct receive_queue *rq, void *buf, unsigned int len)
441 struct virtnet_info *vi = rq->vq->vdev->priv;
442 struct net_device *dev = vi->dev;
443 struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
445 struct skb_vnet_hdr *hdr;
447 if (unlikely(len < sizeof(struct virtio_net_hdr) + ETH_HLEN)) {
448 pr_debug("%s: short packet %i\n", dev->name, len);
449 dev->stats.rx_length_errors++;
450 if (vi->mergeable_rx_bufs) {
451 unsigned long ctx = (unsigned long)buf;
452 void *base = mergeable_ctx_to_buf_address(ctx);
453 put_page(virt_to_head_page(base));
454 } else if (vi->big_packets) {
462 if (vi->mergeable_rx_bufs)
463 skb = receive_mergeable(dev, rq, (unsigned long)buf, len);
464 else if (vi->big_packets)
465 skb = receive_big(dev, rq, buf, len);
467 skb = receive_small(buf, len);
472 hdr = skb_vnet_hdr(skb);
474 u64_stats_update_begin(&stats->rx_syncp);
475 stats->rx_bytes += skb->len;
477 u64_stats_update_end(&stats->rx_syncp);
479 if (hdr->hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
480 pr_debug("Needs csum!\n");
481 if (!skb_partial_csum_set(skb,
483 hdr->hdr.csum_offset))
485 } else if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID) {
486 skb->ip_summed = CHECKSUM_UNNECESSARY;
489 skb->protocol = eth_type_trans(skb, dev);
490 pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
491 ntohs(skb->protocol), skb->len, skb->pkt_type);
493 if (hdr->hdr.gso_type != VIRTIO_NET_HDR_GSO_NONE) {
495 switch (hdr->hdr.gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
496 case VIRTIO_NET_HDR_GSO_TCPV4:
497 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
499 case VIRTIO_NET_HDR_GSO_UDP:
500 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
502 case VIRTIO_NET_HDR_GSO_TCPV6:
503 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
506 net_warn_ratelimited("%s: bad gso type %u.\n",
507 dev->name, hdr->hdr.gso_type);
511 if (hdr->hdr.gso_type & VIRTIO_NET_HDR_GSO_ECN)
512 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
514 skb_shinfo(skb)->gso_size = hdr->hdr.gso_size;
515 if (skb_shinfo(skb)->gso_size == 0) {
516 net_warn_ratelimited("%s: zero gso size.\n", dev->name);
520 /* Header must be checked, and gso_segs computed. */
521 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
522 skb_shinfo(skb)->gso_segs = 0;
525 skb_mark_napi_id(skb, &rq->napi);
527 netif_receive_skb(skb);
531 dev->stats.rx_frame_errors++;
535 static int add_recvbuf_small(struct receive_queue *rq, gfp_t gfp)
537 struct virtnet_info *vi = rq->vq->vdev->priv;
539 struct skb_vnet_hdr *hdr;
542 skb = __netdev_alloc_skb_ip_align(vi->dev, GOOD_PACKET_LEN, gfp);
546 skb_put(skb, GOOD_PACKET_LEN);
548 hdr = skb_vnet_hdr(skb);
549 sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
550 sg_set_buf(rq->sg, &hdr->hdr, sizeof hdr->hdr);
551 skb_to_sgvec(skb, rq->sg + 1, 0, skb->len);
553 err = virtqueue_add_inbuf(rq->vq, rq->sg, 2, skb, gfp);
560 static int add_recvbuf_big(struct receive_queue *rq, gfp_t gfp)
562 struct page *first, *list = NULL;
566 sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
568 /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
569 for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
570 first = get_a_page(rq, gfp);
573 give_pages(rq, list);
576 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
578 /* chain new page in list head to match sg */
579 first->private = (unsigned long)list;
583 first = get_a_page(rq, gfp);
585 give_pages(rq, list);
588 p = page_address(first);
590 /* rq->sg[0], rq->sg[1] share the same page */
591 /* a separated rq->sg[0] for virtio_net_hdr only due to QEMU bug */
592 sg_set_buf(&rq->sg[0], p, sizeof(struct virtio_net_hdr));
594 /* rq->sg[1] for data packet, from offset */
595 offset = sizeof(struct padded_vnet_hdr);
596 sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
598 /* chain first in list head */
599 first->private = (unsigned long)list;
600 err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
603 give_pages(rq, first);
608 static unsigned int get_mergeable_buf_len(struct ewma *avg_pkt_len)
610 const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
613 len = hdr_len + clamp_t(unsigned int, ewma_read(avg_pkt_len),
614 GOOD_PACKET_LEN, PAGE_SIZE - hdr_len);
615 return ALIGN(len, MERGEABLE_BUFFER_ALIGN);
618 static int add_recvbuf_mergeable(struct receive_queue *rq, gfp_t gfp)
620 struct page_frag *alloc_frag = &rq->alloc_frag;
624 unsigned int len, hole;
626 len = get_mergeable_buf_len(&rq->mrg_avg_pkt_len);
627 if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
630 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
631 ctx = mergeable_buf_to_ctx(buf, len);
632 get_page(alloc_frag->page);
633 alloc_frag->offset += len;
634 hole = alloc_frag->size - alloc_frag->offset;
636 /* To avoid internal fragmentation, if there is very likely not
637 * enough space for another buffer, add the remaining space to
638 * the current buffer. This extra space is not included in
639 * the truesize stored in ctx.
642 alloc_frag->offset += hole;
645 sg_init_one(rq->sg, buf, len);
646 err = virtqueue_add_inbuf(rq->vq, rq->sg, 1, (void *)ctx, gfp);
648 put_page(virt_to_head_page(buf));
654 * Returns false if we couldn't fill entirely (OOM).
656 * Normally run in the receive path, but can also be run from ndo_open
657 * before we're receiving packets, or from refill_work which is
658 * careful to disable receiving (using napi_disable).
660 static bool try_fill_recv(struct receive_queue *rq, gfp_t gfp)
662 struct virtnet_info *vi = rq->vq->vdev->priv;
668 if (vi->mergeable_rx_bufs)
669 err = add_recvbuf_mergeable(rq, gfp);
670 else if (vi->big_packets)
671 err = add_recvbuf_big(rq, gfp);
673 err = add_recvbuf_small(rq, gfp);
675 oom = err == -ENOMEM;
678 } while (rq->vq->num_free);
679 virtqueue_kick(rq->vq);
683 static void skb_recv_done(struct virtqueue *rvq)
685 struct virtnet_info *vi = rvq->vdev->priv;
686 struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
688 /* Schedule NAPI, Suppress further interrupts if successful. */
689 if (napi_schedule_prep(&rq->napi)) {
690 virtqueue_disable_cb(rvq);
691 __napi_schedule(&rq->napi);
695 static void virtnet_napi_enable(struct receive_queue *rq)
697 napi_enable(&rq->napi);
699 /* If all buffers were filled by other side before we napi_enabled, we
700 * won't get another interrupt, so process any outstanding packets
701 * now. virtnet_poll wants re-enable the queue, so we disable here.
702 * We synchronize against interrupts via NAPI_STATE_SCHED */
703 if (napi_schedule_prep(&rq->napi)) {
704 virtqueue_disable_cb(rq->vq);
706 __napi_schedule(&rq->napi);
711 static void refill_work(struct work_struct *work)
713 struct virtnet_info *vi =
714 container_of(work, struct virtnet_info, refill.work);
718 for (i = 0; i < vi->curr_queue_pairs; i++) {
719 struct receive_queue *rq = &vi->rq[i];
721 napi_disable(&rq->napi);
722 still_empty = !try_fill_recv(rq, GFP_KERNEL);
723 virtnet_napi_enable(rq);
725 /* In theory, this can happen: if we don't get any buffers in
726 * we will *never* try to fill again.
729 schedule_delayed_work(&vi->refill, HZ/2);
733 static int virtnet_receive(struct receive_queue *rq, int budget)
735 struct virtnet_info *vi = rq->vq->vdev->priv;
736 unsigned int len, received = 0;
739 while (received < budget &&
740 (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
741 receive_buf(rq, buf, len);
745 if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) {
746 if (!try_fill_recv(rq, GFP_ATOMIC))
747 schedule_delayed_work(&vi->refill, 0);
753 static int virtnet_poll(struct napi_struct *napi, int budget)
755 struct receive_queue *rq =
756 container_of(napi, struct receive_queue, napi);
757 unsigned int r, received = 0;
760 received += virtnet_receive(rq, budget - received);
762 /* Out of packets? */
763 if (received < budget) {
764 r = virtqueue_enable_cb_prepare(rq->vq);
766 if (unlikely(virtqueue_poll(rq->vq, r)) &&
767 napi_schedule_prep(napi)) {
768 virtqueue_disable_cb(rq->vq);
769 __napi_schedule(napi);
777 #ifdef CONFIG_NET_RX_BUSY_POLL
778 /* must be called with local_bh_disable()d */
779 static int virtnet_busy_poll(struct napi_struct *napi)
781 struct receive_queue *rq =
782 container_of(napi, struct receive_queue, napi);
783 struct virtnet_info *vi = rq->vq->vdev->priv;
784 int r, received = 0, budget = 4;
786 if (!(vi->status & VIRTIO_NET_S_LINK_UP))
787 return LL_FLUSH_FAILED;
789 if (!napi_schedule_prep(napi))
790 return LL_FLUSH_BUSY;
792 virtqueue_disable_cb(rq->vq);
795 received += virtnet_receive(rq, budget);
797 r = virtqueue_enable_cb_prepare(rq->vq);
798 clear_bit(NAPI_STATE_SCHED, &napi->state);
799 if (unlikely(virtqueue_poll(rq->vq, r)) &&
800 napi_schedule_prep(napi)) {
801 virtqueue_disable_cb(rq->vq);
802 if (received < budget) {
806 __napi_schedule(napi);
812 #endif /* CONFIG_NET_RX_BUSY_POLL */
814 static int virtnet_open(struct net_device *dev)
816 struct virtnet_info *vi = netdev_priv(dev);
819 for (i = 0; i < vi->max_queue_pairs; i++) {
820 if (i < vi->curr_queue_pairs)
821 /* Make sure we have some buffers: if oom use wq. */
822 if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
823 schedule_delayed_work(&vi->refill, 0);
824 virtnet_napi_enable(&vi->rq[i]);
830 static void free_old_xmit_skbs(struct send_queue *sq)
834 struct virtnet_info *vi = sq->vq->vdev->priv;
835 struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
837 while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
838 pr_debug("Sent skb %p\n", skb);
840 u64_stats_update_begin(&stats->tx_syncp);
841 stats->tx_bytes += skb->len;
843 u64_stats_update_end(&stats->tx_syncp);
845 dev_kfree_skb_any(skb);
849 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
851 struct skb_vnet_hdr *hdr;
852 const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
853 struct virtnet_info *vi = sq->vq->vdev->priv;
858 pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
859 if (vi->mergeable_rx_bufs)
860 hdr_len = sizeof hdr->mhdr;
862 hdr_len = sizeof hdr->hdr;
864 can_push = vi->any_header_sg &&
865 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
866 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
867 /* Even if we can, don't push here yet as this would skew
868 * csum_start offset below. */
870 hdr = (struct skb_vnet_hdr *)(skb->data - hdr_len);
872 hdr = skb_vnet_hdr(skb);
874 if (skb->ip_summed == CHECKSUM_PARTIAL) {
875 hdr->hdr.flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
876 hdr->hdr.csum_start = skb_checksum_start_offset(skb);
877 hdr->hdr.csum_offset = skb->csum_offset;
880 hdr->hdr.csum_offset = hdr->hdr.csum_start = 0;
883 if (skb_is_gso(skb)) {
884 hdr->hdr.hdr_len = skb_headlen(skb);
885 hdr->hdr.gso_size = skb_shinfo(skb)->gso_size;
886 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
887 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
888 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
889 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
890 else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
891 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_UDP;
894 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCP_ECN)
895 hdr->hdr.gso_type |= VIRTIO_NET_HDR_GSO_ECN;
897 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE;
898 hdr->hdr.gso_size = hdr->hdr.hdr_len = 0;
901 if (vi->mergeable_rx_bufs)
902 hdr->mhdr.num_buffers = 0;
904 sg_init_table(sq->sg, MAX_SKB_FRAGS + 2);
906 __skb_push(skb, hdr_len);
907 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
908 /* Pull header back to avoid skew in tx bytes calculations. */
909 __skb_pull(skb, hdr_len);
911 sg_set_buf(sq->sg, hdr, hdr_len);
912 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
914 return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
917 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
919 struct virtnet_info *vi = netdev_priv(dev);
920 int qnum = skb_get_queue_mapping(skb);
921 struct send_queue *sq = &vi->sq[qnum];
924 /* Free up any pending old buffers before queueing new ones. */
925 free_old_xmit_skbs(sq);
927 /* Try to transmit */
928 err = xmit_skb(sq, skb);
930 /* This should not happen! */
932 dev->stats.tx_fifo_errors++;
935 "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
936 dev->stats.tx_dropped++;
937 dev_kfree_skb_any(skb);
941 /* Don't wait up for transmitted skbs to be freed. */
945 /* Apparently nice girls don't return TX_BUSY; stop the queue
946 * before it gets out of hand. Naturally, this wastes entries. */
947 if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
948 netif_stop_subqueue(dev, qnum);
949 if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
950 /* More just got used, free them then recheck. */
951 free_old_xmit_skbs(sq);
952 if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
953 netif_start_subqueue(dev, qnum);
954 virtqueue_disable_cb(sq->vq);
959 if (__netif_subqueue_stopped(dev, qnum) || !skb->xmit_more)
960 virtqueue_kick(sq->vq);
966 * Send command via the control virtqueue and check status. Commands
967 * supported by the hypervisor, as indicated by feature bits, should
968 * never fail unless improperly formatted.
970 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
971 struct scatterlist *out)
973 struct scatterlist *sgs[4], hdr, stat;
974 struct virtio_net_ctrl_hdr ctrl;
975 virtio_net_ctrl_ack status = ~0;
976 unsigned out_num = 0, tmp;
978 /* Caller should know better */
979 BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
984 sg_init_one(&hdr, &ctrl, sizeof(ctrl));
985 sgs[out_num++] = &hdr;
988 sgs[out_num++] = out;
990 /* Add return status. */
991 sg_init_one(&stat, &status, sizeof(status));
992 sgs[out_num] = &stat;
994 BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
995 virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
997 if (unlikely(!virtqueue_kick(vi->cvq)))
998 return status == VIRTIO_NET_OK;
1000 /* Spin for a response, the kick causes an ioport write, trapping
1001 * into the hypervisor, so the request should be handled immediately.
1003 while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1004 !virtqueue_is_broken(vi->cvq))
1007 return status == VIRTIO_NET_OK;
1010 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1012 struct virtnet_info *vi = netdev_priv(dev);
1013 struct virtio_device *vdev = vi->vdev;
1015 struct sockaddr *addr = p;
1016 struct scatterlist sg;
1018 ret = eth_prepare_mac_addr_change(dev, p);
1022 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1023 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1024 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1025 VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1026 dev_warn(&vdev->dev,
1027 "Failed to set mac address by vq command.\n");
1030 } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
1033 /* Naturally, this has an atomicity problem. */
1034 for (i = 0; i < dev->addr_len; i++)
1035 virtio_cwrite8(vdev,
1036 offsetof(struct virtio_net_config, mac) +
1037 i, addr->sa_data[i]);
1040 eth_commit_mac_addr_change(dev, p);
1045 static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
1046 struct rtnl_link_stats64 *tot)
1048 struct virtnet_info *vi = netdev_priv(dev);
1052 for_each_possible_cpu(cpu) {
1053 struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
1054 u64 tpackets, tbytes, rpackets, rbytes;
1057 start = u64_stats_fetch_begin_irq(&stats->tx_syncp);
1058 tpackets = stats->tx_packets;
1059 tbytes = stats->tx_bytes;
1060 } while (u64_stats_fetch_retry_irq(&stats->tx_syncp, start));
1063 start = u64_stats_fetch_begin_irq(&stats->rx_syncp);
1064 rpackets = stats->rx_packets;
1065 rbytes = stats->rx_bytes;
1066 } while (u64_stats_fetch_retry_irq(&stats->rx_syncp, start));
1068 tot->rx_packets += rpackets;
1069 tot->tx_packets += tpackets;
1070 tot->rx_bytes += rbytes;
1071 tot->tx_bytes += tbytes;
1074 tot->tx_dropped = dev->stats.tx_dropped;
1075 tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1076 tot->rx_dropped = dev->stats.rx_dropped;
1077 tot->rx_length_errors = dev->stats.rx_length_errors;
1078 tot->rx_frame_errors = dev->stats.rx_frame_errors;
1083 #ifdef CONFIG_NET_POLL_CONTROLLER
1084 static void virtnet_netpoll(struct net_device *dev)
1086 struct virtnet_info *vi = netdev_priv(dev);
1089 for (i = 0; i < vi->curr_queue_pairs; i++)
1090 napi_schedule(&vi->rq[i].napi);
1094 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1097 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1098 VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1099 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1103 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1105 struct scatterlist sg;
1106 struct virtio_net_ctrl_mq s;
1107 struct net_device *dev = vi->dev;
1109 if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1112 s.virtqueue_pairs = queue_pairs;
1113 sg_init_one(&sg, &s, sizeof(s));
1115 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1116 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1117 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1121 vi->curr_queue_pairs = queue_pairs;
1122 /* virtnet_open() will refill when device is going to up. */
1123 if (dev->flags & IFF_UP)
1124 schedule_delayed_work(&vi->refill, 0);
1130 static int virtnet_close(struct net_device *dev)
1132 struct virtnet_info *vi = netdev_priv(dev);
1135 /* Make sure refill_work doesn't re-enable napi! */
1136 cancel_delayed_work_sync(&vi->refill);
1138 for (i = 0; i < vi->max_queue_pairs; i++)
1139 napi_disable(&vi->rq[i].napi);
1144 static void virtnet_set_rx_mode(struct net_device *dev)
1146 struct virtnet_info *vi = netdev_priv(dev);
1147 struct scatterlist sg[2];
1148 u8 promisc, allmulti;
1149 struct virtio_net_ctrl_mac *mac_data;
1150 struct netdev_hw_addr *ha;
1156 /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1157 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1160 promisc = ((dev->flags & IFF_PROMISC) != 0);
1161 allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1163 sg_init_one(sg, &promisc, sizeof(promisc));
1165 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1166 VIRTIO_NET_CTRL_RX_PROMISC, sg))
1167 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1168 promisc ? "en" : "dis");
1170 sg_init_one(sg, &allmulti, sizeof(allmulti));
1172 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1173 VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1174 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1175 allmulti ? "en" : "dis");
1177 uc_count = netdev_uc_count(dev);
1178 mc_count = netdev_mc_count(dev);
1179 /* MAC filter - use one buffer for both lists */
1180 buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1181 (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1186 sg_init_table(sg, 2);
1188 /* Store the unicast list and count in the front of the buffer */
1189 mac_data->entries = uc_count;
1191 netdev_for_each_uc_addr(ha, dev)
1192 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1194 sg_set_buf(&sg[0], mac_data,
1195 sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1197 /* multicast list and count fill the end */
1198 mac_data = (void *)&mac_data->macs[uc_count][0];
1200 mac_data->entries = mc_count;
1202 netdev_for_each_mc_addr(ha, dev)
1203 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1205 sg_set_buf(&sg[1], mac_data,
1206 sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1208 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1209 VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1210 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1215 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1216 __be16 proto, u16 vid)
1218 struct virtnet_info *vi = netdev_priv(dev);
1219 struct scatterlist sg;
1221 sg_init_one(&sg, &vid, sizeof(vid));
1223 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1224 VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1225 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1229 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1230 __be16 proto, u16 vid)
1232 struct virtnet_info *vi = netdev_priv(dev);
1233 struct scatterlist sg;
1235 sg_init_one(&sg, &vid, sizeof(vid));
1237 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1238 VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1239 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1243 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1247 if (vi->affinity_hint_set) {
1248 for (i = 0; i < vi->max_queue_pairs; i++) {
1249 virtqueue_set_affinity(vi->rq[i].vq, -1);
1250 virtqueue_set_affinity(vi->sq[i].vq, -1);
1253 vi->affinity_hint_set = false;
1257 static void virtnet_set_affinity(struct virtnet_info *vi)
1262 /* In multiqueue mode, when the number of cpu is equal to the number of
1263 * queue pairs, we let the queue pairs to be private to one cpu by
1264 * setting the affinity hint to eliminate the contention.
1266 if (vi->curr_queue_pairs == 1 ||
1267 vi->max_queue_pairs != num_online_cpus()) {
1268 virtnet_clean_affinity(vi, -1);
1273 for_each_online_cpu(cpu) {
1274 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1275 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1276 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
1280 vi->affinity_hint_set = true;
1283 static int virtnet_cpu_callback(struct notifier_block *nfb,
1284 unsigned long action, void *hcpu)
1286 struct virtnet_info *vi = container_of(nfb, struct virtnet_info, nb);
1288 switch(action & ~CPU_TASKS_FROZEN) {
1290 case CPU_DOWN_FAILED:
1292 virtnet_set_affinity(vi);
1294 case CPU_DOWN_PREPARE:
1295 virtnet_clean_affinity(vi, (long)hcpu);
1304 static void virtnet_get_ringparam(struct net_device *dev,
1305 struct ethtool_ringparam *ring)
1307 struct virtnet_info *vi = netdev_priv(dev);
1309 ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1310 ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1311 ring->rx_pending = ring->rx_max_pending;
1312 ring->tx_pending = ring->tx_max_pending;
1316 static void virtnet_get_drvinfo(struct net_device *dev,
1317 struct ethtool_drvinfo *info)
1319 struct virtnet_info *vi = netdev_priv(dev);
1320 struct virtio_device *vdev = vi->vdev;
1322 strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1323 strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1324 strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1328 /* TODO: Eliminate OOO packets during switching */
1329 static int virtnet_set_channels(struct net_device *dev,
1330 struct ethtool_channels *channels)
1332 struct virtnet_info *vi = netdev_priv(dev);
1333 u16 queue_pairs = channels->combined_count;
1336 /* We don't support separate rx/tx channels.
1337 * We don't allow setting 'other' channels.
1339 if (channels->rx_count || channels->tx_count || channels->other_count)
1342 if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
1346 err = virtnet_set_queues(vi, queue_pairs);
1348 netif_set_real_num_tx_queues(dev, queue_pairs);
1349 netif_set_real_num_rx_queues(dev, queue_pairs);
1351 virtnet_set_affinity(vi);
1358 static void virtnet_get_channels(struct net_device *dev,
1359 struct ethtool_channels *channels)
1361 struct virtnet_info *vi = netdev_priv(dev);
1363 channels->combined_count = vi->curr_queue_pairs;
1364 channels->max_combined = vi->max_queue_pairs;
1365 channels->max_other = 0;
1366 channels->rx_count = 0;
1367 channels->tx_count = 0;
1368 channels->other_count = 0;
1371 static const struct ethtool_ops virtnet_ethtool_ops = {
1372 .get_drvinfo = virtnet_get_drvinfo,
1373 .get_link = ethtool_op_get_link,
1374 .get_ringparam = virtnet_get_ringparam,
1375 .set_channels = virtnet_set_channels,
1376 .get_channels = virtnet_get_channels,
1380 #define MAX_MTU 65535
1382 static int virtnet_change_mtu(struct net_device *dev, int new_mtu)
1384 if (new_mtu < MIN_MTU || new_mtu > MAX_MTU)
1390 static const struct net_device_ops virtnet_netdev = {
1391 .ndo_open = virtnet_open,
1392 .ndo_stop = virtnet_close,
1393 .ndo_start_xmit = start_xmit,
1394 .ndo_validate_addr = eth_validate_addr,
1395 .ndo_set_mac_address = virtnet_set_mac_address,
1396 .ndo_set_rx_mode = virtnet_set_rx_mode,
1397 .ndo_change_mtu = virtnet_change_mtu,
1398 .ndo_get_stats64 = virtnet_stats,
1399 .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
1400 .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
1401 #ifdef CONFIG_NET_POLL_CONTROLLER
1402 .ndo_poll_controller = virtnet_netpoll,
1404 #ifdef CONFIG_NET_RX_BUSY_POLL
1405 .ndo_busy_poll = virtnet_busy_poll,
1409 static void virtnet_config_changed_work(struct work_struct *work)
1411 struct virtnet_info *vi =
1412 container_of(work, struct virtnet_info, config_work);
1415 mutex_lock(&vi->config_lock);
1416 if (!vi->config_enable)
1419 if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
1420 struct virtio_net_config, status, &v) < 0)
1423 if (v & VIRTIO_NET_S_ANNOUNCE) {
1424 netdev_notify_peers(vi->dev);
1425 virtnet_ack_link_announce(vi);
1428 /* Ignore unknown (future) status bits */
1429 v &= VIRTIO_NET_S_LINK_UP;
1431 if (vi->status == v)
1436 if (vi->status & VIRTIO_NET_S_LINK_UP) {
1437 netif_carrier_on(vi->dev);
1438 netif_tx_wake_all_queues(vi->dev);
1440 netif_carrier_off(vi->dev);
1441 netif_tx_stop_all_queues(vi->dev);
1444 mutex_unlock(&vi->config_lock);
1447 static void virtnet_config_changed(struct virtio_device *vdev)
1449 struct virtnet_info *vi = vdev->priv;
1451 schedule_work(&vi->config_work);
1454 static void virtnet_free_queues(struct virtnet_info *vi)
1458 for (i = 0; i < vi->max_queue_pairs; i++)
1459 netif_napi_del(&vi->rq[i].napi);
1465 static void free_receive_bufs(struct virtnet_info *vi)
1469 for (i = 0; i < vi->max_queue_pairs; i++) {
1470 while (vi->rq[i].pages)
1471 __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
1475 static void free_receive_page_frags(struct virtnet_info *vi)
1478 for (i = 0; i < vi->max_queue_pairs; i++)
1479 if (vi->rq[i].alloc_frag.page)
1480 put_page(vi->rq[i].alloc_frag.page);
1483 static void free_unused_bufs(struct virtnet_info *vi)
1488 for (i = 0; i < vi->max_queue_pairs; i++) {
1489 struct virtqueue *vq = vi->sq[i].vq;
1490 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
1494 for (i = 0; i < vi->max_queue_pairs; i++) {
1495 struct virtqueue *vq = vi->rq[i].vq;
1497 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1498 if (vi->mergeable_rx_bufs) {
1499 unsigned long ctx = (unsigned long)buf;
1500 void *base = mergeable_ctx_to_buf_address(ctx);
1501 put_page(virt_to_head_page(base));
1502 } else if (vi->big_packets) {
1503 give_pages(&vi->rq[i], buf);
1511 static void virtnet_del_vqs(struct virtnet_info *vi)
1513 struct virtio_device *vdev = vi->vdev;
1515 virtnet_clean_affinity(vi, -1);
1517 vdev->config->del_vqs(vdev);
1519 virtnet_free_queues(vi);
1522 static int virtnet_find_vqs(struct virtnet_info *vi)
1524 vq_callback_t **callbacks;
1525 struct virtqueue **vqs;
1530 /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
1531 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
1532 * possible control vq.
1534 total_vqs = vi->max_queue_pairs * 2 +
1535 virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
1537 /* Allocate space for find_vqs parameters */
1538 vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
1541 callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
1544 names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
1548 /* Parameters for control virtqueue, if any */
1550 callbacks[total_vqs - 1] = NULL;
1551 names[total_vqs - 1] = "control";
1554 /* Allocate/initialize parameters for send/receive virtqueues */
1555 for (i = 0; i < vi->max_queue_pairs; i++) {
1556 callbacks[rxq2vq(i)] = skb_recv_done;
1557 callbacks[txq2vq(i)] = skb_xmit_done;
1558 sprintf(vi->rq[i].name, "input.%d", i);
1559 sprintf(vi->sq[i].name, "output.%d", i);
1560 names[rxq2vq(i)] = vi->rq[i].name;
1561 names[txq2vq(i)] = vi->sq[i].name;
1564 ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
1570 vi->cvq = vqs[total_vqs - 1];
1571 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
1572 vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
1575 for (i = 0; i < vi->max_queue_pairs; i++) {
1576 vi->rq[i].vq = vqs[rxq2vq(i)];
1577 vi->sq[i].vq = vqs[txq2vq(i)];
1596 static int virtnet_alloc_queues(struct virtnet_info *vi)
1600 vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
1603 vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
1607 INIT_DELAYED_WORK(&vi->refill, refill_work);
1608 for (i = 0; i < vi->max_queue_pairs; i++) {
1609 vi->rq[i].pages = NULL;
1610 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
1612 napi_hash_add(&vi->rq[i].napi);
1614 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
1615 ewma_init(&vi->rq[i].mrg_avg_pkt_len, 1, RECEIVE_AVG_WEIGHT);
1616 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
1627 static int init_vqs(struct virtnet_info *vi)
1631 /* Allocate send & receive queues */
1632 ret = virtnet_alloc_queues(vi);
1636 ret = virtnet_find_vqs(vi);
1641 virtnet_set_affinity(vi);
1647 virtnet_free_queues(vi);
1653 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
1654 struct rx_queue_attribute *attribute, char *buf)
1656 struct virtnet_info *vi = netdev_priv(queue->dev);
1657 unsigned int queue_index = get_netdev_rx_queue_index(queue);
1660 BUG_ON(queue_index >= vi->max_queue_pairs);
1661 avg = &vi->rq[queue_index].mrg_avg_pkt_len;
1662 return sprintf(buf, "%u\n", get_mergeable_buf_len(avg));
1665 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
1666 __ATTR_RO(mergeable_rx_buffer_size);
1668 static struct attribute *virtio_net_mrg_rx_attrs[] = {
1669 &mergeable_rx_buffer_size_attribute.attr,
1673 static const struct attribute_group virtio_net_mrg_rx_group = {
1674 .name = "virtio_net",
1675 .attrs = virtio_net_mrg_rx_attrs
1679 static int virtnet_probe(struct virtio_device *vdev)
1682 struct net_device *dev;
1683 struct virtnet_info *vi;
1684 u16 max_queue_pairs;
1686 /* Find if host supports multiqueue virtio_net device */
1687 err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
1688 struct virtio_net_config,
1689 max_virtqueue_pairs, &max_queue_pairs);
1691 /* We need at least 2 queue's */
1692 if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
1693 max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
1694 !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1695 max_queue_pairs = 1;
1697 /* Allocate ourselves a network device with room for our info */
1698 dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
1702 /* Set up network device as normal. */
1703 dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
1704 dev->netdev_ops = &virtnet_netdev;
1705 dev->features = NETIF_F_HIGHDMA;
1707 dev->ethtool_ops = &virtnet_ethtool_ops;
1708 SET_NETDEV_DEV(dev, &vdev->dev);
1710 /* Do we support "hardware" checksums? */
1711 if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
1712 /* This opens up the world of extra features. */
1713 dev->hw_features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1715 dev->features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1717 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
1718 dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
1719 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
1721 /* Individual feature bits: what can host handle? */
1722 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
1723 dev->hw_features |= NETIF_F_TSO;
1724 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
1725 dev->hw_features |= NETIF_F_TSO6;
1726 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
1727 dev->hw_features |= NETIF_F_TSO_ECN;
1728 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
1729 dev->hw_features |= NETIF_F_UFO;
1732 dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
1733 /* (!csum && gso) case will be fixed by register_netdev() */
1735 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
1736 dev->features |= NETIF_F_RXCSUM;
1738 dev->vlan_features = dev->features;
1740 /* Configuration may specify what MAC to use. Otherwise random. */
1741 if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
1742 virtio_cread_bytes(vdev,
1743 offsetof(struct virtio_net_config, mac),
1744 dev->dev_addr, dev->addr_len);
1746 eth_hw_addr_random(dev);
1748 /* Set up our device-specific information */
1749 vi = netdev_priv(dev);
1753 vi->stats = alloc_percpu(struct virtnet_stats);
1755 if (vi->stats == NULL)
1758 for_each_possible_cpu(i) {
1759 struct virtnet_stats *virtnet_stats;
1760 virtnet_stats = per_cpu_ptr(vi->stats, i);
1761 u64_stats_init(&virtnet_stats->tx_syncp);
1762 u64_stats_init(&virtnet_stats->rx_syncp);
1765 mutex_init(&vi->config_lock);
1766 vi->config_enable = true;
1767 INIT_WORK(&vi->config_work, virtnet_config_changed_work);
1769 /* If we can receive ANY GSO packets, we must allocate large ones. */
1770 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1771 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1772 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
1773 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
1774 vi->big_packets = true;
1776 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
1777 vi->mergeable_rx_bufs = true;
1779 if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT))
1780 vi->any_header_sg = true;
1782 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1785 if (vi->any_header_sg) {
1786 if (vi->mergeable_rx_bufs)
1787 dev->needed_headroom = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1789 dev->needed_headroom = sizeof(struct virtio_net_hdr);
1792 /* Use single tx/rx queue pair as default */
1793 vi->curr_queue_pairs = 1;
1794 vi->max_queue_pairs = max_queue_pairs;
1796 /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
1802 if (vi->mergeable_rx_bufs)
1803 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
1805 netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
1806 netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
1808 err = register_netdev(dev);
1810 pr_debug("virtio_net: registering device failed\n");
1814 /* Last of all, set up some receive buffers. */
1815 for (i = 0; i < vi->curr_queue_pairs; i++) {
1816 try_fill_recv(&vi->rq[i], GFP_KERNEL);
1818 /* If we didn't even get one input buffer, we're useless. */
1819 if (vi->rq[i].vq->num_free ==
1820 virtqueue_get_vring_size(vi->rq[i].vq)) {
1821 free_unused_bufs(vi);
1823 goto free_recv_bufs;
1827 vi->nb.notifier_call = &virtnet_cpu_callback;
1828 err = register_hotcpu_notifier(&vi->nb);
1830 pr_debug("virtio_net: registering cpu notifier failed\n");
1831 goto free_recv_bufs;
1834 /* Assume link up if device can't report link status,
1835 otherwise get link status from config. */
1836 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
1837 netif_carrier_off(dev);
1838 schedule_work(&vi->config_work);
1840 vi->status = VIRTIO_NET_S_LINK_UP;
1841 netif_carrier_on(dev);
1844 pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
1845 dev->name, max_queue_pairs);
1850 free_receive_bufs(vi);
1851 unregister_netdev(dev);
1853 cancel_delayed_work_sync(&vi->refill);
1854 free_receive_page_frags(vi);
1855 virtnet_del_vqs(vi);
1857 free_percpu(vi->stats);
1863 static void remove_vq_common(struct virtnet_info *vi)
1865 vi->vdev->config->reset(vi->vdev);
1867 /* Free unused buffers in both send and recv, if any. */
1868 free_unused_bufs(vi);
1870 free_receive_bufs(vi);
1872 free_receive_page_frags(vi);
1874 virtnet_del_vqs(vi);
1877 static void virtnet_remove(struct virtio_device *vdev)
1879 struct virtnet_info *vi = vdev->priv;
1881 unregister_hotcpu_notifier(&vi->nb);
1883 /* Prevent config work handler from accessing the device. */
1884 mutex_lock(&vi->config_lock);
1885 vi->config_enable = false;
1886 mutex_unlock(&vi->config_lock);
1888 unregister_netdev(vi->dev);
1890 remove_vq_common(vi);
1892 flush_work(&vi->config_work);
1894 free_percpu(vi->stats);
1895 free_netdev(vi->dev);
1898 #ifdef CONFIG_PM_SLEEP
1899 static int virtnet_freeze(struct virtio_device *vdev)
1901 struct virtnet_info *vi = vdev->priv;
1904 unregister_hotcpu_notifier(&vi->nb);
1906 /* Prevent config work handler from accessing the device */
1907 mutex_lock(&vi->config_lock);
1908 vi->config_enable = false;
1909 mutex_unlock(&vi->config_lock);
1911 netif_device_detach(vi->dev);
1912 cancel_delayed_work_sync(&vi->refill);
1914 if (netif_running(vi->dev)) {
1915 for (i = 0; i < vi->max_queue_pairs; i++) {
1916 napi_disable(&vi->rq[i].napi);
1917 napi_hash_del(&vi->rq[i].napi);
1918 netif_napi_del(&vi->rq[i].napi);
1922 remove_vq_common(vi);
1924 flush_work(&vi->config_work);
1929 static int virtnet_restore(struct virtio_device *vdev)
1931 struct virtnet_info *vi = vdev->priv;
1938 if (netif_running(vi->dev)) {
1939 for (i = 0; i < vi->curr_queue_pairs; i++)
1940 if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
1941 schedule_delayed_work(&vi->refill, 0);
1943 for (i = 0; i < vi->max_queue_pairs; i++)
1944 virtnet_napi_enable(&vi->rq[i]);
1947 netif_device_attach(vi->dev);
1949 mutex_lock(&vi->config_lock);
1950 vi->config_enable = true;
1951 mutex_unlock(&vi->config_lock);
1954 virtnet_set_queues(vi, vi->curr_queue_pairs);
1957 err = register_hotcpu_notifier(&vi->nb);
1965 static struct virtio_device_id id_table[] = {
1966 { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
1970 static unsigned int features[] = {
1971 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM,
1972 VIRTIO_NET_F_GSO, VIRTIO_NET_F_MAC,
1973 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6,
1974 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6,
1975 VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO,
1976 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ,
1977 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN,
1978 VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ,
1979 VIRTIO_NET_F_CTRL_MAC_ADDR,
1980 VIRTIO_F_ANY_LAYOUT,
1983 static struct virtio_driver virtio_net_driver = {
1984 .feature_table = features,
1985 .feature_table_size = ARRAY_SIZE(features),
1986 .driver.name = KBUILD_MODNAME,
1987 .driver.owner = THIS_MODULE,
1988 .id_table = id_table,
1989 .probe = virtnet_probe,
1990 .remove = virtnet_remove,
1991 .config_changed = virtnet_config_changed,
1992 #ifdef CONFIG_PM_SLEEP
1993 .freeze = virtnet_freeze,
1994 .restore = virtnet_restore,
1998 module_virtio_driver(virtio_net_driver);
2000 MODULE_DEVICE_TABLE(virtio, id_table);
2001 MODULE_DESCRIPTION("Virtio network driver");
2002 MODULE_LICENSE("GPL");