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 <net/route.h>
36 static int napi_weight = NAPI_POLL_WEIGHT;
37 module_param(napi_weight, int, 0444);
39 static bool csum = true, gso = true, napi_tx;
40 module_param(csum, bool, 0444);
41 module_param(gso, bool, 0444);
42 module_param(napi_tx, bool, 0644);
44 /* FIXME: MTU in config. */
45 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
46 #define GOOD_COPY_LEN 128
48 #define VIRTNET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
50 /* Amount of XDP headroom to prepend to packets for use by xdp_adjust_head */
51 #define VIRTIO_XDP_HEADROOM 256
53 /* RX packet size EWMA. The average packet size is used to determine the packet
54 * buffer size when refilling RX rings. As the entire RX ring may be refilled
55 * at once, the weight is chosen so that the EWMA will be insensitive to short-
56 * term, transient changes in packet size.
58 DECLARE_EWMA(pkt_len, 0, 64)
60 #define VIRTNET_DRIVER_VERSION "1.0.0"
62 static const unsigned long guest_offloads[] = {
63 VIRTIO_NET_F_GUEST_TSO4,
64 VIRTIO_NET_F_GUEST_TSO6,
65 VIRTIO_NET_F_GUEST_ECN,
66 VIRTIO_NET_F_GUEST_UFO
69 struct virtnet_stat_desc {
70 char desc[ETH_GSTRING_LEN];
74 struct virtnet_sq_stats {
75 struct u64_stats_sync syncp;
80 struct virtnet_rq_stats {
81 struct u64_stats_sync syncp;
86 #define VIRTNET_SQ_STAT(m) offsetof(struct virtnet_sq_stats, m)
87 #define VIRTNET_RQ_STAT(m) offsetof(struct virtnet_rq_stats, m)
89 static const struct virtnet_stat_desc virtnet_sq_stats_desc[] = {
90 { "packets", VIRTNET_SQ_STAT(packets) },
91 { "bytes", VIRTNET_SQ_STAT(bytes) },
94 static const struct virtnet_stat_desc virtnet_rq_stats_desc[] = {
95 { "packets", VIRTNET_RQ_STAT(packets) },
96 { "bytes", VIRTNET_RQ_STAT(bytes) },
99 #define VIRTNET_SQ_STATS_LEN ARRAY_SIZE(virtnet_sq_stats_desc)
100 #define VIRTNET_RQ_STATS_LEN ARRAY_SIZE(virtnet_rq_stats_desc)
102 /* Internal representation of a send virtqueue */
104 /* Virtqueue associated with this send _queue */
105 struct virtqueue *vq;
107 /* TX: fragments + linear part + virtio header */
108 struct scatterlist sg[MAX_SKB_FRAGS + 2];
110 /* Name of the send queue: output.$index */
113 struct virtnet_sq_stats stats;
115 struct napi_struct napi;
118 /* Internal representation of a receive virtqueue */
119 struct receive_queue {
120 /* Virtqueue associated with this receive_queue */
121 struct virtqueue *vq;
123 struct napi_struct napi;
125 struct bpf_prog __rcu *xdp_prog;
127 struct virtnet_rq_stats stats;
129 /* Chain pages by the private ptr. */
132 /* Average packet length for mergeable receive buffers. */
133 struct ewma_pkt_len mrg_avg_pkt_len;
135 /* Page frag for packet buffer allocation. */
136 struct page_frag alloc_frag;
138 /* RX: fragments + linear part + virtio header */
139 struct scatterlist sg[MAX_SKB_FRAGS + 2];
141 /* Min single buffer size for mergeable buffers case. */
142 unsigned int min_buf_len;
144 /* Name of this receive queue: input.$index */
147 struct xdp_rxq_info xdp_rxq;
150 struct virtnet_info {
151 struct virtio_device *vdev;
152 struct virtqueue *cvq;
153 struct net_device *dev;
154 struct send_queue *sq;
155 struct receive_queue *rq;
158 /* Max # of queue pairs supported by the device */
161 /* # of queue pairs currently used by the driver */
162 u16 curr_queue_pairs;
164 /* # of XDP queue pairs currently used by the driver */
167 /* I like... big packets and I cannot lie! */
170 /* Host will merge rx buffers for big packets (shake it! shake it!) */
171 bool mergeable_rx_bufs;
173 /* Has control virtqueue */
176 /* Host can handle any s/g split between our header and packet data */
179 /* Packet virtio header size */
182 /* Work struct for refilling if we run low on memory. */
183 struct delayed_work refill;
185 /* Work struct for config space updates */
186 struct work_struct config_work;
188 /* Does the affinity hint is set for virtqueues? */
189 bool affinity_hint_set;
191 /* CPU hotplug instances for online & dead */
192 struct hlist_node node;
193 struct hlist_node node_dead;
195 /* Control VQ buffers: protected by the rtnl lock */
196 struct virtio_net_ctrl_hdr ctrl_hdr;
197 virtio_net_ctrl_ack ctrl_status;
198 struct virtio_net_ctrl_mq ctrl_mq;
204 /* Ethtool settings */
208 unsigned long guest_offloads;
211 struct padded_vnet_hdr {
212 struct virtio_net_hdr_mrg_rxbuf hdr;
214 * hdr is in a separate sg buffer, and data sg buffer shares same page
215 * with this header sg. This padding makes next sg 16 byte aligned
221 /* Converting between virtqueue no. and kernel tx/rx queue no.
222 * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
224 static int vq2txq(struct virtqueue *vq)
226 return (vq->index - 1) / 2;
229 static int txq2vq(int txq)
234 static int vq2rxq(struct virtqueue *vq)
236 return vq->index / 2;
239 static int rxq2vq(int rxq)
244 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
246 return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
250 * private is used to chain pages for big packets, put the whole
251 * most recent used list in the beginning for reuse
253 static void give_pages(struct receive_queue *rq, struct page *page)
257 /* Find end of list, sew whole thing into vi->rq.pages. */
258 for (end = page; end->private; end = (struct page *)end->private);
259 end->private = (unsigned long)rq->pages;
263 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
265 struct page *p = rq->pages;
268 rq->pages = (struct page *)p->private;
269 /* clear private here, it is used to chain pages */
272 p = alloc_page(gfp_mask);
276 static void virtqueue_napi_schedule(struct napi_struct *napi,
277 struct virtqueue *vq)
279 if (napi_schedule_prep(napi)) {
280 virtqueue_disable_cb(vq);
281 __napi_schedule(napi);
285 static void virtqueue_napi_complete(struct napi_struct *napi,
286 struct virtqueue *vq, int processed)
290 opaque = virtqueue_enable_cb_prepare(vq);
291 if (napi_complete_done(napi, processed)) {
292 if (unlikely(virtqueue_poll(vq, opaque)))
293 virtqueue_napi_schedule(napi, vq);
295 virtqueue_disable_cb(vq);
299 static void skb_xmit_done(struct virtqueue *vq)
301 struct virtnet_info *vi = vq->vdev->priv;
302 struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi;
304 /* Suppress further interrupts. */
305 virtqueue_disable_cb(vq);
308 virtqueue_napi_schedule(napi, vq);
310 /* We were probably waiting for more output buffers. */
311 netif_wake_subqueue(vi->dev, vq2txq(vq));
314 #define MRG_CTX_HEADER_SHIFT 22
315 static void *mergeable_len_to_ctx(unsigned int truesize,
316 unsigned int headroom)
318 return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) | truesize);
321 static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx)
323 return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT;
326 static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx)
328 return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) - 1);
331 /* Called from bottom half context */
332 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
333 struct receive_queue *rq,
334 struct page *page, unsigned int offset,
335 unsigned int len, unsigned int truesize)
338 struct virtio_net_hdr_mrg_rxbuf *hdr;
339 unsigned int copy, hdr_len, hdr_padded_len;
342 p = page_address(page) + offset;
344 /* copy small packet so we can reuse these pages for small data */
345 skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
349 hdr = skb_vnet_hdr(skb);
351 hdr_len = vi->hdr_len;
352 if (vi->mergeable_rx_bufs)
353 hdr_padded_len = sizeof(*hdr);
355 hdr_padded_len = sizeof(struct padded_vnet_hdr);
357 memcpy(hdr, p, hdr_len);
360 offset += hdr_padded_len;
364 if (copy > skb_tailroom(skb))
365 copy = skb_tailroom(skb);
366 skb_put_data(skb, p, copy);
371 if (vi->mergeable_rx_bufs) {
373 skb_add_rx_frag(skb, 0, page, offset, len, truesize);
380 * Verify that we can indeed put this data into a skb.
381 * This is here to handle cases when the device erroneously
382 * tries to receive more than is possible. This is usually
383 * the case of a broken device.
385 if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
386 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
390 BUG_ON(offset >= PAGE_SIZE);
392 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
393 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
394 frag_size, truesize);
396 page = (struct page *)page->private;
401 give_pages(rq, page);
406 static void virtnet_xdp_flush(struct net_device *dev)
408 struct virtnet_info *vi = netdev_priv(dev);
409 struct send_queue *sq;
412 qp = vi->curr_queue_pairs - vi->xdp_queue_pairs + smp_processor_id();
415 virtqueue_kick(sq->vq);
418 static int __virtnet_xdp_xmit(struct virtnet_info *vi,
419 struct xdp_buff *xdp)
421 struct virtio_net_hdr_mrg_rxbuf *hdr;
422 struct xdp_frame *xdpf, *xdpf_sent;
423 struct send_queue *sq;
428 qp = vi->curr_queue_pairs - vi->xdp_queue_pairs + smp_processor_id();
431 /* Free up any pending old buffers before queueing new ones. */
432 while ((xdpf_sent = virtqueue_get_buf(sq->vq, &len)) != NULL)
433 xdp_return_frame(xdpf_sent->data, &xdpf_sent->mem);
435 xdpf = convert_to_xdp_frame(xdp);
439 /* virtqueue want to use data area in-front of packet */
440 if (unlikely(xdpf->metasize > 0))
443 if (unlikely(xdpf->headroom < vi->hdr_len))
446 /* Make room for virtqueue hdr (also change xdpf->headroom?) */
447 xdpf->data -= vi->hdr_len;
448 /* Zero header and leave csum up to XDP layers */
450 memset(hdr, 0, vi->hdr_len);
451 xdpf->len += vi->hdr_len;
453 sg_init_one(sq->sg, xdpf->data, xdpf->len);
455 err = virtqueue_add_outbuf(sq->vq, sq->sg, 1, xdpf, GFP_ATOMIC);
457 return -ENOSPC; /* Caller handle free/refcnt */
462 static int virtnet_xdp_xmit(struct net_device *dev, struct xdp_buff *xdp)
464 struct virtnet_info *vi = netdev_priv(dev);
465 struct receive_queue *rq = vi->rq;
466 struct bpf_prog *xdp_prog;
468 /* Only allow ndo_xdp_xmit if XDP is loaded on dev, as this
469 * indicate XDP resources have been successfully allocated.
471 xdp_prog = rcu_dereference(rq->xdp_prog);
475 return __virtnet_xdp_xmit(vi, xdp);
478 static unsigned int virtnet_get_headroom(struct virtnet_info *vi)
480 return vi->xdp_queue_pairs ? VIRTIO_XDP_HEADROOM : 0;
483 /* We copy the packet for XDP in the following cases:
485 * 1) Packet is scattered across multiple rx buffers.
486 * 2) Headroom space is insufficient.
488 * This is inefficient but it's a temporary condition that
489 * we hit right after XDP is enabled and until queue is refilled
490 * with large buffers with sufficient headroom - so it should affect
491 * at most queue size packets.
492 * Afterwards, the conditions to enable
493 * XDP should preclude the underlying device from sending packets
494 * across multiple buffers (num_buf > 1), and we make sure buffers
495 * have enough headroom.
497 static struct page *xdp_linearize_page(struct receive_queue *rq,
504 struct page *page = alloc_page(GFP_ATOMIC);
509 memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
513 int tailroom = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
518 buf = virtqueue_get_buf(rq->vq, &buflen);
522 p = virt_to_head_page(buf);
523 off = buf - page_address(p);
525 /* guard against a misconfigured or uncooperative backend that
526 * is sending packet larger than the MTU.
528 if ((page_off + buflen + tailroom) > PAGE_SIZE) {
533 memcpy(page_address(page) + page_off,
534 page_address(p) + off, buflen);
539 /* Headroom does not contribute to packet length */
540 *len = page_off - VIRTIO_XDP_HEADROOM;
543 __free_pages(page, 0);
547 static struct sk_buff *receive_small(struct net_device *dev,
548 struct virtnet_info *vi,
549 struct receive_queue *rq,
550 void *buf, void *ctx,
555 struct bpf_prog *xdp_prog;
556 unsigned int xdp_headroom = (unsigned long)ctx;
557 unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom;
558 unsigned int headroom = vi->hdr_len + header_offset;
559 unsigned int buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
560 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
561 struct page *page = virt_to_head_page(buf);
562 unsigned int delta = 0;
563 struct page *xdp_page;
569 xdp_prog = rcu_dereference(rq->xdp_prog);
571 struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset;
576 if (unlikely(hdr->hdr.gso_type))
579 if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) {
580 int offset = buf - page_address(page) + header_offset;
581 unsigned int tlen = len + vi->hdr_len;
584 xdp_headroom = virtnet_get_headroom(vi);
585 header_offset = VIRTNET_RX_PAD + xdp_headroom;
586 headroom = vi->hdr_len + header_offset;
587 buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
588 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
589 xdp_page = xdp_linearize_page(rq, &num_buf, page,
590 offset, header_offset,
595 buf = page_address(xdp_page);
600 xdp.data_hard_start = buf + VIRTNET_RX_PAD + vi->hdr_len;
601 xdp.data = xdp.data_hard_start + xdp_headroom;
602 xdp_set_data_meta_invalid(&xdp);
603 xdp.data_end = xdp.data + len;
604 xdp.rxq = &rq->xdp_rxq;
605 orig_data = xdp.data;
606 act = bpf_prog_run_xdp(xdp_prog, &xdp);
610 /* Recalculate length in case bpf program changed it */
611 delta = orig_data - xdp.data;
614 err = __virtnet_xdp_xmit(vi, &xdp);
616 trace_xdp_exception(vi->dev, xdp_prog, act);
623 err = xdp_do_redirect(dev, &xdp, xdp_prog);
630 bpf_warn_invalid_xdp_action(act);
632 trace_xdp_exception(vi->dev, xdp_prog, act);
639 skb = build_skb(buf, buflen);
644 skb_reserve(skb, headroom - delta);
645 skb_put(skb, len + delta);
647 buf += header_offset;
648 memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
649 } /* keep zeroed vnet hdr since packet was changed by bpf */
656 dev->stats.rx_dropped++;
662 static struct sk_buff *receive_big(struct net_device *dev,
663 struct virtnet_info *vi,
664 struct receive_queue *rq,
668 struct page *page = buf;
669 struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len, PAGE_SIZE);
677 dev->stats.rx_dropped++;
678 give_pages(rq, page);
682 static struct sk_buff *receive_mergeable(struct net_device *dev,
683 struct virtnet_info *vi,
684 struct receive_queue *rq,
690 struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
691 u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
692 struct page *page = virt_to_head_page(buf);
693 int offset = buf - page_address(page);
694 struct sk_buff *head_skb, *curr_skb;
695 struct bpf_prog *xdp_prog;
696 unsigned int truesize;
697 unsigned int headroom = mergeable_ctx_to_headroom(ctx);
703 xdp_prog = rcu_dereference(rq->xdp_prog);
705 struct page *xdp_page;
710 /* This happens when rx buffer size is underestimated
711 * or headroom is not enough because of the buffer
712 * was refilled before XDP is set. This should only
713 * happen for the first several packets, so we don't
714 * care much about its performance.
716 if (unlikely(num_buf > 1 ||
717 headroom < virtnet_get_headroom(vi))) {
718 /* linearize data for XDP */
719 xdp_page = xdp_linearize_page(rq, &num_buf,
725 offset = VIRTIO_XDP_HEADROOM;
730 /* Transient failure which in theory could occur if
731 * in-flight packets from before XDP was enabled reach
732 * the receive path after XDP is loaded. In practice I
733 * was not able to create this condition.
735 if (unlikely(hdr->hdr.gso_type))
738 /* Allow consuming headroom but reserve enough space to push
739 * the descriptor on if we get an XDP_TX return code.
741 data = page_address(xdp_page) + offset;
742 xdp.data_hard_start = data - VIRTIO_XDP_HEADROOM + vi->hdr_len;
743 xdp.data = data + vi->hdr_len;
744 xdp_set_data_meta_invalid(&xdp);
745 xdp.data_end = xdp.data + (len - vi->hdr_len);
746 xdp.rxq = &rq->xdp_rxq;
748 act = bpf_prog_run_xdp(xdp_prog, &xdp);
752 /* recalculate offset to account for any header
753 * adjustments. Note other cases do not build an
754 * skb and avoid using offset
757 page_address(xdp_page) - vi->hdr_len;
759 /* We can only create skb based on xdp_page. */
760 if (unlikely(xdp_page != page)) {
763 head_skb = page_to_skb(vi, rq, xdp_page,
764 offset, len, PAGE_SIZE);
769 err = __virtnet_xdp_xmit(vi, &xdp);
771 trace_xdp_exception(vi->dev, xdp_prog, act);
772 if (unlikely(xdp_page != page))
777 if (unlikely(xdp_page != page))
782 err = xdp_do_redirect(dev, &xdp, xdp_prog);
784 if (unlikely(xdp_page != page))
789 if (unlikely(xdp_page != page))
794 bpf_warn_invalid_xdp_action(act);
796 trace_xdp_exception(vi->dev, xdp_prog, act);
798 if (unlikely(xdp_page != page))
799 __free_pages(xdp_page, 0);
805 truesize = mergeable_ctx_to_truesize(ctx);
806 if (unlikely(len > truesize)) {
807 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
808 dev->name, len, (unsigned long)ctx);
809 dev->stats.rx_length_errors++;
813 head_skb = page_to_skb(vi, rq, page, offset, len, truesize);
816 if (unlikely(!curr_skb))
821 buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx);
822 if (unlikely(!buf)) {
823 pr_debug("%s: rx error: %d buffers out of %d missing\n",
825 virtio16_to_cpu(vi->vdev,
827 dev->stats.rx_length_errors++;
831 page = virt_to_head_page(buf);
833 truesize = mergeable_ctx_to_truesize(ctx);
834 if (unlikely(len > truesize)) {
835 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
836 dev->name, len, (unsigned long)ctx);
837 dev->stats.rx_length_errors++;
841 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
842 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
843 struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
847 if (curr_skb == head_skb)
848 skb_shinfo(curr_skb)->frag_list = nskb;
850 curr_skb->next = nskb;
852 head_skb->truesize += nskb->truesize;
855 if (curr_skb != head_skb) {
856 head_skb->data_len += len;
857 head_skb->len += len;
858 head_skb->truesize += truesize;
860 offset = buf - page_address(page);
861 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
863 skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
866 skb_add_rx_frag(curr_skb, num_skb_frags, page,
867 offset, len, truesize);
871 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
879 buf = virtqueue_get_buf(rq->vq, &len);
880 if (unlikely(!buf)) {
881 pr_debug("%s: rx error: %d buffers missing\n",
883 dev->stats.rx_length_errors++;
886 page = virt_to_head_page(buf);
890 dev->stats.rx_dropped++;
891 dev_kfree_skb(head_skb);
896 static int receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
897 void *buf, unsigned int len, void **ctx, bool *xdp_xmit)
899 struct net_device *dev = vi->dev;
901 struct virtio_net_hdr_mrg_rxbuf *hdr;
904 if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
905 pr_debug("%s: short packet %i\n", dev->name, len);
906 dev->stats.rx_length_errors++;
907 if (vi->mergeable_rx_bufs) {
908 put_page(virt_to_head_page(buf));
909 } else if (vi->big_packets) {
912 put_page(virt_to_head_page(buf));
917 if (vi->mergeable_rx_bufs)
918 skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit);
919 else if (vi->big_packets)
920 skb = receive_big(dev, vi, rq, buf, len);
922 skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit);
927 hdr = skb_vnet_hdr(skb);
931 if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
932 skb->ip_summed = CHECKSUM_UNNECESSARY;
934 if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
935 virtio_is_little_endian(vi->vdev))) {
936 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
937 dev->name, hdr->hdr.gso_type,
942 skb->protocol = eth_type_trans(skb, dev);
943 pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
944 ntohs(skb->protocol), skb->len, skb->pkt_type);
946 napi_gro_receive(&rq->napi, skb);
950 dev->stats.rx_frame_errors++;
955 /* Unlike mergeable buffers, all buffers are allocated to the
956 * same size, except for the headroom. For this reason we do
957 * not need to use mergeable_len_to_ctx here - it is enough
958 * to store the headroom as the context ignoring the truesize.
960 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
963 struct page_frag *alloc_frag = &rq->alloc_frag;
965 unsigned int xdp_headroom = virtnet_get_headroom(vi);
966 void *ctx = (void *)(unsigned long)xdp_headroom;
967 int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
970 len = SKB_DATA_ALIGN(len) +
971 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
972 if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
975 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
976 get_page(alloc_frag->page);
977 alloc_frag->offset += len;
978 sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom,
979 vi->hdr_len + GOOD_PACKET_LEN);
980 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
982 put_page(virt_to_head_page(buf));
986 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
989 struct page *first, *list = NULL;
993 sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
995 /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
996 for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
997 first = get_a_page(rq, gfp);
1000 give_pages(rq, list);
1003 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
1005 /* chain new page in list head to match sg */
1006 first->private = (unsigned long)list;
1010 first = get_a_page(rq, gfp);
1012 give_pages(rq, list);
1015 p = page_address(first);
1017 /* rq->sg[0], rq->sg[1] share the same page */
1018 /* a separated rq->sg[0] for header - required in case !any_header_sg */
1019 sg_set_buf(&rq->sg[0], p, vi->hdr_len);
1021 /* rq->sg[1] for data packet, from offset */
1022 offset = sizeof(struct padded_vnet_hdr);
1023 sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
1025 /* chain first in list head */
1026 first->private = (unsigned long)list;
1027 err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
1030 give_pages(rq, first);
1035 static unsigned int get_mergeable_buf_len(struct receive_queue *rq,
1036 struct ewma_pkt_len *avg_pkt_len,
1039 const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1043 return PAGE_SIZE - room;
1045 len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
1046 rq->min_buf_len, PAGE_SIZE - hdr_len);
1048 return ALIGN(len, L1_CACHE_BYTES);
1051 static int add_recvbuf_mergeable(struct virtnet_info *vi,
1052 struct receive_queue *rq, gfp_t gfp)
1054 struct page_frag *alloc_frag = &rq->alloc_frag;
1055 unsigned int headroom = virtnet_get_headroom(vi);
1056 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1057 unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
1061 unsigned int len, hole;
1063 /* Extra tailroom is needed to satisfy XDP's assumption. This
1064 * means rx frags coalescing won't work, but consider we've
1065 * disabled GSO for XDP, it won't be a big issue.
1067 len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len, room);
1068 if (unlikely(!skb_page_frag_refill(len + room, alloc_frag, gfp)))
1071 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1072 buf += headroom; /* advance address leaving hole at front of pkt */
1073 get_page(alloc_frag->page);
1074 alloc_frag->offset += len + room;
1075 hole = alloc_frag->size - alloc_frag->offset;
1076 if (hole < len + room) {
1077 /* To avoid internal fragmentation, if there is very likely not
1078 * enough space for another buffer, add the remaining space to
1079 * the current buffer.
1082 alloc_frag->offset += hole;
1085 sg_init_one(rq->sg, buf, len);
1086 ctx = mergeable_len_to_ctx(len, headroom);
1087 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1089 put_page(virt_to_head_page(buf));
1095 * Returns false if we couldn't fill entirely (OOM).
1097 * Normally run in the receive path, but can also be run from ndo_open
1098 * before we're receiving packets, or from refill_work which is
1099 * careful to disable receiving (using napi_disable).
1101 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
1108 if (vi->mergeable_rx_bufs)
1109 err = add_recvbuf_mergeable(vi, rq, gfp);
1110 else if (vi->big_packets)
1111 err = add_recvbuf_big(vi, rq, gfp);
1113 err = add_recvbuf_small(vi, rq, gfp);
1115 oom = err == -ENOMEM;
1118 } while (rq->vq->num_free);
1119 virtqueue_kick(rq->vq);
1123 static void skb_recv_done(struct virtqueue *rvq)
1125 struct virtnet_info *vi = rvq->vdev->priv;
1126 struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
1128 virtqueue_napi_schedule(&rq->napi, rvq);
1131 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
1135 /* If all buffers were filled by other side before we napi_enabled, we
1136 * won't get another interrupt, so process any outstanding packets now.
1137 * Call local_bh_enable after to trigger softIRQ processing.
1140 virtqueue_napi_schedule(napi, vq);
1144 static void virtnet_napi_tx_enable(struct virtnet_info *vi,
1145 struct virtqueue *vq,
1146 struct napi_struct *napi)
1151 /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
1152 * enable the feature if this is likely affine with the transmit path.
1154 if (!vi->affinity_hint_set) {
1159 return virtnet_napi_enable(vq, napi);
1162 static void virtnet_napi_tx_disable(struct napi_struct *napi)
1168 static void refill_work(struct work_struct *work)
1170 struct virtnet_info *vi =
1171 container_of(work, struct virtnet_info, refill.work);
1175 for (i = 0; i < vi->curr_queue_pairs; i++) {
1176 struct receive_queue *rq = &vi->rq[i];
1178 napi_disable(&rq->napi);
1179 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
1180 virtnet_napi_enable(rq->vq, &rq->napi);
1182 /* In theory, this can happen: if we don't get any buffers in
1183 * we will *never* try to fill again.
1186 schedule_delayed_work(&vi->refill, HZ/2);
1190 static int virtnet_receive(struct receive_queue *rq, int budget, bool *xdp_xmit)
1192 struct virtnet_info *vi = rq->vq->vdev->priv;
1193 unsigned int len, received = 0, bytes = 0;
1196 if (!vi->big_packets || vi->mergeable_rx_bufs) {
1199 while (received < budget &&
1200 (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) {
1201 bytes += receive_buf(vi, rq, buf, len, ctx, xdp_xmit);
1205 while (received < budget &&
1206 (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
1207 bytes += receive_buf(vi, rq, buf, len, NULL, xdp_xmit);
1212 if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) {
1213 if (!try_fill_recv(vi, rq, GFP_ATOMIC))
1214 schedule_delayed_work(&vi->refill, 0);
1217 u64_stats_update_begin(&rq->stats.syncp);
1218 rq->stats.bytes += bytes;
1219 rq->stats.packets += received;
1220 u64_stats_update_end(&rq->stats.syncp);
1225 static void free_old_xmit_skbs(struct send_queue *sq)
1227 struct sk_buff *skb;
1229 unsigned int packets = 0;
1230 unsigned int bytes = 0;
1232 while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
1233 pr_debug("Sent skb %p\n", skb);
1238 dev_consume_skb_any(skb);
1241 /* Avoid overhead when no packets have been processed
1242 * happens when called speculatively from start_xmit.
1247 u64_stats_update_begin(&sq->stats.syncp);
1248 sq->stats.bytes += bytes;
1249 sq->stats.packets += packets;
1250 u64_stats_update_end(&sq->stats.syncp);
1253 static void virtnet_poll_cleantx(struct receive_queue *rq)
1255 struct virtnet_info *vi = rq->vq->vdev->priv;
1256 unsigned int index = vq2rxq(rq->vq);
1257 struct send_queue *sq = &vi->sq[index];
1258 struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index);
1260 if (!sq->napi.weight)
1263 if (__netif_tx_trylock(txq)) {
1264 free_old_xmit_skbs(sq);
1265 __netif_tx_unlock(txq);
1268 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1269 netif_tx_wake_queue(txq);
1272 static int virtnet_poll(struct napi_struct *napi, int budget)
1274 struct receive_queue *rq =
1275 container_of(napi, struct receive_queue, napi);
1276 unsigned int received;
1277 bool xdp_xmit = false;
1279 virtnet_poll_cleantx(rq);
1281 received = virtnet_receive(rq, budget, &xdp_xmit);
1283 /* Out of packets? */
1284 if (received < budget)
1285 virtqueue_napi_complete(napi, rq->vq, received);
1293 static int virtnet_open(struct net_device *dev)
1295 struct virtnet_info *vi = netdev_priv(dev);
1298 for (i = 0; i < vi->max_queue_pairs; i++) {
1299 if (i < vi->curr_queue_pairs)
1300 /* Make sure we have some buffers: if oom use wq. */
1301 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1302 schedule_delayed_work(&vi->refill, 0);
1304 err = xdp_rxq_info_reg(&vi->rq[i].xdp_rxq, dev, i);
1308 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1309 virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi);
1315 static int virtnet_poll_tx(struct napi_struct *napi, int budget)
1317 struct send_queue *sq = container_of(napi, struct send_queue, napi);
1318 struct virtnet_info *vi = sq->vq->vdev->priv;
1319 struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, vq2txq(sq->vq));
1321 __netif_tx_lock(txq, raw_smp_processor_id());
1322 free_old_xmit_skbs(sq);
1323 __netif_tx_unlock(txq);
1325 virtqueue_napi_complete(napi, sq->vq, 0);
1327 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1328 netif_tx_wake_queue(txq);
1333 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
1335 struct virtio_net_hdr_mrg_rxbuf *hdr;
1336 const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
1337 struct virtnet_info *vi = sq->vq->vdev->priv;
1339 unsigned hdr_len = vi->hdr_len;
1342 pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
1344 can_push = vi->any_header_sg &&
1345 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
1346 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
1347 /* Even if we can, don't push here yet as this would skew
1348 * csum_start offset below. */
1350 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1352 hdr = skb_vnet_hdr(skb);
1354 if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1355 virtio_is_little_endian(vi->vdev), false))
1358 if (vi->mergeable_rx_bufs)
1359 hdr->num_buffers = 0;
1361 sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1363 __skb_push(skb, hdr_len);
1364 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1365 if (unlikely(num_sg < 0))
1367 /* Pull header back to avoid skew in tx bytes calculations. */
1368 __skb_pull(skb, hdr_len);
1370 sg_set_buf(sq->sg, hdr, hdr_len);
1371 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
1372 if (unlikely(num_sg < 0))
1376 return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
1379 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
1381 struct virtnet_info *vi = netdev_priv(dev);
1382 int qnum = skb_get_queue_mapping(skb);
1383 struct send_queue *sq = &vi->sq[qnum];
1385 struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1386 bool kick = !skb->xmit_more;
1387 bool use_napi = sq->napi.weight;
1389 /* Free up any pending old buffers before queueing new ones. */
1390 free_old_xmit_skbs(sq);
1392 if (use_napi && kick)
1393 virtqueue_enable_cb_delayed(sq->vq);
1395 /* timestamp packet in software */
1396 skb_tx_timestamp(skb);
1398 /* Try to transmit */
1399 err = xmit_skb(sq, skb);
1401 /* This should not happen! */
1402 if (unlikely(err)) {
1403 dev->stats.tx_fifo_errors++;
1404 if (net_ratelimit())
1406 "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
1407 dev->stats.tx_dropped++;
1408 dev_kfree_skb_any(skb);
1409 return NETDEV_TX_OK;
1412 /* Don't wait up for transmitted skbs to be freed. */
1418 /* If running out of space, stop queue to avoid getting packets that we
1419 * are then unable to transmit.
1420 * An alternative would be to force queuing layer to requeue the skb by
1421 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1422 * returned in a normal path of operation: it means that driver is not
1423 * maintaining the TX queue stop/start state properly, and causes
1424 * the stack to do a non-trivial amount of useless work.
1425 * Since most packets only take 1 or 2 ring slots, stopping the queue
1426 * early means 16 slots are typically wasted.
1428 if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
1429 netif_stop_subqueue(dev, qnum);
1431 unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
1432 /* More just got used, free them then recheck. */
1433 free_old_xmit_skbs(sq);
1434 if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
1435 netif_start_subqueue(dev, qnum);
1436 virtqueue_disable_cb(sq->vq);
1441 if (kick || netif_xmit_stopped(txq))
1442 virtqueue_kick(sq->vq);
1444 return NETDEV_TX_OK;
1448 * Send command via the control virtqueue and check status. Commands
1449 * supported by the hypervisor, as indicated by feature bits, should
1450 * never fail unless improperly formatted.
1452 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
1453 struct scatterlist *out)
1455 struct scatterlist *sgs[4], hdr, stat;
1456 unsigned out_num = 0, tmp;
1458 /* Caller should know better */
1459 BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
1461 vi->ctrl_status = ~0;
1462 vi->ctrl_hdr.class = class;
1463 vi->ctrl_hdr.cmd = cmd;
1465 sg_init_one(&hdr, &vi->ctrl_hdr, sizeof(vi->ctrl_hdr));
1466 sgs[out_num++] = &hdr;
1469 sgs[out_num++] = out;
1471 /* Add return status. */
1472 sg_init_one(&stat, &vi->ctrl_status, sizeof(vi->ctrl_status));
1473 sgs[out_num] = &stat;
1475 BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1476 virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1478 if (unlikely(!virtqueue_kick(vi->cvq)))
1479 return vi->ctrl_status == VIRTIO_NET_OK;
1481 /* Spin for a response, the kick causes an ioport write, trapping
1482 * into the hypervisor, so the request should be handled immediately.
1484 while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1485 !virtqueue_is_broken(vi->cvq))
1488 return vi->ctrl_status == VIRTIO_NET_OK;
1491 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1493 struct virtnet_info *vi = netdev_priv(dev);
1494 struct virtio_device *vdev = vi->vdev;
1496 struct sockaddr *addr;
1497 struct scatterlist sg;
1499 addr = kmemdup(p, sizeof(*addr), GFP_KERNEL);
1503 ret = eth_prepare_mac_addr_change(dev, addr);
1507 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1508 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1509 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1510 VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1511 dev_warn(&vdev->dev,
1512 "Failed to set mac address by vq command.\n");
1516 } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1517 !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1520 /* Naturally, this has an atomicity problem. */
1521 for (i = 0; i < dev->addr_len; i++)
1522 virtio_cwrite8(vdev,
1523 offsetof(struct virtio_net_config, mac) +
1524 i, addr->sa_data[i]);
1527 eth_commit_mac_addr_change(dev, p);
1535 static void virtnet_stats(struct net_device *dev,
1536 struct rtnl_link_stats64 *tot)
1538 struct virtnet_info *vi = netdev_priv(dev);
1542 for (i = 0; i < vi->max_queue_pairs; i++) {
1543 u64 tpackets, tbytes, rpackets, rbytes;
1544 struct receive_queue *rq = &vi->rq[i];
1545 struct send_queue *sq = &vi->sq[i];
1548 start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
1549 tpackets = sq->stats.packets;
1550 tbytes = sq->stats.bytes;
1551 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
1554 start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
1555 rpackets = rq->stats.packets;
1556 rbytes = rq->stats.bytes;
1557 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
1559 tot->rx_packets += rpackets;
1560 tot->tx_packets += tpackets;
1561 tot->rx_bytes += rbytes;
1562 tot->tx_bytes += tbytes;
1565 tot->tx_dropped = dev->stats.tx_dropped;
1566 tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1567 tot->rx_dropped = dev->stats.rx_dropped;
1568 tot->rx_length_errors = dev->stats.rx_length_errors;
1569 tot->rx_frame_errors = dev->stats.rx_frame_errors;
1572 #ifdef CONFIG_NET_POLL_CONTROLLER
1573 static void virtnet_netpoll(struct net_device *dev)
1575 struct virtnet_info *vi = netdev_priv(dev);
1578 for (i = 0; i < vi->curr_queue_pairs; i++)
1579 napi_schedule(&vi->rq[i].napi);
1583 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1586 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1587 VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1588 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1592 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1594 struct scatterlist sg;
1595 struct net_device *dev = vi->dev;
1597 if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1600 vi->ctrl_mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1601 sg_init_one(&sg, &vi->ctrl_mq, sizeof(vi->ctrl_mq));
1603 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1604 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1605 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1609 vi->curr_queue_pairs = queue_pairs;
1610 /* virtnet_open() will refill when device is going to up. */
1611 if (dev->flags & IFF_UP)
1612 schedule_delayed_work(&vi->refill, 0);
1618 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1623 err = _virtnet_set_queues(vi, queue_pairs);
1628 static int virtnet_close(struct net_device *dev)
1630 struct virtnet_info *vi = netdev_priv(dev);
1633 /* Make sure refill_work doesn't re-enable napi! */
1634 cancel_delayed_work_sync(&vi->refill);
1636 for (i = 0; i < vi->max_queue_pairs; i++) {
1637 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1638 napi_disable(&vi->rq[i].napi);
1639 virtnet_napi_tx_disable(&vi->sq[i].napi);
1645 static void virtnet_set_rx_mode(struct net_device *dev)
1647 struct virtnet_info *vi = netdev_priv(dev);
1648 struct scatterlist sg[2];
1649 struct virtio_net_ctrl_mac *mac_data;
1650 struct netdev_hw_addr *ha;
1656 /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1657 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1660 vi->ctrl_promisc = ((dev->flags & IFF_PROMISC) != 0);
1661 vi->ctrl_allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1663 sg_init_one(sg, &vi->ctrl_promisc, sizeof(vi->ctrl_promisc));
1665 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1666 VIRTIO_NET_CTRL_RX_PROMISC, sg))
1667 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1668 vi->ctrl_promisc ? "en" : "dis");
1670 sg_init_one(sg, &vi->ctrl_allmulti, sizeof(vi->ctrl_allmulti));
1672 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1673 VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1674 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1675 vi->ctrl_allmulti ? "en" : "dis");
1677 uc_count = netdev_uc_count(dev);
1678 mc_count = netdev_mc_count(dev);
1679 /* MAC filter - use one buffer for both lists */
1680 buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1681 (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1686 sg_init_table(sg, 2);
1688 /* Store the unicast list and count in the front of the buffer */
1689 mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1691 netdev_for_each_uc_addr(ha, dev)
1692 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1694 sg_set_buf(&sg[0], mac_data,
1695 sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1697 /* multicast list and count fill the end */
1698 mac_data = (void *)&mac_data->macs[uc_count][0];
1700 mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1702 netdev_for_each_mc_addr(ha, dev)
1703 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1705 sg_set_buf(&sg[1], mac_data,
1706 sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1708 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1709 VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1710 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1715 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1716 __be16 proto, u16 vid)
1718 struct virtnet_info *vi = netdev_priv(dev);
1719 struct scatterlist sg;
1722 sg_init_one(&sg, &vi->ctrl_vid, sizeof(vi->ctrl_vid));
1724 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1725 VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1726 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1730 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1731 __be16 proto, u16 vid)
1733 struct virtnet_info *vi = netdev_priv(dev);
1734 struct scatterlist sg;
1737 sg_init_one(&sg, &vi->ctrl_vid, sizeof(vi->ctrl_vid));
1739 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1740 VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1741 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1745 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1749 if (vi->affinity_hint_set) {
1750 for (i = 0; i < vi->max_queue_pairs; i++) {
1751 virtqueue_set_affinity(vi->rq[i].vq, -1);
1752 virtqueue_set_affinity(vi->sq[i].vq, -1);
1755 vi->affinity_hint_set = false;
1759 static void virtnet_set_affinity(struct virtnet_info *vi)
1764 /* In multiqueue mode, when the number of cpu is equal to the number of
1765 * queue pairs, we let the queue pairs to be private to one cpu by
1766 * setting the affinity hint to eliminate the contention.
1768 if (vi->curr_queue_pairs == 1 ||
1769 vi->max_queue_pairs != num_online_cpus()) {
1770 virtnet_clean_affinity(vi, -1);
1775 for_each_online_cpu(cpu) {
1776 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1777 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1778 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
1782 vi->affinity_hint_set = true;
1785 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
1787 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1789 virtnet_set_affinity(vi);
1793 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
1795 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1797 virtnet_set_affinity(vi);
1801 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
1803 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1806 virtnet_clean_affinity(vi, cpu);
1810 static enum cpuhp_state virtionet_online;
1812 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
1816 ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
1819 ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1823 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1827 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
1829 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1830 cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1834 static void virtnet_get_ringparam(struct net_device *dev,
1835 struct ethtool_ringparam *ring)
1837 struct virtnet_info *vi = netdev_priv(dev);
1839 ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1840 ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1841 ring->rx_pending = ring->rx_max_pending;
1842 ring->tx_pending = ring->tx_max_pending;
1846 static void virtnet_get_drvinfo(struct net_device *dev,
1847 struct ethtool_drvinfo *info)
1849 struct virtnet_info *vi = netdev_priv(dev);
1850 struct virtio_device *vdev = vi->vdev;
1852 strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1853 strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1854 strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1858 /* TODO: Eliminate OOO packets during switching */
1859 static int virtnet_set_channels(struct net_device *dev,
1860 struct ethtool_channels *channels)
1862 struct virtnet_info *vi = netdev_priv(dev);
1863 u16 queue_pairs = channels->combined_count;
1866 /* We don't support separate rx/tx channels.
1867 * We don't allow setting 'other' channels.
1869 if (channels->rx_count || channels->tx_count || channels->other_count)
1872 if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
1875 /* For now we don't support modifying channels while XDP is loaded
1876 * also when XDP is loaded all RX queues have XDP programs so we only
1877 * need to check a single RX queue.
1879 if (vi->rq[0].xdp_prog)
1883 err = _virtnet_set_queues(vi, queue_pairs);
1885 netif_set_real_num_tx_queues(dev, queue_pairs);
1886 netif_set_real_num_rx_queues(dev, queue_pairs);
1888 virtnet_set_affinity(vi);
1895 static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *data)
1897 struct virtnet_info *vi = netdev_priv(dev);
1898 char *p = (char *)data;
1901 switch (stringset) {
1903 for (i = 0; i < vi->curr_queue_pairs; i++) {
1904 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
1905 snprintf(p, ETH_GSTRING_LEN, "rx_queue_%u_%s",
1906 i, virtnet_rq_stats_desc[j].desc);
1907 p += ETH_GSTRING_LEN;
1911 for (i = 0; i < vi->curr_queue_pairs; i++) {
1912 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
1913 snprintf(p, ETH_GSTRING_LEN, "tx_queue_%u_%s",
1914 i, virtnet_sq_stats_desc[j].desc);
1915 p += ETH_GSTRING_LEN;
1922 static int virtnet_get_sset_count(struct net_device *dev, int sset)
1924 struct virtnet_info *vi = netdev_priv(dev);
1928 return vi->curr_queue_pairs * (VIRTNET_RQ_STATS_LEN +
1929 VIRTNET_SQ_STATS_LEN);
1935 static void virtnet_get_ethtool_stats(struct net_device *dev,
1936 struct ethtool_stats *stats, u64 *data)
1938 struct virtnet_info *vi = netdev_priv(dev);
1939 unsigned int idx = 0, start, i, j;
1940 const u8 *stats_base;
1943 for (i = 0; i < vi->curr_queue_pairs; i++) {
1944 struct receive_queue *rq = &vi->rq[i];
1946 stats_base = (u8 *)&rq->stats;
1948 start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
1949 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
1950 offset = virtnet_rq_stats_desc[j].offset;
1951 data[idx + j] = *(u64 *)(stats_base + offset);
1953 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
1954 idx += VIRTNET_RQ_STATS_LEN;
1957 for (i = 0; i < vi->curr_queue_pairs; i++) {
1958 struct send_queue *sq = &vi->sq[i];
1960 stats_base = (u8 *)&sq->stats;
1962 start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
1963 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
1964 offset = virtnet_sq_stats_desc[j].offset;
1965 data[idx + j] = *(u64 *)(stats_base + offset);
1967 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
1968 idx += VIRTNET_SQ_STATS_LEN;
1972 static void virtnet_get_channels(struct net_device *dev,
1973 struct ethtool_channels *channels)
1975 struct virtnet_info *vi = netdev_priv(dev);
1977 channels->combined_count = vi->curr_queue_pairs;
1978 channels->max_combined = vi->max_queue_pairs;
1979 channels->max_other = 0;
1980 channels->rx_count = 0;
1981 channels->tx_count = 0;
1982 channels->other_count = 0;
1985 /* Check if the user is trying to change anything besides speed/duplex */
1987 virtnet_validate_ethtool_cmd(const struct ethtool_link_ksettings *cmd)
1989 struct ethtool_link_ksettings diff1 = *cmd;
1990 struct ethtool_link_ksettings diff2 = {};
1992 /* cmd is always set so we need to clear it, validate the port type
1993 * and also without autonegotiation we can ignore advertising
1995 diff1.base.speed = 0;
1996 diff2.base.port = PORT_OTHER;
1997 ethtool_link_ksettings_zero_link_mode(&diff1, advertising);
1998 diff1.base.duplex = 0;
2000 diff1.base.link_mode_masks_nwords = 0;
2002 return !memcmp(&diff1.base, &diff2.base, sizeof(diff1.base)) &&
2003 bitmap_empty(diff1.link_modes.supported,
2004 __ETHTOOL_LINK_MODE_MASK_NBITS) &&
2005 bitmap_empty(diff1.link_modes.advertising,
2006 __ETHTOOL_LINK_MODE_MASK_NBITS) &&
2007 bitmap_empty(diff1.link_modes.lp_advertising,
2008 __ETHTOOL_LINK_MODE_MASK_NBITS);
2011 static int virtnet_set_link_ksettings(struct net_device *dev,
2012 const struct ethtool_link_ksettings *cmd)
2014 struct virtnet_info *vi = netdev_priv(dev);
2017 speed = cmd->base.speed;
2018 /* don't allow custom speed and duplex */
2019 if (!ethtool_validate_speed(speed) ||
2020 !ethtool_validate_duplex(cmd->base.duplex) ||
2021 !virtnet_validate_ethtool_cmd(cmd))
2024 vi->duplex = cmd->base.duplex;
2029 static int virtnet_get_link_ksettings(struct net_device *dev,
2030 struct ethtool_link_ksettings *cmd)
2032 struct virtnet_info *vi = netdev_priv(dev);
2034 cmd->base.speed = vi->speed;
2035 cmd->base.duplex = vi->duplex;
2036 cmd->base.port = PORT_OTHER;
2041 static void virtnet_init_settings(struct net_device *dev)
2043 struct virtnet_info *vi = netdev_priv(dev);
2045 vi->speed = SPEED_UNKNOWN;
2046 vi->duplex = DUPLEX_UNKNOWN;
2049 static void virtnet_update_settings(struct virtnet_info *vi)
2054 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_SPEED_DUPLEX))
2057 speed = virtio_cread32(vi->vdev, offsetof(struct virtio_net_config,
2059 if (ethtool_validate_speed(speed))
2061 duplex = virtio_cread8(vi->vdev, offsetof(struct virtio_net_config,
2063 if (ethtool_validate_duplex(duplex))
2064 vi->duplex = duplex;
2067 static const struct ethtool_ops virtnet_ethtool_ops = {
2068 .get_drvinfo = virtnet_get_drvinfo,
2069 .get_link = ethtool_op_get_link,
2070 .get_ringparam = virtnet_get_ringparam,
2071 .get_strings = virtnet_get_strings,
2072 .get_sset_count = virtnet_get_sset_count,
2073 .get_ethtool_stats = virtnet_get_ethtool_stats,
2074 .set_channels = virtnet_set_channels,
2075 .get_channels = virtnet_get_channels,
2076 .get_ts_info = ethtool_op_get_ts_info,
2077 .get_link_ksettings = virtnet_get_link_ksettings,
2078 .set_link_ksettings = virtnet_set_link_ksettings,
2081 static void virtnet_freeze_down(struct virtio_device *vdev)
2083 struct virtnet_info *vi = vdev->priv;
2086 /* Make sure no work handler is accessing the device */
2087 flush_work(&vi->config_work);
2089 netif_device_detach(vi->dev);
2090 netif_tx_disable(vi->dev);
2091 cancel_delayed_work_sync(&vi->refill);
2093 if (netif_running(vi->dev)) {
2094 for (i = 0; i < vi->max_queue_pairs; i++) {
2095 napi_disable(&vi->rq[i].napi);
2096 virtnet_napi_tx_disable(&vi->sq[i].napi);
2101 static int init_vqs(struct virtnet_info *vi);
2103 static int virtnet_restore_up(struct virtio_device *vdev)
2105 struct virtnet_info *vi = vdev->priv;
2112 virtio_device_ready(vdev);
2114 if (netif_running(vi->dev)) {
2115 for (i = 0; i < vi->curr_queue_pairs; i++)
2116 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
2117 schedule_delayed_work(&vi->refill, 0);
2119 for (i = 0; i < vi->max_queue_pairs; i++) {
2120 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2121 virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2126 netif_device_attach(vi->dev);
2130 static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads)
2132 struct scatterlist sg;
2133 vi->ctrl_offloads = cpu_to_virtio64(vi->vdev, offloads);
2135 sg_init_one(&sg, &vi->ctrl_offloads, sizeof(vi->ctrl_offloads));
2137 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
2138 VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) {
2139 dev_warn(&vi->dev->dev, "Fail to set guest offload. \n");
2146 static int virtnet_clear_guest_offloads(struct virtnet_info *vi)
2150 if (!vi->guest_offloads)
2153 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))
2154 offloads = 1ULL << VIRTIO_NET_F_GUEST_CSUM;
2156 return virtnet_set_guest_offloads(vi, offloads);
2159 static int virtnet_restore_guest_offloads(struct virtnet_info *vi)
2161 u64 offloads = vi->guest_offloads;
2163 if (!vi->guest_offloads)
2165 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))
2166 offloads |= 1ULL << VIRTIO_NET_F_GUEST_CSUM;
2168 return virtnet_set_guest_offloads(vi, offloads);
2171 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
2172 struct netlink_ext_ack *extack)
2174 unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr);
2175 struct virtnet_info *vi = netdev_priv(dev);
2176 struct bpf_prog *old_prog;
2177 u16 xdp_qp = 0, curr_qp;
2180 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)
2181 && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2182 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2183 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
2184 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO))) {
2185 NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing LRO, disable LRO first");
2189 if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
2190 NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required");
2194 if (dev->mtu > max_sz) {
2195 NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP");
2196 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz);
2200 curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
2202 xdp_qp = nr_cpu_ids;
2204 /* XDP requires extra queues for XDP_TX */
2205 if (curr_qp + xdp_qp > vi->max_queue_pairs) {
2206 NL_SET_ERR_MSG_MOD(extack, "Too few free TX rings available");
2207 netdev_warn(dev, "request %i queues but max is %i\n",
2208 curr_qp + xdp_qp, vi->max_queue_pairs);
2213 prog = bpf_prog_add(prog, vi->max_queue_pairs - 1);
2215 return PTR_ERR(prog);
2218 /* Make sure NAPI is not using any XDP TX queues for RX. */
2219 if (netif_running(dev))
2220 for (i = 0; i < vi->max_queue_pairs; i++)
2221 napi_disable(&vi->rq[i].napi);
2223 netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
2224 err = _virtnet_set_queues(vi, curr_qp + xdp_qp);
2227 vi->xdp_queue_pairs = xdp_qp;
2229 for (i = 0; i < vi->max_queue_pairs; i++) {
2230 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2231 rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2234 virtnet_clear_guest_offloads(vi);
2236 virtnet_restore_guest_offloads(vi);
2239 bpf_prog_put(old_prog);
2240 if (netif_running(dev))
2241 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2247 for (i = 0; i < vi->max_queue_pairs; i++)
2248 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2250 bpf_prog_sub(prog, vi->max_queue_pairs - 1);
2254 static u32 virtnet_xdp_query(struct net_device *dev)
2256 struct virtnet_info *vi = netdev_priv(dev);
2257 const struct bpf_prog *xdp_prog;
2260 for (i = 0; i < vi->max_queue_pairs; i++) {
2261 xdp_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2263 return xdp_prog->aux->id;
2268 static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp)
2270 switch (xdp->command) {
2271 case XDP_SETUP_PROG:
2272 return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
2273 case XDP_QUERY_PROG:
2274 xdp->prog_id = virtnet_xdp_query(dev);
2275 xdp->prog_attached = !!xdp->prog_id;
2282 static const struct net_device_ops virtnet_netdev = {
2283 .ndo_open = virtnet_open,
2284 .ndo_stop = virtnet_close,
2285 .ndo_start_xmit = start_xmit,
2286 .ndo_validate_addr = eth_validate_addr,
2287 .ndo_set_mac_address = virtnet_set_mac_address,
2288 .ndo_set_rx_mode = virtnet_set_rx_mode,
2289 .ndo_get_stats64 = virtnet_stats,
2290 .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
2291 .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
2292 #ifdef CONFIG_NET_POLL_CONTROLLER
2293 .ndo_poll_controller = virtnet_netpoll,
2295 .ndo_bpf = virtnet_xdp,
2296 .ndo_xdp_xmit = virtnet_xdp_xmit,
2297 .ndo_xdp_flush = virtnet_xdp_flush,
2298 .ndo_features_check = passthru_features_check,
2301 static void virtnet_config_changed_work(struct work_struct *work)
2303 struct virtnet_info *vi =
2304 container_of(work, struct virtnet_info, config_work);
2307 if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
2308 struct virtio_net_config, status, &v) < 0)
2311 if (v & VIRTIO_NET_S_ANNOUNCE) {
2312 netdev_notify_peers(vi->dev);
2313 virtnet_ack_link_announce(vi);
2316 /* Ignore unknown (future) status bits */
2317 v &= VIRTIO_NET_S_LINK_UP;
2319 if (vi->status == v)
2324 if (vi->status & VIRTIO_NET_S_LINK_UP) {
2325 virtnet_update_settings(vi);
2326 netif_carrier_on(vi->dev);
2327 netif_tx_wake_all_queues(vi->dev);
2329 netif_carrier_off(vi->dev);
2330 netif_tx_stop_all_queues(vi->dev);
2334 static void virtnet_config_changed(struct virtio_device *vdev)
2336 struct virtnet_info *vi = vdev->priv;
2338 schedule_work(&vi->config_work);
2341 static void virtnet_free_queues(struct virtnet_info *vi)
2345 for (i = 0; i < vi->max_queue_pairs; i++) {
2346 napi_hash_del(&vi->rq[i].napi);
2347 netif_napi_del(&vi->rq[i].napi);
2348 netif_napi_del(&vi->sq[i].napi);
2351 /* We called napi_hash_del() before netif_napi_del(),
2352 * we need to respect an RCU grace period before freeing vi->rq
2360 static void _free_receive_bufs(struct virtnet_info *vi)
2362 struct bpf_prog *old_prog;
2365 for (i = 0; i < vi->max_queue_pairs; i++) {
2366 while (vi->rq[i].pages)
2367 __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
2369 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2370 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
2372 bpf_prog_put(old_prog);
2376 static void free_receive_bufs(struct virtnet_info *vi)
2379 _free_receive_bufs(vi);
2383 static void free_receive_page_frags(struct virtnet_info *vi)
2386 for (i = 0; i < vi->max_queue_pairs; i++)
2387 if (vi->rq[i].alloc_frag.page)
2388 put_page(vi->rq[i].alloc_frag.page);
2391 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
2393 if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
2395 else if (q < vi->curr_queue_pairs)
2401 static void free_unused_bufs(struct virtnet_info *vi)
2406 for (i = 0; i < vi->max_queue_pairs; i++) {
2407 struct virtqueue *vq = vi->sq[i].vq;
2408 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2409 if (!is_xdp_raw_buffer_queue(vi, i))
2412 put_page(virt_to_head_page(buf));
2416 for (i = 0; i < vi->max_queue_pairs; i++) {
2417 struct virtqueue *vq = vi->rq[i].vq;
2419 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2420 if (vi->mergeable_rx_bufs) {
2421 put_page(virt_to_head_page(buf));
2422 } else if (vi->big_packets) {
2423 give_pages(&vi->rq[i], buf);
2425 put_page(virt_to_head_page(buf));
2431 static void virtnet_del_vqs(struct virtnet_info *vi)
2433 struct virtio_device *vdev = vi->vdev;
2435 virtnet_clean_affinity(vi, -1);
2437 vdev->config->del_vqs(vdev);
2439 virtnet_free_queues(vi);
2442 /* How large should a single buffer be so a queue full of these can fit at
2443 * least one full packet?
2444 * Logic below assumes the mergeable buffer header is used.
2446 static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq)
2448 const unsigned int hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2449 unsigned int rq_size = virtqueue_get_vring_size(vq);
2450 unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu;
2451 unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len;
2452 unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size);
2454 return max(max(min_buf_len, hdr_len) - hdr_len,
2455 (unsigned int)GOOD_PACKET_LEN);
2458 static int virtnet_find_vqs(struct virtnet_info *vi)
2460 vq_callback_t **callbacks;
2461 struct virtqueue **vqs;
2467 /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
2468 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
2469 * possible control vq.
2471 total_vqs = vi->max_queue_pairs * 2 +
2472 virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
2474 /* Allocate space for find_vqs parameters */
2475 vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
2478 callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
2481 names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
2484 if (!vi->big_packets || vi->mergeable_rx_bufs) {
2485 ctx = kzalloc(total_vqs * sizeof(*ctx), GFP_KERNEL);
2492 /* Parameters for control virtqueue, if any */
2494 callbacks[total_vqs - 1] = NULL;
2495 names[total_vqs - 1] = "control";
2498 /* Allocate/initialize parameters for send/receive virtqueues */
2499 for (i = 0; i < vi->max_queue_pairs; i++) {
2500 callbacks[rxq2vq(i)] = skb_recv_done;
2501 callbacks[txq2vq(i)] = skb_xmit_done;
2502 sprintf(vi->rq[i].name, "input.%d", i);
2503 sprintf(vi->sq[i].name, "output.%d", i);
2504 names[rxq2vq(i)] = vi->rq[i].name;
2505 names[txq2vq(i)] = vi->sq[i].name;
2507 ctx[rxq2vq(i)] = true;
2510 ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
2516 vi->cvq = vqs[total_vqs - 1];
2517 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
2518 vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
2521 for (i = 0; i < vi->max_queue_pairs; i++) {
2522 vi->rq[i].vq = vqs[rxq2vq(i)];
2523 vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq);
2524 vi->sq[i].vq = vqs[txq2vq(i)];
2546 static int virtnet_alloc_queues(struct virtnet_info *vi)
2550 vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
2553 vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
2557 INIT_DELAYED_WORK(&vi->refill, refill_work);
2558 for (i = 0; i < vi->max_queue_pairs; i++) {
2559 vi->rq[i].pages = NULL;
2560 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
2562 netif_tx_napi_add(vi->dev, &vi->sq[i].napi, virtnet_poll_tx,
2563 napi_tx ? napi_weight : 0);
2565 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
2566 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
2567 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
2569 u64_stats_init(&vi->rq[i].stats.syncp);
2570 u64_stats_init(&vi->sq[i].stats.syncp);
2581 static int init_vqs(struct virtnet_info *vi)
2585 /* Allocate send & receive queues */
2586 ret = virtnet_alloc_queues(vi);
2590 ret = virtnet_find_vqs(vi);
2595 virtnet_set_affinity(vi);
2601 virtnet_free_queues(vi);
2607 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
2610 struct virtnet_info *vi = netdev_priv(queue->dev);
2611 unsigned int queue_index = get_netdev_rx_queue_index(queue);
2612 unsigned int headroom = virtnet_get_headroom(vi);
2613 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
2614 struct ewma_pkt_len *avg;
2616 BUG_ON(queue_index >= vi->max_queue_pairs);
2617 avg = &vi->rq[queue_index].mrg_avg_pkt_len;
2618 return sprintf(buf, "%u\n",
2619 get_mergeable_buf_len(&vi->rq[queue_index], avg,
2620 SKB_DATA_ALIGN(headroom + tailroom)));
2623 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
2624 __ATTR_RO(mergeable_rx_buffer_size);
2626 static struct attribute *virtio_net_mrg_rx_attrs[] = {
2627 &mergeable_rx_buffer_size_attribute.attr,
2631 static const struct attribute_group virtio_net_mrg_rx_group = {
2632 .name = "virtio_net",
2633 .attrs = virtio_net_mrg_rx_attrs
2637 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
2639 const char *fname, const char *dname)
2641 if (!virtio_has_feature(vdev, fbit))
2644 dev_err(&vdev->dev, "device advertises feature %s but not %s",
2650 #define VIRTNET_FAIL_ON(vdev, fbit, dbit) \
2651 virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
2653 static bool virtnet_validate_features(struct virtio_device *vdev)
2655 if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
2656 (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
2657 "VIRTIO_NET_F_CTRL_VQ") ||
2658 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
2659 "VIRTIO_NET_F_CTRL_VQ") ||
2660 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
2661 "VIRTIO_NET_F_CTRL_VQ") ||
2662 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
2663 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
2664 "VIRTIO_NET_F_CTRL_VQ"))) {
2671 #define MIN_MTU ETH_MIN_MTU
2672 #define MAX_MTU ETH_MAX_MTU
2674 static int virtnet_validate(struct virtio_device *vdev)
2676 if (!vdev->config->get) {
2677 dev_err(&vdev->dev, "%s failure: config access disabled\n",
2682 if (!virtnet_validate_features(vdev))
2685 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2686 int mtu = virtio_cread16(vdev,
2687 offsetof(struct virtio_net_config,
2690 __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
2696 static int virtnet_probe(struct virtio_device *vdev)
2698 int i, err = -ENOMEM;
2699 struct net_device *dev;
2700 struct virtnet_info *vi;
2701 u16 max_queue_pairs;
2704 /* Find if host supports multiqueue virtio_net device */
2705 err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
2706 struct virtio_net_config,
2707 max_virtqueue_pairs, &max_queue_pairs);
2709 /* We need at least 2 queue's */
2710 if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
2711 max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
2712 !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2713 max_queue_pairs = 1;
2715 /* Allocate ourselves a network device with room for our info */
2716 dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
2720 /* Set up network device as normal. */
2721 dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
2722 dev->netdev_ops = &virtnet_netdev;
2723 dev->features = NETIF_F_HIGHDMA;
2725 dev->ethtool_ops = &virtnet_ethtool_ops;
2726 SET_NETDEV_DEV(dev, &vdev->dev);
2728 /* Do we support "hardware" checksums? */
2729 if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
2730 /* This opens up the world of extra features. */
2731 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2733 dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2735 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
2736 dev->hw_features |= NETIF_F_TSO
2737 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
2739 /* Individual feature bits: what can host handle? */
2740 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
2741 dev->hw_features |= NETIF_F_TSO;
2742 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
2743 dev->hw_features |= NETIF_F_TSO6;
2744 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
2745 dev->hw_features |= NETIF_F_TSO_ECN;
2747 dev->features |= NETIF_F_GSO_ROBUST;
2750 dev->features |= dev->hw_features & NETIF_F_ALL_TSO;
2751 /* (!csum && gso) case will be fixed by register_netdev() */
2753 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
2754 dev->features |= NETIF_F_RXCSUM;
2756 dev->vlan_features = dev->features;
2758 /* MTU range: 68 - 65535 */
2759 dev->min_mtu = MIN_MTU;
2760 dev->max_mtu = MAX_MTU;
2762 /* Configuration may specify what MAC to use. Otherwise random. */
2763 if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
2764 virtio_cread_bytes(vdev,
2765 offsetof(struct virtio_net_config, mac),
2766 dev->dev_addr, dev->addr_len);
2768 eth_hw_addr_random(dev);
2770 /* Set up our device-specific information */
2771 vi = netdev_priv(dev);
2776 INIT_WORK(&vi->config_work, virtnet_config_changed_work);
2778 /* If we can receive ANY GSO packets, we must allocate large ones. */
2779 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2780 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2781 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
2782 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
2783 vi->big_packets = true;
2785 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
2786 vi->mergeable_rx_bufs = true;
2788 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
2789 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2790 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2792 vi->hdr_len = sizeof(struct virtio_net_hdr);
2794 if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
2795 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2796 vi->any_header_sg = true;
2798 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2801 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2802 mtu = virtio_cread16(vdev,
2803 offsetof(struct virtio_net_config,
2805 if (mtu < dev->min_mtu) {
2806 /* Should never trigger: MTU was previously validated
2807 * in virtnet_validate.
2809 dev_err(&vdev->dev, "device MTU appears to have changed "
2810 "it is now %d < %d", mtu, dev->min_mtu);
2817 /* TODO: size buffers correctly in this case. */
2818 if (dev->mtu > ETH_DATA_LEN)
2819 vi->big_packets = true;
2822 if (vi->any_header_sg)
2823 dev->needed_headroom = vi->hdr_len;
2825 /* Enable multiqueue by default */
2826 if (num_online_cpus() >= max_queue_pairs)
2827 vi->curr_queue_pairs = max_queue_pairs;
2829 vi->curr_queue_pairs = num_online_cpus();
2830 vi->max_queue_pairs = max_queue_pairs;
2832 /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
2838 if (vi->mergeable_rx_bufs)
2839 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
2841 netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
2842 netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
2844 virtnet_init_settings(dev);
2846 err = register_netdev(dev);
2848 pr_debug("virtio_net: registering device failed\n");
2852 virtio_device_ready(vdev);
2854 err = virtnet_cpu_notif_add(vi);
2856 pr_debug("virtio_net: registering cpu notifier failed\n");
2857 goto free_unregister_netdev;
2860 virtnet_set_queues(vi, vi->curr_queue_pairs);
2862 /* Assume link up if device can't report link status,
2863 otherwise get link status from config. */
2864 netif_carrier_off(dev);
2865 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
2866 schedule_work(&vi->config_work);
2868 vi->status = VIRTIO_NET_S_LINK_UP;
2869 virtnet_update_settings(vi);
2870 netif_carrier_on(dev);
2873 for (i = 0; i < ARRAY_SIZE(guest_offloads); i++)
2874 if (virtio_has_feature(vi->vdev, guest_offloads[i]))
2875 set_bit(guest_offloads[i], &vi->guest_offloads);
2877 pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
2878 dev->name, max_queue_pairs);
2882 free_unregister_netdev:
2883 vi->vdev->config->reset(vdev);
2885 unregister_netdev(dev);
2887 cancel_delayed_work_sync(&vi->refill);
2888 free_receive_page_frags(vi);
2889 virtnet_del_vqs(vi);
2895 static void remove_vq_common(struct virtnet_info *vi)
2897 vi->vdev->config->reset(vi->vdev);
2899 /* Free unused buffers in both send and recv, if any. */
2900 free_unused_bufs(vi);
2902 free_receive_bufs(vi);
2904 free_receive_page_frags(vi);
2906 virtnet_del_vqs(vi);
2909 static void virtnet_remove(struct virtio_device *vdev)
2911 struct virtnet_info *vi = vdev->priv;
2913 virtnet_cpu_notif_remove(vi);
2915 /* Make sure no work handler is accessing the device. */
2916 flush_work(&vi->config_work);
2918 unregister_netdev(vi->dev);
2920 remove_vq_common(vi);
2922 free_netdev(vi->dev);
2925 static __maybe_unused int virtnet_freeze(struct virtio_device *vdev)
2927 struct virtnet_info *vi = vdev->priv;
2929 virtnet_cpu_notif_remove(vi);
2930 virtnet_freeze_down(vdev);
2931 remove_vq_common(vi);
2936 static __maybe_unused int virtnet_restore(struct virtio_device *vdev)
2938 struct virtnet_info *vi = vdev->priv;
2941 err = virtnet_restore_up(vdev);
2944 virtnet_set_queues(vi, vi->curr_queue_pairs);
2946 err = virtnet_cpu_notif_add(vi);
2953 static struct virtio_device_id id_table[] = {
2954 { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
2958 #define VIRTNET_FEATURES \
2959 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
2961 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
2962 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
2963 VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
2964 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
2965 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
2966 VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
2967 VIRTIO_NET_F_CTRL_MAC_ADDR, \
2968 VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \
2969 VIRTIO_NET_F_SPEED_DUPLEX
2971 static unsigned int features[] = {
2975 static unsigned int features_legacy[] = {
2978 VIRTIO_F_ANY_LAYOUT,
2981 static struct virtio_driver virtio_net_driver = {
2982 .feature_table = features,
2983 .feature_table_size = ARRAY_SIZE(features),
2984 .feature_table_legacy = features_legacy,
2985 .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
2986 .driver.name = KBUILD_MODNAME,
2987 .driver.owner = THIS_MODULE,
2988 .id_table = id_table,
2989 .validate = virtnet_validate,
2990 .probe = virtnet_probe,
2991 .remove = virtnet_remove,
2992 .config_changed = virtnet_config_changed,
2993 #ifdef CONFIG_PM_SLEEP
2994 .freeze = virtnet_freeze,
2995 .restore = virtnet_restore,
2999 static __init int virtio_net_driver_init(void)
3003 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
3005 virtnet_cpu_down_prep);
3008 virtionet_online = ret;
3009 ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
3010 NULL, virtnet_cpu_dead);
3014 ret = register_virtio_driver(&virtio_net_driver);
3019 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3021 cpuhp_remove_multi_state(virtionet_online);
3025 module_init(virtio_net_driver_init);
3027 static __exit void virtio_net_driver_exit(void)
3029 unregister_virtio_driver(&virtio_net_driver);
3030 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3031 cpuhp_remove_multi_state(virtionet_online);
3033 module_exit(virtio_net_driver_exit);
3035 MODULE_DEVICE_TABLE(virtio, id_table);
3036 MODULE_DESCRIPTION("Virtio network driver");
3037 MODULE_LICENSE("GPL");