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 <linux/filter.h>
33 #include <linux/netdevice.h>
34 #include <linux/pci.h>
35 #include <net/route.h>
37 #include <net/net_failover.h>
39 static int napi_weight = NAPI_POLL_WEIGHT;
40 module_param(napi_weight, int, 0444);
42 static bool csum = true, gso = true, napi_tx;
43 module_param(csum, bool, 0444);
44 module_param(gso, bool, 0444);
45 module_param(napi_tx, bool, 0644);
47 /* FIXME: MTU in config. */
48 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
49 #define GOOD_COPY_LEN 128
51 #define VIRTNET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
53 /* Amount of XDP headroom to prepend to packets for use by xdp_adjust_head */
54 #define VIRTIO_XDP_HEADROOM 256
56 /* RX packet size EWMA. The average packet size is used to determine the packet
57 * buffer size when refilling RX rings. As the entire RX ring may be refilled
58 * at once, the weight is chosen so that the EWMA will be insensitive to short-
59 * term, transient changes in packet size.
61 DECLARE_EWMA(pkt_len, 0, 64)
63 #define VIRTNET_DRIVER_VERSION "1.0.0"
65 static const unsigned long guest_offloads[] = {
66 VIRTIO_NET_F_GUEST_TSO4,
67 VIRTIO_NET_F_GUEST_TSO6,
68 VIRTIO_NET_F_GUEST_ECN,
69 VIRTIO_NET_F_GUEST_UFO
72 struct virtnet_stat_desc {
73 char desc[ETH_GSTRING_LEN];
77 struct virtnet_sq_stats {
78 struct u64_stats_sync syncp;
83 struct virtnet_rq_stats {
84 struct u64_stats_sync syncp;
89 #define VIRTNET_SQ_STAT(m) offsetof(struct virtnet_sq_stats, m)
90 #define VIRTNET_RQ_STAT(m) offsetof(struct virtnet_rq_stats, m)
92 static const struct virtnet_stat_desc virtnet_sq_stats_desc[] = {
93 { "packets", VIRTNET_SQ_STAT(packets) },
94 { "bytes", VIRTNET_SQ_STAT(bytes) },
97 static const struct virtnet_stat_desc virtnet_rq_stats_desc[] = {
98 { "packets", VIRTNET_RQ_STAT(packets) },
99 { "bytes", VIRTNET_RQ_STAT(bytes) },
102 #define VIRTNET_SQ_STATS_LEN ARRAY_SIZE(virtnet_sq_stats_desc)
103 #define VIRTNET_RQ_STATS_LEN ARRAY_SIZE(virtnet_rq_stats_desc)
105 /* Internal representation of a send virtqueue */
107 /* Virtqueue associated with this send _queue */
108 struct virtqueue *vq;
110 /* TX: fragments + linear part + virtio header */
111 struct scatterlist sg[MAX_SKB_FRAGS + 2];
113 /* Name of the send queue: output.$index */
116 struct virtnet_sq_stats stats;
118 struct napi_struct napi;
121 /* Internal representation of a receive virtqueue */
122 struct receive_queue {
123 /* Virtqueue associated with this receive_queue */
124 struct virtqueue *vq;
126 struct napi_struct napi;
128 struct bpf_prog __rcu *xdp_prog;
130 struct virtnet_rq_stats stats;
132 /* Chain pages by the private ptr. */
135 /* Average packet length for mergeable receive buffers. */
136 struct ewma_pkt_len mrg_avg_pkt_len;
138 /* Page frag for packet buffer allocation. */
139 struct page_frag alloc_frag;
141 /* RX: fragments + linear part + virtio header */
142 struct scatterlist sg[MAX_SKB_FRAGS + 2];
144 /* Min single buffer size for mergeable buffers case. */
145 unsigned int min_buf_len;
147 /* Name of this receive queue: input.$index */
150 struct xdp_rxq_info xdp_rxq;
153 /* Control VQ buffers: protected by the rtnl lock */
155 struct virtio_net_ctrl_hdr hdr;
156 virtio_net_ctrl_ack status;
157 struct virtio_net_ctrl_mq mq;
164 struct virtnet_info {
165 struct virtio_device *vdev;
166 struct virtqueue *cvq;
167 struct net_device *dev;
168 struct send_queue *sq;
169 struct receive_queue *rq;
172 /* Max # of queue pairs supported by the device */
175 /* # of queue pairs currently used by the driver */
176 u16 curr_queue_pairs;
178 /* # of XDP queue pairs currently used by the driver */
181 /* I like... big packets and I cannot lie! */
184 /* Host will merge rx buffers for big packets (shake it! shake it!) */
185 bool mergeable_rx_bufs;
187 /* Has control virtqueue */
190 /* Host can handle any s/g split between our header and packet data */
193 /* Packet virtio header size */
196 /* Work struct for refilling if we run low on memory. */
197 struct delayed_work refill;
199 /* Work struct for config space updates */
200 struct work_struct config_work;
202 /* Does the affinity hint is set for virtqueues? */
203 bool affinity_hint_set;
205 /* CPU hotplug instances for online & dead */
206 struct hlist_node node;
207 struct hlist_node node_dead;
209 struct control_buf *ctrl;
211 /* Ethtool settings */
215 unsigned long guest_offloads;
217 /* failover when STANDBY feature enabled */
218 struct failover *failover;
221 struct padded_vnet_hdr {
222 struct virtio_net_hdr_mrg_rxbuf hdr;
224 * hdr is in a separate sg buffer, and data sg buffer shares same page
225 * with this header sg. This padding makes next sg 16 byte aligned
231 /* Converting between virtqueue no. and kernel tx/rx queue no.
232 * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
234 static int vq2txq(struct virtqueue *vq)
236 return (vq->index - 1) / 2;
239 static int txq2vq(int txq)
244 static int vq2rxq(struct virtqueue *vq)
246 return vq->index / 2;
249 static int rxq2vq(int rxq)
254 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
256 return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
260 * private is used to chain pages for big packets, put the whole
261 * most recent used list in the beginning for reuse
263 static void give_pages(struct receive_queue *rq, struct page *page)
267 /* Find end of list, sew whole thing into vi->rq.pages. */
268 for (end = page; end->private; end = (struct page *)end->private);
269 end->private = (unsigned long)rq->pages;
273 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
275 struct page *p = rq->pages;
278 rq->pages = (struct page *)p->private;
279 /* clear private here, it is used to chain pages */
282 p = alloc_page(gfp_mask);
286 static void virtqueue_napi_schedule(struct napi_struct *napi,
287 struct virtqueue *vq)
289 if (napi_schedule_prep(napi)) {
290 virtqueue_disable_cb(vq);
291 __napi_schedule(napi);
295 static void virtqueue_napi_complete(struct napi_struct *napi,
296 struct virtqueue *vq, int processed)
300 opaque = virtqueue_enable_cb_prepare(vq);
301 if (napi_complete_done(napi, processed)) {
302 if (unlikely(virtqueue_poll(vq, opaque)))
303 virtqueue_napi_schedule(napi, vq);
305 virtqueue_disable_cb(vq);
309 static void skb_xmit_done(struct virtqueue *vq)
311 struct virtnet_info *vi = vq->vdev->priv;
312 struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi;
314 /* Suppress further interrupts. */
315 virtqueue_disable_cb(vq);
318 virtqueue_napi_schedule(napi, vq);
320 /* We were probably waiting for more output buffers. */
321 netif_wake_subqueue(vi->dev, vq2txq(vq));
324 #define MRG_CTX_HEADER_SHIFT 22
325 static void *mergeable_len_to_ctx(unsigned int truesize,
326 unsigned int headroom)
328 return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) | truesize);
331 static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx)
333 return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT;
336 static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx)
338 return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) - 1);
341 /* Called from bottom half context */
342 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
343 struct receive_queue *rq,
344 struct page *page, unsigned int offset,
345 unsigned int len, unsigned int truesize)
348 struct virtio_net_hdr_mrg_rxbuf *hdr;
349 unsigned int copy, hdr_len, hdr_padded_len;
352 p = page_address(page) + offset;
354 /* copy small packet so we can reuse these pages for small data */
355 skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
359 hdr = skb_vnet_hdr(skb);
361 hdr_len = vi->hdr_len;
362 if (vi->mergeable_rx_bufs)
363 hdr_padded_len = sizeof(*hdr);
365 hdr_padded_len = sizeof(struct padded_vnet_hdr);
367 memcpy(hdr, p, hdr_len);
370 offset += hdr_padded_len;
374 if (copy > skb_tailroom(skb))
375 copy = skb_tailroom(skb);
376 skb_put_data(skb, p, copy);
381 if (vi->mergeable_rx_bufs) {
383 skb_add_rx_frag(skb, 0, page, offset, len, truesize);
390 * Verify that we can indeed put this data into a skb.
391 * This is here to handle cases when the device erroneously
392 * tries to receive more than is possible. This is usually
393 * the case of a broken device.
395 if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
396 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
400 BUG_ON(offset >= PAGE_SIZE);
402 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
403 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
404 frag_size, truesize);
406 page = (struct page *)page->private;
411 give_pages(rq, page);
416 static int __virtnet_xdp_xmit_one(struct virtnet_info *vi,
417 struct send_queue *sq,
418 struct xdp_frame *xdpf)
420 struct virtio_net_hdr_mrg_rxbuf *hdr;
423 /* virtqueue want to use data area in-front of packet */
424 if (unlikely(xdpf->metasize > 0))
427 if (unlikely(xdpf->headroom < vi->hdr_len))
430 /* Make room for virtqueue hdr (also change xdpf->headroom?) */
431 xdpf->data -= vi->hdr_len;
432 /* Zero header and leave csum up to XDP layers */
434 memset(hdr, 0, vi->hdr_len);
435 xdpf->len += vi->hdr_len;
437 sg_init_one(sq->sg, xdpf->data, xdpf->len);
439 err = virtqueue_add_outbuf(sq->vq, sq->sg, 1, xdpf, GFP_ATOMIC);
441 return -ENOSPC; /* Caller handle free/refcnt */
446 static int __virtnet_xdp_tx_xmit(struct virtnet_info *vi,
447 struct xdp_frame *xdpf)
449 struct xdp_frame *xdpf_sent;
450 struct send_queue *sq;
454 qp = vi->curr_queue_pairs - vi->xdp_queue_pairs + smp_processor_id();
457 /* Free up any pending old buffers before queueing new ones. */
458 while ((xdpf_sent = virtqueue_get_buf(sq->vq, &len)) != NULL)
459 xdp_return_frame(xdpf_sent);
461 return __virtnet_xdp_xmit_one(vi, sq, xdpf);
464 static int virtnet_xdp_xmit(struct net_device *dev,
465 int n, struct xdp_frame **frames, u32 flags)
467 struct virtnet_info *vi = netdev_priv(dev);
468 struct receive_queue *rq = vi->rq;
469 struct xdp_frame *xdpf_sent;
470 struct bpf_prog *xdp_prog;
471 struct send_queue *sq;
478 if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK))
481 qp = vi->curr_queue_pairs - vi->xdp_queue_pairs + smp_processor_id();
484 /* Only allow ndo_xdp_xmit if XDP is loaded on dev, as this
485 * indicate XDP resources have been successfully allocated.
487 xdp_prog = rcu_dereference(rq->xdp_prog);
491 /* Free up any pending old buffers before queueing new ones. */
492 while ((xdpf_sent = virtqueue_get_buf(sq->vq, &len)) != NULL)
493 xdp_return_frame(xdpf_sent);
495 for (i = 0; i < n; i++) {
496 struct xdp_frame *xdpf = frames[i];
498 err = __virtnet_xdp_xmit_one(vi, sq, xdpf);
500 xdp_return_frame_rx_napi(xdpf);
505 if (flags & XDP_XMIT_FLUSH)
506 virtqueue_kick(sq->vq);
511 static unsigned int virtnet_get_headroom(struct virtnet_info *vi)
513 return vi->xdp_queue_pairs ? VIRTIO_XDP_HEADROOM : 0;
516 /* We copy the packet for XDP in the following cases:
518 * 1) Packet is scattered across multiple rx buffers.
519 * 2) Headroom space is insufficient.
521 * This is inefficient but it's a temporary condition that
522 * we hit right after XDP is enabled and until queue is refilled
523 * with large buffers with sufficient headroom - so it should affect
524 * at most queue size packets.
525 * Afterwards, the conditions to enable
526 * XDP should preclude the underlying device from sending packets
527 * across multiple buffers (num_buf > 1), and we make sure buffers
528 * have enough headroom.
530 static struct page *xdp_linearize_page(struct receive_queue *rq,
537 struct page *page = alloc_page(GFP_ATOMIC);
542 memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
546 int tailroom = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
551 buf = virtqueue_get_buf(rq->vq, &buflen);
555 p = virt_to_head_page(buf);
556 off = buf - page_address(p);
558 /* guard against a misconfigured or uncooperative backend that
559 * is sending packet larger than the MTU.
561 if ((page_off + buflen + tailroom) > PAGE_SIZE) {
566 memcpy(page_address(page) + page_off,
567 page_address(p) + off, buflen);
572 /* Headroom does not contribute to packet length */
573 *len = page_off - VIRTIO_XDP_HEADROOM;
576 __free_pages(page, 0);
580 static struct sk_buff *receive_small(struct net_device *dev,
581 struct virtnet_info *vi,
582 struct receive_queue *rq,
583 void *buf, void *ctx,
588 struct bpf_prog *xdp_prog;
589 unsigned int xdp_headroom = (unsigned long)ctx;
590 unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom;
591 unsigned int headroom = vi->hdr_len + header_offset;
592 unsigned int buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
593 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
594 struct page *page = virt_to_head_page(buf);
595 unsigned int delta = 0;
596 struct page *xdp_page;
602 xdp_prog = rcu_dereference(rq->xdp_prog);
604 struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset;
605 struct xdp_frame *xdpf;
610 if (unlikely(hdr->hdr.gso_type))
613 if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) {
614 int offset = buf - page_address(page) + header_offset;
615 unsigned int tlen = len + vi->hdr_len;
618 xdp_headroom = virtnet_get_headroom(vi);
619 header_offset = VIRTNET_RX_PAD + xdp_headroom;
620 headroom = vi->hdr_len + header_offset;
621 buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
622 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
623 xdp_page = xdp_linearize_page(rq, &num_buf, page,
624 offset, header_offset,
629 buf = page_address(xdp_page);
634 xdp.data_hard_start = buf + VIRTNET_RX_PAD + vi->hdr_len;
635 xdp.data = xdp.data_hard_start + xdp_headroom;
636 xdp_set_data_meta_invalid(&xdp);
637 xdp.data_end = xdp.data + len;
638 xdp.rxq = &rq->xdp_rxq;
639 orig_data = xdp.data;
640 act = bpf_prog_run_xdp(xdp_prog, &xdp);
644 /* Recalculate length in case bpf program changed it */
645 delta = orig_data - xdp.data;
646 len = xdp.data_end - xdp.data;
649 xdpf = convert_to_xdp_frame(&xdp);
652 err = __virtnet_xdp_tx_xmit(vi, xdpf);
654 trace_xdp_exception(vi->dev, xdp_prog, act);
661 err = xdp_do_redirect(dev, &xdp, xdp_prog);
668 bpf_warn_invalid_xdp_action(act);
670 trace_xdp_exception(vi->dev, xdp_prog, act);
677 skb = build_skb(buf, buflen);
682 skb_reserve(skb, headroom - delta);
685 buf += header_offset;
686 memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
687 } /* keep zeroed vnet hdr since packet was changed by bpf */
694 dev->stats.rx_dropped++;
700 static struct sk_buff *receive_big(struct net_device *dev,
701 struct virtnet_info *vi,
702 struct receive_queue *rq,
706 struct page *page = buf;
707 struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len, PAGE_SIZE);
715 dev->stats.rx_dropped++;
716 give_pages(rq, page);
720 static struct sk_buff *receive_mergeable(struct net_device *dev,
721 struct virtnet_info *vi,
722 struct receive_queue *rq,
728 struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
729 u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
730 struct page *page = virt_to_head_page(buf);
731 int offset = buf - page_address(page);
732 struct sk_buff *head_skb, *curr_skb;
733 struct bpf_prog *xdp_prog;
734 unsigned int truesize;
735 unsigned int headroom = mergeable_ctx_to_headroom(ctx);
741 xdp_prog = rcu_dereference(rq->xdp_prog);
743 struct xdp_frame *xdpf;
744 struct page *xdp_page;
749 /* Transient failure which in theory could occur if
750 * in-flight packets from before XDP was enabled reach
751 * the receive path after XDP is loaded.
753 if (unlikely(hdr->hdr.gso_type))
756 /* This happens when rx buffer size is underestimated
757 * or headroom is not enough because of the buffer
758 * was refilled before XDP is set. This should only
759 * happen for the first several packets, so we don't
760 * care much about its performance.
762 if (unlikely(num_buf > 1 ||
763 headroom < virtnet_get_headroom(vi))) {
764 /* linearize data for XDP */
765 xdp_page = xdp_linearize_page(rq, &num_buf,
771 offset = VIRTIO_XDP_HEADROOM;
776 /* Allow consuming headroom but reserve enough space to push
777 * the descriptor on if we get an XDP_TX return code.
779 data = page_address(xdp_page) + offset;
780 xdp.data_hard_start = data - VIRTIO_XDP_HEADROOM + vi->hdr_len;
781 xdp.data = data + vi->hdr_len;
782 xdp_set_data_meta_invalid(&xdp);
783 xdp.data_end = xdp.data + (len - vi->hdr_len);
784 xdp.rxq = &rq->xdp_rxq;
786 act = bpf_prog_run_xdp(xdp_prog, &xdp);
790 /* recalculate offset to account for any header
791 * adjustments. Note other cases do not build an
792 * skb and avoid using offset
795 page_address(xdp_page) - vi->hdr_len;
797 /* recalculate len if xdp.data or xdp.data_end were
800 len = xdp.data_end - xdp.data + vi->hdr_len;
801 /* We can only create skb based on xdp_page. */
802 if (unlikely(xdp_page != page)) {
805 head_skb = page_to_skb(vi, rq, xdp_page,
806 offset, len, PAGE_SIZE);
811 xdpf = convert_to_xdp_frame(&xdp);
814 err = __virtnet_xdp_tx_xmit(vi, xdpf);
816 trace_xdp_exception(vi->dev, xdp_prog, act);
817 if (unlikely(xdp_page != page))
822 if (unlikely(xdp_page != page))
827 err = xdp_do_redirect(dev, &xdp, xdp_prog);
829 if (unlikely(xdp_page != page))
834 if (unlikely(xdp_page != page))
839 bpf_warn_invalid_xdp_action(act);
841 trace_xdp_exception(vi->dev, xdp_prog, act);
843 if (unlikely(xdp_page != page))
844 __free_pages(xdp_page, 0);
850 truesize = mergeable_ctx_to_truesize(ctx);
851 if (unlikely(len > truesize)) {
852 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
853 dev->name, len, (unsigned long)ctx);
854 dev->stats.rx_length_errors++;
858 head_skb = page_to_skb(vi, rq, page, offset, len, truesize);
861 if (unlikely(!curr_skb))
866 buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx);
867 if (unlikely(!buf)) {
868 pr_debug("%s: rx error: %d buffers out of %d missing\n",
870 virtio16_to_cpu(vi->vdev,
872 dev->stats.rx_length_errors++;
876 page = virt_to_head_page(buf);
878 truesize = mergeable_ctx_to_truesize(ctx);
879 if (unlikely(len > truesize)) {
880 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
881 dev->name, len, (unsigned long)ctx);
882 dev->stats.rx_length_errors++;
886 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
887 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
888 struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
892 if (curr_skb == head_skb)
893 skb_shinfo(curr_skb)->frag_list = nskb;
895 curr_skb->next = nskb;
897 head_skb->truesize += nskb->truesize;
900 if (curr_skb != head_skb) {
901 head_skb->data_len += len;
902 head_skb->len += len;
903 head_skb->truesize += truesize;
905 offset = buf - page_address(page);
906 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
908 skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
911 skb_add_rx_frag(curr_skb, num_skb_frags, page,
912 offset, len, truesize);
916 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
923 while (num_buf-- > 1) {
924 buf = virtqueue_get_buf(rq->vq, &len);
925 if (unlikely(!buf)) {
926 pr_debug("%s: rx error: %d buffers missing\n",
928 dev->stats.rx_length_errors++;
931 page = virt_to_head_page(buf);
935 dev->stats.rx_dropped++;
936 dev_kfree_skb(head_skb);
941 static int receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
942 void *buf, unsigned int len, void **ctx, bool *xdp_xmit)
944 struct net_device *dev = vi->dev;
946 struct virtio_net_hdr_mrg_rxbuf *hdr;
949 if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
950 pr_debug("%s: short packet %i\n", dev->name, len);
951 dev->stats.rx_length_errors++;
952 if (vi->mergeable_rx_bufs) {
953 put_page(virt_to_head_page(buf));
954 } else if (vi->big_packets) {
957 put_page(virt_to_head_page(buf));
962 if (vi->mergeable_rx_bufs)
963 skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit);
964 else if (vi->big_packets)
965 skb = receive_big(dev, vi, rq, buf, len);
967 skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit);
972 hdr = skb_vnet_hdr(skb);
976 if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
977 skb->ip_summed = CHECKSUM_UNNECESSARY;
979 if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
980 virtio_is_little_endian(vi->vdev))) {
981 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
982 dev->name, hdr->hdr.gso_type,
987 skb->protocol = eth_type_trans(skb, dev);
988 pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
989 ntohs(skb->protocol), skb->len, skb->pkt_type);
991 napi_gro_receive(&rq->napi, skb);
995 dev->stats.rx_frame_errors++;
1000 /* Unlike mergeable buffers, all buffers are allocated to the
1001 * same size, except for the headroom. For this reason we do
1002 * not need to use mergeable_len_to_ctx here - it is enough
1003 * to store the headroom as the context ignoring the truesize.
1005 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
1008 struct page_frag *alloc_frag = &rq->alloc_frag;
1010 unsigned int xdp_headroom = virtnet_get_headroom(vi);
1011 void *ctx = (void *)(unsigned long)xdp_headroom;
1012 int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
1015 len = SKB_DATA_ALIGN(len) +
1016 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1017 if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
1020 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1021 get_page(alloc_frag->page);
1022 alloc_frag->offset += len;
1023 sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom,
1024 vi->hdr_len + GOOD_PACKET_LEN);
1025 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1027 put_page(virt_to_head_page(buf));
1031 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
1034 struct page *first, *list = NULL;
1038 sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
1040 /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
1041 for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
1042 first = get_a_page(rq, gfp);
1045 give_pages(rq, list);
1048 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
1050 /* chain new page in list head to match sg */
1051 first->private = (unsigned long)list;
1055 first = get_a_page(rq, gfp);
1057 give_pages(rq, list);
1060 p = page_address(first);
1062 /* rq->sg[0], rq->sg[1] share the same page */
1063 /* a separated rq->sg[0] for header - required in case !any_header_sg */
1064 sg_set_buf(&rq->sg[0], p, vi->hdr_len);
1066 /* rq->sg[1] for data packet, from offset */
1067 offset = sizeof(struct padded_vnet_hdr);
1068 sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
1070 /* chain first in list head */
1071 first->private = (unsigned long)list;
1072 err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
1075 give_pages(rq, first);
1080 static unsigned int get_mergeable_buf_len(struct receive_queue *rq,
1081 struct ewma_pkt_len *avg_pkt_len,
1084 const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1088 return PAGE_SIZE - room;
1090 len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
1091 rq->min_buf_len, PAGE_SIZE - hdr_len);
1093 return ALIGN(len, L1_CACHE_BYTES);
1096 static int add_recvbuf_mergeable(struct virtnet_info *vi,
1097 struct receive_queue *rq, gfp_t gfp)
1099 struct page_frag *alloc_frag = &rq->alloc_frag;
1100 unsigned int headroom = virtnet_get_headroom(vi);
1101 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1102 unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
1106 unsigned int len, hole;
1108 /* Extra tailroom is needed to satisfy XDP's assumption. This
1109 * means rx frags coalescing won't work, but consider we've
1110 * disabled GSO for XDP, it won't be a big issue.
1112 len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len, room);
1113 if (unlikely(!skb_page_frag_refill(len + room, alloc_frag, gfp)))
1116 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1117 buf += headroom; /* advance address leaving hole at front of pkt */
1118 get_page(alloc_frag->page);
1119 alloc_frag->offset += len + room;
1120 hole = alloc_frag->size - alloc_frag->offset;
1121 if (hole < len + room) {
1122 /* To avoid internal fragmentation, if there is very likely not
1123 * enough space for another buffer, add the remaining space to
1124 * the current buffer.
1127 alloc_frag->offset += hole;
1130 sg_init_one(rq->sg, buf, len);
1131 ctx = mergeable_len_to_ctx(len, headroom);
1132 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1134 put_page(virt_to_head_page(buf));
1140 * Returns false if we couldn't fill entirely (OOM).
1142 * Normally run in the receive path, but can also be run from ndo_open
1143 * before we're receiving packets, or from refill_work which is
1144 * careful to disable receiving (using napi_disable).
1146 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
1153 if (vi->mergeable_rx_bufs)
1154 err = add_recvbuf_mergeable(vi, rq, gfp);
1155 else if (vi->big_packets)
1156 err = add_recvbuf_big(vi, rq, gfp);
1158 err = add_recvbuf_small(vi, rq, gfp);
1160 oom = err == -ENOMEM;
1163 } while (rq->vq->num_free);
1164 virtqueue_kick(rq->vq);
1168 static void skb_recv_done(struct virtqueue *rvq)
1170 struct virtnet_info *vi = rvq->vdev->priv;
1171 struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
1173 virtqueue_napi_schedule(&rq->napi, rvq);
1176 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
1180 /* If all buffers were filled by other side before we napi_enabled, we
1181 * won't get another interrupt, so process any outstanding packets now.
1182 * Call local_bh_enable after to trigger softIRQ processing.
1185 virtqueue_napi_schedule(napi, vq);
1189 static void virtnet_napi_tx_enable(struct virtnet_info *vi,
1190 struct virtqueue *vq,
1191 struct napi_struct *napi)
1196 /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
1197 * enable the feature if this is likely affine with the transmit path.
1199 if (!vi->affinity_hint_set) {
1204 return virtnet_napi_enable(vq, napi);
1207 static void virtnet_napi_tx_disable(struct napi_struct *napi)
1213 static void refill_work(struct work_struct *work)
1215 struct virtnet_info *vi =
1216 container_of(work, struct virtnet_info, refill.work);
1220 for (i = 0; i < vi->curr_queue_pairs; i++) {
1221 struct receive_queue *rq = &vi->rq[i];
1223 napi_disable(&rq->napi);
1224 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
1225 virtnet_napi_enable(rq->vq, &rq->napi);
1227 /* In theory, this can happen: if we don't get any buffers in
1228 * we will *never* try to fill again.
1231 schedule_delayed_work(&vi->refill, HZ/2);
1235 static int virtnet_receive(struct receive_queue *rq, int budget, bool *xdp_xmit)
1237 struct virtnet_info *vi = rq->vq->vdev->priv;
1238 unsigned int len, received = 0, bytes = 0;
1241 if (!vi->big_packets || vi->mergeable_rx_bufs) {
1244 while (received < budget &&
1245 (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) {
1246 bytes += receive_buf(vi, rq, buf, len, ctx, xdp_xmit);
1250 while (received < budget &&
1251 (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
1252 bytes += receive_buf(vi, rq, buf, len, NULL, xdp_xmit);
1257 if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) {
1258 if (!try_fill_recv(vi, rq, GFP_ATOMIC))
1259 schedule_delayed_work(&vi->refill, 0);
1262 u64_stats_update_begin(&rq->stats.syncp);
1263 rq->stats.bytes += bytes;
1264 rq->stats.packets += received;
1265 u64_stats_update_end(&rq->stats.syncp);
1270 static void free_old_xmit_skbs(struct send_queue *sq)
1272 struct sk_buff *skb;
1274 unsigned int packets = 0;
1275 unsigned int bytes = 0;
1277 while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
1278 pr_debug("Sent skb %p\n", skb);
1283 dev_consume_skb_any(skb);
1286 /* Avoid overhead when no packets have been processed
1287 * happens when called speculatively from start_xmit.
1292 u64_stats_update_begin(&sq->stats.syncp);
1293 sq->stats.bytes += bytes;
1294 sq->stats.packets += packets;
1295 u64_stats_update_end(&sq->stats.syncp);
1298 static void virtnet_poll_cleantx(struct receive_queue *rq)
1300 struct virtnet_info *vi = rq->vq->vdev->priv;
1301 unsigned int index = vq2rxq(rq->vq);
1302 struct send_queue *sq = &vi->sq[index];
1303 struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index);
1305 if (!sq->napi.weight)
1308 if (__netif_tx_trylock(txq)) {
1309 free_old_xmit_skbs(sq);
1310 __netif_tx_unlock(txq);
1313 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1314 netif_tx_wake_queue(txq);
1317 static int virtnet_poll(struct napi_struct *napi, int budget)
1319 struct receive_queue *rq =
1320 container_of(napi, struct receive_queue, napi);
1321 struct virtnet_info *vi = rq->vq->vdev->priv;
1322 struct send_queue *sq;
1323 unsigned int received, qp;
1324 bool xdp_xmit = false;
1326 virtnet_poll_cleantx(rq);
1328 received = virtnet_receive(rq, budget, &xdp_xmit);
1330 /* Out of packets? */
1331 if (received < budget)
1332 virtqueue_napi_complete(napi, rq->vq, received);
1335 qp = vi->curr_queue_pairs - vi->xdp_queue_pairs +
1338 virtqueue_kick(sq->vq);
1345 static int virtnet_open(struct net_device *dev)
1347 struct virtnet_info *vi = netdev_priv(dev);
1350 for (i = 0; i < vi->max_queue_pairs; i++) {
1351 if (i < vi->curr_queue_pairs)
1352 /* Make sure we have some buffers: if oom use wq. */
1353 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1354 schedule_delayed_work(&vi->refill, 0);
1356 err = xdp_rxq_info_reg(&vi->rq[i].xdp_rxq, dev, i);
1360 err = xdp_rxq_info_reg_mem_model(&vi->rq[i].xdp_rxq,
1361 MEM_TYPE_PAGE_SHARED, NULL);
1363 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1367 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1368 virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi);
1374 static int virtnet_poll_tx(struct napi_struct *napi, int budget)
1376 struct send_queue *sq = container_of(napi, struct send_queue, napi);
1377 struct virtnet_info *vi = sq->vq->vdev->priv;
1378 struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, vq2txq(sq->vq));
1380 __netif_tx_lock(txq, raw_smp_processor_id());
1381 free_old_xmit_skbs(sq);
1382 __netif_tx_unlock(txq);
1384 virtqueue_napi_complete(napi, sq->vq, 0);
1386 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1387 netif_tx_wake_queue(txq);
1392 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
1394 struct virtio_net_hdr_mrg_rxbuf *hdr;
1395 const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
1396 struct virtnet_info *vi = sq->vq->vdev->priv;
1398 unsigned hdr_len = vi->hdr_len;
1401 pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
1403 can_push = vi->any_header_sg &&
1404 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
1405 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
1406 /* Even if we can, don't push here yet as this would skew
1407 * csum_start offset below. */
1409 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1411 hdr = skb_vnet_hdr(skb);
1413 if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1414 virtio_is_little_endian(vi->vdev), false))
1417 if (vi->mergeable_rx_bufs)
1418 hdr->num_buffers = 0;
1420 sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1422 __skb_push(skb, hdr_len);
1423 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1424 if (unlikely(num_sg < 0))
1426 /* Pull header back to avoid skew in tx bytes calculations. */
1427 __skb_pull(skb, hdr_len);
1429 sg_set_buf(sq->sg, hdr, hdr_len);
1430 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
1431 if (unlikely(num_sg < 0))
1435 return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
1438 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
1440 struct virtnet_info *vi = netdev_priv(dev);
1441 int qnum = skb_get_queue_mapping(skb);
1442 struct send_queue *sq = &vi->sq[qnum];
1444 struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1445 bool kick = !skb->xmit_more;
1446 bool use_napi = sq->napi.weight;
1448 /* Free up any pending old buffers before queueing new ones. */
1449 free_old_xmit_skbs(sq);
1451 if (use_napi && kick)
1452 virtqueue_enable_cb_delayed(sq->vq);
1454 /* timestamp packet in software */
1455 skb_tx_timestamp(skb);
1457 /* Try to transmit */
1458 err = xmit_skb(sq, skb);
1460 /* This should not happen! */
1461 if (unlikely(err)) {
1462 dev->stats.tx_fifo_errors++;
1463 if (net_ratelimit())
1465 "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
1466 dev->stats.tx_dropped++;
1467 dev_kfree_skb_any(skb);
1468 return NETDEV_TX_OK;
1471 /* Don't wait up for transmitted skbs to be freed. */
1477 /* If running out of space, stop queue to avoid getting packets that we
1478 * are then unable to transmit.
1479 * An alternative would be to force queuing layer to requeue the skb by
1480 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1481 * returned in a normal path of operation: it means that driver is not
1482 * maintaining the TX queue stop/start state properly, and causes
1483 * the stack to do a non-trivial amount of useless work.
1484 * Since most packets only take 1 or 2 ring slots, stopping the queue
1485 * early means 16 slots are typically wasted.
1487 if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
1488 netif_stop_subqueue(dev, qnum);
1490 unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
1491 /* More just got used, free them then recheck. */
1492 free_old_xmit_skbs(sq);
1493 if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
1494 netif_start_subqueue(dev, qnum);
1495 virtqueue_disable_cb(sq->vq);
1500 if (kick || netif_xmit_stopped(txq))
1501 virtqueue_kick(sq->vq);
1503 return NETDEV_TX_OK;
1507 * Send command via the control virtqueue and check status. Commands
1508 * supported by the hypervisor, as indicated by feature bits, should
1509 * never fail unless improperly formatted.
1511 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
1512 struct scatterlist *out)
1514 struct scatterlist *sgs[4], hdr, stat;
1515 unsigned out_num = 0, tmp;
1517 /* Caller should know better */
1518 BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
1520 vi->ctrl->status = ~0;
1521 vi->ctrl->hdr.class = class;
1522 vi->ctrl->hdr.cmd = cmd;
1524 sg_init_one(&hdr, &vi->ctrl->hdr, sizeof(vi->ctrl->hdr));
1525 sgs[out_num++] = &hdr;
1528 sgs[out_num++] = out;
1530 /* Add return status. */
1531 sg_init_one(&stat, &vi->ctrl->status, sizeof(vi->ctrl->status));
1532 sgs[out_num] = &stat;
1534 BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1535 virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1537 if (unlikely(!virtqueue_kick(vi->cvq)))
1538 return vi->ctrl->status == VIRTIO_NET_OK;
1540 /* Spin for a response, the kick causes an ioport write, trapping
1541 * into the hypervisor, so the request should be handled immediately.
1543 while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1544 !virtqueue_is_broken(vi->cvq))
1547 return vi->ctrl->status == VIRTIO_NET_OK;
1550 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1552 struct virtnet_info *vi = netdev_priv(dev);
1553 struct virtio_device *vdev = vi->vdev;
1555 struct sockaddr *addr;
1556 struct scatterlist sg;
1558 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
1561 addr = kmemdup(p, sizeof(*addr), GFP_KERNEL);
1565 ret = eth_prepare_mac_addr_change(dev, addr);
1569 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1570 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1571 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1572 VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1573 dev_warn(&vdev->dev,
1574 "Failed to set mac address by vq command.\n");
1578 } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1579 !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1582 /* Naturally, this has an atomicity problem. */
1583 for (i = 0; i < dev->addr_len; i++)
1584 virtio_cwrite8(vdev,
1585 offsetof(struct virtio_net_config, mac) +
1586 i, addr->sa_data[i]);
1589 eth_commit_mac_addr_change(dev, p);
1597 static void virtnet_stats(struct net_device *dev,
1598 struct rtnl_link_stats64 *tot)
1600 struct virtnet_info *vi = netdev_priv(dev);
1604 for (i = 0; i < vi->max_queue_pairs; i++) {
1605 u64 tpackets, tbytes, rpackets, rbytes;
1606 struct receive_queue *rq = &vi->rq[i];
1607 struct send_queue *sq = &vi->sq[i];
1610 start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
1611 tpackets = sq->stats.packets;
1612 tbytes = sq->stats.bytes;
1613 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
1616 start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
1617 rpackets = rq->stats.packets;
1618 rbytes = rq->stats.bytes;
1619 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
1621 tot->rx_packets += rpackets;
1622 tot->tx_packets += tpackets;
1623 tot->rx_bytes += rbytes;
1624 tot->tx_bytes += tbytes;
1627 tot->tx_dropped = dev->stats.tx_dropped;
1628 tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1629 tot->rx_dropped = dev->stats.rx_dropped;
1630 tot->rx_length_errors = dev->stats.rx_length_errors;
1631 tot->rx_frame_errors = dev->stats.rx_frame_errors;
1634 #ifdef CONFIG_NET_POLL_CONTROLLER
1635 static void virtnet_netpoll(struct net_device *dev)
1637 struct virtnet_info *vi = netdev_priv(dev);
1640 for (i = 0; i < vi->curr_queue_pairs; i++)
1641 napi_schedule(&vi->rq[i].napi);
1645 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1648 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1649 VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1650 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1654 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1656 struct scatterlist sg;
1657 struct net_device *dev = vi->dev;
1659 if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1662 vi->ctrl->mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1663 sg_init_one(&sg, &vi->ctrl->mq, sizeof(vi->ctrl->mq));
1665 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1666 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1667 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1671 vi->curr_queue_pairs = queue_pairs;
1672 /* virtnet_open() will refill when device is going to up. */
1673 if (dev->flags & IFF_UP)
1674 schedule_delayed_work(&vi->refill, 0);
1680 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1685 err = _virtnet_set_queues(vi, queue_pairs);
1690 static int virtnet_close(struct net_device *dev)
1692 struct virtnet_info *vi = netdev_priv(dev);
1695 /* Make sure refill_work doesn't re-enable napi! */
1696 cancel_delayed_work_sync(&vi->refill);
1698 for (i = 0; i < vi->max_queue_pairs; i++) {
1699 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1700 napi_disable(&vi->rq[i].napi);
1701 virtnet_napi_tx_disable(&vi->sq[i].napi);
1707 static void virtnet_set_rx_mode(struct net_device *dev)
1709 struct virtnet_info *vi = netdev_priv(dev);
1710 struct scatterlist sg[2];
1711 struct virtio_net_ctrl_mac *mac_data;
1712 struct netdev_hw_addr *ha;
1718 /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1719 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1722 vi->ctrl->promisc = ((dev->flags & IFF_PROMISC) != 0);
1723 vi->ctrl->allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1725 sg_init_one(sg, &vi->ctrl->promisc, sizeof(vi->ctrl->promisc));
1727 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1728 VIRTIO_NET_CTRL_RX_PROMISC, sg))
1729 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1730 vi->ctrl->promisc ? "en" : "dis");
1732 sg_init_one(sg, &vi->ctrl->allmulti, sizeof(vi->ctrl->allmulti));
1734 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1735 VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1736 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1737 vi->ctrl->allmulti ? "en" : "dis");
1739 uc_count = netdev_uc_count(dev);
1740 mc_count = netdev_mc_count(dev);
1741 /* MAC filter - use one buffer for both lists */
1742 buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1743 (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1748 sg_init_table(sg, 2);
1750 /* Store the unicast list and count in the front of the buffer */
1751 mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1753 netdev_for_each_uc_addr(ha, dev)
1754 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1756 sg_set_buf(&sg[0], mac_data,
1757 sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1759 /* multicast list and count fill the end */
1760 mac_data = (void *)&mac_data->macs[uc_count][0];
1762 mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1764 netdev_for_each_mc_addr(ha, dev)
1765 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1767 sg_set_buf(&sg[1], mac_data,
1768 sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1770 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1771 VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1772 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1777 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1778 __be16 proto, u16 vid)
1780 struct virtnet_info *vi = netdev_priv(dev);
1781 struct scatterlist sg;
1783 vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
1784 sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
1786 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1787 VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1788 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1792 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1793 __be16 proto, u16 vid)
1795 struct virtnet_info *vi = netdev_priv(dev);
1796 struct scatterlist sg;
1798 vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
1799 sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
1801 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1802 VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1803 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1807 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1811 if (vi->affinity_hint_set) {
1812 for (i = 0; i < vi->max_queue_pairs; i++) {
1813 virtqueue_set_affinity(vi->rq[i].vq, -1);
1814 virtqueue_set_affinity(vi->sq[i].vq, -1);
1817 vi->affinity_hint_set = false;
1821 static void virtnet_set_affinity(struct virtnet_info *vi)
1826 /* In multiqueue mode, when the number of cpu is equal to the number of
1827 * queue pairs, we let the queue pairs to be private to one cpu by
1828 * setting the affinity hint to eliminate the contention.
1830 if (vi->curr_queue_pairs == 1 ||
1831 vi->max_queue_pairs != num_online_cpus()) {
1832 virtnet_clean_affinity(vi, -1);
1837 for_each_online_cpu(cpu) {
1838 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1839 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1840 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
1844 vi->affinity_hint_set = true;
1847 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
1849 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1851 virtnet_set_affinity(vi);
1855 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
1857 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1859 virtnet_set_affinity(vi);
1863 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
1865 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1868 virtnet_clean_affinity(vi, cpu);
1872 static enum cpuhp_state virtionet_online;
1874 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
1878 ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
1881 ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1885 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1889 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
1891 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1892 cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1896 static void virtnet_get_ringparam(struct net_device *dev,
1897 struct ethtool_ringparam *ring)
1899 struct virtnet_info *vi = netdev_priv(dev);
1901 ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1902 ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1903 ring->rx_pending = ring->rx_max_pending;
1904 ring->tx_pending = ring->tx_max_pending;
1908 static void virtnet_get_drvinfo(struct net_device *dev,
1909 struct ethtool_drvinfo *info)
1911 struct virtnet_info *vi = netdev_priv(dev);
1912 struct virtio_device *vdev = vi->vdev;
1914 strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1915 strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1916 strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1920 /* TODO: Eliminate OOO packets during switching */
1921 static int virtnet_set_channels(struct net_device *dev,
1922 struct ethtool_channels *channels)
1924 struct virtnet_info *vi = netdev_priv(dev);
1925 u16 queue_pairs = channels->combined_count;
1928 /* We don't support separate rx/tx channels.
1929 * We don't allow setting 'other' channels.
1931 if (channels->rx_count || channels->tx_count || channels->other_count)
1934 if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
1937 /* For now we don't support modifying channels while XDP is loaded
1938 * also when XDP is loaded all RX queues have XDP programs so we only
1939 * need to check a single RX queue.
1941 if (vi->rq[0].xdp_prog)
1945 err = _virtnet_set_queues(vi, queue_pairs);
1947 netif_set_real_num_tx_queues(dev, queue_pairs);
1948 netif_set_real_num_rx_queues(dev, queue_pairs);
1950 virtnet_set_affinity(vi);
1957 static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *data)
1959 struct virtnet_info *vi = netdev_priv(dev);
1960 char *p = (char *)data;
1963 switch (stringset) {
1965 for (i = 0; i < vi->curr_queue_pairs; i++) {
1966 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
1967 snprintf(p, ETH_GSTRING_LEN, "rx_queue_%u_%s",
1968 i, virtnet_rq_stats_desc[j].desc);
1969 p += ETH_GSTRING_LEN;
1973 for (i = 0; i < vi->curr_queue_pairs; i++) {
1974 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
1975 snprintf(p, ETH_GSTRING_LEN, "tx_queue_%u_%s",
1976 i, virtnet_sq_stats_desc[j].desc);
1977 p += ETH_GSTRING_LEN;
1984 static int virtnet_get_sset_count(struct net_device *dev, int sset)
1986 struct virtnet_info *vi = netdev_priv(dev);
1990 return vi->curr_queue_pairs * (VIRTNET_RQ_STATS_LEN +
1991 VIRTNET_SQ_STATS_LEN);
1997 static void virtnet_get_ethtool_stats(struct net_device *dev,
1998 struct ethtool_stats *stats, u64 *data)
2000 struct virtnet_info *vi = netdev_priv(dev);
2001 unsigned int idx = 0, start, i, j;
2002 const u8 *stats_base;
2005 for (i = 0; i < vi->curr_queue_pairs; i++) {
2006 struct receive_queue *rq = &vi->rq[i];
2008 stats_base = (u8 *)&rq->stats;
2010 start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
2011 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
2012 offset = virtnet_rq_stats_desc[j].offset;
2013 data[idx + j] = *(u64 *)(stats_base + offset);
2015 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
2016 idx += VIRTNET_RQ_STATS_LEN;
2019 for (i = 0; i < vi->curr_queue_pairs; i++) {
2020 struct send_queue *sq = &vi->sq[i];
2022 stats_base = (u8 *)&sq->stats;
2024 start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
2025 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
2026 offset = virtnet_sq_stats_desc[j].offset;
2027 data[idx + j] = *(u64 *)(stats_base + offset);
2029 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
2030 idx += VIRTNET_SQ_STATS_LEN;
2034 static void virtnet_get_channels(struct net_device *dev,
2035 struct ethtool_channels *channels)
2037 struct virtnet_info *vi = netdev_priv(dev);
2039 channels->combined_count = vi->curr_queue_pairs;
2040 channels->max_combined = vi->max_queue_pairs;
2041 channels->max_other = 0;
2042 channels->rx_count = 0;
2043 channels->tx_count = 0;
2044 channels->other_count = 0;
2047 /* Check if the user is trying to change anything besides speed/duplex */
2049 virtnet_validate_ethtool_cmd(const struct ethtool_link_ksettings *cmd)
2051 struct ethtool_link_ksettings diff1 = *cmd;
2052 struct ethtool_link_ksettings diff2 = {};
2054 /* cmd is always set so we need to clear it, validate the port type
2055 * and also without autonegotiation we can ignore advertising
2057 diff1.base.speed = 0;
2058 diff2.base.port = PORT_OTHER;
2059 ethtool_link_ksettings_zero_link_mode(&diff1, advertising);
2060 diff1.base.duplex = 0;
2062 diff1.base.link_mode_masks_nwords = 0;
2064 return !memcmp(&diff1.base, &diff2.base, sizeof(diff1.base)) &&
2065 bitmap_empty(diff1.link_modes.supported,
2066 __ETHTOOL_LINK_MODE_MASK_NBITS) &&
2067 bitmap_empty(diff1.link_modes.advertising,
2068 __ETHTOOL_LINK_MODE_MASK_NBITS) &&
2069 bitmap_empty(diff1.link_modes.lp_advertising,
2070 __ETHTOOL_LINK_MODE_MASK_NBITS);
2073 static int virtnet_set_link_ksettings(struct net_device *dev,
2074 const struct ethtool_link_ksettings *cmd)
2076 struct virtnet_info *vi = netdev_priv(dev);
2079 speed = cmd->base.speed;
2080 /* don't allow custom speed and duplex */
2081 if (!ethtool_validate_speed(speed) ||
2082 !ethtool_validate_duplex(cmd->base.duplex) ||
2083 !virtnet_validate_ethtool_cmd(cmd))
2086 vi->duplex = cmd->base.duplex;
2091 static int virtnet_get_link_ksettings(struct net_device *dev,
2092 struct ethtool_link_ksettings *cmd)
2094 struct virtnet_info *vi = netdev_priv(dev);
2096 cmd->base.speed = vi->speed;
2097 cmd->base.duplex = vi->duplex;
2098 cmd->base.port = PORT_OTHER;
2103 static void virtnet_init_settings(struct net_device *dev)
2105 struct virtnet_info *vi = netdev_priv(dev);
2107 vi->speed = SPEED_UNKNOWN;
2108 vi->duplex = DUPLEX_UNKNOWN;
2111 static void virtnet_update_settings(struct virtnet_info *vi)
2116 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_SPEED_DUPLEX))
2119 speed = virtio_cread32(vi->vdev, offsetof(struct virtio_net_config,
2121 if (ethtool_validate_speed(speed))
2123 duplex = virtio_cread8(vi->vdev, offsetof(struct virtio_net_config,
2125 if (ethtool_validate_duplex(duplex))
2126 vi->duplex = duplex;
2129 static const struct ethtool_ops virtnet_ethtool_ops = {
2130 .get_drvinfo = virtnet_get_drvinfo,
2131 .get_link = ethtool_op_get_link,
2132 .get_ringparam = virtnet_get_ringparam,
2133 .get_strings = virtnet_get_strings,
2134 .get_sset_count = virtnet_get_sset_count,
2135 .get_ethtool_stats = virtnet_get_ethtool_stats,
2136 .set_channels = virtnet_set_channels,
2137 .get_channels = virtnet_get_channels,
2138 .get_ts_info = ethtool_op_get_ts_info,
2139 .get_link_ksettings = virtnet_get_link_ksettings,
2140 .set_link_ksettings = virtnet_set_link_ksettings,
2143 static void virtnet_freeze_down(struct virtio_device *vdev)
2145 struct virtnet_info *vi = vdev->priv;
2148 /* Make sure no work handler is accessing the device */
2149 flush_work(&vi->config_work);
2151 netif_device_detach(vi->dev);
2152 netif_tx_disable(vi->dev);
2153 cancel_delayed_work_sync(&vi->refill);
2155 if (netif_running(vi->dev)) {
2156 for (i = 0; i < vi->max_queue_pairs; i++) {
2157 napi_disable(&vi->rq[i].napi);
2158 virtnet_napi_tx_disable(&vi->sq[i].napi);
2163 static int init_vqs(struct virtnet_info *vi);
2165 static int virtnet_restore_up(struct virtio_device *vdev)
2167 struct virtnet_info *vi = vdev->priv;
2174 virtio_device_ready(vdev);
2176 if (netif_running(vi->dev)) {
2177 for (i = 0; i < vi->curr_queue_pairs; i++)
2178 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
2179 schedule_delayed_work(&vi->refill, 0);
2181 for (i = 0; i < vi->max_queue_pairs; i++) {
2182 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2183 virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2188 netif_device_attach(vi->dev);
2192 static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads)
2194 struct scatterlist sg;
2195 vi->ctrl->offloads = cpu_to_virtio64(vi->vdev, offloads);
2197 sg_init_one(&sg, &vi->ctrl->offloads, sizeof(vi->ctrl->offloads));
2199 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
2200 VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) {
2201 dev_warn(&vi->dev->dev, "Fail to set guest offload. \n");
2208 static int virtnet_clear_guest_offloads(struct virtnet_info *vi)
2212 if (!vi->guest_offloads)
2215 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))
2216 offloads = 1ULL << VIRTIO_NET_F_GUEST_CSUM;
2218 return virtnet_set_guest_offloads(vi, offloads);
2221 static int virtnet_restore_guest_offloads(struct virtnet_info *vi)
2223 u64 offloads = vi->guest_offloads;
2225 if (!vi->guest_offloads)
2227 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))
2228 offloads |= 1ULL << VIRTIO_NET_F_GUEST_CSUM;
2230 return virtnet_set_guest_offloads(vi, offloads);
2233 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
2234 struct netlink_ext_ack *extack)
2236 unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr);
2237 struct virtnet_info *vi = netdev_priv(dev);
2238 struct bpf_prog *old_prog;
2239 u16 xdp_qp = 0, curr_qp;
2242 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)
2243 && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2244 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2245 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
2246 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO))) {
2247 NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing LRO, disable LRO first");
2251 if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
2252 NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required");
2256 if (dev->mtu > max_sz) {
2257 NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP");
2258 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz);
2262 curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
2264 xdp_qp = nr_cpu_ids;
2266 /* XDP requires extra queues for XDP_TX */
2267 if (curr_qp + xdp_qp > vi->max_queue_pairs) {
2268 NL_SET_ERR_MSG_MOD(extack, "Too few free TX rings available");
2269 netdev_warn(dev, "request %i queues but max is %i\n",
2270 curr_qp + xdp_qp, vi->max_queue_pairs);
2275 prog = bpf_prog_add(prog, vi->max_queue_pairs - 1);
2277 return PTR_ERR(prog);
2280 /* Make sure NAPI is not using any XDP TX queues for RX. */
2281 if (netif_running(dev))
2282 for (i = 0; i < vi->max_queue_pairs; i++)
2283 napi_disable(&vi->rq[i].napi);
2285 netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
2286 err = _virtnet_set_queues(vi, curr_qp + xdp_qp);
2289 vi->xdp_queue_pairs = xdp_qp;
2291 for (i = 0; i < vi->max_queue_pairs; i++) {
2292 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2293 rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2296 virtnet_clear_guest_offloads(vi);
2298 virtnet_restore_guest_offloads(vi);
2301 bpf_prog_put(old_prog);
2302 if (netif_running(dev))
2303 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2309 for (i = 0; i < vi->max_queue_pairs; i++)
2310 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2312 bpf_prog_sub(prog, vi->max_queue_pairs - 1);
2316 static u32 virtnet_xdp_query(struct net_device *dev)
2318 struct virtnet_info *vi = netdev_priv(dev);
2319 const struct bpf_prog *xdp_prog;
2322 for (i = 0; i < vi->max_queue_pairs; i++) {
2323 xdp_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2325 return xdp_prog->aux->id;
2330 static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp)
2332 switch (xdp->command) {
2333 case XDP_SETUP_PROG:
2334 return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
2335 case XDP_QUERY_PROG:
2336 xdp->prog_id = virtnet_xdp_query(dev);
2337 xdp->prog_attached = !!xdp->prog_id;
2344 static int virtnet_get_phys_port_name(struct net_device *dev, char *buf,
2347 struct virtnet_info *vi = netdev_priv(dev);
2350 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
2353 ret = snprintf(buf, len, "sby");
2360 static const struct net_device_ops virtnet_netdev = {
2361 .ndo_open = virtnet_open,
2362 .ndo_stop = virtnet_close,
2363 .ndo_start_xmit = start_xmit,
2364 .ndo_validate_addr = eth_validate_addr,
2365 .ndo_set_mac_address = virtnet_set_mac_address,
2366 .ndo_set_rx_mode = virtnet_set_rx_mode,
2367 .ndo_get_stats64 = virtnet_stats,
2368 .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
2369 .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
2370 #ifdef CONFIG_NET_POLL_CONTROLLER
2371 .ndo_poll_controller = virtnet_netpoll,
2373 .ndo_bpf = virtnet_xdp,
2374 .ndo_xdp_xmit = virtnet_xdp_xmit,
2375 .ndo_features_check = passthru_features_check,
2376 .ndo_get_phys_port_name = virtnet_get_phys_port_name,
2379 static void virtnet_config_changed_work(struct work_struct *work)
2381 struct virtnet_info *vi =
2382 container_of(work, struct virtnet_info, config_work);
2385 if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
2386 struct virtio_net_config, status, &v) < 0)
2389 if (v & VIRTIO_NET_S_ANNOUNCE) {
2390 netdev_notify_peers(vi->dev);
2391 virtnet_ack_link_announce(vi);
2394 /* Ignore unknown (future) status bits */
2395 v &= VIRTIO_NET_S_LINK_UP;
2397 if (vi->status == v)
2402 if (vi->status & VIRTIO_NET_S_LINK_UP) {
2403 virtnet_update_settings(vi);
2404 netif_carrier_on(vi->dev);
2405 netif_tx_wake_all_queues(vi->dev);
2407 netif_carrier_off(vi->dev);
2408 netif_tx_stop_all_queues(vi->dev);
2412 static void virtnet_config_changed(struct virtio_device *vdev)
2414 struct virtnet_info *vi = vdev->priv;
2416 schedule_work(&vi->config_work);
2419 static void virtnet_free_queues(struct virtnet_info *vi)
2423 for (i = 0; i < vi->max_queue_pairs; i++) {
2424 napi_hash_del(&vi->rq[i].napi);
2425 netif_napi_del(&vi->rq[i].napi);
2426 netif_napi_del(&vi->sq[i].napi);
2429 /* We called napi_hash_del() before netif_napi_del(),
2430 * we need to respect an RCU grace period before freeing vi->rq
2439 static void _free_receive_bufs(struct virtnet_info *vi)
2441 struct bpf_prog *old_prog;
2444 for (i = 0; i < vi->max_queue_pairs; i++) {
2445 while (vi->rq[i].pages)
2446 __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
2448 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2449 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
2451 bpf_prog_put(old_prog);
2455 static void free_receive_bufs(struct virtnet_info *vi)
2458 _free_receive_bufs(vi);
2462 static void free_receive_page_frags(struct virtnet_info *vi)
2465 for (i = 0; i < vi->max_queue_pairs; i++)
2466 if (vi->rq[i].alloc_frag.page)
2467 put_page(vi->rq[i].alloc_frag.page);
2470 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
2472 if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
2474 else if (q < vi->curr_queue_pairs)
2480 static void free_unused_bufs(struct virtnet_info *vi)
2485 for (i = 0; i < vi->max_queue_pairs; i++) {
2486 struct virtqueue *vq = vi->sq[i].vq;
2487 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2488 if (!is_xdp_raw_buffer_queue(vi, i))
2491 put_page(virt_to_head_page(buf));
2495 for (i = 0; i < vi->max_queue_pairs; i++) {
2496 struct virtqueue *vq = vi->rq[i].vq;
2498 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2499 if (vi->mergeable_rx_bufs) {
2500 put_page(virt_to_head_page(buf));
2501 } else if (vi->big_packets) {
2502 give_pages(&vi->rq[i], buf);
2504 put_page(virt_to_head_page(buf));
2510 static void virtnet_del_vqs(struct virtnet_info *vi)
2512 struct virtio_device *vdev = vi->vdev;
2514 virtnet_clean_affinity(vi, -1);
2516 vdev->config->del_vqs(vdev);
2518 virtnet_free_queues(vi);
2521 /* How large should a single buffer be so a queue full of these can fit at
2522 * least one full packet?
2523 * Logic below assumes the mergeable buffer header is used.
2525 static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq)
2527 const unsigned int hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2528 unsigned int rq_size = virtqueue_get_vring_size(vq);
2529 unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu;
2530 unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len;
2531 unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size);
2533 return max(max(min_buf_len, hdr_len) - hdr_len,
2534 (unsigned int)GOOD_PACKET_LEN);
2537 static int virtnet_find_vqs(struct virtnet_info *vi)
2539 vq_callback_t **callbacks;
2540 struct virtqueue **vqs;
2546 /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
2547 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
2548 * possible control vq.
2550 total_vqs = vi->max_queue_pairs * 2 +
2551 virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
2553 /* Allocate space for find_vqs parameters */
2554 vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
2557 callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
2560 names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
2563 if (!vi->big_packets || vi->mergeable_rx_bufs) {
2564 ctx = kzalloc(total_vqs * sizeof(*ctx), GFP_KERNEL);
2571 /* Parameters for control virtqueue, if any */
2573 callbacks[total_vqs - 1] = NULL;
2574 names[total_vqs - 1] = "control";
2577 /* Allocate/initialize parameters for send/receive virtqueues */
2578 for (i = 0; i < vi->max_queue_pairs; i++) {
2579 callbacks[rxq2vq(i)] = skb_recv_done;
2580 callbacks[txq2vq(i)] = skb_xmit_done;
2581 sprintf(vi->rq[i].name, "input.%d", i);
2582 sprintf(vi->sq[i].name, "output.%d", i);
2583 names[rxq2vq(i)] = vi->rq[i].name;
2584 names[txq2vq(i)] = vi->sq[i].name;
2586 ctx[rxq2vq(i)] = true;
2589 ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
2595 vi->cvq = vqs[total_vqs - 1];
2596 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
2597 vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
2600 for (i = 0; i < vi->max_queue_pairs; i++) {
2601 vi->rq[i].vq = vqs[rxq2vq(i)];
2602 vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq);
2603 vi->sq[i].vq = vqs[txq2vq(i)];
2606 /* run here: ret == 0. */
2621 static int virtnet_alloc_queues(struct virtnet_info *vi)
2625 vi->ctrl = kzalloc(sizeof(*vi->ctrl), GFP_KERNEL);
2628 vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
2631 vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
2635 INIT_DELAYED_WORK(&vi->refill, refill_work);
2636 for (i = 0; i < vi->max_queue_pairs; i++) {
2637 vi->rq[i].pages = NULL;
2638 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
2640 netif_tx_napi_add(vi->dev, &vi->sq[i].napi, virtnet_poll_tx,
2641 napi_tx ? napi_weight : 0);
2643 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
2644 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
2645 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
2647 u64_stats_init(&vi->rq[i].stats.syncp);
2648 u64_stats_init(&vi->sq[i].stats.syncp);
2661 static int init_vqs(struct virtnet_info *vi)
2665 /* Allocate send & receive queues */
2666 ret = virtnet_alloc_queues(vi);
2670 ret = virtnet_find_vqs(vi);
2675 virtnet_set_affinity(vi);
2681 virtnet_free_queues(vi);
2687 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
2690 struct virtnet_info *vi = netdev_priv(queue->dev);
2691 unsigned int queue_index = get_netdev_rx_queue_index(queue);
2692 unsigned int headroom = virtnet_get_headroom(vi);
2693 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
2694 struct ewma_pkt_len *avg;
2696 BUG_ON(queue_index >= vi->max_queue_pairs);
2697 avg = &vi->rq[queue_index].mrg_avg_pkt_len;
2698 return sprintf(buf, "%u\n",
2699 get_mergeable_buf_len(&vi->rq[queue_index], avg,
2700 SKB_DATA_ALIGN(headroom + tailroom)));
2703 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
2704 __ATTR_RO(mergeable_rx_buffer_size);
2706 static struct attribute *virtio_net_mrg_rx_attrs[] = {
2707 &mergeable_rx_buffer_size_attribute.attr,
2711 static const struct attribute_group virtio_net_mrg_rx_group = {
2712 .name = "virtio_net",
2713 .attrs = virtio_net_mrg_rx_attrs
2717 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
2719 const char *fname, const char *dname)
2721 if (!virtio_has_feature(vdev, fbit))
2724 dev_err(&vdev->dev, "device advertises feature %s but not %s",
2730 #define VIRTNET_FAIL_ON(vdev, fbit, dbit) \
2731 virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
2733 static bool virtnet_validate_features(struct virtio_device *vdev)
2735 if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
2736 (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
2737 "VIRTIO_NET_F_CTRL_VQ") ||
2738 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
2739 "VIRTIO_NET_F_CTRL_VQ") ||
2740 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
2741 "VIRTIO_NET_F_CTRL_VQ") ||
2742 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
2743 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
2744 "VIRTIO_NET_F_CTRL_VQ"))) {
2751 #define MIN_MTU ETH_MIN_MTU
2752 #define MAX_MTU ETH_MAX_MTU
2754 static int virtnet_validate(struct virtio_device *vdev)
2756 if (!vdev->config->get) {
2757 dev_err(&vdev->dev, "%s failure: config access disabled\n",
2762 if (!virtnet_validate_features(vdev))
2765 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2766 int mtu = virtio_cread16(vdev,
2767 offsetof(struct virtio_net_config,
2770 __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
2776 static int virtnet_probe(struct virtio_device *vdev)
2778 int i, err = -ENOMEM;
2779 struct net_device *dev;
2780 struct virtnet_info *vi;
2781 u16 max_queue_pairs;
2784 /* Find if host supports multiqueue virtio_net device */
2785 err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
2786 struct virtio_net_config,
2787 max_virtqueue_pairs, &max_queue_pairs);
2789 /* We need at least 2 queue's */
2790 if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
2791 max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
2792 !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2793 max_queue_pairs = 1;
2795 /* Allocate ourselves a network device with room for our info */
2796 dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
2800 /* Set up network device as normal. */
2801 dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
2802 dev->netdev_ops = &virtnet_netdev;
2803 dev->features = NETIF_F_HIGHDMA;
2805 dev->ethtool_ops = &virtnet_ethtool_ops;
2806 SET_NETDEV_DEV(dev, &vdev->dev);
2808 /* Do we support "hardware" checksums? */
2809 if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
2810 /* This opens up the world of extra features. */
2811 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2813 dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2815 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
2816 dev->hw_features |= NETIF_F_TSO
2817 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
2819 /* Individual feature bits: what can host handle? */
2820 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
2821 dev->hw_features |= NETIF_F_TSO;
2822 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
2823 dev->hw_features |= NETIF_F_TSO6;
2824 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
2825 dev->hw_features |= NETIF_F_TSO_ECN;
2827 dev->features |= NETIF_F_GSO_ROBUST;
2830 dev->features |= dev->hw_features & NETIF_F_ALL_TSO;
2831 /* (!csum && gso) case will be fixed by register_netdev() */
2833 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
2834 dev->features |= NETIF_F_RXCSUM;
2836 dev->vlan_features = dev->features;
2838 /* MTU range: 68 - 65535 */
2839 dev->min_mtu = MIN_MTU;
2840 dev->max_mtu = MAX_MTU;
2842 /* Configuration may specify what MAC to use. Otherwise random. */
2843 if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
2844 virtio_cread_bytes(vdev,
2845 offsetof(struct virtio_net_config, mac),
2846 dev->dev_addr, dev->addr_len);
2848 eth_hw_addr_random(dev);
2850 /* Set up our device-specific information */
2851 vi = netdev_priv(dev);
2856 INIT_WORK(&vi->config_work, virtnet_config_changed_work);
2858 /* If we can receive ANY GSO packets, we must allocate large ones. */
2859 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2860 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2861 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
2862 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
2863 vi->big_packets = true;
2865 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
2866 vi->mergeable_rx_bufs = true;
2868 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
2869 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2870 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2872 vi->hdr_len = sizeof(struct virtio_net_hdr);
2874 if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
2875 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2876 vi->any_header_sg = true;
2878 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2881 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2882 mtu = virtio_cread16(vdev,
2883 offsetof(struct virtio_net_config,
2885 if (mtu < dev->min_mtu) {
2886 /* Should never trigger: MTU was previously validated
2887 * in virtnet_validate.
2889 dev_err(&vdev->dev, "device MTU appears to have changed "
2890 "it is now %d < %d", mtu, dev->min_mtu);
2897 /* TODO: size buffers correctly in this case. */
2898 if (dev->mtu > ETH_DATA_LEN)
2899 vi->big_packets = true;
2902 if (vi->any_header_sg)
2903 dev->needed_headroom = vi->hdr_len;
2905 /* Enable multiqueue by default */
2906 if (num_online_cpus() >= max_queue_pairs)
2907 vi->curr_queue_pairs = max_queue_pairs;
2909 vi->curr_queue_pairs = num_online_cpus();
2910 vi->max_queue_pairs = max_queue_pairs;
2912 /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
2918 if (vi->mergeable_rx_bufs)
2919 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
2921 netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
2922 netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
2924 virtnet_init_settings(dev);
2926 if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY)) {
2927 vi->failover = net_failover_create(vi->dev);
2928 if (IS_ERR(vi->failover)) {
2929 err = PTR_ERR(vi->failover);
2934 err = register_netdev(dev);
2936 pr_debug("virtio_net: registering device failed\n");
2940 virtio_device_ready(vdev);
2942 err = virtnet_cpu_notif_add(vi);
2944 pr_debug("virtio_net: registering cpu notifier failed\n");
2945 goto free_unregister_netdev;
2948 virtnet_set_queues(vi, vi->curr_queue_pairs);
2950 /* Assume link up if device can't report link status,
2951 otherwise get link status from config. */
2952 netif_carrier_off(dev);
2953 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
2954 schedule_work(&vi->config_work);
2956 vi->status = VIRTIO_NET_S_LINK_UP;
2957 virtnet_update_settings(vi);
2958 netif_carrier_on(dev);
2961 for (i = 0; i < ARRAY_SIZE(guest_offloads); i++)
2962 if (virtio_has_feature(vi->vdev, guest_offloads[i]))
2963 set_bit(guest_offloads[i], &vi->guest_offloads);
2965 pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
2966 dev->name, max_queue_pairs);
2970 free_unregister_netdev:
2971 vi->vdev->config->reset(vdev);
2973 unregister_netdev(dev);
2975 net_failover_destroy(vi->failover);
2977 cancel_delayed_work_sync(&vi->refill);
2978 free_receive_page_frags(vi);
2979 virtnet_del_vqs(vi);
2985 static void remove_vq_common(struct virtnet_info *vi)
2987 vi->vdev->config->reset(vi->vdev);
2989 /* Free unused buffers in both send and recv, if any. */
2990 free_unused_bufs(vi);
2992 free_receive_bufs(vi);
2994 free_receive_page_frags(vi);
2996 virtnet_del_vqs(vi);
2999 static void virtnet_remove(struct virtio_device *vdev)
3001 struct virtnet_info *vi = vdev->priv;
3003 virtnet_cpu_notif_remove(vi);
3005 /* Make sure no work handler is accessing the device. */
3006 flush_work(&vi->config_work);
3008 unregister_netdev(vi->dev);
3010 net_failover_destroy(vi->failover);
3012 remove_vq_common(vi);
3014 free_netdev(vi->dev);
3017 static __maybe_unused int virtnet_freeze(struct virtio_device *vdev)
3019 struct virtnet_info *vi = vdev->priv;
3021 virtnet_cpu_notif_remove(vi);
3022 virtnet_freeze_down(vdev);
3023 remove_vq_common(vi);
3028 static __maybe_unused int virtnet_restore(struct virtio_device *vdev)
3030 struct virtnet_info *vi = vdev->priv;
3033 err = virtnet_restore_up(vdev);
3036 virtnet_set_queues(vi, vi->curr_queue_pairs);
3038 err = virtnet_cpu_notif_add(vi);
3045 static struct virtio_device_id id_table[] = {
3046 { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
3050 #define VIRTNET_FEATURES \
3051 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
3053 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
3054 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
3055 VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
3056 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
3057 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
3058 VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
3059 VIRTIO_NET_F_CTRL_MAC_ADDR, \
3060 VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \
3061 VIRTIO_NET_F_SPEED_DUPLEX, VIRTIO_NET_F_STANDBY
3063 static unsigned int features[] = {
3067 static unsigned int features_legacy[] = {
3070 VIRTIO_F_ANY_LAYOUT,
3073 static struct virtio_driver virtio_net_driver = {
3074 .feature_table = features,
3075 .feature_table_size = ARRAY_SIZE(features),
3076 .feature_table_legacy = features_legacy,
3077 .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
3078 .driver.name = KBUILD_MODNAME,
3079 .driver.owner = THIS_MODULE,
3080 .id_table = id_table,
3081 .validate = virtnet_validate,
3082 .probe = virtnet_probe,
3083 .remove = virtnet_remove,
3084 .config_changed = virtnet_config_changed,
3085 #ifdef CONFIG_PM_SLEEP
3086 .freeze = virtnet_freeze,
3087 .restore = virtnet_restore,
3091 static __init int virtio_net_driver_init(void)
3095 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
3097 virtnet_cpu_down_prep);
3100 virtionet_online = ret;
3101 ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
3102 NULL, virtnet_cpu_dead);
3106 ret = register_virtio_driver(&virtio_net_driver);
3111 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3113 cpuhp_remove_multi_state(virtionet_online);
3117 module_init(virtio_net_driver_init);
3119 static __exit void virtio_net_driver_exit(void)
3121 unregister_virtio_driver(&virtio_net_driver);
3122 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3123 cpuhp_remove_multi_state(virtionet_online);
3125 module_exit(virtio_net_driver_exit);
3127 MODULE_DEVICE_TABLE(virtio, id_table);
3128 MODULE_DESCRIPTION("Virtio network driver");
3129 MODULE_LICENSE("GPL");