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 /* Active statistics */
127 struct virtnet_stats __percpu *stats;
129 /* Work struct for refilling if we run low on memory. */
130 struct delayed_work refill;
132 /* Work struct for config space updates */
133 struct work_struct config_work;
135 /* Does the affinity hint is set for virtqueues? */
136 bool affinity_hint_set;
138 /* CPU hot plug notifier */
139 struct notifier_block nb;
142 struct skb_vnet_hdr {
144 struct virtio_net_hdr hdr;
145 struct virtio_net_hdr_mrg_rxbuf mhdr;
149 struct padded_vnet_hdr {
150 struct virtio_net_hdr hdr;
152 * virtio_net_hdr should be in a separated sg buffer because of a
153 * QEMU bug, and data sg buffer shares same page with this header sg.
154 * This padding makes next sg 16 byte aligned after virtio_net_hdr.
159 /* Converting between virtqueue no. and kernel tx/rx queue no.
160 * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
162 static int vq2txq(struct virtqueue *vq)
164 return (vq->index - 1) / 2;
167 static int txq2vq(int txq)
172 static int vq2rxq(struct virtqueue *vq)
174 return vq->index / 2;
177 static int rxq2vq(int rxq)
182 static inline struct skb_vnet_hdr *skb_vnet_hdr(struct sk_buff *skb)
184 return (struct skb_vnet_hdr *)skb->cb;
188 * private is used to chain pages for big packets, put the whole
189 * most recent used list in the beginning for reuse
191 static void give_pages(struct receive_queue *rq, struct page *page)
195 /* Find end of list, sew whole thing into vi->rq.pages. */
196 for (end = page; end->private; end = (struct page *)end->private);
197 end->private = (unsigned long)rq->pages;
201 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
203 struct page *p = rq->pages;
206 rq->pages = (struct page *)p->private;
207 /* clear private here, it is used to chain pages */
210 p = alloc_page(gfp_mask);
214 static void skb_xmit_done(struct virtqueue *vq)
216 struct virtnet_info *vi = vq->vdev->priv;
218 /* Suppress further interrupts. */
219 virtqueue_disable_cb(vq);
221 /* We were probably waiting for more output buffers. */
222 netif_wake_subqueue(vi->dev, vq2txq(vq));
225 static unsigned int mergeable_ctx_to_buf_truesize(unsigned long mrg_ctx)
227 unsigned int truesize = mrg_ctx & (MERGEABLE_BUFFER_ALIGN - 1);
228 return (truesize + 1) * MERGEABLE_BUFFER_ALIGN;
231 static void *mergeable_ctx_to_buf_address(unsigned long mrg_ctx)
233 return (void *)(mrg_ctx & -MERGEABLE_BUFFER_ALIGN);
237 static unsigned long mergeable_buf_to_ctx(void *buf, unsigned int truesize)
239 unsigned int size = truesize / MERGEABLE_BUFFER_ALIGN;
240 return (unsigned long)buf | (size - 1);
243 /* Called from bottom half context */
244 static struct sk_buff *page_to_skb(struct receive_queue *rq,
245 struct page *page, unsigned int offset,
246 unsigned int len, unsigned int truesize)
248 struct virtnet_info *vi = rq->vq->vdev->priv;
250 struct skb_vnet_hdr *hdr;
251 unsigned int copy, hdr_len, hdr_padded_len;
254 p = page_address(page) + offset;
256 /* copy small packet so we can reuse these pages for small data */
257 skb = netdev_alloc_skb_ip_align(vi->dev, GOOD_COPY_LEN);
261 hdr = skb_vnet_hdr(skb);
263 if (vi->mergeable_rx_bufs) {
264 hdr_len = sizeof hdr->mhdr;
265 hdr_padded_len = sizeof hdr->mhdr;
267 hdr_len = sizeof hdr->hdr;
268 hdr_padded_len = sizeof(struct padded_vnet_hdr);
271 memcpy(hdr, p, hdr_len);
274 offset += hdr_padded_len;
278 if (copy > skb_tailroom(skb))
279 copy = skb_tailroom(skb);
280 memcpy(skb_put(skb, copy), p, copy);
285 if (vi->mergeable_rx_bufs) {
287 skb_add_rx_frag(skb, 0, page, offset, len, truesize);
294 * Verify that we can indeed put this data into a skb.
295 * This is here to handle cases when the device erroneously
296 * tries to receive more than is possible. This is usually
297 * the case of a broken device.
299 if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
300 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
304 BUG_ON(offset >= PAGE_SIZE);
306 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
307 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
308 frag_size, truesize);
310 page = (struct page *)page->private;
315 give_pages(rq, page);
320 static struct sk_buff *receive_small(void *buf, unsigned int len)
322 struct sk_buff * skb = buf;
324 len -= sizeof(struct virtio_net_hdr);
330 static struct sk_buff *receive_big(struct net_device *dev,
331 struct receive_queue *rq,
335 struct page *page = buf;
336 struct sk_buff *skb = page_to_skb(rq, page, 0, len, PAGE_SIZE);
344 dev->stats.rx_dropped++;
345 give_pages(rq, page);
349 static struct sk_buff *receive_mergeable(struct net_device *dev,
350 struct receive_queue *rq,
354 void *buf = mergeable_ctx_to_buf_address(ctx);
355 struct skb_vnet_hdr *hdr = buf;
356 int num_buf = hdr->mhdr.num_buffers;
357 struct page *page = virt_to_head_page(buf);
358 int offset = buf - page_address(page);
359 unsigned int truesize = max(len, mergeable_ctx_to_buf_truesize(ctx));
361 struct sk_buff *head_skb = page_to_skb(rq, page, offset, len, truesize);
362 struct sk_buff *curr_skb = head_skb;
364 if (unlikely(!curr_skb))
369 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &len);
370 if (unlikely(!ctx)) {
371 pr_debug("%s: rx error: %d buffers out of %d missing\n",
372 dev->name, num_buf, hdr->mhdr.num_buffers);
373 dev->stats.rx_length_errors++;
377 buf = mergeable_ctx_to_buf_address(ctx);
378 page = virt_to_head_page(buf);
380 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
381 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
382 struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
386 if (curr_skb == head_skb)
387 skb_shinfo(curr_skb)->frag_list = nskb;
389 curr_skb->next = nskb;
391 head_skb->truesize += nskb->truesize;
394 truesize = max(len, mergeable_ctx_to_buf_truesize(ctx));
395 if (curr_skb != head_skb) {
396 head_skb->data_len += len;
397 head_skb->len += len;
398 head_skb->truesize += truesize;
400 offset = buf - page_address(page);
401 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
403 skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
406 skb_add_rx_frag(curr_skb, num_skb_frags, page,
407 offset, len, truesize);
411 ewma_add(&rq->mrg_avg_pkt_len, head_skb->len);
417 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &len);
418 if (unlikely(!ctx)) {
419 pr_debug("%s: rx error: %d buffers missing\n",
421 dev->stats.rx_length_errors++;
424 page = virt_to_head_page(mergeable_ctx_to_buf_address(ctx));
428 dev->stats.rx_dropped++;
429 dev_kfree_skb(head_skb);
433 static void receive_buf(struct receive_queue *rq, void *buf, unsigned int len)
435 struct virtnet_info *vi = rq->vq->vdev->priv;
436 struct net_device *dev = vi->dev;
437 struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
439 struct skb_vnet_hdr *hdr;
441 if (unlikely(len < sizeof(struct virtio_net_hdr) + ETH_HLEN)) {
442 pr_debug("%s: short packet %i\n", dev->name, len);
443 dev->stats.rx_length_errors++;
444 if (vi->mergeable_rx_bufs) {
445 unsigned long ctx = (unsigned long)buf;
446 void *base = mergeable_ctx_to_buf_address(ctx);
447 put_page(virt_to_head_page(base));
448 } else if (vi->big_packets) {
456 if (vi->mergeable_rx_bufs)
457 skb = receive_mergeable(dev, rq, (unsigned long)buf, len);
458 else if (vi->big_packets)
459 skb = receive_big(dev, rq, buf, len);
461 skb = receive_small(buf, len);
466 hdr = skb_vnet_hdr(skb);
468 u64_stats_update_begin(&stats->rx_syncp);
469 stats->rx_bytes += skb->len;
471 u64_stats_update_end(&stats->rx_syncp);
473 if (hdr->hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
474 pr_debug("Needs csum!\n");
475 if (!skb_partial_csum_set(skb,
477 hdr->hdr.csum_offset))
479 } else if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID) {
480 skb->ip_summed = CHECKSUM_UNNECESSARY;
483 skb->protocol = eth_type_trans(skb, dev);
484 pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
485 ntohs(skb->protocol), skb->len, skb->pkt_type);
487 if (hdr->hdr.gso_type != VIRTIO_NET_HDR_GSO_NONE) {
489 switch (hdr->hdr.gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
490 case VIRTIO_NET_HDR_GSO_TCPV4:
491 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
493 case VIRTIO_NET_HDR_GSO_UDP:
494 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
496 case VIRTIO_NET_HDR_GSO_TCPV6:
497 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
500 net_warn_ratelimited("%s: bad gso type %u.\n",
501 dev->name, hdr->hdr.gso_type);
505 if (hdr->hdr.gso_type & VIRTIO_NET_HDR_GSO_ECN)
506 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
508 skb_shinfo(skb)->gso_size = hdr->hdr.gso_size;
509 if (skb_shinfo(skb)->gso_size == 0) {
510 net_warn_ratelimited("%s: zero gso size.\n", dev->name);
514 /* Header must be checked, and gso_segs computed. */
515 skb_shinfo(skb)->gso_type |= SKB_GSO_DODGY;
516 skb_shinfo(skb)->gso_segs = 0;
519 skb_mark_napi_id(skb, &rq->napi);
521 netif_receive_skb(skb);
525 dev->stats.rx_frame_errors++;
529 static int add_recvbuf_small(struct receive_queue *rq, gfp_t gfp)
531 struct virtnet_info *vi = rq->vq->vdev->priv;
533 struct skb_vnet_hdr *hdr;
536 skb = __netdev_alloc_skb_ip_align(vi->dev, GOOD_PACKET_LEN, gfp);
540 skb_put(skb, GOOD_PACKET_LEN);
542 hdr = skb_vnet_hdr(skb);
543 sg_set_buf(rq->sg, &hdr->hdr, sizeof hdr->hdr);
545 skb_to_sgvec(skb, rq->sg + 1, 0, skb->len);
547 err = virtqueue_add_inbuf(rq->vq, rq->sg, 2, skb, gfp);
554 static int add_recvbuf_big(struct receive_queue *rq, gfp_t gfp)
556 struct page *first, *list = NULL;
560 /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
561 for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
562 first = get_a_page(rq, gfp);
565 give_pages(rq, list);
568 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
570 /* chain new page in list head to match sg */
571 first->private = (unsigned long)list;
575 first = get_a_page(rq, gfp);
577 give_pages(rq, list);
580 p = page_address(first);
582 /* rq->sg[0], rq->sg[1] share the same page */
583 /* a separated rq->sg[0] for virtio_net_hdr only due to QEMU bug */
584 sg_set_buf(&rq->sg[0], p, sizeof(struct virtio_net_hdr));
586 /* rq->sg[1] for data packet, from offset */
587 offset = sizeof(struct padded_vnet_hdr);
588 sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
590 /* chain first in list head */
591 first->private = (unsigned long)list;
592 err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
595 give_pages(rq, first);
600 static unsigned int get_mergeable_buf_len(struct ewma *avg_pkt_len)
602 const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
605 len = hdr_len + clamp_t(unsigned int, ewma_read(avg_pkt_len),
606 GOOD_PACKET_LEN, PAGE_SIZE - hdr_len);
607 return ALIGN(len, MERGEABLE_BUFFER_ALIGN);
610 static int add_recvbuf_mergeable(struct receive_queue *rq, gfp_t gfp)
612 struct page_frag *alloc_frag = &rq->alloc_frag;
616 unsigned int len, hole;
618 len = get_mergeable_buf_len(&rq->mrg_avg_pkt_len);
619 if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
622 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
623 ctx = mergeable_buf_to_ctx(buf, len);
624 get_page(alloc_frag->page);
625 alloc_frag->offset += len;
626 hole = alloc_frag->size - alloc_frag->offset;
628 /* To avoid internal fragmentation, if there is very likely not
629 * enough space for another buffer, add the remaining space to
630 * the current buffer. This extra space is not included in
631 * the truesize stored in ctx.
634 alloc_frag->offset += hole;
637 sg_init_one(rq->sg, buf, len);
638 err = virtqueue_add_inbuf(rq->vq, rq->sg, 1, (void *)ctx, gfp);
640 put_page(virt_to_head_page(buf));
646 * Returns false if we couldn't fill entirely (OOM).
648 * Normally run in the receive path, but can also be run from ndo_open
649 * before we're receiving packets, or from refill_work which is
650 * careful to disable receiving (using napi_disable).
652 static bool try_fill_recv(struct receive_queue *rq, gfp_t gfp)
654 struct virtnet_info *vi = rq->vq->vdev->priv;
660 if (vi->mergeable_rx_bufs)
661 err = add_recvbuf_mergeable(rq, gfp);
662 else if (vi->big_packets)
663 err = add_recvbuf_big(rq, gfp);
665 err = add_recvbuf_small(rq, gfp);
667 oom = err == -ENOMEM;
670 } while (rq->vq->num_free);
671 virtqueue_kick(rq->vq);
675 static void skb_recv_done(struct virtqueue *rvq)
677 struct virtnet_info *vi = rvq->vdev->priv;
678 struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
680 /* Schedule NAPI, Suppress further interrupts if successful. */
681 if (napi_schedule_prep(&rq->napi)) {
682 virtqueue_disable_cb(rvq);
683 __napi_schedule(&rq->napi);
687 static void virtnet_napi_enable(struct receive_queue *rq)
689 napi_enable(&rq->napi);
691 /* If all buffers were filled by other side before we napi_enabled, we
692 * won't get another interrupt, so process any outstanding packets
693 * now. virtnet_poll wants re-enable the queue, so we disable here.
694 * We synchronize against interrupts via NAPI_STATE_SCHED */
695 if (napi_schedule_prep(&rq->napi)) {
696 virtqueue_disable_cb(rq->vq);
698 __napi_schedule(&rq->napi);
703 static void refill_work(struct work_struct *work)
705 struct virtnet_info *vi =
706 container_of(work, struct virtnet_info, refill.work);
710 for (i = 0; i < vi->curr_queue_pairs; i++) {
711 struct receive_queue *rq = &vi->rq[i];
713 napi_disable(&rq->napi);
714 still_empty = !try_fill_recv(rq, GFP_KERNEL);
715 virtnet_napi_enable(rq);
717 /* In theory, this can happen: if we don't get any buffers in
718 * we will *never* try to fill again.
721 schedule_delayed_work(&vi->refill, HZ/2);
725 static int virtnet_receive(struct receive_queue *rq, int budget)
727 struct virtnet_info *vi = rq->vq->vdev->priv;
728 unsigned int len, received = 0;
731 while (received < budget &&
732 (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
733 receive_buf(rq, buf, len);
737 if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) {
738 if (!try_fill_recv(rq, GFP_ATOMIC))
739 schedule_delayed_work(&vi->refill, 0);
745 static int virtnet_poll(struct napi_struct *napi, int budget)
747 struct receive_queue *rq =
748 container_of(napi, struct receive_queue, napi);
749 unsigned int r, received = 0;
752 received += virtnet_receive(rq, budget - received);
754 /* Out of packets? */
755 if (received < budget) {
756 r = virtqueue_enable_cb_prepare(rq->vq);
758 if (unlikely(virtqueue_poll(rq->vq, r)) &&
759 napi_schedule_prep(napi)) {
760 virtqueue_disable_cb(rq->vq);
761 __napi_schedule(napi);
769 #ifdef CONFIG_NET_RX_BUSY_POLL
770 /* must be called with local_bh_disable()d */
771 static int virtnet_busy_poll(struct napi_struct *napi)
773 struct receive_queue *rq =
774 container_of(napi, struct receive_queue, napi);
775 struct virtnet_info *vi = rq->vq->vdev->priv;
776 int r, received = 0, budget = 4;
778 if (!(vi->status & VIRTIO_NET_S_LINK_UP))
779 return LL_FLUSH_FAILED;
781 if (!napi_schedule_prep(napi))
782 return LL_FLUSH_BUSY;
784 virtqueue_disable_cb(rq->vq);
787 received += virtnet_receive(rq, budget);
789 r = virtqueue_enable_cb_prepare(rq->vq);
790 clear_bit(NAPI_STATE_SCHED, &napi->state);
791 if (unlikely(virtqueue_poll(rq->vq, r)) &&
792 napi_schedule_prep(napi)) {
793 virtqueue_disable_cb(rq->vq);
794 if (received < budget) {
798 __napi_schedule(napi);
804 #endif /* CONFIG_NET_RX_BUSY_POLL */
806 static int virtnet_open(struct net_device *dev)
808 struct virtnet_info *vi = netdev_priv(dev);
811 for (i = 0; i < vi->max_queue_pairs; i++) {
812 if (i < vi->curr_queue_pairs)
813 /* Make sure we have some buffers: if oom use wq. */
814 if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
815 schedule_delayed_work(&vi->refill, 0);
816 virtnet_napi_enable(&vi->rq[i]);
822 static void free_old_xmit_skbs(struct send_queue *sq)
826 struct virtnet_info *vi = sq->vq->vdev->priv;
827 struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
829 while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
830 pr_debug("Sent skb %p\n", skb);
832 u64_stats_update_begin(&stats->tx_syncp);
833 stats->tx_bytes += skb->len;
835 u64_stats_update_end(&stats->tx_syncp);
837 dev_kfree_skb_any(skb);
841 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
843 struct skb_vnet_hdr *hdr;
844 const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
845 struct virtnet_info *vi = sq->vq->vdev->priv;
850 pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
851 if (vi->mergeable_rx_bufs)
852 hdr_len = sizeof hdr->mhdr;
854 hdr_len = sizeof hdr->hdr;
856 can_push = vi->any_header_sg &&
857 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
858 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
859 /* Even if we can, don't push here yet as this would skew
860 * csum_start offset below. */
862 hdr = (struct skb_vnet_hdr *)(skb->data - hdr_len);
864 hdr = skb_vnet_hdr(skb);
866 if (skb->ip_summed == CHECKSUM_PARTIAL) {
867 hdr->hdr.flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
868 hdr->hdr.csum_start = skb_checksum_start_offset(skb);
869 hdr->hdr.csum_offset = skb->csum_offset;
872 hdr->hdr.csum_offset = hdr->hdr.csum_start = 0;
875 if (skb_is_gso(skb)) {
876 hdr->hdr.hdr_len = skb_headlen(skb);
877 hdr->hdr.gso_size = skb_shinfo(skb)->gso_size;
878 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
879 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
880 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
881 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
882 else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
883 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_UDP;
886 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCP_ECN)
887 hdr->hdr.gso_type |= VIRTIO_NET_HDR_GSO_ECN;
889 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE;
890 hdr->hdr.gso_size = hdr->hdr.hdr_len = 0;
893 if (vi->mergeable_rx_bufs)
894 hdr->mhdr.num_buffers = 0;
897 __skb_push(skb, hdr_len);
898 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
899 /* Pull header back to avoid skew in tx bytes calculations. */
900 __skb_pull(skb, hdr_len);
902 sg_set_buf(sq->sg, hdr, hdr_len);
903 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
905 return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
908 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
910 struct virtnet_info *vi = netdev_priv(dev);
911 int qnum = skb_get_queue_mapping(skb);
912 struct send_queue *sq = &vi->sq[qnum];
915 /* Free up any pending old buffers before queueing new ones. */
916 free_old_xmit_skbs(sq);
918 /* Try to transmit */
919 err = xmit_skb(sq, skb);
921 /* This should not happen! */
923 dev->stats.tx_fifo_errors++;
926 "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
927 dev->stats.tx_dropped++;
928 dev_kfree_skb_any(skb);
931 virtqueue_kick(sq->vq);
933 /* Don't wait up for transmitted skbs to be freed. */
937 /* Apparently nice girls don't return TX_BUSY; stop the queue
938 * before it gets out of hand. Naturally, this wastes entries. */
939 if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
940 netif_stop_subqueue(dev, qnum);
941 if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
942 /* More just got used, free them then recheck. */
943 free_old_xmit_skbs(sq);
944 if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
945 netif_start_subqueue(dev, qnum);
946 virtqueue_disable_cb(sq->vq);
955 * Send command via the control virtqueue and check status. Commands
956 * supported by the hypervisor, as indicated by feature bits, should
957 * never fail unless improperly formatted.
959 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
960 struct scatterlist *out)
962 struct scatterlist *sgs[4], hdr, stat;
963 struct virtio_net_ctrl_hdr ctrl;
964 virtio_net_ctrl_ack status = ~0;
965 unsigned out_num = 0, tmp;
967 /* Caller should know better */
968 BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
973 sg_init_one(&hdr, &ctrl, sizeof(ctrl));
974 sgs[out_num++] = &hdr;
977 sgs[out_num++] = out;
979 /* Add return status. */
980 sg_init_one(&stat, &status, sizeof(status));
981 sgs[out_num] = &stat;
983 BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
984 virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
986 if (unlikely(!virtqueue_kick(vi->cvq)))
987 return status == VIRTIO_NET_OK;
989 /* Spin for a response, the kick causes an ioport write, trapping
990 * into the hypervisor, so the request should be handled immediately.
992 while (!virtqueue_get_buf(vi->cvq, &tmp) &&
993 !virtqueue_is_broken(vi->cvq))
996 return status == VIRTIO_NET_OK;
999 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1001 struct virtnet_info *vi = netdev_priv(dev);
1002 struct virtio_device *vdev = vi->vdev;
1004 struct sockaddr *addr = p;
1005 struct scatterlist sg;
1007 ret = eth_prepare_mac_addr_change(dev, p);
1011 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1012 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1013 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1014 VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1015 dev_warn(&vdev->dev,
1016 "Failed to set mac address by vq command.\n");
1019 } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
1022 /* Naturally, this has an atomicity problem. */
1023 for (i = 0; i < dev->addr_len; i++)
1024 virtio_cwrite8(vdev,
1025 offsetof(struct virtio_net_config, mac) +
1026 i, addr->sa_data[i]);
1029 eth_commit_mac_addr_change(dev, p);
1034 static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
1035 struct rtnl_link_stats64 *tot)
1037 struct virtnet_info *vi = netdev_priv(dev);
1041 for_each_possible_cpu(cpu) {
1042 struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
1043 u64 tpackets, tbytes, rpackets, rbytes;
1046 start = u64_stats_fetch_begin_irq(&stats->tx_syncp);
1047 tpackets = stats->tx_packets;
1048 tbytes = stats->tx_bytes;
1049 } while (u64_stats_fetch_retry_irq(&stats->tx_syncp, start));
1052 start = u64_stats_fetch_begin_irq(&stats->rx_syncp);
1053 rpackets = stats->rx_packets;
1054 rbytes = stats->rx_bytes;
1055 } while (u64_stats_fetch_retry_irq(&stats->rx_syncp, start));
1057 tot->rx_packets += rpackets;
1058 tot->tx_packets += tpackets;
1059 tot->rx_bytes += rbytes;
1060 tot->tx_bytes += tbytes;
1063 tot->tx_dropped = dev->stats.tx_dropped;
1064 tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1065 tot->rx_dropped = dev->stats.rx_dropped;
1066 tot->rx_length_errors = dev->stats.rx_length_errors;
1067 tot->rx_frame_errors = dev->stats.rx_frame_errors;
1072 #ifdef CONFIG_NET_POLL_CONTROLLER
1073 static void virtnet_netpoll(struct net_device *dev)
1075 struct virtnet_info *vi = netdev_priv(dev);
1078 for (i = 0; i < vi->curr_queue_pairs; i++)
1079 napi_schedule(&vi->rq[i].napi);
1083 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1086 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1087 VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1088 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1092 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1094 struct scatterlist sg;
1095 struct virtio_net_ctrl_mq s;
1096 struct net_device *dev = vi->dev;
1098 if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1101 s.virtqueue_pairs = queue_pairs;
1102 sg_init_one(&sg, &s, sizeof(s));
1104 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1105 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1106 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1110 vi->curr_queue_pairs = queue_pairs;
1111 /* virtnet_open() will refill when device is going to up. */
1112 if (dev->flags & IFF_UP)
1113 schedule_delayed_work(&vi->refill, 0);
1119 static int virtnet_close(struct net_device *dev)
1121 struct virtnet_info *vi = netdev_priv(dev);
1124 /* Make sure refill_work doesn't re-enable napi! */
1125 cancel_delayed_work_sync(&vi->refill);
1127 for (i = 0; i < vi->max_queue_pairs; i++)
1128 napi_disable(&vi->rq[i].napi);
1133 static void virtnet_set_rx_mode(struct net_device *dev)
1135 struct virtnet_info *vi = netdev_priv(dev);
1136 struct scatterlist sg[2];
1137 u8 promisc, allmulti;
1138 struct virtio_net_ctrl_mac *mac_data;
1139 struct netdev_hw_addr *ha;
1145 /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1146 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1149 promisc = ((dev->flags & IFF_PROMISC) != 0);
1150 allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1152 sg_init_one(sg, &promisc, sizeof(promisc));
1154 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1155 VIRTIO_NET_CTRL_RX_PROMISC, sg))
1156 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1157 promisc ? "en" : "dis");
1159 sg_init_one(sg, &allmulti, sizeof(allmulti));
1161 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1162 VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1163 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1164 allmulti ? "en" : "dis");
1166 uc_count = netdev_uc_count(dev);
1167 mc_count = netdev_mc_count(dev);
1168 /* MAC filter - use one buffer for both lists */
1169 buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1170 (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1175 sg_init_table(sg, 2);
1177 /* Store the unicast list and count in the front of the buffer */
1178 mac_data->entries = uc_count;
1180 netdev_for_each_uc_addr(ha, dev)
1181 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1183 sg_set_buf(&sg[0], mac_data,
1184 sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1186 /* multicast list and count fill the end */
1187 mac_data = (void *)&mac_data->macs[uc_count][0];
1189 mac_data->entries = mc_count;
1191 netdev_for_each_mc_addr(ha, dev)
1192 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1194 sg_set_buf(&sg[1], mac_data,
1195 sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1197 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1198 VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1199 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1204 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1205 __be16 proto, u16 vid)
1207 struct virtnet_info *vi = netdev_priv(dev);
1208 struct scatterlist sg;
1210 sg_init_one(&sg, &vid, sizeof(vid));
1212 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1213 VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1214 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1218 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1219 __be16 proto, u16 vid)
1221 struct virtnet_info *vi = netdev_priv(dev);
1222 struct scatterlist sg;
1224 sg_init_one(&sg, &vid, sizeof(vid));
1226 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1227 VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1228 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1232 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1236 if (vi->affinity_hint_set) {
1237 for (i = 0; i < vi->max_queue_pairs; i++) {
1238 virtqueue_set_affinity(vi->rq[i].vq, -1);
1239 virtqueue_set_affinity(vi->sq[i].vq, -1);
1242 vi->affinity_hint_set = false;
1246 static void virtnet_set_affinity(struct virtnet_info *vi)
1251 /* In multiqueue mode, when the number of cpu is equal to the number of
1252 * queue pairs, we let the queue pairs to be private to one cpu by
1253 * setting the affinity hint to eliminate the contention.
1255 if (vi->curr_queue_pairs == 1 ||
1256 vi->max_queue_pairs != num_online_cpus()) {
1257 virtnet_clean_affinity(vi, -1);
1262 for_each_online_cpu(cpu) {
1263 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1264 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1265 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
1269 vi->affinity_hint_set = true;
1272 static int virtnet_cpu_callback(struct notifier_block *nfb,
1273 unsigned long action, void *hcpu)
1275 struct virtnet_info *vi = container_of(nfb, struct virtnet_info, nb);
1277 switch(action & ~CPU_TASKS_FROZEN) {
1279 case CPU_DOWN_FAILED:
1281 virtnet_set_affinity(vi);
1283 case CPU_DOWN_PREPARE:
1284 virtnet_clean_affinity(vi, (long)hcpu);
1293 static void virtnet_get_ringparam(struct net_device *dev,
1294 struct ethtool_ringparam *ring)
1296 struct virtnet_info *vi = netdev_priv(dev);
1298 ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1299 ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1300 ring->rx_pending = ring->rx_max_pending;
1301 ring->tx_pending = ring->tx_max_pending;
1305 static void virtnet_get_drvinfo(struct net_device *dev,
1306 struct ethtool_drvinfo *info)
1308 struct virtnet_info *vi = netdev_priv(dev);
1309 struct virtio_device *vdev = vi->vdev;
1311 strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1312 strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1313 strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1317 /* TODO: Eliminate OOO packets during switching */
1318 static int virtnet_set_channels(struct net_device *dev,
1319 struct ethtool_channels *channels)
1321 struct virtnet_info *vi = netdev_priv(dev);
1322 u16 queue_pairs = channels->combined_count;
1325 /* We don't support separate rx/tx channels.
1326 * We don't allow setting 'other' channels.
1328 if (channels->rx_count || channels->tx_count || channels->other_count)
1331 if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
1335 err = virtnet_set_queues(vi, queue_pairs);
1337 netif_set_real_num_tx_queues(dev, queue_pairs);
1338 netif_set_real_num_rx_queues(dev, queue_pairs);
1340 virtnet_set_affinity(vi);
1347 static void virtnet_get_channels(struct net_device *dev,
1348 struct ethtool_channels *channels)
1350 struct virtnet_info *vi = netdev_priv(dev);
1352 channels->combined_count = vi->curr_queue_pairs;
1353 channels->max_combined = vi->max_queue_pairs;
1354 channels->max_other = 0;
1355 channels->rx_count = 0;
1356 channels->tx_count = 0;
1357 channels->other_count = 0;
1360 static const struct ethtool_ops virtnet_ethtool_ops = {
1361 .get_drvinfo = virtnet_get_drvinfo,
1362 .get_link = ethtool_op_get_link,
1363 .get_ringparam = virtnet_get_ringparam,
1364 .set_channels = virtnet_set_channels,
1365 .get_channels = virtnet_get_channels,
1369 #define MAX_MTU 65535
1371 static int virtnet_change_mtu(struct net_device *dev, int new_mtu)
1373 if (new_mtu < MIN_MTU || new_mtu > MAX_MTU)
1379 static const struct net_device_ops virtnet_netdev = {
1380 .ndo_open = virtnet_open,
1381 .ndo_stop = virtnet_close,
1382 .ndo_start_xmit = start_xmit,
1383 .ndo_validate_addr = eth_validate_addr,
1384 .ndo_set_mac_address = virtnet_set_mac_address,
1385 .ndo_set_rx_mode = virtnet_set_rx_mode,
1386 .ndo_change_mtu = virtnet_change_mtu,
1387 .ndo_get_stats64 = virtnet_stats,
1388 .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
1389 .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
1390 #ifdef CONFIG_NET_POLL_CONTROLLER
1391 .ndo_poll_controller = virtnet_netpoll,
1393 #ifdef CONFIG_NET_RX_BUSY_POLL
1394 .ndo_busy_poll = virtnet_busy_poll,
1398 static void virtnet_config_changed_work(struct work_struct *work)
1400 struct virtnet_info *vi =
1401 container_of(work, struct virtnet_info, config_work);
1404 if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
1405 struct virtio_net_config, status, &v) < 0)
1408 if (v & VIRTIO_NET_S_ANNOUNCE) {
1409 netdev_notify_peers(vi->dev);
1410 virtnet_ack_link_announce(vi);
1413 /* Ignore unknown (future) status bits */
1414 v &= VIRTIO_NET_S_LINK_UP;
1416 if (vi->status == v)
1421 if (vi->status & VIRTIO_NET_S_LINK_UP) {
1422 netif_carrier_on(vi->dev);
1423 netif_tx_wake_all_queues(vi->dev);
1425 netif_carrier_off(vi->dev);
1426 netif_tx_stop_all_queues(vi->dev);
1432 static void virtnet_config_changed(struct virtio_device *vdev)
1434 struct virtnet_info *vi = vdev->priv;
1436 schedule_work(&vi->config_work);
1439 static void virtnet_free_queues(struct virtnet_info *vi)
1443 for (i = 0; i < vi->max_queue_pairs; i++)
1444 netif_napi_del(&vi->rq[i].napi);
1450 static void free_receive_bufs(struct virtnet_info *vi)
1454 for (i = 0; i < vi->max_queue_pairs; i++) {
1455 while (vi->rq[i].pages)
1456 __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
1460 static void free_receive_page_frags(struct virtnet_info *vi)
1463 for (i = 0; i < vi->max_queue_pairs; i++)
1464 if (vi->rq[i].alloc_frag.page)
1465 put_page(vi->rq[i].alloc_frag.page);
1468 static void free_unused_bufs(struct virtnet_info *vi)
1473 for (i = 0; i < vi->max_queue_pairs; i++) {
1474 struct virtqueue *vq = vi->sq[i].vq;
1475 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
1479 for (i = 0; i < vi->max_queue_pairs; i++) {
1480 struct virtqueue *vq = vi->rq[i].vq;
1482 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1483 if (vi->mergeable_rx_bufs) {
1484 unsigned long ctx = (unsigned long)buf;
1485 void *base = mergeable_ctx_to_buf_address(ctx);
1486 put_page(virt_to_head_page(base));
1487 } else if (vi->big_packets) {
1488 give_pages(&vi->rq[i], buf);
1496 static void virtnet_del_vqs(struct virtnet_info *vi)
1498 struct virtio_device *vdev = vi->vdev;
1500 virtnet_clean_affinity(vi, -1);
1502 vdev->config->del_vqs(vdev);
1504 virtnet_free_queues(vi);
1507 static int virtnet_find_vqs(struct virtnet_info *vi)
1509 vq_callback_t **callbacks;
1510 struct virtqueue **vqs;
1515 /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
1516 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
1517 * possible control vq.
1519 total_vqs = vi->max_queue_pairs * 2 +
1520 virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
1522 /* Allocate space for find_vqs parameters */
1523 vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
1526 callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
1529 names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
1533 /* Parameters for control virtqueue, if any */
1535 callbacks[total_vqs - 1] = NULL;
1536 names[total_vqs - 1] = "control";
1539 /* Allocate/initialize parameters for send/receive virtqueues */
1540 for (i = 0; i < vi->max_queue_pairs; i++) {
1541 callbacks[rxq2vq(i)] = skb_recv_done;
1542 callbacks[txq2vq(i)] = skb_xmit_done;
1543 sprintf(vi->rq[i].name, "input.%d", i);
1544 sprintf(vi->sq[i].name, "output.%d", i);
1545 names[rxq2vq(i)] = vi->rq[i].name;
1546 names[txq2vq(i)] = vi->sq[i].name;
1549 ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
1555 vi->cvq = vqs[total_vqs - 1];
1556 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
1557 vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
1560 for (i = 0; i < vi->max_queue_pairs; i++) {
1561 vi->rq[i].vq = vqs[rxq2vq(i)];
1562 vi->sq[i].vq = vqs[txq2vq(i)];
1581 static int virtnet_alloc_queues(struct virtnet_info *vi)
1585 vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
1588 vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
1592 INIT_DELAYED_WORK(&vi->refill, refill_work);
1593 for (i = 0; i < vi->max_queue_pairs; i++) {
1594 vi->rq[i].pages = NULL;
1595 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
1597 napi_hash_add(&vi->rq[i].napi);
1599 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
1600 ewma_init(&vi->rq[i].mrg_avg_pkt_len, 1, RECEIVE_AVG_WEIGHT);
1601 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
1612 static int init_vqs(struct virtnet_info *vi)
1616 /* Allocate send & receive queues */
1617 ret = virtnet_alloc_queues(vi);
1621 ret = virtnet_find_vqs(vi);
1626 virtnet_set_affinity(vi);
1632 virtnet_free_queues(vi);
1638 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
1639 struct rx_queue_attribute *attribute, char *buf)
1641 struct virtnet_info *vi = netdev_priv(queue->dev);
1642 unsigned int queue_index = get_netdev_rx_queue_index(queue);
1645 BUG_ON(queue_index >= vi->max_queue_pairs);
1646 avg = &vi->rq[queue_index].mrg_avg_pkt_len;
1647 return sprintf(buf, "%u\n", get_mergeable_buf_len(avg));
1650 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
1651 __ATTR_RO(mergeable_rx_buffer_size);
1653 static struct attribute *virtio_net_mrg_rx_attrs[] = {
1654 &mergeable_rx_buffer_size_attribute.attr,
1658 static const struct attribute_group virtio_net_mrg_rx_group = {
1659 .name = "virtio_net",
1660 .attrs = virtio_net_mrg_rx_attrs
1664 static int virtnet_probe(struct virtio_device *vdev)
1667 struct net_device *dev;
1668 struct virtnet_info *vi;
1669 u16 max_queue_pairs;
1671 /* Find if host supports multiqueue virtio_net device */
1672 err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
1673 struct virtio_net_config,
1674 max_virtqueue_pairs, &max_queue_pairs);
1676 /* We need at least 2 queue's */
1677 if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
1678 max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
1679 !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1680 max_queue_pairs = 1;
1682 /* Allocate ourselves a network device with room for our info */
1683 dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
1687 /* Set up network device as normal. */
1688 dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
1689 dev->netdev_ops = &virtnet_netdev;
1690 dev->features = NETIF_F_HIGHDMA;
1692 dev->ethtool_ops = &virtnet_ethtool_ops;
1693 SET_NETDEV_DEV(dev, &vdev->dev);
1695 /* Do we support "hardware" checksums? */
1696 if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
1697 /* This opens up the world of extra features. */
1698 dev->hw_features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1700 dev->features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1702 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
1703 dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
1704 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
1706 /* Individual feature bits: what can host handle? */
1707 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
1708 dev->hw_features |= NETIF_F_TSO;
1709 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
1710 dev->hw_features |= NETIF_F_TSO6;
1711 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
1712 dev->hw_features |= NETIF_F_TSO_ECN;
1713 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
1714 dev->hw_features |= NETIF_F_UFO;
1717 dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
1718 /* (!csum && gso) case will be fixed by register_netdev() */
1720 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
1721 dev->features |= NETIF_F_RXCSUM;
1723 dev->vlan_features = dev->features;
1725 /* Configuration may specify what MAC to use. Otherwise random. */
1726 if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
1727 virtio_cread_bytes(vdev,
1728 offsetof(struct virtio_net_config, mac),
1729 dev->dev_addr, dev->addr_len);
1731 eth_hw_addr_random(dev);
1733 /* Set up our device-specific information */
1734 vi = netdev_priv(dev);
1738 vi->stats = alloc_percpu(struct virtnet_stats);
1740 if (vi->stats == NULL)
1743 for_each_possible_cpu(i) {
1744 struct virtnet_stats *virtnet_stats;
1745 virtnet_stats = per_cpu_ptr(vi->stats, i);
1746 u64_stats_init(&virtnet_stats->tx_syncp);
1747 u64_stats_init(&virtnet_stats->rx_syncp);
1750 INIT_WORK(&vi->config_work, virtnet_config_changed_work);
1752 /* If we can receive ANY GSO packets, we must allocate large ones. */
1753 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1754 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1755 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
1756 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
1757 vi->big_packets = true;
1759 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
1760 vi->mergeable_rx_bufs = true;
1762 if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT))
1763 vi->any_header_sg = true;
1765 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1768 if (vi->any_header_sg) {
1769 if (vi->mergeable_rx_bufs)
1770 dev->needed_headroom = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1772 dev->needed_headroom = sizeof(struct virtio_net_hdr);
1775 /* Use single tx/rx queue pair as default */
1776 vi->curr_queue_pairs = 1;
1777 vi->max_queue_pairs = max_queue_pairs;
1779 /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
1785 if (vi->mergeable_rx_bufs)
1786 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
1788 netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
1789 netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
1791 err = register_netdev(dev);
1793 pr_debug("virtio_net: registering device failed\n");
1797 /* Last of all, set up some receive buffers. */
1798 for (i = 0; i < vi->curr_queue_pairs; i++) {
1799 try_fill_recv(&vi->rq[i], GFP_KERNEL);
1801 /* If we didn't even get one input buffer, we're useless. */
1802 if (vi->rq[i].vq->num_free ==
1803 virtqueue_get_vring_size(vi->rq[i].vq)) {
1804 free_unused_bufs(vi);
1806 goto free_recv_bufs;
1810 vi->nb.notifier_call = &virtnet_cpu_callback;
1811 err = register_hotcpu_notifier(&vi->nb);
1813 pr_debug("virtio_net: registering cpu notifier failed\n");
1814 goto free_recv_bufs;
1817 /* Assume link up if device can't report link status,
1818 otherwise get link status from config. */
1819 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
1820 netif_carrier_off(dev);
1821 schedule_work(&vi->config_work);
1823 vi->status = VIRTIO_NET_S_LINK_UP;
1824 netif_carrier_on(dev);
1827 pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
1828 dev->name, max_queue_pairs);
1833 free_receive_bufs(vi);
1834 unregister_netdev(dev);
1836 cancel_delayed_work_sync(&vi->refill);
1837 free_receive_page_frags(vi);
1838 virtnet_del_vqs(vi);
1840 free_percpu(vi->stats);
1846 static void remove_vq_common(struct virtnet_info *vi)
1848 vi->vdev->config->reset(vi->vdev);
1850 /* Free unused buffers in both send and recv, if any. */
1851 free_unused_bufs(vi);
1853 free_receive_bufs(vi);
1855 free_receive_page_frags(vi);
1857 virtnet_del_vqs(vi);
1860 static void virtnet_remove(struct virtio_device *vdev)
1862 struct virtnet_info *vi = vdev->priv;
1864 unregister_hotcpu_notifier(&vi->nb);
1866 /* Make sure no work handler is accessing the device. */
1867 flush_work(&vi->config_work);
1869 unregister_netdev(vi->dev);
1871 remove_vq_common(vi);
1873 free_percpu(vi->stats);
1874 free_netdev(vi->dev);
1877 #ifdef CONFIG_PM_SLEEP
1878 static int virtnet_freeze(struct virtio_device *vdev)
1880 struct virtnet_info *vi = vdev->priv;
1883 unregister_hotcpu_notifier(&vi->nb);
1885 /* Make sure no work handler is accessing the device */
1886 flush_work(&vi->config_work);
1888 netif_device_detach(vi->dev);
1889 cancel_delayed_work_sync(&vi->refill);
1891 if (netif_running(vi->dev)) {
1892 for (i = 0; i < vi->max_queue_pairs; i++) {
1893 napi_disable(&vi->rq[i].napi);
1894 napi_hash_del(&vi->rq[i].napi);
1895 netif_napi_del(&vi->rq[i].napi);
1899 remove_vq_common(vi);
1904 static int virtnet_restore(struct virtio_device *vdev)
1906 struct virtnet_info *vi = vdev->priv;
1913 if (netif_running(vi->dev)) {
1914 for (i = 0; i < vi->curr_queue_pairs; i++)
1915 if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
1916 schedule_delayed_work(&vi->refill, 0);
1918 for (i = 0; i < vi->max_queue_pairs; i++)
1919 virtnet_napi_enable(&vi->rq[i]);
1922 netif_device_attach(vi->dev);
1925 virtnet_set_queues(vi, vi->curr_queue_pairs);
1928 err = register_hotcpu_notifier(&vi->nb);
1936 static struct virtio_device_id id_table[] = {
1937 { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
1941 static unsigned int features[] = {
1942 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM,
1943 VIRTIO_NET_F_GSO, VIRTIO_NET_F_MAC,
1944 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6,
1945 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6,
1946 VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO,
1947 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ,
1948 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN,
1949 VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ,
1950 VIRTIO_NET_F_CTRL_MAC_ADDR,
1951 VIRTIO_F_ANY_LAYOUT,
1954 static struct virtio_driver virtio_net_driver = {
1955 .feature_table = features,
1956 .feature_table_size = ARRAY_SIZE(features),
1957 .driver.name = KBUILD_MODNAME,
1958 .driver.owner = THIS_MODULE,
1959 .id_table = id_table,
1960 .probe = virtnet_probe,
1961 .remove = virtnet_remove,
1962 .config_changed = virtnet_config_changed,
1963 #ifdef CONFIG_PM_SLEEP
1964 .freeze = virtnet_freeze,
1965 .restore = virtnet_restore,
1969 module_virtio_driver(virtio_net_driver);
1971 MODULE_DEVICE_TABLE(virtio, id_table);
1972 MODULE_DESCRIPTION("Virtio network driver");
1973 MODULE_LICENSE("GPL");