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_init_table(rq->sg, MAX_SKB_FRAGS + 2);
544 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 sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
562 /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
563 for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
564 first = get_a_page(rq, gfp);
567 give_pages(rq, list);
570 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
572 /* chain new page in list head to match sg */
573 first->private = (unsigned long)list;
577 first = get_a_page(rq, gfp);
579 give_pages(rq, list);
582 p = page_address(first);
584 /* rq->sg[0], rq->sg[1] share the same page */
585 /* a separated rq->sg[0] for virtio_net_hdr only due to QEMU bug */
586 sg_set_buf(&rq->sg[0], p, sizeof(struct virtio_net_hdr));
588 /* rq->sg[1] for data packet, from offset */
589 offset = sizeof(struct padded_vnet_hdr);
590 sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
592 /* chain first in list head */
593 first->private = (unsigned long)list;
594 err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
597 give_pages(rq, first);
602 static unsigned int get_mergeable_buf_len(struct ewma *avg_pkt_len)
604 const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
607 len = hdr_len + clamp_t(unsigned int, ewma_read(avg_pkt_len),
608 GOOD_PACKET_LEN, PAGE_SIZE - hdr_len);
609 return ALIGN(len, MERGEABLE_BUFFER_ALIGN);
612 static int add_recvbuf_mergeable(struct receive_queue *rq, gfp_t gfp)
614 struct page_frag *alloc_frag = &rq->alloc_frag;
618 unsigned int len, hole;
620 len = get_mergeable_buf_len(&rq->mrg_avg_pkt_len);
621 if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
624 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
625 ctx = mergeable_buf_to_ctx(buf, len);
626 get_page(alloc_frag->page);
627 alloc_frag->offset += len;
628 hole = alloc_frag->size - alloc_frag->offset;
630 /* To avoid internal fragmentation, if there is very likely not
631 * enough space for another buffer, add the remaining space to
632 * the current buffer. This extra space is not included in
633 * the truesize stored in ctx.
636 alloc_frag->offset += hole;
639 sg_init_one(rq->sg, buf, len);
640 err = virtqueue_add_inbuf(rq->vq, rq->sg, 1, (void *)ctx, gfp);
642 put_page(virt_to_head_page(buf));
648 * Returns false if we couldn't fill entirely (OOM).
650 * Normally run in the receive path, but can also be run from ndo_open
651 * before we're receiving packets, or from refill_work which is
652 * careful to disable receiving (using napi_disable).
654 static bool try_fill_recv(struct receive_queue *rq, gfp_t gfp)
656 struct virtnet_info *vi = rq->vq->vdev->priv;
662 if (vi->mergeable_rx_bufs)
663 err = add_recvbuf_mergeable(rq, gfp);
664 else if (vi->big_packets)
665 err = add_recvbuf_big(rq, gfp);
667 err = add_recvbuf_small(rq, gfp);
669 oom = err == -ENOMEM;
672 } while (rq->vq->num_free);
673 virtqueue_kick(rq->vq);
677 static void skb_recv_done(struct virtqueue *rvq)
679 struct virtnet_info *vi = rvq->vdev->priv;
680 struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
682 /* Schedule NAPI, Suppress further interrupts if successful. */
683 if (napi_schedule_prep(&rq->napi)) {
684 virtqueue_disable_cb(rvq);
685 __napi_schedule(&rq->napi);
689 static void virtnet_napi_enable(struct receive_queue *rq)
691 napi_enable(&rq->napi);
693 /* If all buffers were filled by other side before we napi_enabled, we
694 * won't get another interrupt, so process any outstanding packets
695 * now. virtnet_poll wants re-enable the queue, so we disable here.
696 * We synchronize against interrupts via NAPI_STATE_SCHED */
697 if (napi_schedule_prep(&rq->napi)) {
698 virtqueue_disable_cb(rq->vq);
700 __napi_schedule(&rq->napi);
705 static void refill_work(struct work_struct *work)
707 struct virtnet_info *vi =
708 container_of(work, struct virtnet_info, refill.work);
712 for (i = 0; i < vi->curr_queue_pairs; i++) {
713 struct receive_queue *rq = &vi->rq[i];
715 napi_disable(&rq->napi);
716 still_empty = !try_fill_recv(rq, GFP_KERNEL);
717 virtnet_napi_enable(rq);
719 /* In theory, this can happen: if we don't get any buffers in
720 * we will *never* try to fill again.
723 schedule_delayed_work(&vi->refill, HZ/2);
727 static int virtnet_receive(struct receive_queue *rq, int budget)
729 struct virtnet_info *vi = rq->vq->vdev->priv;
730 unsigned int len, received = 0;
733 while (received < budget &&
734 (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
735 receive_buf(rq, buf, len);
739 if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) {
740 if (!try_fill_recv(rq, GFP_ATOMIC))
741 schedule_delayed_work(&vi->refill, 0);
747 static int virtnet_poll(struct napi_struct *napi, int budget)
749 struct receive_queue *rq =
750 container_of(napi, struct receive_queue, napi);
751 unsigned int r, received = 0;
754 received += virtnet_receive(rq, budget - received);
756 /* Out of packets? */
757 if (received < budget) {
758 r = virtqueue_enable_cb_prepare(rq->vq);
760 if (unlikely(virtqueue_poll(rq->vq, r)) &&
761 napi_schedule_prep(napi)) {
762 virtqueue_disable_cb(rq->vq);
763 __napi_schedule(napi);
771 #ifdef CONFIG_NET_RX_BUSY_POLL
772 /* must be called with local_bh_disable()d */
773 static int virtnet_busy_poll(struct napi_struct *napi)
775 struct receive_queue *rq =
776 container_of(napi, struct receive_queue, napi);
777 struct virtnet_info *vi = rq->vq->vdev->priv;
778 int r, received = 0, budget = 4;
780 if (!(vi->status & VIRTIO_NET_S_LINK_UP))
781 return LL_FLUSH_FAILED;
783 if (!napi_schedule_prep(napi))
784 return LL_FLUSH_BUSY;
786 virtqueue_disable_cb(rq->vq);
789 received += virtnet_receive(rq, budget);
791 r = virtqueue_enable_cb_prepare(rq->vq);
792 clear_bit(NAPI_STATE_SCHED, &napi->state);
793 if (unlikely(virtqueue_poll(rq->vq, r)) &&
794 napi_schedule_prep(napi)) {
795 virtqueue_disable_cb(rq->vq);
796 if (received < budget) {
800 __napi_schedule(napi);
806 #endif /* CONFIG_NET_RX_BUSY_POLL */
808 static int virtnet_open(struct net_device *dev)
810 struct virtnet_info *vi = netdev_priv(dev);
813 for (i = 0; i < vi->max_queue_pairs; i++) {
814 if (i < vi->curr_queue_pairs)
815 /* Make sure we have some buffers: if oom use wq. */
816 if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
817 schedule_delayed_work(&vi->refill, 0);
818 virtnet_napi_enable(&vi->rq[i]);
824 static void free_old_xmit_skbs(struct send_queue *sq)
828 struct virtnet_info *vi = sq->vq->vdev->priv;
829 struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
831 while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
832 pr_debug("Sent skb %p\n", skb);
834 u64_stats_update_begin(&stats->tx_syncp);
835 stats->tx_bytes += skb->len;
837 u64_stats_update_end(&stats->tx_syncp);
839 dev_kfree_skb_any(skb);
843 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
845 struct skb_vnet_hdr *hdr;
846 const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
847 struct virtnet_info *vi = sq->vq->vdev->priv;
852 pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
853 if (vi->mergeable_rx_bufs)
854 hdr_len = sizeof hdr->mhdr;
856 hdr_len = sizeof hdr->hdr;
858 can_push = vi->any_header_sg &&
859 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
860 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
861 /* Even if we can, don't push here yet as this would skew
862 * csum_start offset below. */
864 hdr = (struct skb_vnet_hdr *)(skb->data - hdr_len);
866 hdr = skb_vnet_hdr(skb);
868 if (skb->ip_summed == CHECKSUM_PARTIAL) {
869 hdr->hdr.flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
870 hdr->hdr.csum_start = skb_checksum_start_offset(skb);
871 hdr->hdr.csum_offset = skb->csum_offset;
874 hdr->hdr.csum_offset = hdr->hdr.csum_start = 0;
877 if (skb_is_gso(skb)) {
878 hdr->hdr.hdr_len = skb_headlen(skb);
879 hdr->hdr.gso_size = skb_shinfo(skb)->gso_size;
880 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV4)
881 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV4;
882 else if (skb_shinfo(skb)->gso_type & SKB_GSO_TCPV6)
883 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_TCPV6;
884 else if (skb_shinfo(skb)->gso_type & SKB_GSO_UDP)
885 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_UDP;
888 if (skb_shinfo(skb)->gso_type & SKB_GSO_TCP_ECN)
889 hdr->hdr.gso_type |= VIRTIO_NET_HDR_GSO_ECN;
891 hdr->hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE;
892 hdr->hdr.gso_size = hdr->hdr.hdr_len = 0;
895 if (vi->mergeable_rx_bufs)
896 hdr->mhdr.num_buffers = 0;
898 sg_init_table(sq->sg, MAX_SKB_FRAGS + 2);
900 __skb_push(skb, hdr_len);
901 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
902 /* Pull header back to avoid skew in tx bytes calculations. */
903 __skb_pull(skb, hdr_len);
905 sg_set_buf(sq->sg, hdr, hdr_len);
906 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
908 return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
911 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
913 struct virtnet_info *vi = netdev_priv(dev);
914 int qnum = skb_get_queue_mapping(skb);
915 struct send_queue *sq = &vi->sq[qnum];
917 struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
918 bool kick = !skb->xmit_more;
920 /* Free up any pending old buffers before queueing new ones. */
921 free_old_xmit_skbs(sq);
923 /* Try to transmit */
924 err = xmit_skb(sq, skb);
926 /* This should not happen! */
928 dev->stats.tx_fifo_errors++;
931 "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
932 dev->stats.tx_dropped++;
933 dev_kfree_skb_any(skb);
937 /* Don't wait up for transmitted skbs to be freed. */
941 /* Apparently nice girls don't return TX_BUSY; stop the queue
942 * before it gets out of hand. Naturally, this wastes entries. */
943 if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
944 netif_stop_subqueue(dev, qnum);
945 if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
946 /* More just got used, free them then recheck. */
947 free_old_xmit_skbs(sq);
948 if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
949 netif_start_subqueue(dev, qnum);
950 virtqueue_disable_cb(sq->vq);
955 if (kick || netif_xmit_stopped(txq))
956 virtqueue_kick(sq->vq);
962 * Send command via the control virtqueue and check status. Commands
963 * supported by the hypervisor, as indicated by feature bits, should
964 * never fail unless improperly formatted.
966 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
967 struct scatterlist *out)
969 struct scatterlist *sgs[4], hdr, stat;
970 struct virtio_net_ctrl_hdr ctrl;
971 virtio_net_ctrl_ack status = ~0;
972 unsigned out_num = 0, tmp;
974 /* Caller should know better */
975 BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
980 sg_init_one(&hdr, &ctrl, sizeof(ctrl));
981 sgs[out_num++] = &hdr;
984 sgs[out_num++] = out;
986 /* Add return status. */
987 sg_init_one(&stat, &status, sizeof(status));
988 sgs[out_num] = &stat;
990 BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
991 virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
993 if (unlikely(!virtqueue_kick(vi->cvq)))
994 return status == VIRTIO_NET_OK;
996 /* Spin for a response, the kick causes an ioport write, trapping
997 * into the hypervisor, so the request should be handled immediately.
999 while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1000 !virtqueue_is_broken(vi->cvq))
1003 return status == VIRTIO_NET_OK;
1006 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1008 struct virtnet_info *vi = netdev_priv(dev);
1009 struct virtio_device *vdev = vi->vdev;
1011 struct sockaddr *addr = p;
1012 struct scatterlist sg;
1014 ret = eth_prepare_mac_addr_change(dev, p);
1018 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1019 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1020 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1021 VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1022 dev_warn(&vdev->dev,
1023 "Failed to set mac address by vq command.\n");
1026 } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
1029 /* Naturally, this has an atomicity problem. */
1030 for (i = 0; i < dev->addr_len; i++)
1031 virtio_cwrite8(vdev,
1032 offsetof(struct virtio_net_config, mac) +
1033 i, addr->sa_data[i]);
1036 eth_commit_mac_addr_change(dev, p);
1041 static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
1042 struct rtnl_link_stats64 *tot)
1044 struct virtnet_info *vi = netdev_priv(dev);
1048 for_each_possible_cpu(cpu) {
1049 struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
1050 u64 tpackets, tbytes, rpackets, rbytes;
1053 start = u64_stats_fetch_begin_irq(&stats->tx_syncp);
1054 tpackets = stats->tx_packets;
1055 tbytes = stats->tx_bytes;
1056 } while (u64_stats_fetch_retry_irq(&stats->tx_syncp, start));
1059 start = u64_stats_fetch_begin_irq(&stats->rx_syncp);
1060 rpackets = stats->rx_packets;
1061 rbytes = stats->rx_bytes;
1062 } while (u64_stats_fetch_retry_irq(&stats->rx_syncp, start));
1064 tot->rx_packets += rpackets;
1065 tot->tx_packets += tpackets;
1066 tot->rx_bytes += rbytes;
1067 tot->tx_bytes += tbytes;
1070 tot->tx_dropped = dev->stats.tx_dropped;
1071 tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1072 tot->rx_dropped = dev->stats.rx_dropped;
1073 tot->rx_length_errors = dev->stats.rx_length_errors;
1074 tot->rx_frame_errors = dev->stats.rx_frame_errors;
1079 #ifdef CONFIG_NET_POLL_CONTROLLER
1080 static void virtnet_netpoll(struct net_device *dev)
1082 struct virtnet_info *vi = netdev_priv(dev);
1085 for (i = 0; i < vi->curr_queue_pairs; i++)
1086 napi_schedule(&vi->rq[i].napi);
1090 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1093 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1094 VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1095 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1099 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1101 struct scatterlist sg;
1102 struct virtio_net_ctrl_mq s;
1103 struct net_device *dev = vi->dev;
1105 if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1108 s.virtqueue_pairs = queue_pairs;
1109 sg_init_one(&sg, &s, sizeof(s));
1111 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1112 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1113 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1117 vi->curr_queue_pairs = queue_pairs;
1118 /* virtnet_open() will refill when device is going to up. */
1119 if (dev->flags & IFF_UP)
1120 schedule_delayed_work(&vi->refill, 0);
1126 static int virtnet_close(struct net_device *dev)
1128 struct virtnet_info *vi = netdev_priv(dev);
1131 /* Make sure refill_work doesn't re-enable napi! */
1132 cancel_delayed_work_sync(&vi->refill);
1134 for (i = 0; i < vi->max_queue_pairs; i++)
1135 napi_disable(&vi->rq[i].napi);
1140 static void virtnet_set_rx_mode(struct net_device *dev)
1142 struct virtnet_info *vi = netdev_priv(dev);
1143 struct scatterlist sg[2];
1144 u8 promisc, allmulti;
1145 struct virtio_net_ctrl_mac *mac_data;
1146 struct netdev_hw_addr *ha;
1152 /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1153 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1156 promisc = ((dev->flags & IFF_PROMISC) != 0);
1157 allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1159 sg_init_one(sg, &promisc, sizeof(promisc));
1161 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1162 VIRTIO_NET_CTRL_RX_PROMISC, sg))
1163 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1164 promisc ? "en" : "dis");
1166 sg_init_one(sg, &allmulti, sizeof(allmulti));
1168 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1169 VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1170 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1171 allmulti ? "en" : "dis");
1173 uc_count = netdev_uc_count(dev);
1174 mc_count = netdev_mc_count(dev);
1175 /* MAC filter - use one buffer for both lists */
1176 buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1177 (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1182 sg_init_table(sg, 2);
1184 /* Store the unicast list and count in the front of the buffer */
1185 mac_data->entries = uc_count;
1187 netdev_for_each_uc_addr(ha, dev)
1188 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1190 sg_set_buf(&sg[0], mac_data,
1191 sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1193 /* multicast list and count fill the end */
1194 mac_data = (void *)&mac_data->macs[uc_count][0];
1196 mac_data->entries = mc_count;
1198 netdev_for_each_mc_addr(ha, dev)
1199 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1201 sg_set_buf(&sg[1], mac_data,
1202 sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1204 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1205 VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1206 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1211 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1212 __be16 proto, u16 vid)
1214 struct virtnet_info *vi = netdev_priv(dev);
1215 struct scatterlist sg;
1217 sg_init_one(&sg, &vid, sizeof(vid));
1219 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1220 VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1221 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1225 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1226 __be16 proto, u16 vid)
1228 struct virtnet_info *vi = netdev_priv(dev);
1229 struct scatterlist sg;
1231 sg_init_one(&sg, &vid, sizeof(vid));
1233 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1234 VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1235 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1239 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1243 if (vi->affinity_hint_set) {
1244 for (i = 0; i < vi->max_queue_pairs; i++) {
1245 virtqueue_set_affinity(vi->rq[i].vq, -1);
1246 virtqueue_set_affinity(vi->sq[i].vq, -1);
1249 vi->affinity_hint_set = false;
1253 static void virtnet_set_affinity(struct virtnet_info *vi)
1258 /* In multiqueue mode, when the number of cpu is equal to the number of
1259 * queue pairs, we let the queue pairs to be private to one cpu by
1260 * setting the affinity hint to eliminate the contention.
1262 if (vi->curr_queue_pairs == 1 ||
1263 vi->max_queue_pairs != num_online_cpus()) {
1264 virtnet_clean_affinity(vi, -1);
1269 for_each_online_cpu(cpu) {
1270 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1271 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1272 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
1276 vi->affinity_hint_set = true;
1279 static int virtnet_cpu_callback(struct notifier_block *nfb,
1280 unsigned long action, void *hcpu)
1282 struct virtnet_info *vi = container_of(nfb, struct virtnet_info, nb);
1284 switch(action & ~CPU_TASKS_FROZEN) {
1286 case CPU_DOWN_FAILED:
1288 virtnet_set_affinity(vi);
1290 case CPU_DOWN_PREPARE:
1291 virtnet_clean_affinity(vi, (long)hcpu);
1300 static void virtnet_get_ringparam(struct net_device *dev,
1301 struct ethtool_ringparam *ring)
1303 struct virtnet_info *vi = netdev_priv(dev);
1305 ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1306 ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1307 ring->rx_pending = ring->rx_max_pending;
1308 ring->tx_pending = ring->tx_max_pending;
1312 static void virtnet_get_drvinfo(struct net_device *dev,
1313 struct ethtool_drvinfo *info)
1315 struct virtnet_info *vi = netdev_priv(dev);
1316 struct virtio_device *vdev = vi->vdev;
1318 strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1319 strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1320 strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1324 /* TODO: Eliminate OOO packets during switching */
1325 static int virtnet_set_channels(struct net_device *dev,
1326 struct ethtool_channels *channels)
1328 struct virtnet_info *vi = netdev_priv(dev);
1329 u16 queue_pairs = channels->combined_count;
1332 /* We don't support separate rx/tx channels.
1333 * We don't allow setting 'other' channels.
1335 if (channels->rx_count || channels->tx_count || channels->other_count)
1338 if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
1342 err = virtnet_set_queues(vi, queue_pairs);
1344 netif_set_real_num_tx_queues(dev, queue_pairs);
1345 netif_set_real_num_rx_queues(dev, queue_pairs);
1347 virtnet_set_affinity(vi);
1354 static void virtnet_get_channels(struct net_device *dev,
1355 struct ethtool_channels *channels)
1357 struct virtnet_info *vi = netdev_priv(dev);
1359 channels->combined_count = vi->curr_queue_pairs;
1360 channels->max_combined = vi->max_queue_pairs;
1361 channels->max_other = 0;
1362 channels->rx_count = 0;
1363 channels->tx_count = 0;
1364 channels->other_count = 0;
1367 static const struct ethtool_ops virtnet_ethtool_ops = {
1368 .get_drvinfo = virtnet_get_drvinfo,
1369 .get_link = ethtool_op_get_link,
1370 .get_ringparam = virtnet_get_ringparam,
1371 .set_channels = virtnet_set_channels,
1372 .get_channels = virtnet_get_channels,
1376 #define MAX_MTU 65535
1378 static int virtnet_change_mtu(struct net_device *dev, int new_mtu)
1380 if (new_mtu < MIN_MTU || new_mtu > MAX_MTU)
1386 static const struct net_device_ops virtnet_netdev = {
1387 .ndo_open = virtnet_open,
1388 .ndo_stop = virtnet_close,
1389 .ndo_start_xmit = start_xmit,
1390 .ndo_validate_addr = eth_validate_addr,
1391 .ndo_set_mac_address = virtnet_set_mac_address,
1392 .ndo_set_rx_mode = virtnet_set_rx_mode,
1393 .ndo_change_mtu = virtnet_change_mtu,
1394 .ndo_get_stats64 = virtnet_stats,
1395 .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
1396 .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
1397 #ifdef CONFIG_NET_POLL_CONTROLLER
1398 .ndo_poll_controller = virtnet_netpoll,
1400 #ifdef CONFIG_NET_RX_BUSY_POLL
1401 .ndo_busy_poll = virtnet_busy_poll,
1405 static void virtnet_config_changed_work(struct work_struct *work)
1407 struct virtnet_info *vi =
1408 container_of(work, struct virtnet_info, config_work);
1411 if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
1412 struct virtio_net_config, status, &v) < 0)
1415 if (v & VIRTIO_NET_S_ANNOUNCE) {
1416 netdev_notify_peers(vi->dev);
1417 virtnet_ack_link_announce(vi);
1420 /* Ignore unknown (future) status bits */
1421 v &= VIRTIO_NET_S_LINK_UP;
1423 if (vi->status == v)
1428 if (vi->status & VIRTIO_NET_S_LINK_UP) {
1429 netif_carrier_on(vi->dev);
1430 netif_tx_wake_all_queues(vi->dev);
1432 netif_carrier_off(vi->dev);
1433 netif_tx_stop_all_queues(vi->dev);
1437 static void virtnet_config_changed(struct virtio_device *vdev)
1439 struct virtnet_info *vi = vdev->priv;
1441 schedule_work(&vi->config_work);
1444 static void virtnet_free_queues(struct virtnet_info *vi)
1448 for (i = 0; i < vi->max_queue_pairs; i++)
1449 netif_napi_del(&vi->rq[i].napi);
1455 static void free_receive_bufs(struct virtnet_info *vi)
1459 for (i = 0; i < vi->max_queue_pairs; i++) {
1460 while (vi->rq[i].pages)
1461 __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
1465 static void free_receive_page_frags(struct virtnet_info *vi)
1468 for (i = 0; i < vi->max_queue_pairs; i++)
1469 if (vi->rq[i].alloc_frag.page)
1470 put_page(vi->rq[i].alloc_frag.page);
1473 static void free_unused_bufs(struct virtnet_info *vi)
1478 for (i = 0; i < vi->max_queue_pairs; i++) {
1479 struct virtqueue *vq = vi->sq[i].vq;
1480 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
1484 for (i = 0; i < vi->max_queue_pairs; i++) {
1485 struct virtqueue *vq = vi->rq[i].vq;
1487 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
1488 if (vi->mergeable_rx_bufs) {
1489 unsigned long ctx = (unsigned long)buf;
1490 void *base = mergeable_ctx_to_buf_address(ctx);
1491 put_page(virt_to_head_page(base));
1492 } else if (vi->big_packets) {
1493 give_pages(&vi->rq[i], buf);
1501 static void virtnet_del_vqs(struct virtnet_info *vi)
1503 struct virtio_device *vdev = vi->vdev;
1505 virtnet_clean_affinity(vi, -1);
1507 vdev->config->del_vqs(vdev);
1509 virtnet_free_queues(vi);
1512 static int virtnet_find_vqs(struct virtnet_info *vi)
1514 vq_callback_t **callbacks;
1515 struct virtqueue **vqs;
1520 /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
1521 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
1522 * possible control vq.
1524 total_vqs = vi->max_queue_pairs * 2 +
1525 virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
1527 /* Allocate space for find_vqs parameters */
1528 vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
1531 callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
1534 names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
1538 /* Parameters for control virtqueue, if any */
1540 callbacks[total_vqs - 1] = NULL;
1541 names[total_vqs - 1] = "control";
1544 /* Allocate/initialize parameters for send/receive virtqueues */
1545 for (i = 0; i < vi->max_queue_pairs; i++) {
1546 callbacks[rxq2vq(i)] = skb_recv_done;
1547 callbacks[txq2vq(i)] = skb_xmit_done;
1548 sprintf(vi->rq[i].name, "input.%d", i);
1549 sprintf(vi->sq[i].name, "output.%d", i);
1550 names[rxq2vq(i)] = vi->rq[i].name;
1551 names[txq2vq(i)] = vi->sq[i].name;
1554 ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
1560 vi->cvq = vqs[total_vqs - 1];
1561 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
1562 vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
1565 for (i = 0; i < vi->max_queue_pairs; i++) {
1566 vi->rq[i].vq = vqs[rxq2vq(i)];
1567 vi->sq[i].vq = vqs[txq2vq(i)];
1586 static int virtnet_alloc_queues(struct virtnet_info *vi)
1590 vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
1593 vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
1597 INIT_DELAYED_WORK(&vi->refill, refill_work);
1598 for (i = 0; i < vi->max_queue_pairs; i++) {
1599 vi->rq[i].pages = NULL;
1600 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
1602 napi_hash_add(&vi->rq[i].napi);
1604 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
1605 ewma_init(&vi->rq[i].mrg_avg_pkt_len, 1, RECEIVE_AVG_WEIGHT);
1606 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
1617 static int init_vqs(struct virtnet_info *vi)
1621 /* Allocate send & receive queues */
1622 ret = virtnet_alloc_queues(vi);
1626 ret = virtnet_find_vqs(vi);
1631 virtnet_set_affinity(vi);
1637 virtnet_free_queues(vi);
1643 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
1644 struct rx_queue_attribute *attribute, char *buf)
1646 struct virtnet_info *vi = netdev_priv(queue->dev);
1647 unsigned int queue_index = get_netdev_rx_queue_index(queue);
1650 BUG_ON(queue_index >= vi->max_queue_pairs);
1651 avg = &vi->rq[queue_index].mrg_avg_pkt_len;
1652 return sprintf(buf, "%u\n", get_mergeable_buf_len(avg));
1655 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
1656 __ATTR_RO(mergeable_rx_buffer_size);
1658 static struct attribute *virtio_net_mrg_rx_attrs[] = {
1659 &mergeable_rx_buffer_size_attribute.attr,
1663 static const struct attribute_group virtio_net_mrg_rx_group = {
1664 .name = "virtio_net",
1665 .attrs = virtio_net_mrg_rx_attrs
1669 static int virtnet_probe(struct virtio_device *vdev)
1672 struct net_device *dev;
1673 struct virtnet_info *vi;
1674 u16 max_queue_pairs;
1676 /* Find if host supports multiqueue virtio_net device */
1677 err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
1678 struct virtio_net_config,
1679 max_virtqueue_pairs, &max_queue_pairs);
1681 /* We need at least 2 queue's */
1682 if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
1683 max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
1684 !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1685 max_queue_pairs = 1;
1687 /* Allocate ourselves a network device with room for our info */
1688 dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
1692 /* Set up network device as normal. */
1693 dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
1694 dev->netdev_ops = &virtnet_netdev;
1695 dev->features = NETIF_F_HIGHDMA;
1697 dev->ethtool_ops = &virtnet_ethtool_ops;
1698 SET_NETDEV_DEV(dev, &vdev->dev);
1700 /* Do we support "hardware" checksums? */
1701 if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
1702 /* This opens up the world of extra features. */
1703 dev->hw_features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1705 dev->features |= NETIF_F_HW_CSUM|NETIF_F_SG|NETIF_F_FRAGLIST;
1707 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
1708 dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
1709 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
1711 /* Individual feature bits: what can host handle? */
1712 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
1713 dev->hw_features |= NETIF_F_TSO;
1714 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
1715 dev->hw_features |= NETIF_F_TSO6;
1716 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
1717 dev->hw_features |= NETIF_F_TSO_ECN;
1718 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
1719 dev->hw_features |= NETIF_F_UFO;
1722 dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
1723 /* (!csum && gso) case will be fixed by register_netdev() */
1725 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
1726 dev->features |= NETIF_F_RXCSUM;
1728 dev->vlan_features = dev->features;
1730 /* Configuration may specify what MAC to use. Otherwise random. */
1731 if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
1732 virtio_cread_bytes(vdev,
1733 offsetof(struct virtio_net_config, mac),
1734 dev->dev_addr, dev->addr_len);
1736 eth_hw_addr_random(dev);
1738 /* Set up our device-specific information */
1739 vi = netdev_priv(dev);
1743 vi->stats = alloc_percpu(struct virtnet_stats);
1745 if (vi->stats == NULL)
1748 for_each_possible_cpu(i) {
1749 struct virtnet_stats *virtnet_stats;
1750 virtnet_stats = per_cpu_ptr(vi->stats, i);
1751 u64_stats_init(&virtnet_stats->tx_syncp);
1752 u64_stats_init(&virtnet_stats->rx_syncp);
1755 INIT_WORK(&vi->config_work, virtnet_config_changed_work);
1757 /* If we can receive ANY GSO packets, we must allocate large ones. */
1758 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1759 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1760 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
1761 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
1762 vi->big_packets = true;
1764 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
1765 vi->mergeable_rx_bufs = true;
1767 if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT))
1768 vi->any_header_sg = true;
1770 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
1773 if (vi->any_header_sg) {
1774 if (vi->mergeable_rx_bufs)
1775 dev->needed_headroom = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1777 dev->needed_headroom = sizeof(struct virtio_net_hdr);
1780 /* Use single tx/rx queue pair as default */
1781 vi->curr_queue_pairs = 1;
1782 vi->max_queue_pairs = max_queue_pairs;
1784 /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
1790 if (vi->mergeable_rx_bufs)
1791 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
1793 netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
1794 netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
1796 err = register_netdev(dev);
1798 pr_debug("virtio_net: registering device failed\n");
1802 virtio_device_ready(vdev);
1804 /* Last of all, set up some receive buffers. */
1805 for (i = 0; i < vi->curr_queue_pairs; i++) {
1806 try_fill_recv(&vi->rq[i], GFP_KERNEL);
1808 /* If we didn't even get one input buffer, we're useless. */
1809 if (vi->rq[i].vq->num_free ==
1810 virtqueue_get_vring_size(vi->rq[i].vq)) {
1811 free_unused_bufs(vi);
1813 goto free_recv_bufs;
1817 vi->nb.notifier_call = &virtnet_cpu_callback;
1818 err = register_hotcpu_notifier(&vi->nb);
1820 pr_debug("virtio_net: registering cpu notifier failed\n");
1821 goto free_recv_bufs;
1824 /* Assume link up if device can't report link status,
1825 otherwise get link status from config. */
1826 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
1827 netif_carrier_off(dev);
1828 schedule_work(&vi->config_work);
1830 vi->status = VIRTIO_NET_S_LINK_UP;
1831 netif_carrier_on(dev);
1834 pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
1835 dev->name, max_queue_pairs);
1840 vi->vdev->config->reset(vdev);
1842 free_receive_bufs(vi);
1843 unregister_netdev(dev);
1845 cancel_delayed_work_sync(&vi->refill);
1846 free_receive_page_frags(vi);
1847 virtnet_del_vqs(vi);
1849 free_percpu(vi->stats);
1855 static void remove_vq_common(struct virtnet_info *vi)
1857 vi->vdev->config->reset(vi->vdev);
1859 /* Free unused buffers in both send and recv, if any. */
1860 free_unused_bufs(vi);
1862 free_receive_bufs(vi);
1864 free_receive_page_frags(vi);
1866 virtnet_del_vqs(vi);
1869 static void virtnet_remove(struct virtio_device *vdev)
1871 struct virtnet_info *vi = vdev->priv;
1873 unregister_hotcpu_notifier(&vi->nb);
1875 /* Make sure no work handler is accessing the device. */
1876 flush_work(&vi->config_work);
1878 unregister_netdev(vi->dev);
1880 remove_vq_common(vi);
1882 free_percpu(vi->stats);
1883 free_netdev(vi->dev);
1886 #ifdef CONFIG_PM_SLEEP
1887 static int virtnet_freeze(struct virtio_device *vdev)
1889 struct virtnet_info *vi = vdev->priv;
1892 unregister_hotcpu_notifier(&vi->nb);
1894 /* Make sure no work handler is accessing the device */
1895 flush_work(&vi->config_work);
1897 netif_device_detach(vi->dev);
1898 cancel_delayed_work_sync(&vi->refill);
1900 if (netif_running(vi->dev)) {
1901 for (i = 0; i < vi->max_queue_pairs; i++) {
1902 napi_disable(&vi->rq[i].napi);
1903 napi_hash_del(&vi->rq[i].napi);
1904 netif_napi_del(&vi->rq[i].napi);
1908 remove_vq_common(vi);
1913 static int virtnet_restore(struct virtio_device *vdev)
1915 struct virtnet_info *vi = vdev->priv;
1922 virtio_device_ready(vdev);
1924 if (netif_running(vi->dev)) {
1925 for (i = 0; i < vi->curr_queue_pairs; i++)
1926 if (!try_fill_recv(&vi->rq[i], GFP_KERNEL))
1927 schedule_delayed_work(&vi->refill, 0);
1929 for (i = 0; i < vi->max_queue_pairs; i++)
1930 virtnet_napi_enable(&vi->rq[i]);
1933 netif_device_attach(vi->dev);
1936 virtnet_set_queues(vi, vi->curr_queue_pairs);
1939 err = register_hotcpu_notifier(&vi->nb);
1947 static struct virtio_device_id id_table[] = {
1948 { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
1952 static unsigned int features[] = {
1953 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM,
1954 VIRTIO_NET_F_GSO, VIRTIO_NET_F_MAC,
1955 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6,
1956 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6,
1957 VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO,
1958 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ,
1959 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN,
1960 VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ,
1961 VIRTIO_NET_F_CTRL_MAC_ADDR,
1962 VIRTIO_F_ANY_LAYOUT,
1965 static struct virtio_driver virtio_net_driver = {
1966 .feature_table = features,
1967 .feature_table_size = ARRAY_SIZE(features),
1968 .driver.name = KBUILD_MODNAME,
1969 .driver.owner = THIS_MODULE,
1970 .id_table = id_table,
1971 .probe = virtnet_probe,
1972 .remove = virtnet_remove,
1973 .config_changed = virtnet_config_changed,
1974 #ifdef CONFIG_PM_SLEEP
1975 .freeze = virtnet_freeze,
1976 .restore = virtnet_restore,
1980 module_virtio_driver(virtio_net_driver);
1982 MODULE_DEVICE_TABLE(virtio, id_table);
1983 MODULE_DESCRIPTION("Virtio network driver");
1984 MODULE_LICENSE("GPL");