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 /* RX packet size EWMA. The average packet size is used to determine the packet
44 * buffer size when refilling RX rings. As the entire RX ring may be refilled
45 * at once, the weight is chosen so that the EWMA will be insensitive to short-
46 * term, transient changes in packet size.
48 DECLARE_EWMA(pkt_len, 1, 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.$index */
77 /* Internal representation of a receive virtqueue */
78 struct receive_queue {
79 /* Virtqueue associated with this receive_queue */
82 struct napi_struct napi;
84 /* Chain pages by the private ptr. */
87 /* Average packet length for mergeable receive buffers. */
88 struct ewma_pkt_len mrg_avg_pkt_len;
90 /* Page frag for packet buffer allocation. */
91 struct page_frag alloc_frag;
93 /* RX: fragments + linear part + virtio header */
94 struct scatterlist sg[MAX_SKB_FRAGS + 2];
96 /* Name of this receive queue: input.$index */
100 struct virtnet_info {
101 struct virtio_device *vdev;
102 struct virtqueue *cvq;
103 struct net_device *dev;
104 struct send_queue *sq;
105 struct receive_queue *rq;
108 /* Max # of queue pairs supported by the device */
111 /* # of queue pairs currently used by the driver */
112 u16 curr_queue_pairs;
114 /* I like... big packets and I cannot lie! */
117 /* Host will merge rx buffers for big packets (shake it! shake it!) */
118 bool mergeable_rx_bufs;
120 /* Has control virtqueue */
123 /* Host can handle any s/g split between our header and packet data */
126 /* Packet virtio header size */
129 /* Active statistics */
130 struct virtnet_stats __percpu *stats;
132 /* Work struct for refilling if we run low on memory. */
133 struct delayed_work refill;
135 /* Work struct for config space updates */
136 struct work_struct config_work;
138 /* Does the affinity hint is set for virtqueues? */
139 bool affinity_hint_set;
141 /* CPU hot plug notifier */
142 struct notifier_block nb;
144 /* Control VQ buffers: protected by the rtnl lock */
145 struct virtio_net_ctrl_hdr ctrl_hdr;
146 virtio_net_ctrl_ack ctrl_status;
151 struct padded_vnet_hdr {
152 struct virtio_net_hdr_mrg_rxbuf hdr;
154 * hdr is in a separate sg buffer, and data sg buffer shares same page
155 * with this header sg. This padding makes next sg 16 byte aligned
161 /* Converting between virtqueue no. and kernel tx/rx queue no.
162 * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
164 static int vq2txq(struct virtqueue *vq)
166 return (vq->index - 1) / 2;
169 static int txq2vq(int txq)
174 static int vq2rxq(struct virtqueue *vq)
176 return vq->index / 2;
179 static int rxq2vq(int rxq)
184 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
186 return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
190 * private is used to chain pages for big packets, put the whole
191 * most recent used list in the beginning for reuse
193 static void give_pages(struct receive_queue *rq, struct page *page)
197 /* Find end of list, sew whole thing into vi->rq.pages. */
198 for (end = page; end->private; end = (struct page *)end->private);
199 end->private = (unsigned long)rq->pages;
203 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
205 struct page *p = rq->pages;
208 rq->pages = (struct page *)p->private;
209 /* clear private here, it is used to chain pages */
212 p = alloc_page(gfp_mask);
216 static void skb_xmit_done(struct virtqueue *vq)
218 struct virtnet_info *vi = vq->vdev->priv;
220 /* Suppress further interrupts. */
221 virtqueue_disable_cb(vq);
223 /* We were probably waiting for more output buffers. */
224 netif_wake_subqueue(vi->dev, vq2txq(vq));
227 static unsigned int mergeable_ctx_to_buf_truesize(unsigned long mrg_ctx)
229 unsigned int truesize = mrg_ctx & (MERGEABLE_BUFFER_ALIGN - 1);
230 return (truesize + 1) * MERGEABLE_BUFFER_ALIGN;
233 static void *mergeable_ctx_to_buf_address(unsigned long mrg_ctx)
235 return (void *)(mrg_ctx & -MERGEABLE_BUFFER_ALIGN);
239 static unsigned long mergeable_buf_to_ctx(void *buf, unsigned int truesize)
241 unsigned int size = truesize / MERGEABLE_BUFFER_ALIGN;
242 return (unsigned long)buf | (size - 1);
245 /* Called from bottom half context */
246 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
247 struct receive_queue *rq,
248 struct page *page, unsigned int offset,
249 unsigned int len, unsigned int truesize)
252 struct virtio_net_hdr_mrg_rxbuf *hdr;
253 unsigned int copy, hdr_len, hdr_padded_len;
256 p = page_address(page) + offset;
258 /* copy small packet so we can reuse these pages for small data */
259 skb = netdev_alloc_skb_ip_align(vi->dev, GOOD_COPY_LEN);
263 hdr = skb_vnet_hdr(skb);
265 hdr_len = vi->hdr_len;
266 if (vi->mergeable_rx_bufs)
267 hdr_padded_len = sizeof *hdr;
269 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(struct virtnet_info *vi, void *buf, unsigned int len)
322 struct sk_buff * skb = buf;
330 static struct sk_buff *receive_big(struct net_device *dev,
331 struct virtnet_info *vi,
332 struct receive_queue *rq,
336 struct page *page = buf;
337 struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len, PAGE_SIZE);
345 dev->stats.rx_dropped++;
346 give_pages(rq, page);
350 static struct sk_buff *receive_mergeable(struct net_device *dev,
351 struct virtnet_info *vi,
352 struct receive_queue *rq,
356 void *buf = mergeable_ctx_to_buf_address(ctx);
357 struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
358 u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
359 struct page *page = virt_to_head_page(buf);
360 int offset = buf - page_address(page);
361 unsigned int truesize = max(len, mergeable_ctx_to_buf_truesize(ctx));
363 struct sk_buff *head_skb = page_to_skb(vi, rq, page, offset, len,
365 struct sk_buff *curr_skb = head_skb;
367 if (unlikely(!curr_skb))
372 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &len);
373 if (unlikely(!ctx)) {
374 pr_debug("%s: rx error: %d buffers out of %d missing\n",
376 virtio16_to_cpu(vi->vdev,
378 dev->stats.rx_length_errors++;
382 buf = mergeable_ctx_to_buf_address(ctx);
383 page = virt_to_head_page(buf);
385 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
386 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
387 struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
391 if (curr_skb == head_skb)
392 skb_shinfo(curr_skb)->frag_list = nskb;
394 curr_skb->next = nskb;
396 head_skb->truesize += nskb->truesize;
399 truesize = max(len, mergeable_ctx_to_buf_truesize(ctx));
400 if (curr_skb != head_skb) {
401 head_skb->data_len += len;
402 head_skb->len += len;
403 head_skb->truesize += truesize;
405 offset = buf - page_address(page);
406 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
408 skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
411 skb_add_rx_frag(curr_skb, num_skb_frags, page,
412 offset, len, truesize);
416 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
422 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &len);
423 if (unlikely(!ctx)) {
424 pr_debug("%s: rx error: %d buffers missing\n",
426 dev->stats.rx_length_errors++;
429 page = virt_to_head_page(mergeable_ctx_to_buf_address(ctx));
433 dev->stats.rx_dropped++;
434 dev_kfree_skb(head_skb);
438 static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
439 void *buf, unsigned int len)
441 struct net_device *dev = vi->dev;
442 struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
444 struct virtio_net_hdr_mrg_rxbuf *hdr;
446 if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
447 pr_debug("%s: short packet %i\n", dev->name, len);
448 dev->stats.rx_length_errors++;
449 if (vi->mergeable_rx_bufs) {
450 unsigned long ctx = (unsigned long)buf;
451 void *base = mergeable_ctx_to_buf_address(ctx);
452 put_page(virt_to_head_page(base));
453 } else if (vi->big_packets) {
461 if (vi->mergeable_rx_bufs)
462 skb = receive_mergeable(dev, vi, rq, (unsigned long)buf, len);
463 else if (vi->big_packets)
464 skb = receive_big(dev, vi, rq, buf, len);
466 skb = receive_small(vi, buf, len);
471 hdr = skb_vnet_hdr(skb);
473 u64_stats_update_begin(&stats->rx_syncp);
474 stats->rx_bytes += skb->len;
476 u64_stats_update_end(&stats->rx_syncp);
478 if (hdr->hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
479 pr_debug("Needs csum!\n");
480 if (!skb_partial_csum_set(skb,
481 virtio16_to_cpu(vi->vdev, hdr->hdr.csum_start),
482 virtio16_to_cpu(vi->vdev, hdr->hdr.csum_offset)))
484 } else if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID) {
485 skb->ip_summed = CHECKSUM_UNNECESSARY;
488 skb->protocol = eth_type_trans(skb, dev);
489 pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
490 ntohs(skb->protocol), skb->len, skb->pkt_type);
492 if (hdr->hdr.gso_type != VIRTIO_NET_HDR_GSO_NONE) {
494 switch (hdr->hdr.gso_type & ~VIRTIO_NET_HDR_GSO_ECN) {
495 case VIRTIO_NET_HDR_GSO_TCPV4:
496 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV4;
498 case VIRTIO_NET_HDR_GSO_UDP:
499 skb_shinfo(skb)->gso_type = SKB_GSO_UDP;
501 case VIRTIO_NET_HDR_GSO_TCPV6:
502 skb_shinfo(skb)->gso_type = SKB_GSO_TCPV6;
505 net_warn_ratelimited("%s: bad gso type %u.\n",
506 dev->name, hdr->hdr.gso_type);
510 if (hdr->hdr.gso_type & VIRTIO_NET_HDR_GSO_ECN)
511 skb_shinfo(skb)->gso_type |= SKB_GSO_TCP_ECN;
513 skb_shinfo(skb)->gso_size = virtio16_to_cpu(vi->vdev,
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 napi_gro_receive(&rq->napi, skb);
529 dev->stats.rx_frame_errors++;
533 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
537 struct virtio_net_hdr_mrg_rxbuf *hdr;
540 skb = __netdev_alloc_skb_ip_align(vi->dev, GOOD_PACKET_LEN, gfp);
544 skb_put(skb, GOOD_PACKET_LEN);
546 hdr = skb_vnet_hdr(skb);
547 sg_init_table(rq->sg, 2);
548 sg_set_buf(rq->sg, hdr, vi->hdr_len);
549 skb_to_sgvec(skb, rq->sg + 1, 0, skb->len);
551 err = virtqueue_add_inbuf(rq->vq, rq->sg, 2, skb, gfp);
558 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
561 struct page *first, *list = NULL;
565 sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
567 /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
568 for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
569 first = get_a_page(rq, gfp);
572 give_pages(rq, list);
575 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
577 /* chain new page in list head to match sg */
578 first->private = (unsigned long)list;
582 first = get_a_page(rq, gfp);
584 give_pages(rq, list);
587 p = page_address(first);
589 /* rq->sg[0], rq->sg[1] share the same page */
590 /* a separated rq->sg[0] for header - required in case !any_header_sg */
591 sg_set_buf(&rq->sg[0], p, vi->hdr_len);
593 /* rq->sg[1] for data packet, from offset */
594 offset = sizeof(struct padded_vnet_hdr);
595 sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
597 /* chain first in list head */
598 first->private = (unsigned long)list;
599 err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
602 give_pages(rq, first);
607 static unsigned int get_mergeable_buf_len(struct ewma_pkt_len *avg_pkt_len)
609 const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
612 len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
613 GOOD_PACKET_LEN, PAGE_SIZE - hdr_len);
614 return ALIGN(len, MERGEABLE_BUFFER_ALIGN);
617 static int add_recvbuf_mergeable(struct receive_queue *rq, gfp_t gfp)
619 struct page_frag *alloc_frag = &rq->alloc_frag;
623 unsigned int len, hole;
625 len = get_mergeable_buf_len(&rq->mrg_avg_pkt_len);
626 if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
629 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
630 ctx = mergeable_buf_to_ctx(buf, len);
631 get_page(alloc_frag->page);
632 alloc_frag->offset += len;
633 hole = alloc_frag->size - alloc_frag->offset;
635 /* To avoid internal fragmentation, if there is very likely not
636 * enough space for another buffer, add the remaining space to
637 * the current buffer. This extra space is not included in
638 * the truesize stored in ctx.
641 alloc_frag->offset += hole;
644 sg_init_one(rq->sg, buf, len);
645 err = virtqueue_add_inbuf(rq->vq, rq->sg, 1, (void *)ctx, gfp);
647 put_page(virt_to_head_page(buf));
653 * Returns false if we couldn't fill entirely (OOM).
655 * Normally run in the receive path, but can also be run from ndo_open
656 * before we're receiving packets, or from refill_work which is
657 * careful to disable receiving (using napi_disable).
659 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
667 if (vi->mergeable_rx_bufs)
668 err = add_recvbuf_mergeable(rq, gfp);
669 else if (vi->big_packets)
670 err = add_recvbuf_big(vi, rq, gfp);
672 err = add_recvbuf_small(vi, rq, gfp);
674 oom = err == -ENOMEM;
677 } while (rq->vq->num_free);
678 virtqueue_kick(rq->vq);
682 static void skb_recv_done(struct virtqueue *rvq)
684 struct virtnet_info *vi = rvq->vdev->priv;
685 struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
687 /* Schedule NAPI, Suppress further interrupts if successful. */
688 if (napi_schedule_prep(&rq->napi)) {
689 virtqueue_disable_cb(rvq);
690 __napi_schedule(&rq->napi);
694 static void virtnet_napi_enable(struct receive_queue *rq)
696 napi_enable(&rq->napi);
698 /* If all buffers were filled by other side before we napi_enabled, we
699 * won't get another interrupt, so process any outstanding packets
700 * now. virtnet_poll wants re-enable the queue, so we disable here.
701 * We synchronize against interrupts via NAPI_STATE_SCHED */
702 if (napi_schedule_prep(&rq->napi)) {
703 virtqueue_disable_cb(rq->vq);
705 __napi_schedule(&rq->napi);
710 static void refill_work(struct work_struct *work)
712 struct virtnet_info *vi =
713 container_of(work, struct virtnet_info, refill.work);
717 for (i = 0; i < vi->curr_queue_pairs; i++) {
718 struct receive_queue *rq = &vi->rq[i];
720 napi_disable(&rq->napi);
721 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
722 virtnet_napi_enable(rq);
724 /* In theory, this can happen: if we don't get any buffers in
725 * we will *never* try to fill again.
728 schedule_delayed_work(&vi->refill, HZ/2);
732 static int virtnet_receive(struct receive_queue *rq, int budget)
734 struct virtnet_info *vi = rq->vq->vdev->priv;
735 unsigned int len, received = 0;
738 while (received < budget &&
739 (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
740 receive_buf(vi, rq, buf, len);
744 if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) {
745 if (!try_fill_recv(vi, rq, GFP_ATOMIC))
746 schedule_delayed_work(&vi->refill, 0);
752 static int virtnet_poll(struct napi_struct *napi, int budget)
754 struct receive_queue *rq =
755 container_of(napi, struct receive_queue, napi);
756 unsigned int r, received;
758 received = virtnet_receive(rq, budget);
760 /* Out of packets? */
761 if (received < budget) {
762 r = virtqueue_enable_cb_prepare(rq->vq);
763 napi_complete_done(napi, received);
764 if (unlikely(virtqueue_poll(rq->vq, r)) &&
765 napi_schedule_prep(napi)) {
766 virtqueue_disable_cb(rq->vq);
767 __napi_schedule(napi);
774 #ifdef CONFIG_NET_RX_BUSY_POLL
775 /* must be called with local_bh_disable()d */
776 static int virtnet_busy_poll(struct napi_struct *napi)
778 struct receive_queue *rq =
779 container_of(napi, struct receive_queue, napi);
780 struct virtnet_info *vi = rq->vq->vdev->priv;
781 int r, received = 0, budget = 4;
783 if (!(vi->status & VIRTIO_NET_S_LINK_UP))
784 return LL_FLUSH_FAILED;
786 if (!napi_schedule_prep(napi))
787 return LL_FLUSH_BUSY;
789 virtqueue_disable_cb(rq->vq);
792 received += virtnet_receive(rq, budget);
794 r = virtqueue_enable_cb_prepare(rq->vq);
795 clear_bit(NAPI_STATE_SCHED, &napi->state);
796 if (unlikely(virtqueue_poll(rq->vq, r)) &&
797 napi_schedule_prep(napi)) {
798 virtqueue_disable_cb(rq->vq);
799 if (received < budget) {
803 __napi_schedule(napi);
809 #endif /* CONFIG_NET_RX_BUSY_POLL */
811 static int virtnet_open(struct net_device *dev)
813 struct virtnet_info *vi = netdev_priv(dev);
816 for (i = 0; i < vi->max_queue_pairs; i++) {
817 if (i < vi->curr_queue_pairs)
818 /* Make sure we have some buffers: if oom use wq. */
819 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
820 schedule_delayed_work(&vi->refill, 0);
821 virtnet_napi_enable(&vi->rq[i]);
827 static void free_old_xmit_skbs(struct send_queue *sq)
831 struct virtnet_info *vi = sq->vq->vdev->priv;
832 struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
834 while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
835 pr_debug("Sent skb %p\n", skb);
837 u64_stats_update_begin(&stats->tx_syncp);
838 stats->tx_bytes += skb->len;
840 u64_stats_update_end(&stats->tx_syncp);
842 dev_kfree_skb_any(skb);
846 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
848 struct virtio_net_hdr_mrg_rxbuf *hdr;
849 const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
850 struct virtnet_info *vi = sq->vq->vdev->priv;
852 unsigned hdr_len = vi->hdr_len;
855 pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
857 can_push = vi->any_header_sg &&
858 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
859 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
860 /* Even if we can, don't push here yet as this would skew
861 * csum_start offset below. */
863 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
865 hdr = skb_vnet_hdr(skb);
867 if (skb->ip_summed == CHECKSUM_PARTIAL) {
868 hdr->hdr.flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
869 hdr->hdr.csum_start = cpu_to_virtio16(vi->vdev,
870 skb_checksum_start_offset(skb));
871 hdr->hdr.csum_offset = cpu_to_virtio16(vi->vdev,
875 hdr->hdr.csum_offset = hdr->hdr.csum_start = 0;
878 if (skb_is_gso(skb)) {
879 hdr->hdr.hdr_len = cpu_to_virtio16(vi->vdev, skb_headlen(skb));
880 hdr->hdr.gso_size = cpu_to_virtio16(vi->vdev,
881 skb_shinfo(skb)->gso_size);
882 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
883 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
884 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
885 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
886 else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
887 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_UDP;
890 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCP_ECN)
891 hdr->hdr.gso_type |= VIRTIO_NET_HDR_GSO_ECN;
893 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE;
894 hdr->hdr.gso_size = hdr->hdr.hdr_len = 0;
897 if (vi->mergeable_rx_bufs)
898 hdr->num_buffers = 0;
900 sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
902 __skb_push(skb, hdr_len);
903 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
904 /* Pull header back to avoid skew in tx bytes calculations. */
905 __skb_pull(skb, hdr_len);
907 sg_set_buf(sq->sg, hdr, hdr_len);
908 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
910 return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
913 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
915 struct virtnet_info *vi = netdev_priv(dev);
916 int qnum = skb_get_queue_mapping(skb);
917 struct send_queue *sq = &vi->sq[qnum];
919 struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
920 bool kick = !skb->xmit_more;
922 /* Free up any pending old buffers before queueing new ones. */
923 free_old_xmit_skbs(sq);
925 /* timestamp packet in software */
926 skb_tx_timestamp(skb);
928 /* Try to transmit */
929 err = xmit_skb(sq, skb);
931 /* This should not happen! */
933 dev->stats.tx_fifo_errors++;
936 "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
937 dev->stats.tx_dropped++;
938 dev_kfree_skb_any(skb);
942 /* Don't wait up for transmitted skbs to be freed. */
946 /* If running out of space, stop queue to avoid getting packets that we
947 * are then unable to transmit.
948 * An alternative would be to force queuing layer to requeue the skb by
949 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
950 * returned in a normal path of operation: it means that driver is not
951 * maintaining the TX queue stop/start state properly, and causes
952 * the stack to do a non-trivial amount of useless work.
953 * Since most packets only take 1 or 2 ring slots, stopping the queue
954 * early means 16 slots are typically wasted.
956 if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
957 netif_stop_subqueue(dev, qnum);
958 if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
959 /* More just got used, free them then recheck. */
960 free_old_xmit_skbs(sq);
961 if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
962 netif_start_subqueue(dev, qnum);
963 virtqueue_disable_cb(sq->vq);
968 if (kick || netif_xmit_stopped(txq))
969 virtqueue_kick(sq->vq);
975 * Send command via the control virtqueue and check status. Commands
976 * supported by the hypervisor, as indicated by feature bits, should
977 * never fail unless improperly formatted.
979 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
980 struct scatterlist *out)
982 struct scatterlist *sgs[4], hdr, stat;
983 unsigned out_num = 0, tmp;
985 /* Caller should know better */
986 BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
988 vi->ctrl_status = ~0;
989 vi->ctrl_hdr.class = class;
990 vi->ctrl_hdr.cmd = cmd;
992 sg_init_one(&hdr, &vi->ctrl_hdr, sizeof(vi->ctrl_hdr));
993 sgs[out_num++] = &hdr;
996 sgs[out_num++] = out;
998 /* Add return status. */
999 sg_init_one(&stat, &vi->ctrl_status, sizeof(vi->ctrl_status));
1000 sgs[out_num] = &stat;
1002 BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1003 virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1005 if (unlikely(!virtqueue_kick(vi->cvq)))
1006 return vi->ctrl_status == VIRTIO_NET_OK;
1008 /* Spin for a response, the kick causes an ioport write, trapping
1009 * into the hypervisor, so the request should be handled immediately.
1011 while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1012 !virtqueue_is_broken(vi->cvq))
1015 return vi->ctrl_status == VIRTIO_NET_OK;
1018 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1020 struct virtnet_info *vi = netdev_priv(dev);
1021 struct virtio_device *vdev = vi->vdev;
1023 struct sockaddr *addr = p;
1024 struct scatterlist sg;
1026 ret = eth_prepare_mac_addr_change(dev, p);
1030 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1031 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1032 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1033 VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1034 dev_warn(&vdev->dev,
1035 "Failed to set mac address by vq command.\n");
1038 } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1039 !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1042 /* Naturally, this has an atomicity problem. */
1043 for (i = 0; i < dev->addr_len; i++)
1044 virtio_cwrite8(vdev,
1045 offsetof(struct virtio_net_config, mac) +
1046 i, addr->sa_data[i]);
1049 eth_commit_mac_addr_change(dev, p);
1054 static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
1055 struct rtnl_link_stats64 *tot)
1057 struct virtnet_info *vi = netdev_priv(dev);
1061 for_each_possible_cpu(cpu) {
1062 struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
1063 u64 tpackets, tbytes, rpackets, rbytes;
1066 start = u64_stats_fetch_begin_irq(&stats->tx_syncp);
1067 tpackets = stats->tx_packets;
1068 tbytes = stats->tx_bytes;
1069 } while (u64_stats_fetch_retry_irq(&stats->tx_syncp, start));
1072 start = u64_stats_fetch_begin_irq(&stats->rx_syncp);
1073 rpackets = stats->rx_packets;
1074 rbytes = stats->rx_bytes;
1075 } while (u64_stats_fetch_retry_irq(&stats->rx_syncp, start));
1077 tot->rx_packets += rpackets;
1078 tot->tx_packets += tpackets;
1079 tot->rx_bytes += rbytes;
1080 tot->tx_bytes += tbytes;
1083 tot->tx_dropped = dev->stats.tx_dropped;
1084 tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1085 tot->rx_dropped = dev->stats.rx_dropped;
1086 tot->rx_length_errors = dev->stats.rx_length_errors;
1087 tot->rx_frame_errors = dev->stats.rx_frame_errors;
1092 #ifdef CONFIG_NET_POLL_CONTROLLER
1093 static void virtnet_netpoll(struct net_device *dev)
1095 struct virtnet_info *vi = netdev_priv(dev);
1098 for (i = 0; i < vi->curr_queue_pairs; i++)
1099 napi_schedule(&vi->rq[i].napi);
1103 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1106 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1107 VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1108 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1112 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1114 struct scatterlist sg;
1115 struct virtio_net_ctrl_mq s;
1116 struct net_device *dev = vi->dev;
1118 if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1121 s.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1122 sg_init_one(&sg, &s, sizeof(s));
1124 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1125 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1126 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1130 vi->curr_queue_pairs = queue_pairs;
1131 /* virtnet_open() will refill when device is going to up. */
1132 if (dev->flags & IFF_UP)
1133 schedule_delayed_work(&vi->refill, 0);
1139 static int virtnet_close(struct net_device *dev)
1141 struct virtnet_info *vi = netdev_priv(dev);
1144 /* Make sure refill_work doesn't re-enable napi! */
1145 cancel_delayed_work_sync(&vi->refill);
1147 for (i = 0; i < vi->max_queue_pairs; i++)
1148 napi_disable(&vi->rq[i].napi);
1153 static void virtnet_set_rx_mode(struct net_device *dev)
1155 struct virtnet_info *vi = netdev_priv(dev);
1156 struct scatterlist sg[2];
1157 struct virtio_net_ctrl_mac *mac_data;
1158 struct netdev_hw_addr *ha;
1164 /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1165 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1168 vi->ctrl_promisc = ((dev->flags & IFF_PROMISC) != 0);
1169 vi->ctrl_allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1171 sg_init_one(sg, &vi->ctrl_promisc, sizeof(vi->ctrl_promisc));
1173 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1174 VIRTIO_NET_CTRL_RX_PROMISC, sg))
1175 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1176 vi->ctrl_promisc ? "en" : "dis");
1178 sg_init_one(sg, &vi->ctrl_allmulti, sizeof(vi->ctrl_allmulti));
1180 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1181 VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1182 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1183 vi->ctrl_allmulti ? "en" : "dis");
1185 uc_count = netdev_uc_count(dev);
1186 mc_count = netdev_mc_count(dev);
1187 /* MAC filter - use one buffer for both lists */
1188 buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1189 (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1194 sg_init_table(sg, 2);
1196 /* Store the unicast list and count in the front of the buffer */
1197 mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1199 netdev_for_each_uc_addr(ha, dev)
1200 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1202 sg_set_buf(&sg[0], mac_data,
1203 sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1205 /* multicast list and count fill the end */
1206 mac_data = (void *)&mac_data->macs[uc_count][0];
1208 mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1210 netdev_for_each_mc_addr(ha, dev)
1211 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1213 sg_set_buf(&sg[1], mac_data,
1214 sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1216 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1217 VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1218 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1223 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1224 __be16 proto, u16 vid)
1226 struct virtnet_info *vi = netdev_priv(dev);
1227 struct scatterlist sg;
1229 sg_init_one(&sg, &vid, sizeof(vid));
1231 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1232 VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1233 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1237 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1238 __be16 proto, u16 vid)
1240 struct virtnet_info *vi = netdev_priv(dev);
1241 struct scatterlist sg;
1243 sg_init_one(&sg, &vid, sizeof(vid));
1245 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1246 VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1247 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1251 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1255 if (vi->affinity_hint_set) {
1256 for (i = 0; i < vi->max_queue_pairs; i++) {
1257 virtqueue_set_affinity(vi->rq[i].vq, -1);
1258 virtqueue_set_affinity(vi->sq[i].vq, -1);
1261 vi->affinity_hint_set = false;
1265 static void virtnet_set_affinity(struct virtnet_info *vi)
1270 /* In multiqueue mode, when the number of cpu is equal to the number of
1271 * queue pairs, we let the queue pairs to be private to one cpu by
1272 * setting the affinity hint to eliminate the contention.
1274 if (vi->curr_queue_pairs == 1 ||
1275 vi->max_queue_pairs != num_online_cpus()) {
1276 virtnet_clean_affinity(vi, -1);
1281 for_each_online_cpu(cpu) {
1282 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1283 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1284 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
1288 vi->affinity_hint_set = true;
1291 static int virtnet_cpu_callback(struct notifier_block *nfb,
1292 unsigned long action, void *hcpu)
1294 struct virtnet_info *vi = container_of(nfb, struct virtnet_info, nb);
1296 switch(action & ~CPU_TASKS_FROZEN) {
1298 case CPU_DOWN_FAILED:
1300 virtnet_set_affinity(vi);
1302 case CPU_DOWN_PREPARE:
1303 virtnet_clean_affinity(vi, (long)hcpu);
1312 static void virtnet_get_ringparam(struct net_device *dev,
1313 struct ethtool_ringparam *ring)
1315 struct virtnet_info *vi = netdev_priv(dev);
1317 ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1318 ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1319 ring->rx_pending = ring->rx_max_pending;
1320 ring->tx_pending = ring->tx_max_pending;
1324 static void virtnet_get_drvinfo(struct net_device *dev,
1325 struct ethtool_drvinfo *info)
1327 struct virtnet_info *vi = netdev_priv(dev);
1328 struct virtio_device *vdev = vi->vdev;
1330 strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1331 strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1332 strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1336 /* TODO: Eliminate OOO packets during switching */
1337 static int virtnet_set_channels(struct net_device *dev,
1338 struct ethtool_channels *channels)
1340 struct virtnet_info *vi = netdev_priv(dev);
1341 u16 queue_pairs = channels->combined_count;
1344 /* We don't support separate rx/tx channels.
1345 * We don't allow setting 'other' channels.
1347 if (channels->rx_count || channels->tx_count || channels->other_count)
1350 if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
1354 err = virtnet_set_queues(vi, queue_pairs);
1356 netif_set_real_num_tx_queues(dev, queue_pairs);
1357 netif_set_real_num_rx_queues(dev, queue_pairs);
1359 virtnet_set_affinity(vi);
1366 static void virtnet_get_channels(struct net_device *dev,
1367 struct ethtool_channels *channels)
1369 struct virtnet_info *vi = netdev_priv(dev);
1371 channels->combined_count = vi->curr_queue_pairs;
1372 channels->max_combined = vi->max_queue_pairs;
1373 channels->max_other = 0;
1374 channels->rx_count = 0;
1375 channels->tx_count = 0;
1376 channels->other_count = 0;
1379 static const struct ethtool_ops virtnet_ethtool_ops = {
1380 .get_drvinfo = virtnet_get_drvinfo,
1381 .get_link = ethtool_op_get_link,
1382 .get_ringparam = virtnet_get_ringparam,
1383 .set_channels = virtnet_set_channels,
1384 .get_channels = virtnet_get_channels,
1385 .get_ts_info = ethtool_op_get_ts_info,
1389 #define MAX_MTU 65535
1391 static int virtnet_change_mtu(struct net_device *dev, int new_mtu)
1393 if (new_mtu < MIN_MTU || new_mtu > MAX_MTU)
1399 static const struct net_device_ops virtnet_netdev = {
1400 .ndo_open = virtnet_open,
1401 .ndo_stop = virtnet_close,
1402 .ndo_start_xmit = start_xmit,
1403 .ndo_validate_addr = eth_validate_addr,
1404 .ndo_set_mac_address = virtnet_set_mac_address,
1405 .ndo_set_rx_mode = virtnet_set_rx_mode,
1406 .ndo_change_mtu = virtnet_change_mtu,
1407 .ndo_get_stats64 = virtnet_stats,
1408 .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
1409 .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
1410 #ifdef CONFIG_NET_POLL_CONTROLLER
1411 .ndo_poll_controller = virtnet_netpoll,
1413 #ifdef CONFIG_NET_RX_BUSY_POLL
1414 .ndo_busy_poll = virtnet_busy_poll,
1418 static void virtnet_config_changed_work(struct work_struct *work)
1420 struct virtnet_info *vi =
1421 container_of(work, struct virtnet_info, config_work);
1424 if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
1425 struct virtio_net_config, status, &v) < 0)
1428 if (v & VIRTIO_NET_S_ANNOUNCE) {
1429 netdev_notify_peers(vi->dev);
1430 virtnet_ack_link_announce(vi);
1433 /* Ignore unknown (future) status bits */
1434 v &= VIRTIO_NET_S_LINK_UP;
1436 if (vi->status == v)
1441 if (vi->status & VIRTIO_NET_S_LINK_UP) {
1442 netif_carrier_on(vi->dev);
1443 netif_tx_wake_all_queues(vi->dev);
1445 netif_carrier_off(vi->dev);
1446 netif_tx_stop_all_queues(vi->dev);
1450 static void virtnet_config_changed(struct virtio_device *vdev)
1452 struct virtnet_info *vi = vdev->priv;
1454 schedule_work(&vi->config_work);
1457 static void virtnet_free_queues(struct virtnet_info *vi)
1461 for (i = 0; i < vi->max_queue_pairs; i++) {
1462 napi_hash_del(&vi->rq[i].napi);
1463 netif_napi_del(&vi->rq[i].napi);
1470 static void free_receive_bufs(struct virtnet_info *vi)
1474 for (i = 0; i < vi->max_queue_pairs; i++) {
1475 while (vi->rq[i].pages)
1476 __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
1480 static void free_receive_page_frags(struct virtnet_info *vi)
1483 for (i = 0; i < vi->max_queue_pairs; i++)
1484 if (vi->rq[i].alloc_frag.page)
1485 put_page(vi->rq[i].alloc_frag.page);
1488 static void free_unused_bufs(struct virtnet_info *vi)
1493 for (i = 0; i < vi->max_queue_pairs; i++) {
1494 struct virtqueue *vq = vi->sq[i].vq;
1495 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
1499 for (i = 0; i < vi->max_queue_pairs; i++) {
1500 struct virtqueue *vq = vi->rq[i].vq;
1502 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1503 if (vi->mergeable_rx_bufs) {
1504 unsigned long ctx = (unsigned long)buf;
1505 void *base = mergeable_ctx_to_buf_address(ctx);
1506 put_page(virt_to_head_page(base));
1507 } else if (vi->big_packets) {
1508 give_pages(&vi->rq[i], buf);
1516 static void virtnet_del_vqs(struct virtnet_info *vi)
1518 struct virtio_device *vdev = vi->vdev;
1520 virtnet_clean_affinity(vi, -1);
1522 vdev->config->del_vqs(vdev);
1524 virtnet_free_queues(vi);
1527 static int virtnet_find_vqs(struct virtnet_info *vi)
1529 vq_callback_t **callbacks;
1530 struct virtqueue **vqs;
1535 /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
1536 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
1537 * possible control vq.
1539 total_vqs = vi->max_queue_pairs * 2 +
1540 virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
1542 /* Allocate space for find_vqs parameters */
1543 vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
1546 callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
1549 names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
1553 /* Parameters for control virtqueue, if any */
1555 callbacks[total_vqs - 1] = NULL;
1556 names[total_vqs - 1] = "control";
1559 /* Allocate/initialize parameters for send/receive virtqueues */
1560 for (i = 0; i < vi->max_queue_pairs; i++) {
1561 callbacks[rxq2vq(i)] = skb_recv_done;
1562 callbacks[txq2vq(i)] = skb_xmit_done;
1563 sprintf(vi->rq[i].name, "input.%d", i);
1564 sprintf(vi->sq[i].name, "output.%d", i);
1565 names[rxq2vq(i)] = vi->rq[i].name;
1566 names[txq2vq(i)] = vi->sq[i].name;
1569 ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
1575 vi->cvq = vqs[total_vqs - 1];
1576 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
1577 vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
1580 for (i = 0; i < vi->max_queue_pairs; i++) {
1581 vi->rq[i].vq = vqs[rxq2vq(i)];
1582 vi->sq[i].vq = vqs[txq2vq(i)];
1601 static int virtnet_alloc_queues(struct virtnet_info *vi)
1605 vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
1608 vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
1612 INIT_DELAYED_WORK(&vi->refill, refill_work);
1613 for (i = 0; i < vi->max_queue_pairs; i++) {
1614 vi->rq[i].pages = NULL;
1615 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
1618 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
1619 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
1620 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
1631 static int init_vqs(struct virtnet_info *vi)
1635 /* Allocate send & receive queues */
1636 ret = virtnet_alloc_queues(vi);
1640 ret = virtnet_find_vqs(vi);
1645 virtnet_set_affinity(vi);
1651 virtnet_free_queues(vi);
1657 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
1658 struct rx_queue_attribute *attribute, char *buf)
1660 struct virtnet_info *vi = netdev_priv(queue->dev);
1661 unsigned int queue_index = get_netdev_rx_queue_index(queue);
1662 struct ewma_pkt_len *avg;
1664 BUG_ON(queue_index >= vi->max_queue_pairs);
1665 avg = &vi->rq[queue_index].mrg_avg_pkt_len;
1666 return sprintf(buf, "%u\n", get_mergeable_buf_len(avg));
1669 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
1670 __ATTR_RO(mergeable_rx_buffer_size);
1672 static struct attribute *virtio_net_mrg_rx_attrs[] = {
1673 &mergeable_rx_buffer_size_attribute.attr,
1677 static const struct attribute_group virtio_net_mrg_rx_group = {
1678 .name = "virtio_net",
1679 .attrs = virtio_net_mrg_rx_attrs
1683 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
1685 const char *fname, const char *dname)
1687 if (!virtio_has_feature(vdev, fbit))
1690 dev_err(&vdev->dev, "device advertises feature %s but not %s",
1696 #define VIRTNET_FAIL_ON(vdev, fbit, dbit) \
1697 virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
1699 static bool virtnet_validate_features(struct virtio_device *vdev)
1701 if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
1702 (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
1703 "VIRTIO_NET_F_CTRL_VQ") ||
1704 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
1705 "VIRTIO_NET_F_CTRL_VQ") ||
1706 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
1707 "VIRTIO_NET_F_CTRL_VQ") ||
1708 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
1709 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
1710 "VIRTIO_NET_F_CTRL_VQ"))) {
1717 static int virtnet_probe(struct virtio_device *vdev)
1720 struct net_device *dev;
1721 struct virtnet_info *vi;
1722 u16 max_queue_pairs;
1724 if (!vdev->config->get) {
1725 dev_err(&vdev->dev, "%s failure: config access disabled\n",
1730 if (!virtnet_validate_features(vdev))
1733 /* Find if host supports multiqueue virtio_net device */
1734 err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
1735 struct virtio_net_config,
1736 max_virtqueue_pairs, &max_queue_pairs);
1738 /* We need at least 2 queue's */
1739 if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
1740 max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
1741 !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1742 max_queue_pairs = 1;
1744 /* Allocate ourselves a network device with room for our info */
1745 dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
1749 /* Set up network device as normal. */
1750 dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
1751 dev->netdev_ops = &virtnet_netdev;
1752 dev->features = NETIF_F_HIGHDMA;
1754 dev->ethtool_ops = &virtnet_ethtool_ops;
1755 SET_NETDEV_DEV(dev, &vdev->dev);
1757 /* Do we support "hardware" checksums? */
1758 if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
1759 /* This opens up the world of extra features. */
1760 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
1762 dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
1764 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
1765 dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
1766 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
1768 /* Individual feature bits: what can host handle? */
1769 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
1770 dev->hw_features |= NETIF_F_TSO;
1771 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
1772 dev->hw_features |= NETIF_F_TSO6;
1773 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
1774 dev->hw_features |= NETIF_F_TSO_ECN;
1775 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
1776 dev->hw_features |= NETIF_F_UFO;
1778 dev->features |= NETIF_F_GSO_ROBUST;
1781 dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
1782 /* (!csum && gso) case will be fixed by register_netdev() */
1784 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
1785 dev->features |= NETIF_F_RXCSUM;
1787 dev->vlan_features = dev->features;
1789 /* Configuration may specify what MAC to use. Otherwise random. */
1790 if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
1791 virtio_cread_bytes(vdev,
1792 offsetof(struct virtio_net_config, mac),
1793 dev->dev_addr, dev->addr_len);
1795 eth_hw_addr_random(dev);
1797 /* Set up our device-specific information */
1798 vi = netdev_priv(dev);
1802 vi->stats = alloc_percpu(struct virtnet_stats);
1804 if (vi->stats == NULL)
1807 for_each_possible_cpu(i) {
1808 struct virtnet_stats *virtnet_stats;
1809 virtnet_stats = per_cpu_ptr(vi->stats, i);
1810 u64_stats_init(&virtnet_stats->tx_syncp);
1811 u64_stats_init(&virtnet_stats->rx_syncp);
1814 INIT_WORK(&vi->config_work, virtnet_config_changed_work);
1816 /* If we can receive ANY GSO packets, we must allocate large ones. */
1817 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1818 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1819 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
1820 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
1821 vi->big_packets = true;
1823 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
1824 vi->mergeable_rx_bufs = true;
1826 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
1827 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
1828 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1830 vi->hdr_len = sizeof(struct virtio_net_hdr);
1832 if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
1833 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
1834 vi->any_header_sg = true;
1836 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1839 if (vi->any_header_sg)
1840 dev->needed_headroom = vi->hdr_len;
1842 /* Use single tx/rx queue pair as default */
1843 vi->curr_queue_pairs = 1;
1844 vi->max_queue_pairs = max_queue_pairs;
1846 /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
1852 if (vi->mergeable_rx_bufs)
1853 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
1855 netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
1856 netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
1858 err = register_netdev(dev);
1860 pr_debug("virtio_net: registering device failed\n");
1864 virtio_device_ready(vdev);
1866 /* Last of all, set up some receive buffers. */
1867 for (i = 0; i < vi->curr_queue_pairs; i++) {
1868 try_fill_recv(vi, &vi->rq[i], GFP_KERNEL);
1870 /* If we didn't even get one input buffer, we're useless. */
1871 if (vi->rq[i].vq->num_free ==
1872 virtqueue_get_vring_size(vi->rq[i].vq)) {
1873 free_unused_bufs(vi);
1875 goto free_recv_bufs;
1879 vi->nb.notifier_call = &virtnet_cpu_callback;
1880 err = register_hotcpu_notifier(&vi->nb);
1882 pr_debug("virtio_net: registering cpu notifier failed\n");
1883 goto free_recv_bufs;
1886 /* Assume link up if device can't report link status,
1887 otherwise get link status from config. */
1888 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
1889 netif_carrier_off(dev);
1890 schedule_work(&vi->config_work);
1892 vi->status = VIRTIO_NET_S_LINK_UP;
1893 netif_carrier_on(dev);
1896 pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
1897 dev->name, max_queue_pairs);
1902 vi->vdev->config->reset(vdev);
1904 free_receive_bufs(vi);
1905 unregister_netdev(dev);
1907 cancel_delayed_work_sync(&vi->refill);
1908 free_receive_page_frags(vi);
1909 virtnet_del_vqs(vi);
1911 free_percpu(vi->stats);
1917 static void remove_vq_common(struct virtnet_info *vi)
1919 vi->vdev->config->reset(vi->vdev);
1921 /* Free unused buffers in both send and recv, if any. */
1922 free_unused_bufs(vi);
1924 free_receive_bufs(vi);
1926 free_receive_page_frags(vi);
1928 virtnet_del_vqs(vi);
1931 static void virtnet_remove(struct virtio_device *vdev)
1933 struct virtnet_info *vi = vdev->priv;
1935 unregister_hotcpu_notifier(&vi->nb);
1937 /* Make sure no work handler is accessing the device. */
1938 flush_work(&vi->config_work);
1940 unregister_netdev(vi->dev);
1942 remove_vq_common(vi);
1944 free_percpu(vi->stats);
1945 free_netdev(vi->dev);
1948 #ifdef CONFIG_PM_SLEEP
1949 static int virtnet_freeze(struct virtio_device *vdev)
1951 struct virtnet_info *vi = vdev->priv;
1954 unregister_hotcpu_notifier(&vi->nb);
1956 /* Make sure no work handler is accessing the device */
1957 flush_work(&vi->config_work);
1959 netif_device_detach(vi->dev);
1960 cancel_delayed_work_sync(&vi->refill);
1962 if (netif_running(vi->dev)) {
1963 for (i = 0; i < vi->max_queue_pairs; i++)
1964 napi_disable(&vi->rq[i].napi);
1967 remove_vq_common(vi);
1972 static int virtnet_restore(struct virtio_device *vdev)
1974 struct virtnet_info *vi = vdev->priv;
1981 virtio_device_ready(vdev);
1983 if (netif_running(vi->dev)) {
1984 for (i = 0; i < vi->curr_queue_pairs; i++)
1985 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1986 schedule_delayed_work(&vi->refill, 0);
1988 for (i = 0; i < vi->max_queue_pairs; i++)
1989 virtnet_napi_enable(&vi->rq[i]);
1992 netif_device_attach(vi->dev);
1995 virtnet_set_queues(vi, vi->curr_queue_pairs);
1998 err = register_hotcpu_notifier(&vi->nb);
2006 static struct virtio_device_id id_table[] = {
2007 { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
2011 static unsigned int features[] = {
2012 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM,
2013 VIRTIO_NET_F_GSO, VIRTIO_NET_F_MAC,
2014 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6,
2015 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6,
2016 VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO,
2017 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ,
2018 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN,
2019 VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ,
2020 VIRTIO_NET_F_CTRL_MAC_ADDR,
2021 VIRTIO_F_ANY_LAYOUT,
2024 static struct virtio_driver virtio_net_driver = {
2025 .feature_table = features,
2026 .feature_table_size = ARRAY_SIZE(features),
2027 .driver.name = KBUILD_MODNAME,
2028 .driver.owner = THIS_MODULE,
2029 .id_table = id_table,
2030 .probe = virtnet_probe,
2031 .remove = virtnet_remove,
2032 .config_changed = virtnet_config_changed,
2033 #ifdef CONFIG_PM_SLEEP
2034 .freeze = virtnet_freeze,
2035 .restore = virtnet_restore,
2039 module_virtio_driver(virtio_net_driver);
2041 MODULE_DEVICE_TABLE(virtio, id_table);
2042 MODULE_DESCRIPTION("Virtio network driver");
2043 MODULE_LICENSE("GPL");