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 struct page *page = alloc_page(GFP_ATOMIC);
460 memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
468 buf = virtqueue_get_buf(rq->vq, &buflen);
472 p = virt_to_head_page(buf);
473 off = buf - page_address(p);
475 /* guard against a misconfigured or uncooperative backend that
476 * is sending packet larger than the MTU.
478 if ((page_off + buflen) > PAGE_SIZE) {
483 memcpy(page_address(page) + page_off,
484 page_address(p) + off, buflen);
489 /* Headroom does not contribute to packet length */
490 *len = page_off - VIRTIO_XDP_HEADROOM;
493 __free_pages(page, 0);
497 static struct sk_buff *receive_small(struct net_device *dev,
498 struct virtnet_info *vi,
499 struct receive_queue *rq,
500 void *buf, void *ctx,
504 struct bpf_prog *xdp_prog;
505 unsigned int xdp_headroom = (unsigned long)ctx;
506 unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom;
507 unsigned int headroom = vi->hdr_len + header_offset;
508 unsigned int buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
509 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
510 struct page *page = virt_to_head_page(buf);
511 unsigned int delta = 0;
512 struct page *xdp_page;
516 xdp_prog = rcu_dereference(rq->xdp_prog);
518 struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset;
523 if (unlikely(hdr->hdr.gso_type))
526 if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) {
527 int offset = buf - page_address(page) + header_offset;
528 unsigned int tlen = len + vi->hdr_len;
531 xdp_headroom = virtnet_get_headroom(vi);
532 header_offset = VIRTNET_RX_PAD + xdp_headroom;
533 headroom = vi->hdr_len + header_offset;
534 buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
535 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
536 xdp_page = xdp_linearize_page(rq, &num_buf, page,
537 offset, header_offset,
542 buf = page_address(xdp_page);
547 xdp.data_hard_start = buf + VIRTNET_RX_PAD + vi->hdr_len;
548 xdp.data = xdp.data_hard_start + xdp_headroom;
549 xdp.data_end = xdp.data + len;
550 orig_data = xdp.data;
551 act = bpf_prog_run_xdp(xdp_prog, &xdp);
555 /* Recalculate length in case bpf program changed it */
556 delta = orig_data - xdp.data;
559 if (unlikely(!virtnet_xdp_xmit(vi, rq, &xdp)))
560 trace_xdp_exception(vi->dev, xdp_prog, act);
564 bpf_warn_invalid_xdp_action(act);
566 trace_xdp_exception(vi->dev, xdp_prog, act);
573 skb = build_skb(buf, buflen);
578 skb_reserve(skb, headroom - delta);
579 skb_put(skb, len + delta);
581 buf += header_offset;
582 memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
583 } /* keep zeroed vnet hdr since packet was changed by bpf */
590 dev->stats.rx_dropped++;
596 static struct sk_buff *receive_big(struct net_device *dev,
597 struct virtnet_info *vi,
598 struct receive_queue *rq,
602 struct page *page = buf;
603 struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len,
612 dev->stats.rx_dropped++;
613 give_pages(rq, page);
617 static struct sk_buff *receive_mergeable(struct net_device *dev,
618 struct virtnet_info *vi,
619 struct receive_queue *rq,
624 struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
625 u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
626 struct page *page = virt_to_head_page(buf);
627 int offset = buf - page_address(page);
628 struct sk_buff *head_skb, *curr_skb;
629 struct bpf_prog *xdp_prog;
630 unsigned int truesize;
631 unsigned int headroom = mergeable_ctx_to_headroom(ctx);
636 xdp_prog = rcu_dereference(rq->xdp_prog);
638 struct page *xdp_page;
643 /* Transient failure which in theory could occur if
644 * in-flight packets from before XDP was enabled reach
645 * the receive path after XDP is loaded.
647 if (unlikely(hdr->hdr.gso_type))
650 /* This happens when rx buffer size is underestimated */
651 if (unlikely(num_buf > 1 ||
652 headroom < virtnet_get_headroom(vi))) {
653 /* linearize data for XDP */
654 xdp_page = xdp_linearize_page(rq, &num_buf,
660 offset = VIRTIO_XDP_HEADROOM;
665 /* Allow consuming headroom but reserve enough space to push
666 * the descriptor on if we get an XDP_TX return code.
668 data = page_address(xdp_page) + offset;
669 xdp.data_hard_start = data - VIRTIO_XDP_HEADROOM + vi->hdr_len;
670 xdp.data = data + vi->hdr_len;
671 xdp.data_end = xdp.data + (len - vi->hdr_len);
672 act = bpf_prog_run_xdp(xdp_prog, &xdp);
676 /* recalculate offset to account for any header
677 * adjustments. Note other cases do not build an
678 * skb and avoid using offset
681 page_address(xdp_page) - vi->hdr_len;
683 /* We can only create skb based on xdp_page. */
684 if (unlikely(xdp_page != page)) {
687 head_skb = page_to_skb(vi, rq, xdp_page,
690 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, len);
695 if (unlikely(!virtnet_xdp_xmit(vi, rq, &xdp)))
696 trace_xdp_exception(vi->dev, xdp_prog, act);
697 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, len);
698 if (unlikely(xdp_page != page))
703 bpf_warn_invalid_xdp_action(act);
705 trace_xdp_exception(vi->dev, xdp_prog, act);
707 if (unlikely(xdp_page != page))
708 __free_pages(xdp_page, 0);
709 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, len);
715 truesize = mergeable_ctx_to_truesize(ctx);
716 if (unlikely(len > truesize)) {
717 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
718 dev->name, len, (unsigned long)ctx);
719 dev->stats.rx_length_errors++;
723 head_skb = page_to_skb(vi, rq, page, offset, len, truesize, !xdp_prog);
726 if (unlikely(!curr_skb))
731 buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx);
732 if (unlikely(!buf)) {
733 pr_debug("%s: rx error: %d buffers out of %d missing\n",
735 virtio16_to_cpu(vi->vdev,
737 dev->stats.rx_length_errors++;
741 page = virt_to_head_page(buf);
743 truesize = mergeable_ctx_to_truesize(ctx);
744 if (unlikely(len > truesize)) {
745 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
746 dev->name, len, (unsigned long)ctx);
747 dev->stats.rx_length_errors++;
751 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
752 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
753 struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
757 if (curr_skb == head_skb)
758 skb_shinfo(curr_skb)->frag_list = nskb;
760 curr_skb->next = nskb;
762 head_skb->truesize += nskb->truesize;
765 if (curr_skb != head_skb) {
766 head_skb->data_len += len;
767 head_skb->len += len;
768 head_skb->truesize += truesize;
770 offset = buf - page_address(page);
771 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
773 skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
776 skb_add_rx_frag(curr_skb, num_skb_frags, page,
777 offset, len, truesize);
781 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
788 while (num_buf-- > 1) {
789 buf = virtqueue_get_buf(rq->vq, &len);
790 if (unlikely(!buf)) {
791 pr_debug("%s: rx error: %d buffers missing\n",
793 dev->stats.rx_length_errors++;
796 page = virt_to_head_page(buf);
800 dev->stats.rx_dropped++;
801 dev_kfree_skb(head_skb);
806 static int receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
807 void *buf, unsigned int len, void **ctx)
809 struct net_device *dev = vi->dev;
811 struct virtio_net_hdr_mrg_rxbuf *hdr;
814 if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
815 pr_debug("%s: short packet %i\n", dev->name, len);
816 dev->stats.rx_length_errors++;
817 if (vi->mergeable_rx_bufs) {
818 put_page(virt_to_head_page(buf));
819 } else if (vi->big_packets) {
822 put_page(virt_to_head_page(buf));
827 if (vi->mergeable_rx_bufs)
828 skb = receive_mergeable(dev, vi, rq, buf, ctx, len);
829 else if (vi->big_packets)
830 skb = receive_big(dev, vi, rq, buf, len);
832 skb = receive_small(dev, vi, rq, buf, ctx, len);
837 hdr = skb_vnet_hdr(skb);
841 if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
842 skb->ip_summed = CHECKSUM_UNNECESSARY;
844 if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
845 virtio_is_little_endian(vi->vdev))) {
846 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
847 dev->name, hdr->hdr.gso_type,
852 skb->protocol = eth_type_trans(skb, dev);
853 pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
854 ntohs(skb->protocol), skb->len, skb->pkt_type);
856 napi_gro_receive(&rq->napi, skb);
860 dev->stats.rx_frame_errors++;
865 /* Unlike mergeable buffers, all buffers are allocated to the
866 * same size, except for the headroom. For this reason we do
867 * not need to use mergeable_len_to_ctx here - it is enough
868 * to store the headroom as the context ignoring the truesize.
870 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
873 struct page_frag *alloc_frag = &rq->alloc_frag;
875 unsigned int xdp_headroom = virtnet_get_headroom(vi);
876 void *ctx = (void *)(unsigned long)xdp_headroom;
877 int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
880 len = SKB_DATA_ALIGN(len) +
881 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
882 if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
885 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
886 get_page(alloc_frag->page);
887 alloc_frag->offset += len;
888 sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom,
889 vi->hdr_len + GOOD_PACKET_LEN);
890 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
892 put_page(virt_to_head_page(buf));
896 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
899 struct page *first, *list = NULL;
903 sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
905 /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
906 for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
907 first = get_a_page(rq, gfp);
910 give_pages(rq, list);
913 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
915 /* chain new page in list head to match sg */
916 first->private = (unsigned long)list;
920 first = get_a_page(rq, gfp);
922 give_pages(rq, list);
925 p = page_address(first);
927 /* rq->sg[0], rq->sg[1] share the same page */
928 /* a separated rq->sg[0] for header - required in case !any_header_sg */
929 sg_set_buf(&rq->sg[0], p, vi->hdr_len);
931 /* rq->sg[1] for data packet, from offset */
932 offset = sizeof(struct padded_vnet_hdr);
933 sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
935 /* chain first in list head */
936 first->private = (unsigned long)list;
937 err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
940 give_pages(rq, first);
945 static unsigned int get_mergeable_buf_len(struct receive_queue *rq,
946 struct ewma_pkt_len *avg_pkt_len)
948 const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
951 len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
952 rq->min_buf_len, PAGE_SIZE - hdr_len);
953 return ALIGN(len, L1_CACHE_BYTES);
956 static int add_recvbuf_mergeable(struct virtnet_info *vi,
957 struct receive_queue *rq, gfp_t gfp)
959 struct page_frag *alloc_frag = &rq->alloc_frag;
960 unsigned int headroom = virtnet_get_headroom(vi);
964 unsigned int len, hole;
966 len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len);
967 if (unlikely(!skb_page_frag_refill(len + headroom, alloc_frag, gfp)))
970 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
971 buf += headroom; /* advance address leaving hole at front of pkt */
972 get_page(alloc_frag->page);
973 alloc_frag->offset += len + headroom;
974 hole = alloc_frag->size - alloc_frag->offset;
975 if (hole < len + headroom) {
976 /* To avoid internal fragmentation, if there is very likely not
977 * enough space for another buffer, add the remaining space to
978 * the current buffer.
981 alloc_frag->offset += hole;
984 sg_init_one(rq->sg, buf, len);
985 ctx = mergeable_len_to_ctx(len, headroom);
986 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
988 put_page(virt_to_head_page(buf));
994 * Returns false if we couldn't fill entirely (OOM).
996 * Normally run in the receive path, but can also be run from ndo_open
997 * before we're receiving packets, or from refill_work which is
998 * careful to disable receiving (using napi_disable).
1000 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
1008 if (vi->mergeable_rx_bufs)
1009 err = add_recvbuf_mergeable(vi, rq, gfp);
1010 else if (vi->big_packets)
1011 err = add_recvbuf_big(vi, rq, gfp);
1013 err = add_recvbuf_small(vi, rq, gfp);
1015 oom = err == -ENOMEM;
1018 } while (rq->vq->num_free);
1019 virtqueue_kick(rq->vq);
1023 static void skb_recv_done(struct virtqueue *rvq)
1025 struct virtnet_info *vi = rvq->vdev->priv;
1026 struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
1028 virtqueue_napi_schedule(&rq->napi, rvq);
1031 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
1035 /* If all buffers were filled by other side before we napi_enabled, we
1036 * won't get another interrupt, so process any outstanding packets now.
1037 * Call local_bh_enable after to trigger softIRQ processing.
1040 virtqueue_napi_schedule(napi, vq);
1044 static void virtnet_napi_tx_enable(struct virtnet_info *vi,
1045 struct virtqueue *vq,
1046 struct napi_struct *napi)
1051 /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
1052 * enable the feature if this is likely affine with the transmit path.
1054 if (!vi->affinity_hint_set) {
1059 return virtnet_napi_enable(vq, napi);
1062 static void virtnet_napi_tx_disable(struct napi_struct *napi)
1068 static void refill_work(struct work_struct *work)
1070 struct virtnet_info *vi =
1071 container_of(work, struct virtnet_info, refill.work);
1075 for (i = 0; i < vi->curr_queue_pairs; i++) {
1076 struct receive_queue *rq = &vi->rq[i];
1078 napi_disable(&rq->napi);
1079 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
1080 virtnet_napi_enable(rq->vq, &rq->napi);
1082 /* In theory, this can happen: if we don't get any buffers in
1083 * we will *never* try to fill again.
1086 schedule_delayed_work(&vi->refill, HZ/2);
1090 static int virtnet_receive(struct receive_queue *rq, int budget)
1092 struct virtnet_info *vi = rq->vq->vdev->priv;
1093 unsigned int len, received = 0, bytes = 0;
1095 struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
1097 if (!vi->big_packets || vi->mergeable_rx_bufs) {
1100 while (received < budget &&
1101 (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) {
1102 bytes += receive_buf(vi, rq, buf, len, ctx);
1106 while (received < budget &&
1107 (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
1108 bytes += receive_buf(vi, rq, buf, len, NULL);
1113 if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) {
1114 if (!try_fill_recv(vi, rq, GFP_ATOMIC))
1115 schedule_delayed_work(&vi->refill, 0);
1118 u64_stats_update_begin(&stats->rx_syncp);
1119 stats->rx_bytes += bytes;
1120 stats->rx_packets += received;
1121 u64_stats_update_end(&stats->rx_syncp);
1126 static void free_old_xmit_skbs(struct send_queue *sq)
1128 struct sk_buff *skb;
1130 struct virtnet_info *vi = sq->vq->vdev->priv;
1131 struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
1132 unsigned int packets = 0;
1133 unsigned int bytes = 0;
1135 while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
1136 pr_debug("Sent skb %p\n", skb);
1141 dev_consume_skb_any(skb);
1144 /* Avoid overhead when no packets have been processed
1145 * happens when called speculatively from start_xmit.
1150 u64_stats_update_begin(&stats->tx_syncp);
1151 stats->tx_bytes += bytes;
1152 stats->tx_packets += packets;
1153 u64_stats_update_end(&stats->tx_syncp);
1156 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
1158 if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
1160 else if (q < vi->curr_queue_pairs)
1166 static void virtnet_poll_cleantx(struct receive_queue *rq)
1168 struct virtnet_info *vi = rq->vq->vdev->priv;
1169 unsigned int index = vq2rxq(rq->vq);
1170 struct send_queue *sq = &vi->sq[index];
1171 struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index);
1173 if (!sq->napi.weight || is_xdp_raw_buffer_queue(vi, index))
1176 if (__netif_tx_trylock(txq)) {
1177 free_old_xmit_skbs(sq);
1178 __netif_tx_unlock(txq);
1181 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1182 netif_tx_wake_queue(txq);
1185 static int virtnet_poll(struct napi_struct *napi, int budget)
1187 struct receive_queue *rq =
1188 container_of(napi, struct receive_queue, napi);
1189 unsigned int received;
1191 virtnet_poll_cleantx(rq);
1193 received = virtnet_receive(rq, budget);
1195 /* Out of packets? */
1196 if (received < budget)
1197 virtqueue_napi_complete(napi, rq->vq, received);
1202 static int virtnet_open(struct net_device *dev)
1204 struct virtnet_info *vi = netdev_priv(dev);
1207 for (i = 0; i < vi->max_queue_pairs; i++) {
1208 if (i < vi->curr_queue_pairs)
1209 /* Make sure we have some buffers: if oom use wq. */
1210 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1211 schedule_delayed_work(&vi->refill, 0);
1212 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1213 virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi);
1219 static int virtnet_poll_tx(struct napi_struct *napi, int budget)
1221 struct send_queue *sq = container_of(napi, struct send_queue, napi);
1222 struct virtnet_info *vi = sq->vq->vdev->priv;
1223 unsigned int index = vq2txq(sq->vq);
1224 struct netdev_queue *txq;
1226 if (unlikely(is_xdp_raw_buffer_queue(vi, index))) {
1227 /* We don't need to enable cb for XDP */
1228 napi_complete_done(napi, 0);
1232 txq = netdev_get_tx_queue(vi->dev, index);
1233 __netif_tx_lock(txq, raw_smp_processor_id());
1234 free_old_xmit_skbs(sq);
1235 __netif_tx_unlock(txq);
1237 virtqueue_napi_complete(napi, sq->vq, 0);
1239 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1240 netif_tx_wake_queue(txq);
1245 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
1247 struct virtio_net_hdr_mrg_rxbuf *hdr;
1248 const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
1249 struct virtnet_info *vi = sq->vq->vdev->priv;
1251 unsigned hdr_len = vi->hdr_len;
1254 pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
1256 can_push = vi->any_header_sg &&
1257 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
1258 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
1259 /* Even if we can, don't push here yet as this would skew
1260 * csum_start offset below. */
1262 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1264 hdr = skb_vnet_hdr(skb);
1266 if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1267 virtio_is_little_endian(vi->vdev), false,
1271 if (vi->mergeable_rx_bufs)
1272 hdr->num_buffers = 0;
1274 sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1276 __skb_push(skb, hdr_len);
1277 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1278 if (unlikely(num_sg < 0))
1280 /* Pull header back to avoid skew in tx bytes calculations. */
1281 __skb_pull(skb, hdr_len);
1283 sg_set_buf(sq->sg, hdr, hdr_len);
1284 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
1285 if (unlikely(num_sg < 0))
1289 return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
1292 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
1294 struct virtnet_info *vi = netdev_priv(dev);
1295 int qnum = skb_get_queue_mapping(skb);
1296 struct send_queue *sq = &vi->sq[qnum];
1298 struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1299 bool kick = !skb->xmit_more;
1300 bool use_napi = sq->napi.weight;
1302 /* Free up any pending old buffers before queueing new ones. */
1303 free_old_xmit_skbs(sq);
1305 if (use_napi && kick)
1306 virtqueue_enable_cb_delayed(sq->vq);
1308 /* timestamp packet in software */
1309 skb_tx_timestamp(skb);
1311 /* Try to transmit */
1312 err = xmit_skb(sq, skb);
1314 /* This should not happen! */
1315 if (unlikely(err)) {
1316 dev->stats.tx_fifo_errors++;
1317 if (net_ratelimit())
1319 "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
1320 dev->stats.tx_dropped++;
1321 dev_kfree_skb_any(skb);
1322 return NETDEV_TX_OK;
1325 /* Don't wait up for transmitted skbs to be freed. */
1331 /* If running out of space, stop queue to avoid getting packets that we
1332 * are then unable to transmit.
1333 * An alternative would be to force queuing layer to requeue the skb by
1334 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1335 * returned in a normal path of operation: it means that driver is not
1336 * maintaining the TX queue stop/start state properly, and causes
1337 * the stack to do a non-trivial amount of useless work.
1338 * Since most packets only take 1 or 2 ring slots, stopping the queue
1339 * early means 16 slots are typically wasted.
1341 if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
1342 netif_stop_subqueue(dev, qnum);
1344 unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
1345 /* More just got used, free them then recheck. */
1346 free_old_xmit_skbs(sq);
1347 if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
1348 netif_start_subqueue(dev, qnum);
1349 virtqueue_disable_cb(sq->vq);
1354 if (kick || netif_xmit_stopped(txq))
1355 virtqueue_kick(sq->vq);
1357 return NETDEV_TX_OK;
1361 * Send command via the control virtqueue and check status. Commands
1362 * supported by the hypervisor, as indicated by feature bits, should
1363 * never fail unless improperly formatted.
1365 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
1366 struct scatterlist *out)
1368 struct scatterlist *sgs[4], hdr, stat;
1369 unsigned out_num = 0, tmp;
1371 /* Caller should know better */
1372 BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
1374 vi->ctrl->status = ~0;
1375 vi->ctrl->hdr.class = class;
1376 vi->ctrl->hdr.cmd = cmd;
1378 sg_init_one(&hdr, &vi->ctrl->hdr, sizeof(vi->ctrl->hdr));
1379 sgs[out_num++] = &hdr;
1382 sgs[out_num++] = out;
1384 /* Add return status. */
1385 sg_init_one(&stat, &vi->ctrl->status, sizeof(vi->ctrl->status));
1386 sgs[out_num] = &stat;
1388 BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1389 virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1391 if (unlikely(!virtqueue_kick(vi->cvq)))
1392 return vi->ctrl->status == VIRTIO_NET_OK;
1394 /* Spin for a response, the kick causes an ioport write, trapping
1395 * into the hypervisor, so the request should be handled immediately.
1397 while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1398 !virtqueue_is_broken(vi->cvq))
1401 return vi->ctrl->status == VIRTIO_NET_OK;
1404 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1406 struct virtnet_info *vi = netdev_priv(dev);
1407 struct virtio_device *vdev = vi->vdev;
1409 struct sockaddr *addr;
1410 struct scatterlist sg;
1412 addr = kmemdup(p, sizeof(*addr), GFP_KERNEL);
1416 ret = eth_prepare_mac_addr_change(dev, addr);
1420 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1421 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1422 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1423 VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1424 dev_warn(&vdev->dev,
1425 "Failed to set mac address by vq command.\n");
1429 } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1430 !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1433 /* Naturally, this has an atomicity problem. */
1434 for (i = 0; i < dev->addr_len; i++)
1435 virtio_cwrite8(vdev,
1436 offsetof(struct virtio_net_config, mac) +
1437 i, addr->sa_data[i]);
1440 eth_commit_mac_addr_change(dev, p);
1448 static void virtnet_stats(struct net_device *dev,
1449 struct rtnl_link_stats64 *tot)
1451 struct virtnet_info *vi = netdev_priv(dev);
1455 for_each_possible_cpu(cpu) {
1456 struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
1457 u64 tpackets, tbytes, rpackets, rbytes;
1460 start = u64_stats_fetch_begin_irq(&stats->tx_syncp);
1461 tpackets = stats->tx_packets;
1462 tbytes = stats->tx_bytes;
1463 } while (u64_stats_fetch_retry_irq(&stats->tx_syncp, start));
1466 start = u64_stats_fetch_begin_irq(&stats->rx_syncp);
1467 rpackets = stats->rx_packets;
1468 rbytes = stats->rx_bytes;
1469 } while (u64_stats_fetch_retry_irq(&stats->rx_syncp, start));
1471 tot->rx_packets += rpackets;
1472 tot->tx_packets += tpackets;
1473 tot->rx_bytes += rbytes;
1474 tot->tx_bytes += tbytes;
1477 tot->tx_dropped = dev->stats.tx_dropped;
1478 tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1479 tot->rx_dropped = dev->stats.rx_dropped;
1480 tot->rx_length_errors = dev->stats.rx_length_errors;
1481 tot->rx_frame_errors = dev->stats.rx_frame_errors;
1484 #ifdef CONFIG_NET_POLL_CONTROLLER
1485 static void virtnet_netpoll(struct net_device *dev)
1487 struct virtnet_info *vi = netdev_priv(dev);
1490 for (i = 0; i < vi->curr_queue_pairs; i++)
1491 napi_schedule(&vi->rq[i].napi);
1495 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1498 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1499 VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1500 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1504 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1506 struct scatterlist sg;
1507 struct net_device *dev = vi->dev;
1509 if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1512 vi->ctrl->mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1513 sg_init_one(&sg, &vi->ctrl->mq, sizeof(vi->ctrl->mq));
1515 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1516 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1517 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1521 vi->curr_queue_pairs = queue_pairs;
1522 /* virtnet_open() will refill when device is going to up. */
1523 if (dev->flags & IFF_UP)
1524 schedule_delayed_work(&vi->refill, 0);
1530 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1535 err = _virtnet_set_queues(vi, queue_pairs);
1540 static int virtnet_close(struct net_device *dev)
1542 struct virtnet_info *vi = netdev_priv(dev);
1545 /* Make sure refill_work doesn't re-enable napi! */
1546 cancel_delayed_work_sync(&vi->refill);
1548 for (i = 0; i < vi->max_queue_pairs; i++) {
1549 napi_disable(&vi->rq[i].napi);
1550 virtnet_napi_tx_disable(&vi->sq[i].napi);
1556 static void virtnet_set_rx_mode(struct net_device *dev)
1558 struct virtnet_info *vi = netdev_priv(dev);
1559 struct scatterlist sg[2];
1560 struct virtio_net_ctrl_mac *mac_data;
1561 struct netdev_hw_addr *ha;
1567 /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1568 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1571 vi->ctrl->promisc = ((dev->flags & IFF_PROMISC) != 0);
1572 vi->ctrl->allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1574 sg_init_one(sg, &vi->ctrl->promisc, sizeof(vi->ctrl->promisc));
1576 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1577 VIRTIO_NET_CTRL_RX_PROMISC, sg))
1578 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1579 vi->ctrl->promisc ? "en" : "dis");
1581 sg_init_one(sg, &vi->ctrl->allmulti, sizeof(vi->ctrl->allmulti));
1583 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1584 VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1585 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1586 vi->ctrl->allmulti ? "en" : "dis");
1588 uc_count = netdev_uc_count(dev);
1589 mc_count = netdev_mc_count(dev);
1590 /* MAC filter - use one buffer for both lists */
1591 buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1592 (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1597 sg_init_table(sg, 2);
1599 /* Store the unicast list and count in the front of the buffer */
1600 mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1602 netdev_for_each_uc_addr(ha, dev)
1603 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1605 sg_set_buf(&sg[0], mac_data,
1606 sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1608 /* multicast list and count fill the end */
1609 mac_data = (void *)&mac_data->macs[uc_count][0];
1611 mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1613 netdev_for_each_mc_addr(ha, dev)
1614 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1616 sg_set_buf(&sg[1], mac_data,
1617 sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1619 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1620 VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1621 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1626 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1627 __be16 proto, u16 vid)
1629 struct virtnet_info *vi = netdev_priv(dev);
1630 struct scatterlist sg;
1632 vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
1633 sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
1635 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1636 VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1637 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1641 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1642 __be16 proto, u16 vid)
1644 struct virtnet_info *vi = netdev_priv(dev);
1645 struct scatterlist sg;
1647 vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
1648 sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
1650 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1651 VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1652 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1656 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1660 if (vi->affinity_hint_set) {
1661 for (i = 0; i < vi->max_queue_pairs; i++) {
1662 virtqueue_set_affinity(vi->rq[i].vq, -1);
1663 virtqueue_set_affinity(vi->sq[i].vq, -1);
1666 vi->affinity_hint_set = false;
1670 static void virtnet_set_affinity(struct virtnet_info *vi)
1675 /* In multiqueue mode, when the number of cpu is equal to the number of
1676 * queue pairs, we let the queue pairs to be private to one cpu by
1677 * setting the affinity hint to eliminate the contention.
1679 if (vi->curr_queue_pairs == 1 ||
1680 vi->max_queue_pairs != num_online_cpus()) {
1681 virtnet_clean_affinity(vi, -1);
1686 for_each_online_cpu(cpu) {
1687 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1688 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1689 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
1693 vi->affinity_hint_set = true;
1696 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
1698 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1700 virtnet_set_affinity(vi);
1704 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
1706 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1708 virtnet_set_affinity(vi);
1712 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
1714 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1717 virtnet_clean_affinity(vi, cpu);
1721 static enum cpuhp_state virtionet_online;
1723 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
1727 ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
1730 ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1734 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1738 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
1740 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1741 cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1745 static void virtnet_get_ringparam(struct net_device *dev,
1746 struct ethtool_ringparam *ring)
1748 struct virtnet_info *vi = netdev_priv(dev);
1750 ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1751 ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1752 ring->rx_pending = ring->rx_max_pending;
1753 ring->tx_pending = ring->tx_max_pending;
1757 static void virtnet_get_drvinfo(struct net_device *dev,
1758 struct ethtool_drvinfo *info)
1760 struct virtnet_info *vi = netdev_priv(dev);
1761 struct virtio_device *vdev = vi->vdev;
1763 strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1764 strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1765 strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1769 /* TODO: Eliminate OOO packets during switching */
1770 static int virtnet_set_channels(struct net_device *dev,
1771 struct ethtool_channels *channels)
1773 struct virtnet_info *vi = netdev_priv(dev);
1774 u16 queue_pairs = channels->combined_count;
1777 /* We don't support separate rx/tx channels.
1778 * We don't allow setting 'other' channels.
1780 if (channels->rx_count || channels->tx_count || channels->other_count)
1783 if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
1786 /* For now we don't support modifying channels while XDP is loaded
1787 * also when XDP is loaded all RX queues have XDP programs so we only
1788 * need to check a single RX queue.
1790 if (vi->rq[0].xdp_prog)
1794 err = _virtnet_set_queues(vi, queue_pairs);
1799 virtnet_set_affinity(vi);
1802 netif_set_real_num_tx_queues(dev, queue_pairs);
1803 netif_set_real_num_rx_queues(dev, queue_pairs);
1808 static void virtnet_get_channels(struct net_device *dev,
1809 struct ethtool_channels *channels)
1811 struct virtnet_info *vi = netdev_priv(dev);
1813 channels->combined_count = vi->curr_queue_pairs;
1814 channels->max_combined = vi->max_queue_pairs;
1815 channels->max_other = 0;
1816 channels->rx_count = 0;
1817 channels->tx_count = 0;
1818 channels->other_count = 0;
1821 /* Check if the user is trying to change anything besides speed/duplex */
1823 virtnet_validate_ethtool_cmd(const struct ethtool_link_ksettings *cmd)
1825 struct ethtool_link_ksettings diff1 = *cmd;
1826 struct ethtool_link_ksettings diff2 = {};
1828 /* cmd is always set so we need to clear it, validate the port type
1829 * and also without autonegotiation we can ignore advertising
1831 diff1.base.speed = 0;
1832 diff2.base.port = PORT_OTHER;
1833 ethtool_link_ksettings_zero_link_mode(&diff1, advertising);
1834 diff1.base.duplex = 0;
1836 diff1.base.link_mode_masks_nwords = 0;
1838 return !memcmp(&diff1.base, &diff2.base, sizeof(diff1.base)) &&
1839 bitmap_empty(diff1.link_modes.supported,
1840 __ETHTOOL_LINK_MODE_MASK_NBITS) &&
1841 bitmap_empty(diff1.link_modes.advertising,
1842 __ETHTOOL_LINK_MODE_MASK_NBITS) &&
1843 bitmap_empty(diff1.link_modes.lp_advertising,
1844 __ETHTOOL_LINK_MODE_MASK_NBITS);
1847 static int virtnet_set_link_ksettings(struct net_device *dev,
1848 const struct ethtool_link_ksettings *cmd)
1850 struct virtnet_info *vi = netdev_priv(dev);
1853 speed = cmd->base.speed;
1854 /* don't allow custom speed and duplex */
1855 if (!ethtool_validate_speed(speed) ||
1856 !ethtool_validate_duplex(cmd->base.duplex) ||
1857 !virtnet_validate_ethtool_cmd(cmd))
1860 vi->duplex = cmd->base.duplex;
1865 static int virtnet_get_link_ksettings(struct net_device *dev,
1866 struct ethtool_link_ksettings *cmd)
1868 struct virtnet_info *vi = netdev_priv(dev);
1870 cmd->base.speed = vi->speed;
1871 cmd->base.duplex = vi->duplex;
1872 cmd->base.port = PORT_OTHER;
1877 static void virtnet_init_settings(struct net_device *dev)
1879 struct virtnet_info *vi = netdev_priv(dev);
1881 vi->speed = SPEED_UNKNOWN;
1882 vi->duplex = DUPLEX_UNKNOWN;
1885 static const struct ethtool_ops virtnet_ethtool_ops = {
1886 .get_drvinfo = virtnet_get_drvinfo,
1887 .get_link = ethtool_op_get_link,
1888 .get_ringparam = virtnet_get_ringparam,
1889 .set_channels = virtnet_set_channels,
1890 .get_channels = virtnet_get_channels,
1891 .get_ts_info = ethtool_op_get_ts_info,
1892 .get_link_ksettings = virtnet_get_link_ksettings,
1893 .set_link_ksettings = virtnet_set_link_ksettings,
1896 static void virtnet_freeze_down(struct virtio_device *vdev)
1898 struct virtnet_info *vi = vdev->priv;
1901 /* Make sure no work handler is accessing the device */
1902 flush_work(&vi->config_work);
1904 netif_tx_lock_bh(vi->dev);
1905 netif_device_detach(vi->dev);
1906 netif_tx_unlock_bh(vi->dev);
1907 cancel_delayed_work_sync(&vi->refill);
1909 if (netif_running(vi->dev)) {
1910 for (i = 0; i < vi->max_queue_pairs; i++) {
1911 napi_disable(&vi->rq[i].napi);
1912 virtnet_napi_tx_disable(&vi->sq[i].napi);
1917 static int init_vqs(struct virtnet_info *vi);
1919 static int virtnet_restore_up(struct virtio_device *vdev)
1921 struct virtnet_info *vi = vdev->priv;
1928 virtio_device_ready(vdev);
1930 if (netif_running(vi->dev)) {
1931 for (i = 0; i < vi->curr_queue_pairs; i++)
1932 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1933 schedule_delayed_work(&vi->refill, 0);
1935 for (i = 0; i < vi->max_queue_pairs; i++) {
1936 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1937 virtnet_napi_tx_enable(vi, vi->sq[i].vq,
1942 netif_tx_lock_bh(vi->dev);
1943 netif_device_attach(vi->dev);
1944 netif_tx_unlock_bh(vi->dev);
1948 static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads)
1950 struct scatterlist sg;
1951 vi->ctrl->offloads = cpu_to_virtio64(vi->vdev, offloads);
1953 sg_init_one(&sg, &vi->ctrl->offloads, sizeof(vi->ctrl->offloads));
1955 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
1956 VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) {
1957 dev_warn(&vi->dev->dev, "Fail to set guest offload. \n");
1964 static int virtnet_clear_guest_offloads(struct virtnet_info *vi)
1968 if (!vi->guest_offloads)
1971 return virtnet_set_guest_offloads(vi, offloads);
1974 static int virtnet_restore_guest_offloads(struct virtnet_info *vi)
1976 u64 offloads = vi->guest_offloads;
1978 if (!vi->guest_offloads)
1981 return virtnet_set_guest_offloads(vi, offloads);
1984 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
1985 struct netlink_ext_ack *extack)
1987 unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr);
1988 struct virtnet_info *vi = netdev_priv(dev);
1989 struct bpf_prog *old_prog;
1990 u16 xdp_qp = 0, curr_qp;
1993 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)
1994 && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1995 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1996 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
1997 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) ||
1998 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))) {
1999 NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing LRO/CSUM, disable LRO/CSUM first");
2003 if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
2004 NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required");
2008 if (dev->mtu > max_sz) {
2009 NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP");
2010 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz);
2014 curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
2016 xdp_qp = nr_cpu_ids;
2018 /* XDP requires extra queues for XDP_TX */
2019 if (curr_qp + xdp_qp > vi->max_queue_pairs) {
2020 NL_SET_ERR_MSG_MOD(extack, "Too few free TX rings available");
2021 netdev_warn(dev, "request %i queues but max is %i\n",
2022 curr_qp + xdp_qp, vi->max_queue_pairs);
2027 prog = bpf_prog_add(prog, vi->max_queue_pairs - 1);
2029 return PTR_ERR(prog);
2032 /* Make sure NAPI is not using any XDP TX queues for RX. */
2033 if (netif_running(dev)) {
2034 for (i = 0; i < vi->max_queue_pairs; i++) {
2035 napi_disable(&vi->rq[i].napi);
2036 virtnet_napi_tx_disable(&vi->sq[i].napi);
2040 err = _virtnet_set_queues(vi, curr_qp + xdp_qp);
2043 netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
2044 vi->xdp_queue_pairs = xdp_qp;
2046 for (i = 0; i < vi->max_queue_pairs; i++) {
2047 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2048 rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2051 virtnet_clear_guest_offloads(vi);
2053 virtnet_restore_guest_offloads(vi);
2056 bpf_prog_put(old_prog);
2057 if (netif_running(dev)) {
2058 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2059 virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2067 if (netif_running(dev)) {
2068 for (i = 0; i < vi->max_queue_pairs; i++) {
2069 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2070 virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2075 bpf_prog_sub(prog, vi->max_queue_pairs - 1);
2079 static u32 virtnet_xdp_query(struct net_device *dev)
2081 struct virtnet_info *vi = netdev_priv(dev);
2082 const struct bpf_prog *xdp_prog;
2085 for (i = 0; i < vi->max_queue_pairs; i++) {
2086 xdp_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2088 return xdp_prog->aux->id;
2093 static int virtnet_xdp(struct net_device *dev, struct netdev_xdp *xdp)
2095 switch (xdp->command) {
2096 case XDP_SETUP_PROG:
2097 return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
2098 case XDP_QUERY_PROG:
2099 xdp->prog_id = virtnet_xdp_query(dev);
2100 xdp->prog_attached = !!xdp->prog_id;
2107 static const struct net_device_ops virtnet_netdev = {
2108 .ndo_open = virtnet_open,
2109 .ndo_stop = virtnet_close,
2110 .ndo_start_xmit = start_xmit,
2111 .ndo_validate_addr = eth_validate_addr,
2112 .ndo_set_mac_address = virtnet_set_mac_address,
2113 .ndo_set_rx_mode = virtnet_set_rx_mode,
2114 .ndo_get_stats64 = virtnet_stats,
2115 .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
2116 .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
2117 #ifdef CONFIG_NET_POLL_CONTROLLER
2118 .ndo_poll_controller = virtnet_netpoll,
2120 .ndo_xdp = virtnet_xdp,
2121 .ndo_features_check = passthru_features_check,
2124 static void virtnet_config_changed_work(struct work_struct *work)
2126 struct virtnet_info *vi =
2127 container_of(work, struct virtnet_info, config_work);
2130 if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
2131 struct virtio_net_config, status, &v) < 0)
2134 if (v & VIRTIO_NET_S_ANNOUNCE) {
2135 netdev_notify_peers(vi->dev);
2136 virtnet_ack_link_announce(vi);
2139 /* Ignore unknown (future) status bits */
2140 v &= VIRTIO_NET_S_LINK_UP;
2142 if (vi->status == v)
2147 if (vi->status & VIRTIO_NET_S_LINK_UP) {
2148 netif_carrier_on(vi->dev);
2149 netif_tx_wake_all_queues(vi->dev);
2151 netif_carrier_off(vi->dev);
2152 netif_tx_stop_all_queues(vi->dev);
2156 static void virtnet_config_changed(struct virtio_device *vdev)
2158 struct virtnet_info *vi = vdev->priv;
2160 schedule_work(&vi->config_work);
2163 static void virtnet_free_queues(struct virtnet_info *vi)
2167 for (i = 0; i < vi->max_queue_pairs; i++) {
2168 napi_hash_del(&vi->rq[i].napi);
2169 netif_napi_del(&vi->rq[i].napi);
2170 netif_napi_del(&vi->sq[i].napi);
2173 /* We called napi_hash_del() before netif_napi_del(),
2174 * we need to respect an RCU grace period before freeing vi->rq
2183 static void _free_receive_bufs(struct virtnet_info *vi)
2185 struct bpf_prog *old_prog;
2188 for (i = 0; i < vi->max_queue_pairs; i++) {
2189 while (vi->rq[i].pages)
2190 __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
2192 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2193 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
2195 bpf_prog_put(old_prog);
2199 static void free_receive_bufs(struct virtnet_info *vi)
2202 _free_receive_bufs(vi);
2206 static void free_receive_page_frags(struct virtnet_info *vi)
2209 for (i = 0; i < vi->max_queue_pairs; i++)
2210 if (vi->rq[i].alloc_frag.page)
2211 put_page(vi->rq[i].alloc_frag.page);
2214 static void free_unused_bufs(struct virtnet_info *vi)
2219 for (i = 0; i < vi->max_queue_pairs; i++) {
2220 struct virtqueue *vq = vi->sq[i].vq;
2221 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2222 if (!is_xdp_raw_buffer_queue(vi, i))
2225 put_page(virt_to_head_page(buf));
2229 for (i = 0; i < vi->max_queue_pairs; i++) {
2230 struct virtqueue *vq = vi->rq[i].vq;
2232 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2233 if (vi->mergeable_rx_bufs) {
2234 put_page(virt_to_head_page(buf));
2235 } else if (vi->big_packets) {
2236 give_pages(&vi->rq[i], buf);
2238 put_page(virt_to_head_page(buf));
2244 static void virtnet_del_vqs(struct virtnet_info *vi)
2246 struct virtio_device *vdev = vi->vdev;
2248 virtnet_clean_affinity(vi, -1);
2250 vdev->config->del_vqs(vdev);
2252 virtnet_free_queues(vi);
2255 /* How large should a single buffer be so a queue full of these can fit at
2256 * least one full packet?
2257 * Logic below assumes the mergeable buffer header is used.
2259 static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq)
2261 const unsigned int hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2262 unsigned int rq_size = virtqueue_get_vring_size(vq);
2263 unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu;
2264 unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len;
2265 unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size);
2267 return max(max(min_buf_len, hdr_len) - hdr_len,
2268 (unsigned int)GOOD_PACKET_LEN);
2271 static int virtnet_find_vqs(struct virtnet_info *vi)
2273 vq_callback_t **callbacks;
2274 struct virtqueue **vqs;
2280 /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
2281 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
2282 * possible control vq.
2284 total_vqs = vi->max_queue_pairs * 2 +
2285 virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
2287 /* Allocate space for find_vqs parameters */
2288 vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
2291 callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
2294 names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
2297 if (!vi->big_packets || vi->mergeable_rx_bufs) {
2298 ctx = kzalloc(total_vqs * sizeof(*ctx), GFP_KERNEL);
2305 /* Parameters for control virtqueue, if any */
2307 callbacks[total_vqs - 1] = NULL;
2308 names[total_vqs - 1] = "control";
2311 /* Allocate/initialize parameters for send/receive virtqueues */
2312 for (i = 0; i < vi->max_queue_pairs; i++) {
2313 callbacks[rxq2vq(i)] = skb_recv_done;
2314 callbacks[txq2vq(i)] = skb_xmit_done;
2315 sprintf(vi->rq[i].name, "input.%d", i);
2316 sprintf(vi->sq[i].name, "output.%d", i);
2317 names[rxq2vq(i)] = vi->rq[i].name;
2318 names[txq2vq(i)] = vi->sq[i].name;
2320 ctx[rxq2vq(i)] = true;
2323 ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
2329 vi->cvq = vqs[total_vqs - 1];
2330 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
2331 vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
2334 for (i = 0; i < vi->max_queue_pairs; i++) {
2335 vi->rq[i].vq = vqs[rxq2vq(i)];
2336 vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq);
2337 vi->sq[i].vq = vqs[txq2vq(i)];
2359 static int virtnet_alloc_queues(struct virtnet_info *vi)
2363 vi->ctrl = kzalloc(sizeof(*vi->ctrl), GFP_KERNEL);
2366 vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
2369 vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
2373 INIT_DELAYED_WORK(&vi->refill, refill_work);
2374 for (i = 0; i < vi->max_queue_pairs; i++) {
2375 vi->rq[i].pages = NULL;
2376 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
2378 netif_tx_napi_add(vi->dev, &vi->sq[i].napi, virtnet_poll_tx,
2379 napi_tx ? napi_weight : 0);
2381 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
2382 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
2383 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
2396 static int init_vqs(struct virtnet_info *vi)
2400 /* Allocate send & receive queues */
2401 ret = virtnet_alloc_queues(vi);
2405 ret = virtnet_find_vqs(vi);
2410 virtnet_set_affinity(vi);
2416 virtnet_free_queues(vi);
2422 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
2425 struct virtnet_info *vi = netdev_priv(queue->dev);
2426 unsigned int queue_index = get_netdev_rx_queue_index(queue);
2427 struct ewma_pkt_len *avg;
2429 BUG_ON(queue_index >= vi->max_queue_pairs);
2430 avg = &vi->rq[queue_index].mrg_avg_pkt_len;
2431 return sprintf(buf, "%u\n",
2432 get_mergeable_buf_len(&vi->rq[queue_index], avg));
2435 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
2436 __ATTR_RO(mergeable_rx_buffer_size);
2438 static struct attribute *virtio_net_mrg_rx_attrs[] = {
2439 &mergeable_rx_buffer_size_attribute.attr,
2443 static const struct attribute_group virtio_net_mrg_rx_group = {
2444 .name = "virtio_net",
2445 .attrs = virtio_net_mrg_rx_attrs
2449 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
2451 const char *fname, const char *dname)
2453 if (!virtio_has_feature(vdev, fbit))
2456 dev_err(&vdev->dev, "device advertises feature %s but not %s",
2462 #define VIRTNET_FAIL_ON(vdev, fbit, dbit) \
2463 virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
2465 static bool virtnet_validate_features(struct virtio_device *vdev)
2467 if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
2468 (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
2469 "VIRTIO_NET_F_CTRL_VQ") ||
2470 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
2471 "VIRTIO_NET_F_CTRL_VQ") ||
2472 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
2473 "VIRTIO_NET_F_CTRL_VQ") ||
2474 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
2475 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
2476 "VIRTIO_NET_F_CTRL_VQ"))) {
2483 #define MIN_MTU ETH_MIN_MTU
2484 #define MAX_MTU ETH_MAX_MTU
2486 static int virtnet_validate(struct virtio_device *vdev)
2488 if (!vdev->config->get) {
2489 dev_err(&vdev->dev, "%s failure: config access disabled\n",
2494 if (!virtnet_validate_features(vdev))
2497 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2498 int mtu = virtio_cread16(vdev,
2499 offsetof(struct virtio_net_config,
2502 __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
2508 static int virtnet_probe(struct virtio_device *vdev)
2511 struct net_device *dev;
2512 struct virtnet_info *vi;
2513 u16 max_queue_pairs;
2516 /* Find if host supports multiqueue virtio_net device */
2517 err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
2518 struct virtio_net_config,
2519 max_virtqueue_pairs, &max_queue_pairs);
2521 /* We need at least 2 queue's */
2522 if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
2523 max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
2524 !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2525 max_queue_pairs = 1;
2527 /* Allocate ourselves a network device with room for our info */
2528 dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
2532 /* Set up network device as normal. */
2533 dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
2534 dev->netdev_ops = &virtnet_netdev;
2535 dev->features = NETIF_F_HIGHDMA;
2537 dev->ethtool_ops = &virtnet_ethtool_ops;
2538 SET_NETDEV_DEV(dev, &vdev->dev);
2540 /* Do we support "hardware" checksums? */
2541 if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
2542 /* This opens up the world of extra features. */
2543 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2545 dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2547 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
2548 dev->hw_features |= NETIF_F_TSO
2549 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
2551 /* Individual feature bits: what can host handle? */
2552 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
2553 dev->hw_features |= NETIF_F_TSO;
2554 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
2555 dev->hw_features |= NETIF_F_TSO6;
2556 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
2557 dev->hw_features |= NETIF_F_TSO_ECN;
2559 dev->features |= NETIF_F_GSO_ROBUST;
2562 dev->features |= dev->hw_features & NETIF_F_ALL_TSO;
2563 /* (!csum && gso) case will be fixed by register_netdev() */
2565 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
2566 dev->features |= NETIF_F_RXCSUM;
2568 dev->vlan_features = dev->features;
2570 /* MTU range: 68 - 65535 */
2571 dev->min_mtu = MIN_MTU;
2572 dev->max_mtu = MAX_MTU;
2574 /* Configuration may specify what MAC to use. Otherwise random. */
2575 if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
2576 virtio_cread_bytes(vdev,
2577 offsetof(struct virtio_net_config, mac),
2578 dev->dev_addr, dev->addr_len);
2580 eth_hw_addr_random(dev);
2582 /* Set up our device-specific information */
2583 vi = netdev_priv(dev);
2587 vi->stats = alloc_percpu(struct virtnet_stats);
2589 if (vi->stats == NULL)
2592 for_each_possible_cpu(i) {
2593 struct virtnet_stats *virtnet_stats;
2594 virtnet_stats = per_cpu_ptr(vi->stats, i);
2595 u64_stats_init(&virtnet_stats->tx_syncp);
2596 u64_stats_init(&virtnet_stats->rx_syncp);
2599 INIT_WORK(&vi->config_work, virtnet_config_changed_work);
2601 /* If we can receive ANY GSO packets, we must allocate large ones. */
2602 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2603 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2604 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
2605 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
2606 vi->big_packets = true;
2608 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
2609 vi->mergeable_rx_bufs = true;
2611 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
2612 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2613 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2615 vi->hdr_len = sizeof(struct virtio_net_hdr);
2617 if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
2618 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2619 vi->any_header_sg = true;
2621 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2624 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2625 mtu = virtio_cread16(vdev,
2626 offsetof(struct virtio_net_config,
2628 if (mtu < dev->min_mtu) {
2629 /* Should never trigger: MTU was previously validated
2630 * in virtnet_validate.
2632 dev_err(&vdev->dev, "device MTU appears to have changed "
2633 "it is now %d < %d", mtu, dev->min_mtu);
2640 /* TODO: size buffers correctly in this case. */
2641 if (dev->mtu > ETH_DATA_LEN)
2642 vi->big_packets = true;
2645 if (vi->any_header_sg)
2646 dev->needed_headroom = vi->hdr_len;
2648 /* Enable multiqueue by default */
2649 if (num_online_cpus() >= max_queue_pairs)
2650 vi->curr_queue_pairs = max_queue_pairs;
2652 vi->curr_queue_pairs = num_online_cpus();
2653 vi->max_queue_pairs = max_queue_pairs;
2655 /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
2661 if (vi->mergeable_rx_bufs)
2662 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
2664 netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
2665 netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
2667 virtnet_init_settings(dev);
2669 err = register_netdev(dev);
2671 pr_debug("virtio_net: registering device failed\n");
2675 virtio_device_ready(vdev);
2677 err = virtnet_cpu_notif_add(vi);
2679 pr_debug("virtio_net: registering cpu notifier failed\n");
2680 goto free_unregister_netdev;
2683 virtnet_set_queues(vi, vi->curr_queue_pairs);
2685 /* Assume link up if device can't report link status,
2686 otherwise get link status from config. */
2687 netif_carrier_off(dev);
2688 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
2689 schedule_work(&vi->config_work);
2691 vi->status = VIRTIO_NET_S_LINK_UP;
2692 netif_carrier_on(dev);
2695 for (i = 0; i < ARRAY_SIZE(guest_offloads); i++)
2696 if (virtio_has_feature(vi->vdev, guest_offloads[i]))
2697 set_bit(guest_offloads[i], &vi->guest_offloads);
2699 pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
2700 dev->name, max_queue_pairs);
2704 free_unregister_netdev:
2705 vi->vdev->config->reset(vdev);
2707 unregister_netdev(dev);
2709 cancel_delayed_work_sync(&vi->refill);
2710 free_receive_page_frags(vi);
2711 virtnet_del_vqs(vi);
2713 free_percpu(vi->stats);
2719 static void remove_vq_common(struct virtnet_info *vi)
2721 vi->vdev->config->reset(vi->vdev);
2723 /* Free unused buffers in both send and recv, if any. */
2724 free_unused_bufs(vi);
2726 free_receive_bufs(vi);
2728 free_receive_page_frags(vi);
2730 virtnet_del_vqs(vi);
2733 static void virtnet_remove(struct virtio_device *vdev)
2735 struct virtnet_info *vi = vdev->priv;
2737 virtnet_cpu_notif_remove(vi);
2739 /* Make sure no work handler is accessing the device. */
2740 flush_work(&vi->config_work);
2742 unregister_netdev(vi->dev);
2744 remove_vq_common(vi);
2746 free_percpu(vi->stats);
2747 free_netdev(vi->dev);
2750 static __maybe_unused int virtnet_freeze(struct virtio_device *vdev)
2752 struct virtnet_info *vi = vdev->priv;
2754 virtnet_cpu_notif_remove(vi);
2755 virtnet_freeze_down(vdev);
2756 remove_vq_common(vi);
2761 static __maybe_unused int virtnet_restore(struct virtio_device *vdev)
2763 struct virtnet_info *vi = vdev->priv;
2766 err = virtnet_restore_up(vdev);
2769 virtnet_set_queues(vi, vi->curr_queue_pairs);
2771 err = virtnet_cpu_notif_add(vi);
2773 virtnet_freeze_down(vdev);
2774 remove_vq_common(vi);
2781 static struct virtio_device_id id_table[] = {
2782 { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
2786 #define VIRTNET_FEATURES \
2787 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
2789 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
2790 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
2791 VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
2792 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
2793 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
2794 VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
2795 VIRTIO_NET_F_CTRL_MAC_ADDR, \
2796 VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS
2798 static unsigned int features[] = {
2802 static unsigned int features_legacy[] = {
2805 VIRTIO_F_ANY_LAYOUT,
2808 static struct virtio_driver virtio_net_driver = {
2809 .feature_table = features,
2810 .feature_table_size = ARRAY_SIZE(features),
2811 .feature_table_legacy = features_legacy,
2812 .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
2813 .driver.name = KBUILD_MODNAME,
2814 .driver.owner = THIS_MODULE,
2815 .id_table = id_table,
2816 .validate = virtnet_validate,
2817 .probe = virtnet_probe,
2818 .remove = virtnet_remove,
2819 .config_changed = virtnet_config_changed,
2820 #ifdef CONFIG_PM_SLEEP
2821 .freeze = virtnet_freeze,
2822 .restore = virtnet_restore,
2826 static __init int virtio_net_driver_init(void)
2830 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
2832 virtnet_cpu_down_prep);
2835 virtionet_online = ret;
2836 ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
2837 NULL, virtnet_cpu_dead);
2841 ret = register_virtio_driver(&virtio_net_driver);
2846 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
2848 cpuhp_remove_multi_state(virtionet_online);
2852 module_init(virtio_net_driver_init);
2854 static __exit void virtio_net_driver_exit(void)
2856 unregister_virtio_driver(&virtio_net_driver);
2857 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
2858 cpuhp_remove_multi_state(virtionet_online);
2860 module_exit(virtio_net_driver_exit);
2862 MODULE_DEVICE_TABLE(virtio, id_table);
2863 MODULE_DESCRIPTION("Virtio network driver");
2864 MODULE_LICENSE("GPL");