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/bpf.h>
26 #include <linux/bpf_trace.h>
27 #include <linux/scatterlist.h>
28 #include <linux/if_vlan.h>
29 #include <linux/slab.h>
30 #include <linux/cpu.h>
31 #include <linux/average.h>
32 #include <net/route.h>
34 static int napi_weight = NAPI_POLL_WEIGHT;
35 module_param(napi_weight, int, 0444);
37 static bool csum = true, gso = true, napi_tx;
38 module_param(csum, bool, 0444);
39 module_param(gso, bool, 0444);
40 module_param(napi_tx, bool, 0644);
42 /* FIXME: MTU in config. */
43 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
44 #define GOOD_COPY_LEN 128
46 #define VIRTNET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
48 /* Amount of XDP headroom to prepend to packets for use by xdp_adjust_head */
49 #define VIRTIO_XDP_HEADROOM 256
51 /* RX packet size EWMA. The average packet size is used to determine the packet
52 * buffer size when refilling RX rings. As the entire RX ring may be refilled
53 * at once, the weight is chosen so that the EWMA will be insensitive to short-
54 * term, transient changes in packet size.
56 DECLARE_EWMA(pkt_len, 0, 64)
58 #define VIRTNET_DRIVER_VERSION "1.0.0"
60 static const unsigned long guest_offloads[] = {
61 VIRTIO_NET_F_GUEST_TSO4,
62 VIRTIO_NET_F_GUEST_TSO6,
63 VIRTIO_NET_F_GUEST_ECN,
64 VIRTIO_NET_F_GUEST_UFO,
65 VIRTIO_NET_F_GUEST_CSUM
68 struct virtnet_stats {
69 struct u64_stats_sync tx_syncp;
70 struct u64_stats_sync rx_syncp;
78 /* Internal representation of a send virtqueue */
80 /* Virtqueue associated with this send _queue */
83 /* TX: fragments + linear part + virtio header */
84 struct scatterlist sg[MAX_SKB_FRAGS + 2];
86 /* Name of the send queue: output.$index */
89 struct napi_struct napi;
92 /* Internal representation of a receive virtqueue */
93 struct receive_queue {
94 /* Virtqueue associated with this receive_queue */
97 struct napi_struct napi;
99 struct bpf_prog __rcu *xdp_prog;
101 /* Chain pages by the private ptr. */
104 /* Average packet length for mergeable receive buffers. */
105 struct ewma_pkt_len mrg_avg_pkt_len;
107 /* Page frag for packet buffer allocation. */
108 struct page_frag alloc_frag;
110 /* RX: fragments + linear part + virtio header */
111 struct scatterlist sg[MAX_SKB_FRAGS + 2];
113 /* Min single buffer size for mergeable buffers case. */
114 unsigned int min_buf_len;
116 /* Name of this receive queue: input.$index */
120 /* Control VQ buffers: protected by the rtnl lock */
122 struct virtio_net_ctrl_hdr hdr;
123 virtio_net_ctrl_ack status;
124 struct virtio_net_ctrl_mq mq;
131 struct virtnet_info {
132 struct virtio_device *vdev;
133 struct virtqueue *cvq;
134 struct net_device *dev;
135 struct send_queue *sq;
136 struct receive_queue *rq;
139 /* Max # of queue pairs supported by the device */
142 /* # of queue pairs currently used by the driver */
143 u16 curr_queue_pairs;
145 /* # of XDP queue pairs currently used by the driver */
148 /* I like... big packets and I cannot lie! */
151 /* Host will merge rx buffers for big packets (shake it! shake it!) */
152 bool mergeable_rx_bufs;
154 /* Has control virtqueue */
157 /* Host can handle any s/g split between our header and packet data */
160 /* Packet virtio header size */
163 /* Active statistics */
164 struct virtnet_stats __percpu *stats;
166 /* Work struct for refilling if we run low on memory. */
167 struct delayed_work refill;
169 /* Work struct for config space updates */
170 struct work_struct config_work;
172 /* Does the affinity hint is set for virtqueues? */
173 bool affinity_hint_set;
175 /* CPU hotplug instances for online & dead */
176 struct hlist_node node;
177 struct hlist_node node_dead;
179 struct control_buf *ctrl;
181 /* Ethtool settings */
185 unsigned long guest_offloads;
188 struct padded_vnet_hdr {
189 struct virtio_net_hdr_mrg_rxbuf hdr;
191 * hdr is in a separate sg buffer, and data sg buffer shares same page
192 * with this header sg. This padding makes next sg 16 byte aligned
198 /* Converting between virtqueue no. and kernel tx/rx queue no.
199 * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
201 static int vq2txq(struct virtqueue *vq)
203 return (vq->index - 1) / 2;
206 static int txq2vq(int txq)
211 static int vq2rxq(struct virtqueue *vq)
213 return vq->index / 2;
216 static int rxq2vq(int rxq)
221 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
223 return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
227 * private is used to chain pages for big packets, put the whole
228 * most recent used list in the beginning for reuse
230 static void give_pages(struct receive_queue *rq, struct page *page)
234 /* Find end of list, sew whole thing into vi->rq.pages. */
235 for (end = page; end->private; end = (struct page *)end->private);
236 end->private = (unsigned long)rq->pages;
240 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
242 struct page *p = rq->pages;
245 rq->pages = (struct page *)p->private;
246 /* clear private here, it is used to chain pages */
249 p = alloc_page(gfp_mask);
253 static void virtqueue_napi_schedule(struct napi_struct *napi,
254 struct virtqueue *vq)
256 if (napi_schedule_prep(napi)) {
257 virtqueue_disable_cb(vq);
258 __napi_schedule(napi);
262 static void virtqueue_napi_complete(struct napi_struct *napi,
263 struct virtqueue *vq, int processed)
267 opaque = virtqueue_enable_cb_prepare(vq);
268 if (napi_complete_done(napi, processed)) {
269 if (unlikely(virtqueue_poll(vq, opaque)))
270 virtqueue_napi_schedule(napi, vq);
272 virtqueue_disable_cb(vq);
276 static void skb_xmit_done(struct virtqueue *vq)
278 struct virtnet_info *vi = vq->vdev->priv;
279 struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi;
281 /* Suppress further interrupts. */
282 virtqueue_disable_cb(vq);
285 virtqueue_napi_schedule(napi, vq);
287 /* We were probably waiting for more output buffers. */
288 netif_wake_subqueue(vi->dev, vq2txq(vq));
291 #define MRG_CTX_HEADER_SHIFT 22
292 static void *mergeable_len_to_ctx(unsigned int truesize,
293 unsigned int headroom)
295 return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) | truesize);
298 static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx)
300 return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT;
303 static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx)
305 return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) - 1);
308 /* Called from bottom half context */
309 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
310 struct receive_queue *rq,
311 struct page *page, unsigned int offset,
312 unsigned int len, unsigned int truesize,
316 struct virtio_net_hdr_mrg_rxbuf *hdr;
317 unsigned int copy, hdr_len, hdr_padded_len;
320 p = page_address(page) + offset;
322 /* copy small packet so we can reuse these pages for small data */
323 skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
327 hdr = skb_vnet_hdr(skb);
329 hdr_len = vi->hdr_len;
330 if (vi->mergeable_rx_bufs)
331 hdr_padded_len = sizeof(*hdr);
333 hdr_padded_len = sizeof(struct padded_vnet_hdr);
336 memcpy(hdr, p, hdr_len);
339 offset += hdr_padded_len;
342 /* Copy all frame if it fits skb->head, otherwise
343 * we let virtio_net_hdr_to_skb() and GRO pull headers as needed.
345 if (len <= skb_tailroom(skb))
349 skb_put_data(skb, p, copy);
354 if (vi->mergeable_rx_bufs) {
356 skb_add_rx_frag(skb, 0, page, offset, len, truesize);
363 * Verify that we can indeed put this data into a skb.
364 * This is here to handle cases when the device erroneously
365 * tries to receive more than is possible. This is usually
366 * the case of a broken device.
368 if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
369 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
373 BUG_ON(offset >= PAGE_SIZE);
375 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
376 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
377 frag_size, truesize);
379 page = (struct page *)page->private;
384 give_pages(rq, page);
389 static bool virtnet_xdp_xmit(struct virtnet_info *vi,
390 struct receive_queue *rq,
391 struct xdp_buff *xdp)
393 struct virtio_net_hdr_mrg_rxbuf *hdr;
395 struct send_queue *sq;
400 qp = vi->curr_queue_pairs - vi->xdp_queue_pairs + smp_processor_id();
403 /* Free up any pending old buffers before queueing new ones. */
404 while ((xdp_sent = virtqueue_get_buf(sq->vq, &len)) != NULL) {
405 struct page *sent_page = virt_to_head_page(xdp_sent);
410 xdp->data -= vi->hdr_len;
411 /* Zero header and leave csum up to XDP layers */
413 memset(hdr, 0, vi->hdr_len);
415 sg_init_one(sq->sg, xdp->data, xdp->data_end - xdp->data);
417 err = virtqueue_add_outbuf(sq->vq, sq->sg, 1, xdp->data, GFP_ATOMIC);
419 struct page *page = virt_to_head_page(xdp->data);
425 virtqueue_kick(sq->vq);
429 static unsigned int virtnet_get_headroom(struct virtnet_info *vi)
431 return vi->xdp_queue_pairs ? VIRTIO_XDP_HEADROOM : 0;
434 /* We copy the packet for XDP in the following cases:
436 * 1) Packet is scattered across multiple rx buffers.
437 * 2) Headroom space is insufficient.
439 * This is inefficient but it's a temporary condition that
440 * we hit right after XDP is enabled and until queue is refilled
441 * with large buffers with sufficient headroom - so it should affect
442 * at most queue size packets.
443 * Afterwards, the conditions to enable
444 * XDP should preclude the underlying device from sending packets
445 * across multiple buffers (num_buf > 1), and we make sure buffers
446 * have enough headroom.
448 static struct page *xdp_linearize_page(struct receive_queue *rq,
455 int tailroom = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
458 if (page_off + *len + tailroom > PAGE_SIZE)
461 page = alloc_page(GFP_ATOMIC);
465 memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
473 buf = virtqueue_get_buf(rq->vq, &buflen);
477 p = virt_to_head_page(buf);
478 off = buf - page_address(p);
480 /* guard against a misconfigured or uncooperative backend that
481 * is sending packet larger than the MTU.
483 if ((page_off + buflen) > PAGE_SIZE) {
488 memcpy(page_address(page) + page_off,
489 page_address(p) + off, buflen);
494 /* Headroom does not contribute to packet length */
495 *len = page_off - VIRTIO_XDP_HEADROOM;
498 __free_pages(page, 0);
502 static struct sk_buff *receive_small(struct net_device *dev,
503 struct virtnet_info *vi,
504 struct receive_queue *rq,
505 void *buf, void *ctx,
509 struct bpf_prog *xdp_prog;
510 unsigned int xdp_headroom = (unsigned long)ctx;
511 unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom;
512 unsigned int headroom = vi->hdr_len + header_offset;
513 unsigned int buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
514 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
515 struct page *page = virt_to_head_page(buf);
516 unsigned int delta = 0;
517 struct page *xdp_page;
521 xdp_prog = rcu_dereference(rq->xdp_prog);
523 struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset;
528 if (unlikely(hdr->hdr.gso_type))
531 if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) {
532 int offset = buf - page_address(page) + header_offset;
533 unsigned int tlen = len + vi->hdr_len;
536 xdp_headroom = virtnet_get_headroom(vi);
537 header_offset = VIRTNET_RX_PAD + xdp_headroom;
538 headroom = vi->hdr_len + header_offset;
539 buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
540 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
541 xdp_page = xdp_linearize_page(rq, &num_buf, page,
542 offset, header_offset,
547 buf = page_address(xdp_page);
552 xdp.data_hard_start = buf + VIRTNET_RX_PAD + vi->hdr_len;
553 xdp.data = xdp.data_hard_start + xdp_headroom;
554 xdp.data_end = xdp.data + len;
555 orig_data = xdp.data;
556 act = bpf_prog_run_xdp(xdp_prog, &xdp);
560 /* Recalculate length in case bpf program changed it */
561 delta = orig_data - xdp.data;
564 if (unlikely(!virtnet_xdp_xmit(vi, rq, &xdp)))
565 trace_xdp_exception(vi->dev, xdp_prog, act);
569 bpf_warn_invalid_xdp_action(act);
571 trace_xdp_exception(vi->dev, xdp_prog, act);
578 skb = build_skb(buf, buflen);
583 skb_reserve(skb, headroom - delta);
584 skb_put(skb, len + delta);
586 buf += header_offset;
587 memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
588 } /* keep zeroed vnet hdr since packet was changed by bpf */
595 dev->stats.rx_dropped++;
601 static struct sk_buff *receive_big(struct net_device *dev,
602 struct virtnet_info *vi,
603 struct receive_queue *rq,
607 struct page *page = buf;
608 struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len,
617 dev->stats.rx_dropped++;
618 give_pages(rq, page);
622 static struct sk_buff *receive_mergeable(struct net_device *dev,
623 struct virtnet_info *vi,
624 struct receive_queue *rq,
629 struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
630 u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
631 struct page *page = virt_to_head_page(buf);
632 int offset = buf - page_address(page);
633 struct sk_buff *head_skb, *curr_skb;
634 struct bpf_prog *xdp_prog;
635 unsigned int truesize;
636 unsigned int headroom = mergeable_ctx_to_headroom(ctx);
641 xdp_prog = rcu_dereference(rq->xdp_prog);
643 struct page *xdp_page;
648 /* Transient failure which in theory could occur if
649 * in-flight packets from before XDP was enabled reach
650 * the receive path after XDP is loaded.
652 if (unlikely(hdr->hdr.gso_type))
655 /* This happens when rx buffer size is underestimated */
656 if (unlikely(num_buf > 1 ||
657 headroom < virtnet_get_headroom(vi))) {
658 /* linearize data for XDP */
659 xdp_page = xdp_linearize_page(rq, &num_buf,
665 offset = VIRTIO_XDP_HEADROOM;
670 /* Allow consuming headroom but reserve enough space to push
671 * the descriptor on if we get an XDP_TX return code.
673 data = page_address(xdp_page) + offset;
674 xdp.data_hard_start = data - VIRTIO_XDP_HEADROOM + vi->hdr_len;
675 xdp.data = data + vi->hdr_len;
676 xdp.data_end = xdp.data + (len - vi->hdr_len);
677 act = bpf_prog_run_xdp(xdp_prog, &xdp);
681 /* recalculate offset to account for any header
682 * adjustments. Note other cases do not build an
683 * skb and avoid using offset
686 page_address(xdp_page) - vi->hdr_len;
688 /* We can only create skb based on xdp_page. */
689 if (unlikely(xdp_page != page)) {
692 head_skb = page_to_skb(vi, rq, xdp_page,
695 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, len);
700 if (unlikely(!virtnet_xdp_xmit(vi, rq, &xdp)))
701 trace_xdp_exception(vi->dev, xdp_prog, act);
702 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, len);
703 if (unlikely(xdp_page != page))
708 bpf_warn_invalid_xdp_action(act);
710 trace_xdp_exception(vi->dev, xdp_prog, act);
712 if (unlikely(xdp_page != page))
713 __free_pages(xdp_page, 0);
714 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, len);
720 truesize = mergeable_ctx_to_truesize(ctx);
721 if (unlikely(len > truesize)) {
722 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
723 dev->name, len, (unsigned long)ctx);
724 dev->stats.rx_length_errors++;
728 head_skb = page_to_skb(vi, rq, page, offset, len, truesize, !xdp_prog);
731 if (unlikely(!curr_skb))
736 buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx);
737 if (unlikely(!buf)) {
738 pr_debug("%s: rx error: %d buffers out of %d missing\n",
740 virtio16_to_cpu(vi->vdev,
742 dev->stats.rx_length_errors++;
746 page = virt_to_head_page(buf);
748 truesize = mergeable_ctx_to_truesize(ctx);
749 if (unlikely(len > truesize)) {
750 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
751 dev->name, len, (unsigned long)ctx);
752 dev->stats.rx_length_errors++;
756 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
757 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
758 struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
762 if (curr_skb == head_skb)
763 skb_shinfo(curr_skb)->frag_list = nskb;
765 curr_skb->next = nskb;
767 head_skb->truesize += nskb->truesize;
770 if (curr_skb != head_skb) {
771 head_skb->data_len += len;
772 head_skb->len += len;
773 head_skb->truesize += truesize;
775 offset = buf - page_address(page);
776 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
778 skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
781 skb_add_rx_frag(curr_skb, num_skb_frags, page,
782 offset, len, truesize);
786 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
793 while (num_buf-- > 1) {
794 buf = virtqueue_get_buf(rq->vq, &len);
795 if (unlikely(!buf)) {
796 pr_debug("%s: rx error: %d buffers missing\n",
798 dev->stats.rx_length_errors++;
801 page = virt_to_head_page(buf);
805 dev->stats.rx_dropped++;
806 dev_kfree_skb(head_skb);
811 static int receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
812 void *buf, unsigned int len, void **ctx)
814 struct net_device *dev = vi->dev;
816 struct virtio_net_hdr_mrg_rxbuf *hdr;
819 if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
820 pr_debug("%s: short packet %i\n", dev->name, len);
821 dev->stats.rx_length_errors++;
822 if (vi->mergeable_rx_bufs) {
823 put_page(virt_to_head_page(buf));
824 } else if (vi->big_packets) {
827 put_page(virt_to_head_page(buf));
832 if (vi->mergeable_rx_bufs)
833 skb = receive_mergeable(dev, vi, rq, buf, ctx, len);
834 else if (vi->big_packets)
835 skb = receive_big(dev, vi, rq, buf, len);
837 skb = receive_small(dev, vi, rq, buf, ctx, len);
842 hdr = skb_vnet_hdr(skb);
846 if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
847 skb->ip_summed = CHECKSUM_UNNECESSARY;
849 if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
850 virtio_is_little_endian(vi->vdev))) {
851 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
852 dev->name, hdr->hdr.gso_type,
857 skb->protocol = eth_type_trans(skb, dev);
858 pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
859 ntohs(skb->protocol), skb->len, skb->pkt_type);
861 napi_gro_receive(&rq->napi, skb);
865 dev->stats.rx_frame_errors++;
870 /* Unlike mergeable buffers, all buffers are allocated to the
871 * same size, except for the headroom. For this reason we do
872 * not need to use mergeable_len_to_ctx here - it is enough
873 * to store the headroom as the context ignoring the truesize.
875 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
878 struct page_frag *alloc_frag = &rq->alloc_frag;
880 unsigned int xdp_headroom = virtnet_get_headroom(vi);
881 void *ctx = (void *)(unsigned long)xdp_headroom;
882 int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
885 len = SKB_DATA_ALIGN(len) +
886 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
887 if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
890 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
891 get_page(alloc_frag->page);
892 alloc_frag->offset += len;
893 sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom,
894 vi->hdr_len + GOOD_PACKET_LEN);
895 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
897 put_page(virt_to_head_page(buf));
901 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
904 struct page *first, *list = NULL;
908 sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
910 /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
911 for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
912 first = get_a_page(rq, gfp);
915 give_pages(rq, list);
918 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
920 /* chain new page in list head to match sg */
921 first->private = (unsigned long)list;
925 first = get_a_page(rq, gfp);
927 give_pages(rq, list);
930 p = page_address(first);
932 /* rq->sg[0], rq->sg[1] share the same page */
933 /* a separated rq->sg[0] for header - required in case !any_header_sg */
934 sg_set_buf(&rq->sg[0], p, vi->hdr_len);
936 /* rq->sg[1] for data packet, from offset */
937 offset = sizeof(struct padded_vnet_hdr);
938 sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
940 /* chain first in list head */
941 first->private = (unsigned long)list;
942 err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
945 give_pages(rq, first);
950 static unsigned int get_mergeable_buf_len(struct receive_queue *rq,
951 struct ewma_pkt_len *avg_pkt_len)
953 const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
956 len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
957 rq->min_buf_len, PAGE_SIZE - hdr_len);
958 return ALIGN(len, L1_CACHE_BYTES);
961 static int add_recvbuf_mergeable(struct virtnet_info *vi,
962 struct receive_queue *rq, gfp_t gfp)
964 struct page_frag *alloc_frag = &rq->alloc_frag;
965 unsigned int headroom = virtnet_get_headroom(vi);
969 unsigned int len, hole;
971 len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len);
972 if (unlikely(!skb_page_frag_refill(len + headroom, alloc_frag, gfp)))
975 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
976 buf += headroom; /* advance address leaving hole at front of pkt */
977 get_page(alloc_frag->page);
978 alloc_frag->offset += len + headroom;
979 hole = alloc_frag->size - alloc_frag->offset;
980 if (hole < len + headroom) {
981 /* To avoid internal fragmentation, if there is very likely not
982 * enough space for another buffer, add the remaining space to
983 * the current buffer.
986 alloc_frag->offset += hole;
989 sg_init_one(rq->sg, buf, len);
990 ctx = mergeable_len_to_ctx(len, headroom);
991 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
993 put_page(virt_to_head_page(buf));
999 * Returns false if we couldn't fill entirely (OOM).
1001 * Normally run in the receive path, but can also be run from ndo_open
1002 * before we're receiving packets, or from refill_work which is
1003 * careful to disable receiving (using napi_disable).
1005 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
1013 if (vi->mergeable_rx_bufs)
1014 err = add_recvbuf_mergeable(vi, rq, gfp);
1015 else if (vi->big_packets)
1016 err = add_recvbuf_big(vi, rq, gfp);
1018 err = add_recvbuf_small(vi, rq, gfp);
1020 oom = err == -ENOMEM;
1023 } while (rq->vq->num_free);
1024 virtqueue_kick(rq->vq);
1028 static void skb_recv_done(struct virtqueue *rvq)
1030 struct virtnet_info *vi = rvq->vdev->priv;
1031 struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
1033 virtqueue_napi_schedule(&rq->napi, rvq);
1036 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
1040 /* If all buffers were filled by other side before we napi_enabled, we
1041 * won't get another interrupt, so process any outstanding packets now.
1042 * Call local_bh_enable after to trigger softIRQ processing.
1045 virtqueue_napi_schedule(napi, vq);
1049 static void virtnet_napi_tx_enable(struct virtnet_info *vi,
1050 struct virtqueue *vq,
1051 struct napi_struct *napi)
1056 /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
1057 * enable the feature if this is likely affine with the transmit path.
1059 if (!vi->affinity_hint_set) {
1064 return virtnet_napi_enable(vq, napi);
1067 static void virtnet_napi_tx_disable(struct napi_struct *napi)
1073 static void refill_work(struct work_struct *work)
1075 struct virtnet_info *vi =
1076 container_of(work, struct virtnet_info, refill.work);
1080 for (i = 0; i < vi->curr_queue_pairs; i++) {
1081 struct receive_queue *rq = &vi->rq[i];
1083 napi_disable(&rq->napi);
1084 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
1085 virtnet_napi_enable(rq->vq, &rq->napi);
1087 /* In theory, this can happen: if we don't get any buffers in
1088 * we will *never* try to fill again.
1091 schedule_delayed_work(&vi->refill, HZ/2);
1095 static int virtnet_receive(struct receive_queue *rq, int budget)
1097 struct virtnet_info *vi = rq->vq->vdev->priv;
1098 unsigned int len, received = 0, bytes = 0;
1100 struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
1102 if (!vi->big_packets || vi->mergeable_rx_bufs) {
1105 while (received < budget &&
1106 (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) {
1107 bytes += receive_buf(vi, rq, buf, len, ctx);
1111 while (received < budget &&
1112 (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
1113 bytes += receive_buf(vi, rq, buf, len, NULL);
1118 if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) {
1119 if (!try_fill_recv(vi, rq, GFP_ATOMIC))
1120 schedule_delayed_work(&vi->refill, 0);
1123 u64_stats_update_begin(&stats->rx_syncp);
1124 stats->rx_bytes += bytes;
1125 stats->rx_packets += received;
1126 u64_stats_update_end(&stats->rx_syncp);
1131 static void free_old_xmit_skbs(struct send_queue *sq)
1133 struct sk_buff *skb;
1135 struct virtnet_info *vi = sq->vq->vdev->priv;
1136 struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
1137 unsigned int packets = 0;
1138 unsigned int bytes = 0;
1140 while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
1141 pr_debug("Sent skb %p\n", skb);
1146 dev_consume_skb_any(skb);
1149 /* Avoid overhead when no packets have been processed
1150 * happens when called speculatively from start_xmit.
1155 u64_stats_update_begin(&stats->tx_syncp);
1156 stats->tx_bytes += bytes;
1157 stats->tx_packets += packets;
1158 u64_stats_update_end(&stats->tx_syncp);
1161 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
1163 if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
1165 else if (q < vi->curr_queue_pairs)
1171 static void virtnet_poll_cleantx(struct receive_queue *rq)
1173 struct virtnet_info *vi = rq->vq->vdev->priv;
1174 unsigned int index = vq2rxq(rq->vq);
1175 struct send_queue *sq = &vi->sq[index];
1176 struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index);
1178 if (!sq->napi.weight || is_xdp_raw_buffer_queue(vi, index))
1181 if (__netif_tx_trylock(txq)) {
1182 free_old_xmit_skbs(sq);
1183 __netif_tx_unlock(txq);
1186 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1187 netif_tx_wake_queue(txq);
1190 static int virtnet_poll(struct napi_struct *napi, int budget)
1192 struct receive_queue *rq =
1193 container_of(napi, struct receive_queue, napi);
1194 unsigned int received;
1196 virtnet_poll_cleantx(rq);
1198 received = virtnet_receive(rq, budget);
1200 /* Out of packets? */
1201 if (received < budget)
1202 virtqueue_napi_complete(napi, rq->vq, received);
1207 static int virtnet_open(struct net_device *dev)
1209 struct virtnet_info *vi = netdev_priv(dev);
1212 for (i = 0; i < vi->max_queue_pairs; i++) {
1213 if (i < vi->curr_queue_pairs)
1214 /* Make sure we have some buffers: if oom use wq. */
1215 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1216 schedule_delayed_work(&vi->refill, 0);
1217 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1218 virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi);
1224 static int virtnet_poll_tx(struct napi_struct *napi, int budget)
1226 struct send_queue *sq = container_of(napi, struct send_queue, napi);
1227 struct virtnet_info *vi = sq->vq->vdev->priv;
1228 unsigned int index = vq2txq(sq->vq);
1229 struct netdev_queue *txq;
1231 if (unlikely(is_xdp_raw_buffer_queue(vi, index))) {
1232 /* We don't need to enable cb for XDP */
1233 napi_complete_done(napi, 0);
1237 txq = netdev_get_tx_queue(vi->dev, index);
1238 __netif_tx_lock(txq, raw_smp_processor_id());
1239 free_old_xmit_skbs(sq);
1240 __netif_tx_unlock(txq);
1242 virtqueue_napi_complete(napi, sq->vq, 0);
1244 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1245 netif_tx_wake_queue(txq);
1250 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
1252 struct virtio_net_hdr_mrg_rxbuf *hdr;
1253 const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
1254 struct virtnet_info *vi = sq->vq->vdev->priv;
1256 unsigned hdr_len = vi->hdr_len;
1259 pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
1261 can_push = vi->any_header_sg &&
1262 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
1263 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
1264 /* Even if we can, don't push here yet as this would skew
1265 * csum_start offset below. */
1267 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1269 hdr = skb_vnet_hdr(skb);
1271 if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1272 virtio_is_little_endian(vi->vdev), false,
1276 if (vi->mergeable_rx_bufs)
1277 hdr->num_buffers = 0;
1279 sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1281 __skb_push(skb, hdr_len);
1282 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1283 if (unlikely(num_sg < 0))
1285 /* Pull header back to avoid skew in tx bytes calculations. */
1286 __skb_pull(skb, hdr_len);
1288 sg_set_buf(sq->sg, hdr, hdr_len);
1289 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
1290 if (unlikely(num_sg < 0))
1294 return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
1297 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
1299 struct virtnet_info *vi = netdev_priv(dev);
1300 int qnum = skb_get_queue_mapping(skb);
1301 struct send_queue *sq = &vi->sq[qnum];
1303 struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1304 bool kick = !skb->xmit_more;
1305 bool use_napi = sq->napi.weight;
1307 /* Free up any pending old buffers before queueing new ones. */
1308 free_old_xmit_skbs(sq);
1310 if (use_napi && kick)
1311 virtqueue_enable_cb_delayed(sq->vq);
1313 /* timestamp packet in software */
1314 skb_tx_timestamp(skb);
1316 /* Try to transmit */
1317 err = xmit_skb(sq, skb);
1319 /* This should not happen! */
1320 if (unlikely(err)) {
1321 dev->stats.tx_fifo_errors++;
1322 if (net_ratelimit())
1324 "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
1325 dev->stats.tx_dropped++;
1326 dev_kfree_skb_any(skb);
1327 return NETDEV_TX_OK;
1330 /* Don't wait up for transmitted skbs to be freed. */
1336 /* If running out of space, stop queue to avoid getting packets that we
1337 * are then unable to transmit.
1338 * An alternative would be to force queuing layer to requeue the skb by
1339 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1340 * returned in a normal path of operation: it means that driver is not
1341 * maintaining the TX queue stop/start state properly, and causes
1342 * the stack to do a non-trivial amount of useless work.
1343 * Since most packets only take 1 or 2 ring slots, stopping the queue
1344 * early means 16 slots are typically wasted.
1346 if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
1347 netif_stop_subqueue(dev, qnum);
1349 unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
1350 /* More just got used, free them then recheck. */
1351 free_old_xmit_skbs(sq);
1352 if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
1353 netif_start_subqueue(dev, qnum);
1354 virtqueue_disable_cb(sq->vq);
1359 if (kick || netif_xmit_stopped(txq))
1360 virtqueue_kick(sq->vq);
1362 return NETDEV_TX_OK;
1366 * Send command via the control virtqueue and check status. Commands
1367 * supported by the hypervisor, as indicated by feature bits, should
1368 * never fail unless improperly formatted.
1370 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
1371 struct scatterlist *out)
1373 struct scatterlist *sgs[4], hdr, stat;
1374 unsigned out_num = 0, tmp;
1376 /* Caller should know better */
1377 BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
1379 vi->ctrl->status = ~0;
1380 vi->ctrl->hdr.class = class;
1381 vi->ctrl->hdr.cmd = cmd;
1383 sg_init_one(&hdr, &vi->ctrl->hdr, sizeof(vi->ctrl->hdr));
1384 sgs[out_num++] = &hdr;
1387 sgs[out_num++] = out;
1389 /* Add return status. */
1390 sg_init_one(&stat, &vi->ctrl->status, sizeof(vi->ctrl->status));
1391 sgs[out_num] = &stat;
1393 BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1394 virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1396 if (unlikely(!virtqueue_kick(vi->cvq)))
1397 return vi->ctrl->status == VIRTIO_NET_OK;
1399 /* Spin for a response, the kick causes an ioport write, trapping
1400 * into the hypervisor, so the request should be handled immediately.
1402 while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1403 !virtqueue_is_broken(vi->cvq))
1406 return vi->ctrl->status == VIRTIO_NET_OK;
1409 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1411 struct virtnet_info *vi = netdev_priv(dev);
1412 struct virtio_device *vdev = vi->vdev;
1414 struct sockaddr *addr;
1415 struct scatterlist sg;
1417 addr = kmemdup(p, sizeof(*addr), GFP_KERNEL);
1421 ret = eth_prepare_mac_addr_change(dev, addr);
1425 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1426 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1427 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1428 VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1429 dev_warn(&vdev->dev,
1430 "Failed to set mac address by vq command.\n");
1434 } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1435 !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1438 /* Naturally, this has an atomicity problem. */
1439 for (i = 0; i < dev->addr_len; i++)
1440 virtio_cwrite8(vdev,
1441 offsetof(struct virtio_net_config, mac) +
1442 i, addr->sa_data[i]);
1445 eth_commit_mac_addr_change(dev, p);
1453 static void virtnet_stats(struct net_device *dev,
1454 struct rtnl_link_stats64 *tot)
1456 struct virtnet_info *vi = netdev_priv(dev);
1460 for_each_possible_cpu(cpu) {
1461 struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
1462 u64 tpackets, tbytes, rpackets, rbytes;
1465 start = u64_stats_fetch_begin_irq(&stats->tx_syncp);
1466 tpackets = stats->tx_packets;
1467 tbytes = stats->tx_bytes;
1468 } while (u64_stats_fetch_retry_irq(&stats->tx_syncp, start));
1471 start = u64_stats_fetch_begin_irq(&stats->rx_syncp);
1472 rpackets = stats->rx_packets;
1473 rbytes = stats->rx_bytes;
1474 } while (u64_stats_fetch_retry_irq(&stats->rx_syncp, start));
1476 tot->rx_packets += rpackets;
1477 tot->tx_packets += tpackets;
1478 tot->rx_bytes += rbytes;
1479 tot->tx_bytes += tbytes;
1482 tot->tx_dropped = dev->stats.tx_dropped;
1483 tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1484 tot->rx_dropped = dev->stats.rx_dropped;
1485 tot->rx_length_errors = dev->stats.rx_length_errors;
1486 tot->rx_frame_errors = dev->stats.rx_frame_errors;
1489 #ifdef CONFIG_NET_POLL_CONTROLLER
1490 static void virtnet_netpoll(struct net_device *dev)
1492 struct virtnet_info *vi = netdev_priv(dev);
1495 for (i = 0; i < vi->curr_queue_pairs; i++)
1496 napi_schedule(&vi->rq[i].napi);
1500 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1503 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1504 VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1505 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1509 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1511 struct scatterlist sg;
1512 struct net_device *dev = vi->dev;
1514 if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1517 vi->ctrl->mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1518 sg_init_one(&sg, &vi->ctrl->mq, sizeof(vi->ctrl->mq));
1520 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1521 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1522 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1526 vi->curr_queue_pairs = queue_pairs;
1527 /* virtnet_open() will refill when device is going to up. */
1528 if (dev->flags & IFF_UP)
1529 schedule_delayed_work(&vi->refill, 0);
1535 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1540 err = _virtnet_set_queues(vi, queue_pairs);
1545 static int virtnet_close(struct net_device *dev)
1547 struct virtnet_info *vi = netdev_priv(dev);
1550 /* Make sure refill_work doesn't re-enable napi! */
1551 cancel_delayed_work_sync(&vi->refill);
1553 for (i = 0; i < vi->max_queue_pairs; i++) {
1554 napi_disable(&vi->rq[i].napi);
1555 virtnet_napi_tx_disable(&vi->sq[i].napi);
1561 static void virtnet_set_rx_mode(struct net_device *dev)
1563 struct virtnet_info *vi = netdev_priv(dev);
1564 struct scatterlist sg[2];
1565 struct virtio_net_ctrl_mac *mac_data;
1566 struct netdev_hw_addr *ha;
1572 /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1573 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1576 vi->ctrl->promisc = ((dev->flags & IFF_PROMISC) != 0);
1577 vi->ctrl->allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1579 sg_init_one(sg, &vi->ctrl->promisc, sizeof(vi->ctrl->promisc));
1581 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1582 VIRTIO_NET_CTRL_RX_PROMISC, sg))
1583 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1584 vi->ctrl->promisc ? "en" : "dis");
1586 sg_init_one(sg, &vi->ctrl->allmulti, sizeof(vi->ctrl->allmulti));
1588 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1589 VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1590 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1591 vi->ctrl->allmulti ? "en" : "dis");
1593 uc_count = netdev_uc_count(dev);
1594 mc_count = netdev_mc_count(dev);
1595 /* MAC filter - use one buffer for both lists */
1596 buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1597 (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1602 sg_init_table(sg, 2);
1604 /* Store the unicast list and count in the front of the buffer */
1605 mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1607 netdev_for_each_uc_addr(ha, dev)
1608 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1610 sg_set_buf(&sg[0], mac_data,
1611 sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1613 /* multicast list and count fill the end */
1614 mac_data = (void *)&mac_data->macs[uc_count][0];
1616 mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1618 netdev_for_each_mc_addr(ha, dev)
1619 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1621 sg_set_buf(&sg[1], mac_data,
1622 sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1624 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1625 VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1626 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1631 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1632 __be16 proto, u16 vid)
1634 struct virtnet_info *vi = netdev_priv(dev);
1635 struct scatterlist sg;
1637 vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
1638 sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
1640 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1641 VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1642 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1646 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1647 __be16 proto, u16 vid)
1649 struct virtnet_info *vi = netdev_priv(dev);
1650 struct scatterlist sg;
1652 vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
1653 sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
1655 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1656 VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1657 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1661 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1665 if (vi->affinity_hint_set) {
1666 for (i = 0; i < vi->max_queue_pairs; i++) {
1667 virtqueue_set_affinity(vi->rq[i].vq, -1);
1668 virtqueue_set_affinity(vi->sq[i].vq, -1);
1671 vi->affinity_hint_set = false;
1675 static void virtnet_set_affinity(struct virtnet_info *vi)
1680 /* In multiqueue mode, when the number of cpu is equal to the number of
1681 * queue pairs, we let the queue pairs to be private to one cpu by
1682 * setting the affinity hint to eliminate the contention.
1684 if (vi->curr_queue_pairs == 1 ||
1685 vi->max_queue_pairs != num_online_cpus()) {
1686 virtnet_clean_affinity(vi, -1);
1691 for_each_online_cpu(cpu) {
1692 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1693 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1694 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
1698 vi->affinity_hint_set = true;
1701 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
1703 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1705 virtnet_set_affinity(vi);
1709 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
1711 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1713 virtnet_set_affinity(vi);
1717 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
1719 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1722 virtnet_clean_affinity(vi, cpu);
1726 static enum cpuhp_state virtionet_online;
1728 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
1732 ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
1735 ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1739 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1743 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
1745 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1746 cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1750 static void virtnet_get_ringparam(struct net_device *dev,
1751 struct ethtool_ringparam *ring)
1753 struct virtnet_info *vi = netdev_priv(dev);
1755 ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1756 ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1757 ring->rx_pending = ring->rx_max_pending;
1758 ring->tx_pending = ring->tx_max_pending;
1762 static void virtnet_get_drvinfo(struct net_device *dev,
1763 struct ethtool_drvinfo *info)
1765 struct virtnet_info *vi = netdev_priv(dev);
1766 struct virtio_device *vdev = vi->vdev;
1768 strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1769 strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1770 strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1774 /* TODO: Eliminate OOO packets during switching */
1775 static int virtnet_set_channels(struct net_device *dev,
1776 struct ethtool_channels *channels)
1778 struct virtnet_info *vi = netdev_priv(dev);
1779 u16 queue_pairs = channels->combined_count;
1782 /* We don't support separate rx/tx channels.
1783 * We don't allow setting 'other' channels.
1785 if (channels->rx_count || channels->tx_count || channels->other_count)
1788 if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
1791 /* For now we don't support modifying channels while XDP is loaded
1792 * also when XDP is loaded all RX queues have XDP programs so we only
1793 * need to check a single RX queue.
1795 if (vi->rq[0].xdp_prog)
1799 err = _virtnet_set_queues(vi, queue_pairs);
1804 virtnet_set_affinity(vi);
1807 netif_set_real_num_tx_queues(dev, queue_pairs);
1808 netif_set_real_num_rx_queues(dev, queue_pairs);
1813 static void virtnet_get_channels(struct net_device *dev,
1814 struct ethtool_channels *channels)
1816 struct virtnet_info *vi = netdev_priv(dev);
1818 channels->combined_count = vi->curr_queue_pairs;
1819 channels->max_combined = vi->max_queue_pairs;
1820 channels->max_other = 0;
1821 channels->rx_count = 0;
1822 channels->tx_count = 0;
1823 channels->other_count = 0;
1826 /* Check if the user is trying to change anything besides speed/duplex */
1828 virtnet_validate_ethtool_cmd(const struct ethtool_link_ksettings *cmd)
1830 struct ethtool_link_ksettings diff1 = *cmd;
1831 struct ethtool_link_ksettings diff2 = {};
1833 /* cmd is always set so we need to clear it, validate the port type
1834 * and also without autonegotiation we can ignore advertising
1836 diff1.base.speed = 0;
1837 diff2.base.port = PORT_OTHER;
1838 ethtool_link_ksettings_zero_link_mode(&diff1, advertising);
1839 diff1.base.duplex = 0;
1841 diff1.base.link_mode_masks_nwords = 0;
1843 return !memcmp(&diff1.base, &diff2.base, sizeof(diff1.base)) &&
1844 bitmap_empty(diff1.link_modes.supported,
1845 __ETHTOOL_LINK_MODE_MASK_NBITS) &&
1846 bitmap_empty(diff1.link_modes.advertising,
1847 __ETHTOOL_LINK_MODE_MASK_NBITS) &&
1848 bitmap_empty(diff1.link_modes.lp_advertising,
1849 __ETHTOOL_LINK_MODE_MASK_NBITS);
1852 static int virtnet_set_link_ksettings(struct net_device *dev,
1853 const struct ethtool_link_ksettings *cmd)
1855 struct virtnet_info *vi = netdev_priv(dev);
1858 speed = cmd->base.speed;
1859 /* don't allow custom speed and duplex */
1860 if (!ethtool_validate_speed(speed) ||
1861 !ethtool_validate_duplex(cmd->base.duplex) ||
1862 !virtnet_validate_ethtool_cmd(cmd))
1865 vi->duplex = cmd->base.duplex;
1870 static int virtnet_get_link_ksettings(struct net_device *dev,
1871 struct ethtool_link_ksettings *cmd)
1873 struct virtnet_info *vi = netdev_priv(dev);
1875 cmd->base.speed = vi->speed;
1876 cmd->base.duplex = vi->duplex;
1877 cmd->base.port = PORT_OTHER;
1882 static void virtnet_init_settings(struct net_device *dev)
1884 struct virtnet_info *vi = netdev_priv(dev);
1886 vi->speed = SPEED_UNKNOWN;
1887 vi->duplex = DUPLEX_UNKNOWN;
1890 static const struct ethtool_ops virtnet_ethtool_ops = {
1891 .get_drvinfo = virtnet_get_drvinfo,
1892 .get_link = ethtool_op_get_link,
1893 .get_ringparam = virtnet_get_ringparam,
1894 .set_channels = virtnet_set_channels,
1895 .get_channels = virtnet_get_channels,
1896 .get_ts_info = ethtool_op_get_ts_info,
1897 .get_link_ksettings = virtnet_get_link_ksettings,
1898 .set_link_ksettings = virtnet_set_link_ksettings,
1901 static void virtnet_freeze_down(struct virtio_device *vdev)
1903 struct virtnet_info *vi = vdev->priv;
1906 /* Make sure no work handler is accessing the device */
1907 flush_work(&vi->config_work);
1909 netif_tx_lock_bh(vi->dev);
1910 netif_device_detach(vi->dev);
1911 netif_tx_unlock_bh(vi->dev);
1912 cancel_delayed_work_sync(&vi->refill);
1914 if (netif_running(vi->dev)) {
1915 for (i = 0; i < vi->max_queue_pairs; i++) {
1916 napi_disable(&vi->rq[i].napi);
1917 virtnet_napi_tx_disable(&vi->sq[i].napi);
1922 static int init_vqs(struct virtnet_info *vi);
1924 static int virtnet_restore_up(struct virtio_device *vdev)
1926 struct virtnet_info *vi = vdev->priv;
1933 virtio_device_ready(vdev);
1935 if (netif_running(vi->dev)) {
1936 for (i = 0; i < vi->curr_queue_pairs; i++)
1937 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1938 schedule_delayed_work(&vi->refill, 0);
1940 for (i = 0; i < vi->max_queue_pairs; i++) {
1941 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1942 virtnet_napi_tx_enable(vi, vi->sq[i].vq,
1947 netif_tx_lock_bh(vi->dev);
1948 netif_device_attach(vi->dev);
1949 netif_tx_unlock_bh(vi->dev);
1953 static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads)
1955 struct scatterlist sg;
1956 vi->ctrl->offloads = cpu_to_virtio64(vi->vdev, offloads);
1958 sg_init_one(&sg, &vi->ctrl->offloads, sizeof(vi->ctrl->offloads));
1960 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
1961 VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) {
1962 dev_warn(&vi->dev->dev, "Fail to set guest offload. \n");
1969 static int virtnet_clear_guest_offloads(struct virtnet_info *vi)
1973 if (!vi->guest_offloads)
1976 return virtnet_set_guest_offloads(vi, offloads);
1979 static int virtnet_restore_guest_offloads(struct virtnet_info *vi)
1981 u64 offloads = vi->guest_offloads;
1983 if (!vi->guest_offloads)
1986 return virtnet_set_guest_offloads(vi, offloads);
1989 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
1990 struct netlink_ext_ack *extack)
1992 unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr);
1993 struct virtnet_info *vi = netdev_priv(dev);
1994 struct bpf_prog *old_prog;
1995 u16 xdp_qp = 0, curr_qp;
1998 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)
1999 && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2000 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2001 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
2002 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) ||
2003 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))) {
2004 NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing LRO/CSUM, disable LRO/CSUM first");
2008 if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
2009 NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required");
2013 if (dev->mtu > max_sz) {
2014 NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP");
2015 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz);
2019 curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
2021 xdp_qp = nr_cpu_ids;
2023 /* XDP requires extra queues for XDP_TX */
2024 if (curr_qp + xdp_qp > vi->max_queue_pairs) {
2025 NL_SET_ERR_MSG_MOD(extack, "Too few free TX rings available");
2026 netdev_warn(dev, "request %i queues but max is %i\n",
2027 curr_qp + xdp_qp, vi->max_queue_pairs);
2032 prog = bpf_prog_add(prog, vi->max_queue_pairs - 1);
2034 return PTR_ERR(prog);
2037 /* Make sure NAPI is not using any XDP TX queues for RX. */
2038 if (netif_running(dev)) {
2039 for (i = 0; i < vi->max_queue_pairs; i++) {
2040 napi_disable(&vi->rq[i].napi);
2041 virtnet_napi_tx_disable(&vi->sq[i].napi);
2045 err = _virtnet_set_queues(vi, curr_qp + xdp_qp);
2048 netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
2049 vi->xdp_queue_pairs = xdp_qp;
2051 for (i = 0; i < vi->max_queue_pairs; i++) {
2052 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2053 rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2056 virtnet_clear_guest_offloads(vi);
2058 virtnet_restore_guest_offloads(vi);
2061 bpf_prog_put(old_prog);
2062 if (netif_running(dev)) {
2063 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2064 virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2072 if (netif_running(dev)) {
2073 for (i = 0; i < vi->max_queue_pairs; i++) {
2074 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2075 virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2080 bpf_prog_sub(prog, vi->max_queue_pairs - 1);
2084 static u32 virtnet_xdp_query(struct net_device *dev)
2086 struct virtnet_info *vi = netdev_priv(dev);
2087 const struct bpf_prog *xdp_prog;
2090 for (i = 0; i < vi->max_queue_pairs; i++) {
2091 xdp_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2093 return xdp_prog->aux->id;
2098 static int virtnet_xdp(struct net_device *dev, struct netdev_xdp *xdp)
2100 switch (xdp->command) {
2101 case XDP_SETUP_PROG:
2102 return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
2103 case XDP_QUERY_PROG:
2104 xdp->prog_id = virtnet_xdp_query(dev);
2105 xdp->prog_attached = !!xdp->prog_id;
2112 static const struct net_device_ops virtnet_netdev = {
2113 .ndo_open = virtnet_open,
2114 .ndo_stop = virtnet_close,
2115 .ndo_start_xmit = start_xmit,
2116 .ndo_validate_addr = eth_validate_addr,
2117 .ndo_set_mac_address = virtnet_set_mac_address,
2118 .ndo_set_rx_mode = virtnet_set_rx_mode,
2119 .ndo_get_stats64 = virtnet_stats,
2120 .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
2121 .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
2122 #ifdef CONFIG_NET_POLL_CONTROLLER
2123 .ndo_poll_controller = virtnet_netpoll,
2125 .ndo_xdp = virtnet_xdp,
2126 .ndo_features_check = passthru_features_check,
2129 static void virtnet_config_changed_work(struct work_struct *work)
2131 struct virtnet_info *vi =
2132 container_of(work, struct virtnet_info, config_work);
2135 if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
2136 struct virtio_net_config, status, &v) < 0)
2139 if (v & VIRTIO_NET_S_ANNOUNCE) {
2140 netdev_notify_peers(vi->dev);
2141 virtnet_ack_link_announce(vi);
2144 /* Ignore unknown (future) status bits */
2145 v &= VIRTIO_NET_S_LINK_UP;
2147 if (vi->status == v)
2152 if (vi->status & VIRTIO_NET_S_LINK_UP) {
2153 netif_carrier_on(vi->dev);
2154 netif_tx_wake_all_queues(vi->dev);
2156 netif_carrier_off(vi->dev);
2157 netif_tx_stop_all_queues(vi->dev);
2161 static void virtnet_config_changed(struct virtio_device *vdev)
2163 struct virtnet_info *vi = vdev->priv;
2165 schedule_work(&vi->config_work);
2168 static void virtnet_free_queues(struct virtnet_info *vi)
2172 for (i = 0; i < vi->max_queue_pairs; i++) {
2173 napi_hash_del(&vi->rq[i].napi);
2174 netif_napi_del(&vi->rq[i].napi);
2175 netif_napi_del(&vi->sq[i].napi);
2178 /* We called napi_hash_del() before netif_napi_del(),
2179 * we need to respect an RCU grace period before freeing vi->rq
2188 static void _free_receive_bufs(struct virtnet_info *vi)
2190 struct bpf_prog *old_prog;
2193 for (i = 0; i < vi->max_queue_pairs; i++) {
2194 while (vi->rq[i].pages)
2195 __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
2197 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2198 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
2200 bpf_prog_put(old_prog);
2204 static void free_receive_bufs(struct virtnet_info *vi)
2207 _free_receive_bufs(vi);
2211 static void free_receive_page_frags(struct virtnet_info *vi)
2214 for (i = 0; i < vi->max_queue_pairs; i++)
2215 if (vi->rq[i].alloc_frag.page)
2216 put_page(vi->rq[i].alloc_frag.page);
2219 static void free_unused_bufs(struct virtnet_info *vi)
2224 for (i = 0; i < vi->max_queue_pairs; i++) {
2225 struct virtqueue *vq = vi->sq[i].vq;
2226 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2227 if (!is_xdp_raw_buffer_queue(vi, i))
2230 put_page(virt_to_head_page(buf));
2234 for (i = 0; i < vi->max_queue_pairs; i++) {
2235 struct virtqueue *vq = vi->rq[i].vq;
2237 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2238 if (vi->mergeable_rx_bufs) {
2239 put_page(virt_to_head_page(buf));
2240 } else if (vi->big_packets) {
2241 give_pages(&vi->rq[i], buf);
2243 put_page(virt_to_head_page(buf));
2249 static void virtnet_del_vqs(struct virtnet_info *vi)
2251 struct virtio_device *vdev = vi->vdev;
2253 virtnet_clean_affinity(vi, -1);
2255 vdev->config->del_vqs(vdev);
2257 virtnet_free_queues(vi);
2260 /* How large should a single buffer be so a queue full of these can fit at
2261 * least one full packet?
2262 * Logic below assumes the mergeable buffer header is used.
2264 static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq)
2266 const unsigned int hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2267 unsigned int rq_size = virtqueue_get_vring_size(vq);
2268 unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu;
2269 unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len;
2270 unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size);
2272 return max(max(min_buf_len, hdr_len) - hdr_len,
2273 (unsigned int)GOOD_PACKET_LEN);
2276 static int virtnet_find_vqs(struct virtnet_info *vi)
2278 vq_callback_t **callbacks;
2279 struct virtqueue **vqs;
2285 /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
2286 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
2287 * possible control vq.
2289 total_vqs = vi->max_queue_pairs * 2 +
2290 virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
2292 /* Allocate space for find_vqs parameters */
2293 vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
2296 callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
2299 names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
2302 if (!vi->big_packets || vi->mergeable_rx_bufs) {
2303 ctx = kzalloc(total_vqs * sizeof(*ctx), GFP_KERNEL);
2310 /* Parameters for control virtqueue, if any */
2312 callbacks[total_vqs - 1] = NULL;
2313 names[total_vqs - 1] = "control";
2316 /* Allocate/initialize parameters for send/receive virtqueues */
2317 for (i = 0; i < vi->max_queue_pairs; i++) {
2318 callbacks[rxq2vq(i)] = skb_recv_done;
2319 callbacks[txq2vq(i)] = skb_xmit_done;
2320 sprintf(vi->rq[i].name, "input.%d", i);
2321 sprintf(vi->sq[i].name, "output.%d", i);
2322 names[rxq2vq(i)] = vi->rq[i].name;
2323 names[txq2vq(i)] = vi->sq[i].name;
2325 ctx[rxq2vq(i)] = true;
2328 ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
2334 vi->cvq = vqs[total_vqs - 1];
2335 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
2336 vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
2339 for (i = 0; i < vi->max_queue_pairs; i++) {
2340 vi->rq[i].vq = vqs[rxq2vq(i)];
2341 vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq);
2342 vi->sq[i].vq = vqs[txq2vq(i)];
2364 static int virtnet_alloc_queues(struct virtnet_info *vi)
2368 vi->ctrl = kzalloc(sizeof(*vi->ctrl), GFP_KERNEL);
2371 vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
2374 vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
2378 INIT_DELAYED_WORK(&vi->refill, refill_work);
2379 for (i = 0; i < vi->max_queue_pairs; i++) {
2380 vi->rq[i].pages = NULL;
2381 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
2383 netif_tx_napi_add(vi->dev, &vi->sq[i].napi, virtnet_poll_tx,
2384 napi_tx ? napi_weight : 0);
2386 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
2387 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
2388 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
2401 static int init_vqs(struct virtnet_info *vi)
2405 /* Allocate send & receive queues */
2406 ret = virtnet_alloc_queues(vi);
2410 ret = virtnet_find_vqs(vi);
2415 virtnet_set_affinity(vi);
2421 virtnet_free_queues(vi);
2427 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
2430 struct virtnet_info *vi = netdev_priv(queue->dev);
2431 unsigned int queue_index = get_netdev_rx_queue_index(queue);
2432 struct ewma_pkt_len *avg;
2434 BUG_ON(queue_index >= vi->max_queue_pairs);
2435 avg = &vi->rq[queue_index].mrg_avg_pkt_len;
2436 return sprintf(buf, "%u\n",
2437 get_mergeable_buf_len(&vi->rq[queue_index], avg));
2440 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
2441 __ATTR_RO(mergeable_rx_buffer_size);
2443 static struct attribute *virtio_net_mrg_rx_attrs[] = {
2444 &mergeable_rx_buffer_size_attribute.attr,
2448 static const struct attribute_group virtio_net_mrg_rx_group = {
2449 .name = "virtio_net",
2450 .attrs = virtio_net_mrg_rx_attrs
2454 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
2456 const char *fname, const char *dname)
2458 if (!virtio_has_feature(vdev, fbit))
2461 dev_err(&vdev->dev, "device advertises feature %s but not %s",
2467 #define VIRTNET_FAIL_ON(vdev, fbit, dbit) \
2468 virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
2470 static bool virtnet_validate_features(struct virtio_device *vdev)
2472 if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
2473 (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
2474 "VIRTIO_NET_F_CTRL_VQ") ||
2475 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
2476 "VIRTIO_NET_F_CTRL_VQ") ||
2477 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
2478 "VIRTIO_NET_F_CTRL_VQ") ||
2479 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
2480 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
2481 "VIRTIO_NET_F_CTRL_VQ"))) {
2488 #define MIN_MTU ETH_MIN_MTU
2489 #define MAX_MTU ETH_MAX_MTU
2491 static int virtnet_validate(struct virtio_device *vdev)
2493 if (!vdev->config->get) {
2494 dev_err(&vdev->dev, "%s failure: config access disabled\n",
2499 if (!virtnet_validate_features(vdev))
2502 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2503 int mtu = virtio_cread16(vdev,
2504 offsetof(struct virtio_net_config,
2507 __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
2513 static int virtnet_probe(struct virtio_device *vdev)
2516 struct net_device *dev;
2517 struct virtnet_info *vi;
2518 u16 max_queue_pairs;
2521 /* Find if host supports multiqueue virtio_net device */
2522 err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
2523 struct virtio_net_config,
2524 max_virtqueue_pairs, &max_queue_pairs);
2526 /* We need at least 2 queue's */
2527 if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
2528 max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
2529 !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2530 max_queue_pairs = 1;
2532 /* Allocate ourselves a network device with room for our info */
2533 dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
2537 /* Set up network device as normal. */
2538 dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
2539 dev->netdev_ops = &virtnet_netdev;
2540 dev->features = NETIF_F_HIGHDMA;
2542 dev->ethtool_ops = &virtnet_ethtool_ops;
2543 SET_NETDEV_DEV(dev, &vdev->dev);
2545 /* Do we support "hardware" checksums? */
2546 if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
2547 /* This opens up the world of extra features. */
2548 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2550 dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2552 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
2553 dev->hw_features |= NETIF_F_TSO
2554 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
2556 /* Individual feature bits: what can host handle? */
2557 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
2558 dev->hw_features |= NETIF_F_TSO;
2559 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
2560 dev->hw_features |= NETIF_F_TSO6;
2561 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
2562 dev->hw_features |= NETIF_F_TSO_ECN;
2564 dev->features |= NETIF_F_GSO_ROBUST;
2567 dev->features |= dev->hw_features & NETIF_F_ALL_TSO;
2568 /* (!csum && gso) case will be fixed by register_netdev() */
2570 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
2571 dev->features |= NETIF_F_RXCSUM;
2573 dev->vlan_features = dev->features;
2575 /* MTU range: 68 - 65535 */
2576 dev->min_mtu = MIN_MTU;
2577 dev->max_mtu = MAX_MTU;
2579 /* Configuration may specify what MAC to use. Otherwise random. */
2580 if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
2581 virtio_cread_bytes(vdev,
2582 offsetof(struct virtio_net_config, mac),
2583 dev->dev_addr, dev->addr_len);
2585 eth_hw_addr_random(dev);
2587 /* Set up our device-specific information */
2588 vi = netdev_priv(dev);
2592 vi->stats = alloc_percpu(struct virtnet_stats);
2594 if (vi->stats == NULL)
2597 for_each_possible_cpu(i) {
2598 struct virtnet_stats *virtnet_stats;
2599 virtnet_stats = per_cpu_ptr(vi->stats, i);
2600 u64_stats_init(&virtnet_stats->tx_syncp);
2601 u64_stats_init(&virtnet_stats->rx_syncp);
2604 INIT_WORK(&vi->config_work, virtnet_config_changed_work);
2606 /* If we can receive ANY GSO packets, we must allocate large ones. */
2607 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2608 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2609 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
2610 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
2611 vi->big_packets = true;
2613 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
2614 vi->mergeable_rx_bufs = true;
2616 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
2617 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2618 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2620 vi->hdr_len = sizeof(struct virtio_net_hdr);
2622 if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
2623 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2624 vi->any_header_sg = true;
2626 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2629 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2630 mtu = virtio_cread16(vdev,
2631 offsetof(struct virtio_net_config,
2633 if (mtu < dev->min_mtu) {
2634 /* Should never trigger: MTU was previously validated
2635 * in virtnet_validate.
2637 dev_err(&vdev->dev, "device MTU appears to have changed "
2638 "it is now %d < %d", mtu, dev->min_mtu);
2645 /* TODO: size buffers correctly in this case. */
2646 if (dev->mtu > ETH_DATA_LEN)
2647 vi->big_packets = true;
2650 if (vi->any_header_sg)
2651 dev->needed_headroom = vi->hdr_len;
2653 /* Enable multiqueue by default */
2654 if (num_online_cpus() >= max_queue_pairs)
2655 vi->curr_queue_pairs = max_queue_pairs;
2657 vi->curr_queue_pairs = num_online_cpus();
2658 vi->max_queue_pairs = max_queue_pairs;
2660 /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
2666 if (vi->mergeable_rx_bufs)
2667 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
2669 netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
2670 netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
2672 virtnet_init_settings(dev);
2674 err = register_netdev(dev);
2676 pr_debug("virtio_net: registering device failed\n");
2680 virtio_device_ready(vdev);
2682 err = virtnet_cpu_notif_add(vi);
2684 pr_debug("virtio_net: registering cpu notifier failed\n");
2685 goto free_unregister_netdev;
2688 virtnet_set_queues(vi, vi->curr_queue_pairs);
2690 /* Assume link up if device can't report link status,
2691 otherwise get link status from config. */
2692 netif_carrier_off(dev);
2693 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
2694 schedule_work(&vi->config_work);
2696 vi->status = VIRTIO_NET_S_LINK_UP;
2697 netif_carrier_on(dev);
2700 for (i = 0; i < ARRAY_SIZE(guest_offloads); i++)
2701 if (virtio_has_feature(vi->vdev, guest_offloads[i]))
2702 set_bit(guest_offloads[i], &vi->guest_offloads);
2704 pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
2705 dev->name, max_queue_pairs);
2709 free_unregister_netdev:
2710 vi->vdev->config->reset(vdev);
2712 unregister_netdev(dev);
2714 cancel_delayed_work_sync(&vi->refill);
2715 free_receive_page_frags(vi);
2716 virtnet_del_vqs(vi);
2718 free_percpu(vi->stats);
2724 static void remove_vq_common(struct virtnet_info *vi)
2726 vi->vdev->config->reset(vi->vdev);
2728 /* Free unused buffers in both send and recv, if any. */
2729 free_unused_bufs(vi);
2731 free_receive_bufs(vi);
2733 free_receive_page_frags(vi);
2735 virtnet_del_vqs(vi);
2738 static void virtnet_remove(struct virtio_device *vdev)
2740 struct virtnet_info *vi = vdev->priv;
2742 virtnet_cpu_notif_remove(vi);
2744 /* Make sure no work handler is accessing the device. */
2745 flush_work(&vi->config_work);
2747 unregister_netdev(vi->dev);
2749 remove_vq_common(vi);
2751 free_percpu(vi->stats);
2752 free_netdev(vi->dev);
2755 static __maybe_unused int virtnet_freeze(struct virtio_device *vdev)
2757 struct virtnet_info *vi = vdev->priv;
2759 virtnet_cpu_notif_remove(vi);
2760 virtnet_freeze_down(vdev);
2761 remove_vq_common(vi);
2766 static __maybe_unused int virtnet_restore(struct virtio_device *vdev)
2768 struct virtnet_info *vi = vdev->priv;
2771 err = virtnet_restore_up(vdev);
2774 virtnet_set_queues(vi, vi->curr_queue_pairs);
2776 err = virtnet_cpu_notif_add(vi);
2778 virtnet_freeze_down(vdev);
2779 remove_vq_common(vi);
2786 static struct virtio_device_id id_table[] = {
2787 { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
2791 #define VIRTNET_FEATURES \
2792 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
2794 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
2795 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
2796 VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
2797 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
2798 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
2799 VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
2800 VIRTIO_NET_F_CTRL_MAC_ADDR, \
2801 VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS
2803 static unsigned int features[] = {
2807 static unsigned int features_legacy[] = {
2810 VIRTIO_F_ANY_LAYOUT,
2813 static struct virtio_driver virtio_net_driver = {
2814 .feature_table = features,
2815 .feature_table_size = ARRAY_SIZE(features),
2816 .feature_table_legacy = features_legacy,
2817 .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
2818 .driver.name = KBUILD_MODNAME,
2819 .driver.owner = THIS_MODULE,
2820 .id_table = id_table,
2821 .validate = virtnet_validate,
2822 .probe = virtnet_probe,
2823 .remove = virtnet_remove,
2824 .config_changed = virtnet_config_changed,
2825 #ifdef CONFIG_PM_SLEEP
2826 .freeze = virtnet_freeze,
2827 .restore = virtnet_restore,
2831 static __init int virtio_net_driver_init(void)
2835 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
2837 virtnet_cpu_down_prep);
2840 virtionet_online = ret;
2841 ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
2842 NULL, virtnet_cpu_dead);
2846 ret = register_virtio_driver(&virtio_net_driver);
2851 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
2853 cpuhp_remove_multi_state(virtionet_online);
2857 module_init(virtio_net_driver_init);
2859 static __exit void virtio_net_driver_exit(void)
2861 unregister_virtio_driver(&virtio_net_driver);
2862 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
2863 cpuhp_remove_multi_state(virtionet_online);
2865 module_exit(virtio_net_driver_exit);
2867 MODULE_DEVICE_TABLE(virtio, id_table);
2868 MODULE_DESCRIPTION("Virtio network driver");
2869 MODULE_LICENSE("GPL");