1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* A network driver using virtio.
4 * Copyright 2007 Rusty Russell <rusty@rustcorp.com.au> IBM Corporation
7 #include <linux/netdevice.h>
8 #include <linux/etherdevice.h>
9 #include <linux/ethtool.h>
10 #include <linux/module.h>
11 #include <linux/virtio.h>
12 #include <linux/virtio_net.h>
13 #include <linux/bpf.h>
14 #include <linux/bpf_trace.h>
15 #include <linux/scatterlist.h>
16 #include <linux/if_vlan.h>
17 #include <linux/slab.h>
18 #include <linux/cpu.h>
19 #include <linux/average.h>
20 #include <linux/filter.h>
21 #include <linux/kernel.h>
22 #include <net/route.h>
24 #include <net/net_failover.h>
26 static int napi_weight = NAPI_POLL_WEIGHT;
27 module_param(napi_weight, int, 0444);
29 static bool csum = true, gso = true, napi_tx = true;
30 module_param(csum, bool, 0444);
31 module_param(gso, bool, 0444);
32 module_param(napi_tx, bool, 0644);
34 /* FIXME: MTU in config. */
35 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
36 #define GOOD_COPY_LEN 128
38 #define VIRTNET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
40 /* Amount of XDP headroom to prepend to packets for use by xdp_adjust_head */
41 #define VIRTIO_XDP_HEADROOM 256
43 /* Separating two types of XDP xmit */
44 #define VIRTIO_XDP_TX BIT(0)
45 #define VIRTIO_XDP_REDIR BIT(1)
47 #define VIRTIO_XDP_FLAG BIT(0)
49 /* RX packet size EWMA. The average packet size is used to determine the packet
50 * buffer size when refilling RX rings. As the entire RX ring may be refilled
51 * at once, the weight is chosen so that the EWMA will be insensitive to short-
52 * term, transient changes in packet size.
54 DECLARE_EWMA(pkt_len, 0, 64)
56 #define VIRTNET_DRIVER_VERSION "1.0.0"
58 static const unsigned long guest_offloads[] = {
59 VIRTIO_NET_F_GUEST_TSO4,
60 VIRTIO_NET_F_GUEST_TSO6,
61 VIRTIO_NET_F_GUEST_ECN,
62 VIRTIO_NET_F_GUEST_UFO,
63 VIRTIO_NET_F_GUEST_CSUM
66 #define GUEST_OFFLOAD_GRO_HW_MASK ((1ULL << VIRTIO_NET_F_GUEST_TSO4) | \
67 (1ULL << VIRTIO_NET_F_GUEST_TSO6) | \
68 (1ULL << VIRTIO_NET_F_GUEST_ECN) | \
69 (1ULL << VIRTIO_NET_F_GUEST_UFO))
71 struct virtnet_stat_desc {
72 char desc[ETH_GSTRING_LEN];
76 struct virtnet_sq_stats {
77 struct u64_stats_sync syncp;
85 struct virtnet_rq_stats {
86 struct u64_stats_sync syncp;
97 #define VIRTNET_SQ_STAT(m) offsetof(struct virtnet_sq_stats, m)
98 #define VIRTNET_RQ_STAT(m) offsetof(struct virtnet_rq_stats, m)
100 static const struct virtnet_stat_desc virtnet_sq_stats_desc[] = {
101 { "packets", VIRTNET_SQ_STAT(packets) },
102 { "bytes", VIRTNET_SQ_STAT(bytes) },
103 { "xdp_tx", VIRTNET_SQ_STAT(xdp_tx) },
104 { "xdp_tx_drops", VIRTNET_SQ_STAT(xdp_tx_drops) },
105 { "kicks", VIRTNET_SQ_STAT(kicks) },
108 static const struct virtnet_stat_desc virtnet_rq_stats_desc[] = {
109 { "packets", VIRTNET_RQ_STAT(packets) },
110 { "bytes", VIRTNET_RQ_STAT(bytes) },
111 { "drops", VIRTNET_RQ_STAT(drops) },
112 { "xdp_packets", VIRTNET_RQ_STAT(xdp_packets) },
113 { "xdp_tx", VIRTNET_RQ_STAT(xdp_tx) },
114 { "xdp_redirects", VIRTNET_RQ_STAT(xdp_redirects) },
115 { "xdp_drops", VIRTNET_RQ_STAT(xdp_drops) },
116 { "kicks", VIRTNET_RQ_STAT(kicks) },
119 #define VIRTNET_SQ_STATS_LEN ARRAY_SIZE(virtnet_sq_stats_desc)
120 #define VIRTNET_RQ_STATS_LEN ARRAY_SIZE(virtnet_rq_stats_desc)
122 /* Internal representation of a send virtqueue */
124 /* Virtqueue associated with this send _queue */
125 struct virtqueue *vq;
127 /* TX: fragments + linear part + virtio header */
128 struct scatterlist sg[MAX_SKB_FRAGS + 2];
130 /* Name of the send queue: output.$index */
133 struct virtnet_sq_stats stats;
135 struct napi_struct napi;
138 /* Internal representation of a receive virtqueue */
139 struct receive_queue {
140 /* Virtqueue associated with this receive_queue */
141 struct virtqueue *vq;
143 struct napi_struct napi;
145 struct bpf_prog __rcu *xdp_prog;
147 struct virtnet_rq_stats stats;
149 /* Chain pages by the private ptr. */
152 /* Average packet length for mergeable receive buffers. */
153 struct ewma_pkt_len mrg_avg_pkt_len;
155 /* Page frag for packet buffer allocation. */
156 struct page_frag alloc_frag;
158 /* RX: fragments + linear part + virtio header */
159 struct scatterlist sg[MAX_SKB_FRAGS + 2];
161 /* Min single buffer size for mergeable buffers case. */
162 unsigned int min_buf_len;
164 /* Name of this receive queue: input.$index */
167 struct xdp_rxq_info xdp_rxq;
170 /* Control VQ buffers: protected by the rtnl lock */
172 struct virtio_net_ctrl_hdr hdr;
173 virtio_net_ctrl_ack status;
174 struct virtio_net_ctrl_mq mq;
181 struct virtnet_info {
182 struct virtio_device *vdev;
183 struct virtqueue *cvq;
184 struct net_device *dev;
185 struct send_queue *sq;
186 struct receive_queue *rq;
189 /* Max # of queue pairs supported by the device */
192 /* # of queue pairs currently used by the driver */
193 u16 curr_queue_pairs;
195 /* # of XDP queue pairs currently used by the driver */
198 /* xdp_queue_pairs may be 0, when xdp is already loaded. So add this. */
201 /* I like... big packets and I cannot lie! */
204 /* Host will merge rx buffers for big packets (shake it! shake it!) */
205 bool mergeable_rx_bufs;
207 /* Has control virtqueue */
210 /* Host can handle any s/g split between our header and packet data */
213 /* Packet virtio header size */
216 /* Work struct for delayed refilling if we run low on memory. */
217 struct delayed_work refill;
219 /* Is delayed refill enabled? */
222 /* The lock to synchronize the access to refill_enabled */
223 spinlock_t refill_lock;
225 /* Work struct for config space updates */
226 struct work_struct config_work;
228 /* Does the affinity hint is set for virtqueues? */
229 bool affinity_hint_set;
231 /* CPU hotplug instances for online & dead */
232 struct hlist_node node;
233 struct hlist_node node_dead;
235 struct control_buf *ctrl;
237 /* Ethtool settings */
241 unsigned long guest_offloads;
242 unsigned long guest_offloads_capable;
244 /* failover when STANDBY feature enabled */
245 struct failover *failover;
248 struct padded_vnet_hdr {
249 struct virtio_net_hdr_mrg_rxbuf hdr;
251 * hdr is in a separate sg buffer, and data sg buffer shares same page
252 * with this header sg. This padding makes next sg 16 byte aligned
258 static bool is_xdp_frame(void *ptr)
260 return (unsigned long)ptr & VIRTIO_XDP_FLAG;
263 static void *xdp_to_ptr(struct xdp_frame *ptr)
265 return (void *)((unsigned long)ptr | VIRTIO_XDP_FLAG);
268 static struct xdp_frame *ptr_to_xdp(void *ptr)
270 return (struct xdp_frame *)((unsigned long)ptr & ~VIRTIO_XDP_FLAG);
273 /* Converting between virtqueue no. and kernel tx/rx queue no.
274 * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
276 static int vq2txq(struct virtqueue *vq)
278 return (vq->index - 1) / 2;
281 static int txq2vq(int txq)
286 static int vq2rxq(struct virtqueue *vq)
288 return vq->index / 2;
291 static int rxq2vq(int rxq)
296 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
298 return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
302 * private is used to chain pages for big packets, put the whole
303 * most recent used list in the beginning for reuse
305 static void give_pages(struct receive_queue *rq, struct page *page)
309 /* Find end of list, sew whole thing into vi->rq.pages. */
310 for (end = page; end->private; end = (struct page *)end->private);
311 end->private = (unsigned long)rq->pages;
315 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
317 struct page *p = rq->pages;
320 rq->pages = (struct page *)p->private;
321 /* clear private here, it is used to chain pages */
324 p = alloc_page(gfp_mask);
328 static void enable_delayed_refill(struct virtnet_info *vi)
330 spin_lock_bh(&vi->refill_lock);
331 vi->refill_enabled = true;
332 spin_unlock_bh(&vi->refill_lock);
335 static void disable_delayed_refill(struct virtnet_info *vi)
337 spin_lock_bh(&vi->refill_lock);
338 vi->refill_enabled = false;
339 spin_unlock_bh(&vi->refill_lock);
342 static void virtqueue_napi_schedule(struct napi_struct *napi,
343 struct virtqueue *vq)
345 if (napi_schedule_prep(napi)) {
346 virtqueue_disable_cb(vq);
347 __napi_schedule(napi);
351 static void virtqueue_napi_complete(struct napi_struct *napi,
352 struct virtqueue *vq, int processed)
356 opaque = virtqueue_enable_cb_prepare(vq);
357 if (napi_complete_done(napi, processed)) {
358 if (unlikely(virtqueue_poll(vq, opaque)))
359 virtqueue_napi_schedule(napi, vq);
361 virtqueue_disable_cb(vq);
365 static void skb_xmit_done(struct virtqueue *vq)
367 struct virtnet_info *vi = vq->vdev->priv;
368 struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi;
370 /* Suppress further interrupts. */
371 virtqueue_disable_cb(vq);
374 virtqueue_napi_schedule(napi, vq);
376 /* We were probably waiting for more output buffers. */
377 netif_wake_subqueue(vi->dev, vq2txq(vq));
380 #define MRG_CTX_HEADER_SHIFT 22
381 static void *mergeable_len_to_ctx(unsigned int truesize,
382 unsigned int headroom)
384 return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) | truesize);
387 static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx)
389 return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT;
392 static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx)
394 return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) - 1);
397 /* Called from bottom half context */
398 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
399 struct receive_queue *rq,
400 struct page *page, unsigned int offset,
401 unsigned int len, unsigned int truesize,
402 bool hdr_valid, unsigned int metasize,
403 unsigned int headroom)
406 struct virtio_net_hdr_mrg_rxbuf *hdr;
407 unsigned int copy, hdr_len, hdr_padded_len;
408 struct page *page_to_free = NULL;
409 int tailroom, shinfo_size;
410 char *p, *hdr_p, *buf;
412 p = page_address(page) + offset;
415 hdr_len = vi->hdr_len;
416 if (vi->mergeable_rx_bufs)
417 hdr_padded_len = sizeof(*hdr);
419 hdr_padded_len = sizeof(struct padded_vnet_hdr);
421 /* If headroom is not 0, there is an offset between the beginning of the
422 * data and the allocated space, otherwise the data and the allocated
425 * Buffers with headroom use PAGE_SIZE as alloc size, see
426 * add_recvbuf_mergeable() + get_mergeable_buf_len()
428 truesize = headroom ? PAGE_SIZE : truesize;
429 tailroom = truesize - len - headroom - (hdr_padded_len - hdr_len);
433 offset += hdr_padded_len;
436 shinfo_size = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
438 /* copy small packet so we can reuse these pages */
439 if (!NET_IP_ALIGN && len > GOOD_COPY_LEN && tailroom >= shinfo_size) {
440 skb = build_skb(buf, truesize);
444 skb_reserve(skb, p - buf);
447 page = (struct page *)page->private;
449 give_pages(rq, page);
453 /* copy small packet so we can reuse these pages for small data */
454 skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
458 /* Copy all frame if it fits skb->head, otherwise
459 * we let virtio_net_hdr_to_skb() and GRO pull headers as needed.
461 if (len <= skb_tailroom(skb))
464 copy = ETH_HLEN + metasize;
465 skb_put_data(skb, p, copy);
470 if (vi->mergeable_rx_bufs) {
472 skb_add_rx_frag(skb, 0, page, offset, len, truesize);
479 * Verify that we can indeed put this data into a skb.
480 * This is here to handle cases when the device erroneously
481 * tries to receive more than is possible. This is usually
482 * the case of a broken device.
484 if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
485 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
489 BUG_ON(offset >= PAGE_SIZE);
491 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
492 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
493 frag_size, truesize);
495 page = (struct page *)page->private;
500 give_pages(rq, page);
503 /* hdr_valid means no XDP, so we can copy the vnet header */
505 hdr = skb_vnet_hdr(skb);
506 memcpy(hdr, hdr_p, hdr_len);
509 put_page(page_to_free);
512 __skb_pull(skb, metasize);
513 skb_metadata_set(skb, metasize);
519 static int __virtnet_xdp_xmit_one(struct virtnet_info *vi,
520 struct send_queue *sq,
521 struct xdp_frame *xdpf)
523 struct virtio_net_hdr_mrg_rxbuf *hdr;
526 if (unlikely(xdpf->headroom < vi->hdr_len))
529 /* Make room for virtqueue hdr (also change xdpf->headroom?) */
530 xdpf->data -= vi->hdr_len;
531 /* Zero header and leave csum up to XDP layers */
533 memset(hdr, 0, vi->hdr_len);
534 xdpf->len += vi->hdr_len;
536 sg_init_one(sq->sg, xdpf->data, xdpf->len);
538 err = virtqueue_add_outbuf(sq->vq, sq->sg, 1, xdp_to_ptr(xdpf),
541 return -ENOSPC; /* Caller handle free/refcnt */
546 /* when vi->curr_queue_pairs > nr_cpu_ids, the txq/sq is only used for xdp tx on
547 * the current cpu, so it does not need to be locked.
549 * Here we use marco instead of inline functions because we have to deal with
550 * three issues at the same time: 1. the choice of sq. 2. judge and execute the
551 * lock/unlock of txq 3. make sparse happy. It is difficult for two inline
552 * functions to perfectly solve these three problems at the same time.
554 #define virtnet_xdp_get_sq(vi) ({ \
555 int cpu = smp_processor_id(); \
556 struct netdev_queue *txq; \
557 typeof(vi) v = (vi); \
560 if (v->curr_queue_pairs > nr_cpu_ids) { \
561 qp = v->curr_queue_pairs - v->xdp_queue_pairs; \
563 txq = netdev_get_tx_queue(v->dev, qp); \
564 __netif_tx_acquire(txq); \
566 qp = cpu % v->curr_queue_pairs; \
567 txq = netdev_get_tx_queue(v->dev, qp); \
568 __netif_tx_lock(txq, cpu); \
573 #define virtnet_xdp_put_sq(vi, q) { \
574 struct netdev_queue *txq; \
575 typeof(vi) v = (vi); \
577 txq = netdev_get_tx_queue(v->dev, (q) - v->sq); \
578 if (v->curr_queue_pairs > nr_cpu_ids) \
579 __netif_tx_release(txq); \
581 __netif_tx_unlock(txq); \
584 static int virtnet_xdp_xmit(struct net_device *dev,
585 int n, struct xdp_frame **frames, u32 flags)
587 struct virtnet_info *vi = netdev_priv(dev);
588 struct receive_queue *rq = vi->rq;
589 struct bpf_prog *xdp_prog;
590 struct send_queue *sq;
600 /* Only allow ndo_xdp_xmit if XDP is loaded on dev, as this
601 * indicate XDP resources have been successfully allocated.
603 xdp_prog = rcu_access_pointer(rq->xdp_prog);
607 sq = virtnet_xdp_get_sq(vi);
609 if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK)) {
614 /* Free up any pending old buffers before queueing new ones. */
615 while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
616 if (likely(is_xdp_frame(ptr))) {
617 struct xdp_frame *frame = ptr_to_xdp(ptr);
620 xdp_return_frame(frame);
622 struct sk_buff *skb = ptr;
625 napi_consume_skb(skb, false);
630 for (i = 0; i < n; i++) {
631 struct xdp_frame *xdpf = frames[i];
633 if (__virtnet_xdp_xmit_one(vi, sq, xdpf))
639 if (flags & XDP_XMIT_FLUSH) {
640 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq))
644 u64_stats_update_begin(&sq->stats.syncp);
645 sq->stats.bytes += bytes;
646 sq->stats.packets += packets;
647 sq->stats.xdp_tx += n;
648 sq->stats.xdp_tx_drops += n - nxmit;
649 sq->stats.kicks += kicks;
650 u64_stats_update_end(&sq->stats.syncp);
652 virtnet_xdp_put_sq(vi, sq);
656 static unsigned int virtnet_get_headroom(struct virtnet_info *vi)
658 return vi->xdp_enabled ? VIRTIO_XDP_HEADROOM : 0;
661 /* We copy the packet for XDP in the following cases:
663 * 1) Packet is scattered across multiple rx buffers.
664 * 2) Headroom space is insufficient.
666 * This is inefficient but it's a temporary condition that
667 * we hit right after XDP is enabled and until queue is refilled
668 * with large buffers with sufficient headroom - so it should affect
669 * at most queue size packets.
670 * Afterwards, the conditions to enable
671 * XDP should preclude the underlying device from sending packets
672 * across multiple buffers (num_buf > 1), and we make sure buffers
673 * have enough headroom.
675 static struct page *xdp_linearize_page(struct receive_queue *rq,
682 struct page *page = alloc_page(GFP_ATOMIC);
687 memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
691 int tailroom = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
696 buf = virtqueue_get_buf(rq->vq, &buflen);
700 p = virt_to_head_page(buf);
701 off = buf - page_address(p);
703 /* guard against a misconfigured or uncooperative backend that
704 * is sending packet larger than the MTU.
706 if ((page_off + buflen + tailroom) > PAGE_SIZE) {
711 memcpy(page_address(page) + page_off,
712 page_address(p) + off, buflen);
717 /* Headroom does not contribute to packet length */
718 *len = page_off - VIRTIO_XDP_HEADROOM;
721 __free_pages(page, 0);
725 static struct sk_buff *receive_small(struct net_device *dev,
726 struct virtnet_info *vi,
727 struct receive_queue *rq,
728 void *buf, void *ctx,
730 unsigned int *xdp_xmit,
731 struct virtnet_rq_stats *stats)
734 struct bpf_prog *xdp_prog;
735 unsigned int xdp_headroom = (unsigned long)ctx;
736 unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom;
737 unsigned int headroom = vi->hdr_len + header_offset;
738 unsigned int buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
739 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
740 struct page *page = virt_to_head_page(buf);
741 unsigned int delta = 0;
742 struct page *xdp_page;
744 unsigned int metasize = 0;
749 if (unlikely(len > GOOD_PACKET_LEN)) {
750 pr_debug("%s: rx error: len %u exceeds max size %d\n",
751 dev->name, len, GOOD_PACKET_LEN);
752 dev->stats.rx_length_errors++;
756 xdp_prog = rcu_dereference(rq->xdp_prog);
758 struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset;
759 struct xdp_frame *xdpf;
764 if (unlikely(hdr->hdr.gso_type))
767 if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) {
768 int offset = buf - page_address(page) + header_offset;
769 unsigned int tlen = len + vi->hdr_len;
772 xdp_headroom = virtnet_get_headroom(vi);
773 header_offset = VIRTNET_RX_PAD + xdp_headroom;
774 headroom = vi->hdr_len + header_offset;
775 buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
776 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
777 xdp_page = xdp_linearize_page(rq, &num_buf, page,
778 offset, header_offset,
783 buf = page_address(xdp_page);
788 xdp_init_buff(&xdp, buflen, &rq->xdp_rxq);
789 xdp_prepare_buff(&xdp, buf + VIRTNET_RX_PAD + vi->hdr_len,
790 xdp_headroom, len, true);
791 orig_data = xdp.data;
792 act = bpf_prog_run_xdp(xdp_prog, &xdp);
793 stats->xdp_packets++;
797 /* Recalculate length in case bpf program changed it */
798 delta = orig_data - xdp.data;
799 len = xdp.data_end - xdp.data;
800 metasize = xdp.data - xdp.data_meta;
804 xdpf = xdp_convert_buff_to_frame(&xdp);
807 err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
808 if (unlikely(!err)) {
809 xdp_return_frame_rx_napi(xdpf);
810 } else if (unlikely(err < 0)) {
811 trace_xdp_exception(vi->dev, xdp_prog, act);
814 *xdp_xmit |= VIRTIO_XDP_TX;
818 stats->xdp_redirects++;
819 err = xdp_do_redirect(dev, &xdp, xdp_prog);
822 *xdp_xmit |= VIRTIO_XDP_REDIR;
826 bpf_warn_invalid_xdp_action(act);
829 trace_xdp_exception(vi->dev, xdp_prog, act);
837 skb = build_skb(buf, buflen);
842 skb_reserve(skb, headroom - delta);
845 buf += header_offset;
846 memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
847 } /* keep zeroed vnet hdr since XDP is loaded */
850 skb_metadata_set(skb, metasize);
865 static struct sk_buff *receive_big(struct net_device *dev,
866 struct virtnet_info *vi,
867 struct receive_queue *rq,
870 struct virtnet_rq_stats *stats)
872 struct page *page = buf;
873 struct sk_buff *skb =
874 page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, true, 0, 0);
876 stats->bytes += len - vi->hdr_len;
884 give_pages(rq, page);
888 static struct sk_buff *receive_mergeable(struct net_device *dev,
889 struct virtnet_info *vi,
890 struct receive_queue *rq,
894 unsigned int *xdp_xmit,
895 struct virtnet_rq_stats *stats)
897 struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
898 u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
899 struct page *page = virt_to_head_page(buf);
900 int offset = buf - page_address(page);
901 struct sk_buff *head_skb, *curr_skb;
902 struct bpf_prog *xdp_prog;
903 unsigned int truesize = mergeable_ctx_to_truesize(ctx);
904 unsigned int headroom = mergeable_ctx_to_headroom(ctx);
905 unsigned int metasize = 0;
906 unsigned int frame_sz;
910 stats->bytes += len - vi->hdr_len;
912 if (unlikely(len > truesize)) {
913 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
914 dev->name, len, (unsigned long)ctx);
915 dev->stats.rx_length_errors++;
919 xdp_prog = rcu_dereference(rq->xdp_prog);
921 struct xdp_frame *xdpf;
922 struct page *xdp_page;
927 /* Transient failure which in theory could occur if
928 * in-flight packets from before XDP was enabled reach
929 * the receive path after XDP is loaded.
931 if (unlikely(hdr->hdr.gso_type))
934 /* Buffers with headroom use PAGE_SIZE as alloc size,
935 * see add_recvbuf_mergeable() + get_mergeable_buf_len()
937 frame_sz = headroom ? PAGE_SIZE : truesize;
939 /* This happens when rx buffer size is underestimated
940 * or headroom is not enough because of the buffer
941 * was refilled before XDP is set. This should only
942 * happen for the first several packets, so we don't
943 * care much about its performance.
945 if (unlikely(num_buf > 1 ||
946 headroom < virtnet_get_headroom(vi))) {
947 /* linearize data for XDP */
948 xdp_page = xdp_linearize_page(rq, &num_buf,
952 frame_sz = PAGE_SIZE;
956 offset = VIRTIO_XDP_HEADROOM;
961 /* Allow consuming headroom but reserve enough space to push
962 * the descriptor on if we get an XDP_TX return code.
964 data = page_address(xdp_page) + offset;
965 xdp_init_buff(&xdp, frame_sz - vi->hdr_len, &rq->xdp_rxq);
966 xdp_prepare_buff(&xdp, data - VIRTIO_XDP_HEADROOM + vi->hdr_len,
967 VIRTIO_XDP_HEADROOM, len - vi->hdr_len, true);
969 act = bpf_prog_run_xdp(xdp_prog, &xdp);
970 stats->xdp_packets++;
974 metasize = xdp.data - xdp.data_meta;
976 /* recalculate offset to account for any header
977 * adjustments and minus the metasize to copy the
978 * metadata in page_to_skb(). Note other cases do not
979 * build an skb and avoid using offset
981 offset = xdp.data - page_address(xdp_page) -
982 vi->hdr_len - metasize;
984 /* recalculate len if xdp.data, xdp.data_end or
985 * xdp.data_meta were adjusted
987 len = xdp.data_end - xdp.data + vi->hdr_len + metasize;
989 /* recalculate headroom if xdp.data or xdp_data_meta
990 * were adjusted, note that offset should always point
991 * to the start of the reserved bytes for virtio_net
992 * header which are followed by xdp.data, that means
993 * that offset is equal to the headroom (when buf is
994 * starting at the beginning of the page, otherwise
995 * there is a base offset inside the page) but it's used
996 * with a different starting point (buf start) than
997 * xdp.data (buf start + vnet hdr size). If xdp.data or
998 * data_meta were adjusted by the xdp prog then the
999 * headroom size has changed and so has the offset, we
1000 * can use data_hard_start, which points at buf start +
1001 * vnet hdr size, to calculate the new headroom and use
1002 * it later to compute buf start in page_to_skb()
1004 headroom = xdp.data - xdp.data_hard_start - metasize;
1006 /* We can only create skb based on xdp_page. */
1007 if (unlikely(xdp_page != page)) {
1010 head_skb = page_to_skb(vi, rq, xdp_page, offset,
1011 len, PAGE_SIZE, false,
1019 xdpf = xdp_convert_buff_to_frame(&xdp);
1020 if (unlikely(!xdpf)) {
1021 if (unlikely(xdp_page != page))
1025 err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
1026 if (unlikely(!err)) {
1027 xdp_return_frame_rx_napi(xdpf);
1028 } else if (unlikely(err < 0)) {
1029 trace_xdp_exception(vi->dev, xdp_prog, act);
1030 if (unlikely(xdp_page != page))
1034 *xdp_xmit |= VIRTIO_XDP_TX;
1035 if (unlikely(xdp_page != page))
1040 stats->xdp_redirects++;
1041 err = xdp_do_redirect(dev, &xdp, xdp_prog);
1043 if (unlikely(xdp_page != page))
1047 *xdp_xmit |= VIRTIO_XDP_REDIR;
1048 if (unlikely(xdp_page != page))
1053 bpf_warn_invalid_xdp_action(act);
1056 trace_xdp_exception(vi->dev, xdp_prog, act);
1059 if (unlikely(xdp_page != page))
1060 __free_pages(xdp_page, 0);
1066 head_skb = page_to_skb(vi, rq, page, offset, len, truesize, !xdp_prog,
1067 metasize, headroom);
1068 curr_skb = head_skb;
1070 if (unlikely(!curr_skb))
1075 buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx);
1076 if (unlikely(!buf)) {
1077 pr_debug("%s: rx error: %d buffers out of %d missing\n",
1079 virtio16_to_cpu(vi->vdev,
1081 dev->stats.rx_length_errors++;
1085 stats->bytes += len;
1086 page = virt_to_head_page(buf);
1088 truesize = mergeable_ctx_to_truesize(ctx);
1089 if (unlikely(len > truesize)) {
1090 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
1091 dev->name, len, (unsigned long)ctx);
1092 dev->stats.rx_length_errors++;
1096 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
1097 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
1098 struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
1100 if (unlikely(!nskb))
1102 if (curr_skb == head_skb)
1103 skb_shinfo(curr_skb)->frag_list = nskb;
1105 curr_skb->next = nskb;
1107 head_skb->truesize += nskb->truesize;
1110 if (curr_skb != head_skb) {
1111 head_skb->data_len += len;
1112 head_skb->len += len;
1113 head_skb->truesize += truesize;
1115 offset = buf - page_address(page);
1116 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
1118 skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
1121 skb_add_rx_frag(curr_skb, num_skb_frags, page,
1122 offset, len, truesize);
1126 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
1134 while (num_buf-- > 1) {
1135 buf = virtqueue_get_buf(rq->vq, &len);
1136 if (unlikely(!buf)) {
1137 pr_debug("%s: rx error: %d buffers missing\n",
1138 dev->name, num_buf);
1139 dev->stats.rx_length_errors++;
1142 stats->bytes += len;
1143 page = virt_to_head_page(buf);
1148 dev_kfree_skb(head_skb);
1153 static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
1154 void *buf, unsigned int len, void **ctx,
1155 unsigned int *xdp_xmit,
1156 struct virtnet_rq_stats *stats)
1158 struct net_device *dev = vi->dev;
1159 struct sk_buff *skb;
1160 struct virtio_net_hdr_mrg_rxbuf *hdr;
1162 if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
1163 pr_debug("%s: short packet %i\n", dev->name, len);
1164 dev->stats.rx_length_errors++;
1165 if (vi->mergeable_rx_bufs) {
1166 put_page(virt_to_head_page(buf));
1167 } else if (vi->big_packets) {
1168 give_pages(rq, buf);
1170 put_page(virt_to_head_page(buf));
1175 if (vi->mergeable_rx_bufs)
1176 skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit,
1178 else if (vi->big_packets)
1179 skb = receive_big(dev, vi, rq, buf, len, stats);
1181 skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit, stats);
1186 hdr = skb_vnet_hdr(skb);
1188 if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
1189 skb->ip_summed = CHECKSUM_UNNECESSARY;
1191 if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
1192 virtio_is_little_endian(vi->vdev))) {
1193 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
1194 dev->name, hdr->hdr.gso_type,
1199 skb_record_rx_queue(skb, vq2rxq(rq->vq));
1200 skb->protocol = eth_type_trans(skb, dev);
1201 pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
1202 ntohs(skb->protocol), skb->len, skb->pkt_type);
1204 napi_gro_receive(&rq->napi, skb);
1208 dev->stats.rx_frame_errors++;
1212 /* Unlike mergeable buffers, all buffers are allocated to the
1213 * same size, except for the headroom. For this reason we do
1214 * not need to use mergeable_len_to_ctx here - it is enough
1215 * to store the headroom as the context ignoring the truesize.
1217 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
1220 struct page_frag *alloc_frag = &rq->alloc_frag;
1222 unsigned int xdp_headroom = virtnet_get_headroom(vi);
1223 void *ctx = (void *)(unsigned long)xdp_headroom;
1224 int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
1227 len = SKB_DATA_ALIGN(len) +
1228 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1229 if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
1232 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1233 get_page(alloc_frag->page);
1234 alloc_frag->offset += len;
1235 sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom,
1236 vi->hdr_len + GOOD_PACKET_LEN);
1237 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1239 put_page(virt_to_head_page(buf));
1243 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
1246 struct page *first, *list = NULL;
1250 sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
1252 /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
1253 for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
1254 first = get_a_page(rq, gfp);
1257 give_pages(rq, list);
1260 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
1262 /* chain new page in list head to match sg */
1263 first->private = (unsigned long)list;
1267 first = get_a_page(rq, gfp);
1269 give_pages(rq, list);
1272 p = page_address(first);
1274 /* rq->sg[0], rq->sg[1] share the same page */
1275 /* a separated rq->sg[0] for header - required in case !any_header_sg */
1276 sg_set_buf(&rq->sg[0], p, vi->hdr_len);
1278 /* rq->sg[1] for data packet, from offset */
1279 offset = sizeof(struct padded_vnet_hdr);
1280 sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
1282 /* chain first in list head */
1283 first->private = (unsigned long)list;
1284 err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
1287 give_pages(rq, first);
1292 static unsigned int get_mergeable_buf_len(struct receive_queue *rq,
1293 struct ewma_pkt_len *avg_pkt_len,
1296 const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1300 return PAGE_SIZE - room;
1302 len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
1303 rq->min_buf_len, PAGE_SIZE - hdr_len);
1305 return ALIGN(len, L1_CACHE_BYTES);
1308 static int add_recvbuf_mergeable(struct virtnet_info *vi,
1309 struct receive_queue *rq, gfp_t gfp)
1311 struct page_frag *alloc_frag = &rq->alloc_frag;
1312 unsigned int headroom = virtnet_get_headroom(vi);
1313 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1314 unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
1318 unsigned int len, hole;
1320 /* Extra tailroom is needed to satisfy XDP's assumption. This
1321 * means rx frags coalescing won't work, but consider we've
1322 * disabled GSO for XDP, it won't be a big issue.
1324 len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len, room);
1325 if (unlikely(!skb_page_frag_refill(len + room, alloc_frag, gfp)))
1328 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1329 buf += headroom; /* advance address leaving hole at front of pkt */
1330 get_page(alloc_frag->page);
1331 alloc_frag->offset += len + room;
1332 hole = alloc_frag->size - alloc_frag->offset;
1333 if (hole < len + room) {
1334 /* To avoid internal fragmentation, if there is very likely not
1335 * enough space for another buffer, add the remaining space to
1336 * the current buffer.
1339 alloc_frag->offset += hole;
1342 sg_init_one(rq->sg, buf, len);
1343 ctx = mergeable_len_to_ctx(len, headroom);
1344 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1346 put_page(virt_to_head_page(buf));
1352 * Returns false if we couldn't fill entirely (OOM).
1354 * Normally run in the receive path, but can also be run from ndo_open
1355 * before we're receiving packets, or from refill_work which is
1356 * careful to disable receiving (using napi_disable).
1358 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
1365 if (vi->mergeable_rx_bufs)
1366 err = add_recvbuf_mergeable(vi, rq, gfp);
1367 else if (vi->big_packets)
1368 err = add_recvbuf_big(vi, rq, gfp);
1370 err = add_recvbuf_small(vi, rq, gfp);
1372 oom = err == -ENOMEM;
1375 } while (rq->vq->num_free);
1376 if (virtqueue_kick_prepare(rq->vq) && virtqueue_notify(rq->vq)) {
1377 unsigned long flags;
1379 flags = u64_stats_update_begin_irqsave(&rq->stats.syncp);
1381 u64_stats_update_end_irqrestore(&rq->stats.syncp, flags);
1387 static void skb_recv_done(struct virtqueue *rvq)
1389 struct virtnet_info *vi = rvq->vdev->priv;
1390 struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
1392 virtqueue_napi_schedule(&rq->napi, rvq);
1395 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
1399 /* If all buffers were filled by other side before we napi_enabled, we
1400 * won't get another interrupt, so process any outstanding packets now.
1401 * Call local_bh_enable after to trigger softIRQ processing.
1404 virtqueue_napi_schedule(napi, vq);
1408 static void virtnet_napi_tx_enable(struct virtnet_info *vi,
1409 struct virtqueue *vq,
1410 struct napi_struct *napi)
1415 /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
1416 * enable the feature if this is likely affine with the transmit path.
1418 if (!vi->affinity_hint_set) {
1423 return virtnet_napi_enable(vq, napi);
1426 static void virtnet_napi_tx_disable(struct napi_struct *napi)
1432 static void refill_work(struct work_struct *work)
1434 struct virtnet_info *vi =
1435 container_of(work, struct virtnet_info, refill.work);
1439 for (i = 0; i < vi->curr_queue_pairs; i++) {
1440 struct receive_queue *rq = &vi->rq[i];
1442 napi_disable(&rq->napi);
1443 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
1444 virtnet_napi_enable(rq->vq, &rq->napi);
1446 /* In theory, this can happen: if we don't get any buffers in
1447 * we will *never* try to fill again.
1450 schedule_delayed_work(&vi->refill, HZ/2);
1454 static int virtnet_receive(struct receive_queue *rq, int budget,
1455 unsigned int *xdp_xmit)
1457 struct virtnet_info *vi = rq->vq->vdev->priv;
1458 struct virtnet_rq_stats stats = {};
1463 if (!vi->big_packets || vi->mergeable_rx_bufs) {
1466 while (stats.packets < budget &&
1467 (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) {
1468 receive_buf(vi, rq, buf, len, ctx, xdp_xmit, &stats);
1472 while (stats.packets < budget &&
1473 (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
1474 receive_buf(vi, rq, buf, len, NULL, xdp_xmit, &stats);
1479 if (rq->vq->num_free > min((unsigned int)budget, virtqueue_get_vring_size(rq->vq)) / 2) {
1480 if (!try_fill_recv(vi, rq, GFP_ATOMIC)) {
1481 spin_lock(&vi->refill_lock);
1482 if (vi->refill_enabled)
1483 schedule_delayed_work(&vi->refill, 0);
1484 spin_unlock(&vi->refill_lock);
1488 u64_stats_update_begin(&rq->stats.syncp);
1489 for (i = 0; i < VIRTNET_RQ_STATS_LEN; i++) {
1490 size_t offset = virtnet_rq_stats_desc[i].offset;
1493 item = (u64 *)((u8 *)&rq->stats + offset);
1494 *item += *(u64 *)((u8 *)&stats + offset);
1496 u64_stats_update_end(&rq->stats.syncp);
1498 return stats.packets;
1501 static void free_old_xmit_skbs(struct send_queue *sq, bool in_napi)
1504 unsigned int packets = 0;
1505 unsigned int bytes = 0;
1508 while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
1509 if (likely(!is_xdp_frame(ptr))) {
1510 struct sk_buff *skb = ptr;
1512 pr_debug("Sent skb %p\n", skb);
1515 napi_consume_skb(skb, in_napi);
1517 struct xdp_frame *frame = ptr_to_xdp(ptr);
1519 bytes += frame->len;
1520 xdp_return_frame(frame);
1525 /* Avoid overhead when no packets have been processed
1526 * happens when called speculatively from start_xmit.
1531 u64_stats_update_begin(&sq->stats.syncp);
1532 sq->stats.bytes += bytes;
1533 sq->stats.packets += packets;
1534 u64_stats_update_end(&sq->stats.syncp);
1537 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
1539 if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
1541 else if (q < vi->curr_queue_pairs)
1547 static void virtnet_poll_cleantx(struct receive_queue *rq)
1549 struct virtnet_info *vi = rq->vq->vdev->priv;
1550 unsigned int index = vq2rxq(rq->vq);
1551 struct send_queue *sq = &vi->sq[index];
1552 struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index);
1554 if (!sq->napi.weight || is_xdp_raw_buffer_queue(vi, index))
1557 if (__netif_tx_trylock(txq)) {
1559 virtqueue_disable_cb(sq->vq);
1560 free_old_xmit_skbs(sq, true);
1561 } while (unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
1563 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1564 netif_tx_wake_queue(txq);
1566 __netif_tx_unlock(txq);
1570 static int virtnet_poll(struct napi_struct *napi, int budget)
1572 struct receive_queue *rq =
1573 container_of(napi, struct receive_queue, napi);
1574 struct virtnet_info *vi = rq->vq->vdev->priv;
1575 struct send_queue *sq;
1576 unsigned int received;
1577 unsigned int xdp_xmit = 0;
1579 virtnet_poll_cleantx(rq);
1581 received = virtnet_receive(rq, budget, &xdp_xmit);
1583 /* Out of packets? */
1584 if (received < budget)
1585 virtqueue_napi_complete(napi, rq->vq, received);
1587 if (xdp_xmit & VIRTIO_XDP_REDIR)
1590 if (xdp_xmit & VIRTIO_XDP_TX) {
1591 sq = virtnet_xdp_get_sq(vi);
1592 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1593 u64_stats_update_begin(&sq->stats.syncp);
1595 u64_stats_update_end(&sq->stats.syncp);
1597 virtnet_xdp_put_sq(vi, sq);
1603 static int virtnet_open(struct net_device *dev)
1605 struct virtnet_info *vi = netdev_priv(dev);
1608 enable_delayed_refill(vi);
1610 for (i = 0; i < vi->max_queue_pairs; i++) {
1611 if (i < vi->curr_queue_pairs)
1612 /* Make sure we have some buffers: if oom use wq. */
1613 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1614 schedule_delayed_work(&vi->refill, 0);
1616 err = xdp_rxq_info_reg(&vi->rq[i].xdp_rxq, dev, i, vi->rq[i].napi.napi_id);
1620 err = xdp_rxq_info_reg_mem_model(&vi->rq[i].xdp_rxq,
1621 MEM_TYPE_PAGE_SHARED, NULL);
1623 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1627 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1628 virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi);
1634 static int virtnet_poll_tx(struct napi_struct *napi, int budget)
1636 struct send_queue *sq = container_of(napi, struct send_queue, napi);
1637 struct virtnet_info *vi = sq->vq->vdev->priv;
1638 unsigned int index = vq2txq(sq->vq);
1639 struct netdev_queue *txq;
1643 if (unlikely(is_xdp_raw_buffer_queue(vi, index))) {
1644 /* We don't need to enable cb for XDP */
1645 napi_complete_done(napi, 0);
1649 txq = netdev_get_tx_queue(vi->dev, index);
1650 __netif_tx_lock(txq, raw_smp_processor_id());
1651 virtqueue_disable_cb(sq->vq);
1652 free_old_xmit_skbs(sq, true);
1654 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1655 netif_tx_wake_queue(txq);
1657 opaque = virtqueue_enable_cb_prepare(sq->vq);
1659 done = napi_complete_done(napi, 0);
1662 virtqueue_disable_cb(sq->vq);
1664 __netif_tx_unlock(txq);
1667 if (unlikely(virtqueue_poll(sq->vq, opaque))) {
1668 if (napi_schedule_prep(napi)) {
1669 __netif_tx_lock(txq, raw_smp_processor_id());
1670 virtqueue_disable_cb(sq->vq);
1671 __netif_tx_unlock(txq);
1672 __napi_schedule(napi);
1680 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
1682 struct virtio_net_hdr_mrg_rxbuf *hdr;
1683 const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
1684 struct virtnet_info *vi = sq->vq->vdev->priv;
1686 unsigned hdr_len = vi->hdr_len;
1689 pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
1691 can_push = vi->any_header_sg &&
1692 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
1693 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
1694 /* Even if we can, don't push here yet as this would skew
1695 * csum_start offset below. */
1697 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1699 hdr = skb_vnet_hdr(skb);
1701 if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1702 virtio_is_little_endian(vi->vdev), false,
1706 if (vi->mergeable_rx_bufs)
1707 hdr->num_buffers = 0;
1709 sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1711 __skb_push(skb, hdr_len);
1712 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1713 if (unlikely(num_sg < 0))
1715 /* Pull header back to avoid skew in tx bytes calculations. */
1716 __skb_pull(skb, hdr_len);
1718 sg_set_buf(sq->sg, hdr, hdr_len);
1719 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
1720 if (unlikely(num_sg < 0))
1724 return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
1727 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
1729 struct virtnet_info *vi = netdev_priv(dev);
1730 int qnum = skb_get_queue_mapping(skb);
1731 struct send_queue *sq = &vi->sq[qnum];
1733 struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1734 bool kick = !netdev_xmit_more();
1735 bool use_napi = sq->napi.weight;
1737 /* Free up any pending old buffers before queueing new ones. */
1740 virtqueue_disable_cb(sq->vq);
1742 free_old_xmit_skbs(sq, false);
1744 } while (use_napi && kick &&
1745 unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
1747 /* timestamp packet in software */
1748 skb_tx_timestamp(skb);
1750 /* Try to transmit */
1751 err = xmit_skb(sq, skb);
1753 /* This should not happen! */
1754 if (unlikely(err)) {
1755 dev->stats.tx_fifo_errors++;
1756 if (net_ratelimit())
1758 "Unexpected TXQ (%d) queue failure: %d\n",
1760 dev->stats.tx_dropped++;
1761 dev_kfree_skb_any(skb);
1762 return NETDEV_TX_OK;
1765 /* Don't wait up for transmitted skbs to be freed. */
1771 /* If running out of space, stop queue to avoid getting packets that we
1772 * are then unable to transmit.
1773 * An alternative would be to force queuing layer to requeue the skb by
1774 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1775 * returned in a normal path of operation: it means that driver is not
1776 * maintaining the TX queue stop/start state properly, and causes
1777 * the stack to do a non-trivial amount of useless work.
1778 * Since most packets only take 1 or 2 ring slots, stopping the queue
1779 * early means 16 slots are typically wasted.
1781 if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
1782 netif_stop_subqueue(dev, qnum);
1784 unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
1785 /* More just got used, free them then recheck. */
1786 free_old_xmit_skbs(sq, false);
1787 if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
1788 netif_start_subqueue(dev, qnum);
1789 virtqueue_disable_cb(sq->vq);
1794 if (kick || netif_xmit_stopped(txq)) {
1795 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1796 u64_stats_update_begin(&sq->stats.syncp);
1798 u64_stats_update_end(&sq->stats.syncp);
1802 return NETDEV_TX_OK;
1806 * Send command via the control virtqueue and check status. Commands
1807 * supported by the hypervisor, as indicated by feature bits, should
1808 * never fail unless improperly formatted.
1810 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
1811 struct scatterlist *out)
1813 struct scatterlist *sgs[4], hdr, stat;
1814 unsigned out_num = 0, tmp;
1817 /* Caller should know better */
1818 BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
1820 vi->ctrl->status = ~0;
1821 vi->ctrl->hdr.class = class;
1822 vi->ctrl->hdr.cmd = cmd;
1824 sg_init_one(&hdr, &vi->ctrl->hdr, sizeof(vi->ctrl->hdr));
1825 sgs[out_num++] = &hdr;
1828 sgs[out_num++] = out;
1830 /* Add return status. */
1831 sg_init_one(&stat, &vi->ctrl->status, sizeof(vi->ctrl->status));
1832 sgs[out_num] = &stat;
1834 BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1835 ret = virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1837 dev_warn(&vi->vdev->dev,
1838 "Failed to add sgs for command vq: %d\n.", ret);
1842 if (unlikely(!virtqueue_kick(vi->cvq)))
1843 return vi->ctrl->status == VIRTIO_NET_OK;
1845 /* Spin for a response, the kick causes an ioport write, trapping
1846 * into the hypervisor, so the request should be handled immediately.
1848 while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1849 !virtqueue_is_broken(vi->cvq))
1852 return vi->ctrl->status == VIRTIO_NET_OK;
1855 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1857 struct virtnet_info *vi = netdev_priv(dev);
1858 struct virtio_device *vdev = vi->vdev;
1860 struct sockaddr *addr;
1861 struct scatterlist sg;
1863 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
1866 addr = kmemdup(p, sizeof(*addr), GFP_KERNEL);
1870 ret = eth_prepare_mac_addr_change(dev, addr);
1874 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1875 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1876 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1877 VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1878 dev_warn(&vdev->dev,
1879 "Failed to set mac address by vq command.\n");
1883 } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1884 !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1887 /* Naturally, this has an atomicity problem. */
1888 for (i = 0; i < dev->addr_len; i++)
1889 virtio_cwrite8(vdev,
1890 offsetof(struct virtio_net_config, mac) +
1891 i, addr->sa_data[i]);
1894 eth_commit_mac_addr_change(dev, p);
1902 static void virtnet_stats(struct net_device *dev,
1903 struct rtnl_link_stats64 *tot)
1905 struct virtnet_info *vi = netdev_priv(dev);
1909 for (i = 0; i < vi->max_queue_pairs; i++) {
1910 u64 tpackets, tbytes, rpackets, rbytes, rdrops;
1911 struct receive_queue *rq = &vi->rq[i];
1912 struct send_queue *sq = &vi->sq[i];
1915 start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
1916 tpackets = sq->stats.packets;
1917 tbytes = sq->stats.bytes;
1918 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
1921 start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
1922 rpackets = rq->stats.packets;
1923 rbytes = rq->stats.bytes;
1924 rdrops = rq->stats.drops;
1925 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
1927 tot->rx_packets += rpackets;
1928 tot->tx_packets += tpackets;
1929 tot->rx_bytes += rbytes;
1930 tot->tx_bytes += tbytes;
1931 tot->rx_dropped += rdrops;
1934 tot->tx_dropped = dev->stats.tx_dropped;
1935 tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1936 tot->rx_length_errors = dev->stats.rx_length_errors;
1937 tot->rx_frame_errors = dev->stats.rx_frame_errors;
1940 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1943 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1944 VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1945 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1949 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1951 struct scatterlist sg;
1952 struct net_device *dev = vi->dev;
1954 if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1957 vi->ctrl->mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1958 sg_init_one(&sg, &vi->ctrl->mq, sizeof(vi->ctrl->mq));
1960 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1961 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1962 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1966 vi->curr_queue_pairs = queue_pairs;
1967 /* virtnet_open() will refill when device is going to up. */
1968 if (dev->flags & IFF_UP)
1969 schedule_delayed_work(&vi->refill, 0);
1975 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1980 err = _virtnet_set_queues(vi, queue_pairs);
1985 static int virtnet_close(struct net_device *dev)
1987 struct virtnet_info *vi = netdev_priv(dev);
1990 /* Make sure NAPI doesn't schedule refill work */
1991 disable_delayed_refill(vi);
1992 /* Make sure refill_work doesn't re-enable napi! */
1993 cancel_delayed_work_sync(&vi->refill);
1995 for (i = 0; i < vi->max_queue_pairs; i++) {
1996 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1997 napi_disable(&vi->rq[i].napi);
1998 virtnet_napi_tx_disable(&vi->sq[i].napi);
2004 static void virtnet_set_rx_mode(struct net_device *dev)
2006 struct virtnet_info *vi = netdev_priv(dev);
2007 struct scatterlist sg[2];
2008 struct virtio_net_ctrl_mac *mac_data;
2009 struct netdev_hw_addr *ha;
2015 /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
2016 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
2019 vi->ctrl->promisc = ((dev->flags & IFF_PROMISC) != 0);
2020 vi->ctrl->allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
2022 sg_init_one(sg, &vi->ctrl->promisc, sizeof(vi->ctrl->promisc));
2024 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
2025 VIRTIO_NET_CTRL_RX_PROMISC, sg))
2026 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
2027 vi->ctrl->promisc ? "en" : "dis");
2029 sg_init_one(sg, &vi->ctrl->allmulti, sizeof(vi->ctrl->allmulti));
2031 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
2032 VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
2033 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
2034 vi->ctrl->allmulti ? "en" : "dis");
2036 uc_count = netdev_uc_count(dev);
2037 mc_count = netdev_mc_count(dev);
2038 /* MAC filter - use one buffer for both lists */
2039 buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
2040 (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
2045 sg_init_table(sg, 2);
2047 /* Store the unicast list and count in the front of the buffer */
2048 mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
2050 netdev_for_each_uc_addr(ha, dev)
2051 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
2053 sg_set_buf(&sg[0], mac_data,
2054 sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
2056 /* multicast list and count fill the end */
2057 mac_data = (void *)&mac_data->macs[uc_count][0];
2059 mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
2061 netdev_for_each_mc_addr(ha, dev)
2062 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
2064 sg_set_buf(&sg[1], mac_data,
2065 sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
2067 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
2068 VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
2069 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
2074 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
2075 __be16 proto, u16 vid)
2077 struct virtnet_info *vi = netdev_priv(dev);
2078 struct scatterlist sg;
2080 vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
2081 sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
2083 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
2084 VIRTIO_NET_CTRL_VLAN_ADD, &sg))
2085 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
2089 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
2090 __be16 proto, u16 vid)
2092 struct virtnet_info *vi = netdev_priv(dev);
2093 struct scatterlist sg;
2095 vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
2096 sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
2098 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
2099 VIRTIO_NET_CTRL_VLAN_DEL, &sg))
2100 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
2104 static void virtnet_clean_affinity(struct virtnet_info *vi)
2108 if (vi->affinity_hint_set) {
2109 for (i = 0; i < vi->max_queue_pairs; i++) {
2110 virtqueue_set_affinity(vi->rq[i].vq, NULL);
2111 virtqueue_set_affinity(vi->sq[i].vq, NULL);
2114 vi->affinity_hint_set = false;
2118 static void virtnet_set_affinity(struct virtnet_info *vi)
2127 if (!zalloc_cpumask_var(&mask, GFP_KERNEL)) {
2128 virtnet_clean_affinity(vi);
2132 num_cpu = num_online_cpus();
2133 stride = max_t(int, num_cpu / vi->curr_queue_pairs, 1);
2134 stragglers = num_cpu >= vi->curr_queue_pairs ?
2135 num_cpu % vi->curr_queue_pairs :
2137 cpu = cpumask_next(-1, cpu_online_mask);
2139 for (i = 0; i < vi->curr_queue_pairs; i++) {
2140 group_size = stride + (i < stragglers ? 1 : 0);
2142 for (j = 0; j < group_size; j++) {
2143 cpumask_set_cpu(cpu, mask);
2144 cpu = cpumask_next_wrap(cpu, cpu_online_mask,
2147 virtqueue_set_affinity(vi->rq[i].vq, mask);
2148 virtqueue_set_affinity(vi->sq[i].vq, mask);
2149 __netif_set_xps_queue(vi->dev, cpumask_bits(mask), i, XPS_CPUS);
2150 cpumask_clear(mask);
2153 vi->affinity_hint_set = true;
2154 free_cpumask_var(mask);
2157 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
2159 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2161 virtnet_set_affinity(vi);
2165 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
2167 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2169 virtnet_set_affinity(vi);
2173 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
2175 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2178 virtnet_clean_affinity(vi);
2182 static enum cpuhp_state virtionet_online;
2184 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
2188 ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
2191 ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2195 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2199 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
2201 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2202 cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2206 static void virtnet_get_ringparam(struct net_device *dev,
2207 struct ethtool_ringparam *ring)
2209 struct virtnet_info *vi = netdev_priv(dev);
2211 ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
2212 ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
2213 ring->rx_pending = ring->rx_max_pending;
2214 ring->tx_pending = ring->tx_max_pending;
2218 static void virtnet_get_drvinfo(struct net_device *dev,
2219 struct ethtool_drvinfo *info)
2221 struct virtnet_info *vi = netdev_priv(dev);
2222 struct virtio_device *vdev = vi->vdev;
2224 strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
2225 strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
2226 strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
2230 /* TODO: Eliminate OOO packets during switching */
2231 static int virtnet_set_channels(struct net_device *dev,
2232 struct ethtool_channels *channels)
2234 struct virtnet_info *vi = netdev_priv(dev);
2235 u16 queue_pairs = channels->combined_count;
2238 /* We don't support separate rx/tx channels.
2239 * We don't allow setting 'other' channels.
2241 if (channels->rx_count || channels->tx_count || channels->other_count)
2244 if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
2247 /* For now we don't support modifying channels while XDP is loaded
2248 * also when XDP is loaded all RX queues have XDP programs so we only
2249 * need to check a single RX queue.
2251 if (vi->rq[0].xdp_prog)
2255 err = _virtnet_set_queues(vi, queue_pairs);
2260 virtnet_set_affinity(vi);
2263 netif_set_real_num_tx_queues(dev, queue_pairs);
2264 netif_set_real_num_rx_queues(dev, queue_pairs);
2269 static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *data)
2271 struct virtnet_info *vi = netdev_priv(dev);
2275 switch (stringset) {
2277 for (i = 0; i < vi->curr_queue_pairs; i++) {
2278 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++)
2279 ethtool_sprintf(&p, "rx_queue_%u_%s", i,
2280 virtnet_rq_stats_desc[j].desc);
2283 for (i = 0; i < vi->curr_queue_pairs; i++) {
2284 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++)
2285 ethtool_sprintf(&p, "tx_queue_%u_%s", i,
2286 virtnet_sq_stats_desc[j].desc);
2292 static int virtnet_get_sset_count(struct net_device *dev, int sset)
2294 struct virtnet_info *vi = netdev_priv(dev);
2298 return vi->curr_queue_pairs * (VIRTNET_RQ_STATS_LEN +
2299 VIRTNET_SQ_STATS_LEN);
2305 static void virtnet_get_ethtool_stats(struct net_device *dev,
2306 struct ethtool_stats *stats, u64 *data)
2308 struct virtnet_info *vi = netdev_priv(dev);
2309 unsigned int idx = 0, start, i, j;
2310 const u8 *stats_base;
2313 for (i = 0; i < vi->curr_queue_pairs; i++) {
2314 struct receive_queue *rq = &vi->rq[i];
2316 stats_base = (u8 *)&rq->stats;
2318 start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
2319 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
2320 offset = virtnet_rq_stats_desc[j].offset;
2321 data[idx + j] = *(u64 *)(stats_base + offset);
2323 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
2324 idx += VIRTNET_RQ_STATS_LEN;
2327 for (i = 0; i < vi->curr_queue_pairs; i++) {
2328 struct send_queue *sq = &vi->sq[i];
2330 stats_base = (u8 *)&sq->stats;
2332 start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
2333 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
2334 offset = virtnet_sq_stats_desc[j].offset;
2335 data[idx + j] = *(u64 *)(stats_base + offset);
2337 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
2338 idx += VIRTNET_SQ_STATS_LEN;
2342 static void virtnet_get_channels(struct net_device *dev,
2343 struct ethtool_channels *channels)
2345 struct virtnet_info *vi = netdev_priv(dev);
2347 channels->combined_count = vi->curr_queue_pairs;
2348 channels->max_combined = vi->max_queue_pairs;
2349 channels->max_other = 0;
2350 channels->rx_count = 0;
2351 channels->tx_count = 0;
2352 channels->other_count = 0;
2355 static int virtnet_set_link_ksettings(struct net_device *dev,
2356 const struct ethtool_link_ksettings *cmd)
2358 struct virtnet_info *vi = netdev_priv(dev);
2360 return ethtool_virtdev_set_link_ksettings(dev, cmd,
2361 &vi->speed, &vi->duplex);
2364 static int virtnet_get_link_ksettings(struct net_device *dev,
2365 struct ethtool_link_ksettings *cmd)
2367 struct virtnet_info *vi = netdev_priv(dev);
2369 cmd->base.speed = vi->speed;
2370 cmd->base.duplex = vi->duplex;
2371 cmd->base.port = PORT_OTHER;
2376 static int virtnet_set_coalesce(struct net_device *dev,
2377 struct ethtool_coalesce *ec,
2378 struct kernel_ethtool_coalesce *kernel_coal,
2379 struct netlink_ext_ack *extack)
2381 struct virtnet_info *vi = netdev_priv(dev);
2384 if (ec->tx_max_coalesced_frames > 1 ||
2385 ec->rx_max_coalesced_frames != 1)
2388 napi_weight = ec->tx_max_coalesced_frames ? NAPI_POLL_WEIGHT : 0;
2389 if (napi_weight ^ vi->sq[0].napi.weight) {
2390 if (dev->flags & IFF_UP)
2392 for (i = 0; i < vi->max_queue_pairs; i++)
2393 vi->sq[i].napi.weight = napi_weight;
2399 static int virtnet_get_coalesce(struct net_device *dev,
2400 struct ethtool_coalesce *ec,
2401 struct kernel_ethtool_coalesce *kernel_coal,
2402 struct netlink_ext_ack *extack)
2404 struct ethtool_coalesce ec_default = {
2405 .cmd = ETHTOOL_GCOALESCE,
2406 .rx_max_coalesced_frames = 1,
2408 struct virtnet_info *vi = netdev_priv(dev);
2410 memcpy(ec, &ec_default, sizeof(ec_default));
2412 if (vi->sq[0].napi.weight)
2413 ec->tx_max_coalesced_frames = 1;
2418 static void virtnet_init_settings(struct net_device *dev)
2420 struct virtnet_info *vi = netdev_priv(dev);
2422 vi->speed = SPEED_UNKNOWN;
2423 vi->duplex = DUPLEX_UNKNOWN;
2426 static void virtnet_update_settings(struct virtnet_info *vi)
2431 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_SPEED_DUPLEX))
2434 virtio_cread_le(vi->vdev, struct virtio_net_config, speed, &speed);
2436 if (ethtool_validate_speed(speed))
2439 virtio_cread_le(vi->vdev, struct virtio_net_config, duplex, &duplex);
2441 if (ethtool_validate_duplex(duplex))
2442 vi->duplex = duplex;
2445 static const struct ethtool_ops virtnet_ethtool_ops = {
2446 .supported_coalesce_params = ETHTOOL_COALESCE_MAX_FRAMES,
2447 .get_drvinfo = virtnet_get_drvinfo,
2448 .get_link = ethtool_op_get_link,
2449 .get_ringparam = virtnet_get_ringparam,
2450 .get_strings = virtnet_get_strings,
2451 .get_sset_count = virtnet_get_sset_count,
2452 .get_ethtool_stats = virtnet_get_ethtool_stats,
2453 .set_channels = virtnet_set_channels,
2454 .get_channels = virtnet_get_channels,
2455 .get_ts_info = ethtool_op_get_ts_info,
2456 .get_link_ksettings = virtnet_get_link_ksettings,
2457 .set_link_ksettings = virtnet_set_link_ksettings,
2458 .set_coalesce = virtnet_set_coalesce,
2459 .get_coalesce = virtnet_get_coalesce,
2462 static void virtnet_freeze_down(struct virtio_device *vdev)
2464 struct virtnet_info *vi = vdev->priv;
2466 /* Make sure no work handler is accessing the device */
2467 flush_work(&vi->config_work);
2469 netif_tx_lock_bh(vi->dev);
2470 netif_device_detach(vi->dev);
2471 netif_tx_unlock_bh(vi->dev);
2472 if (netif_running(vi->dev))
2473 virtnet_close(vi->dev);
2476 static int init_vqs(struct virtnet_info *vi);
2478 static int virtnet_restore_up(struct virtio_device *vdev)
2480 struct virtnet_info *vi = vdev->priv;
2487 virtio_device_ready(vdev);
2489 enable_delayed_refill(vi);
2491 if (netif_running(vi->dev)) {
2492 err = virtnet_open(vi->dev);
2497 netif_tx_lock_bh(vi->dev);
2498 netif_device_attach(vi->dev);
2499 netif_tx_unlock_bh(vi->dev);
2503 static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads)
2505 struct scatterlist sg;
2506 vi->ctrl->offloads = cpu_to_virtio64(vi->vdev, offloads);
2508 sg_init_one(&sg, &vi->ctrl->offloads, sizeof(vi->ctrl->offloads));
2510 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
2511 VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) {
2512 dev_warn(&vi->dev->dev, "Fail to set guest offload.\n");
2519 static int virtnet_clear_guest_offloads(struct virtnet_info *vi)
2523 if (!vi->guest_offloads)
2526 return virtnet_set_guest_offloads(vi, offloads);
2529 static int virtnet_restore_guest_offloads(struct virtnet_info *vi)
2531 u64 offloads = vi->guest_offloads;
2533 if (!vi->guest_offloads)
2536 return virtnet_set_guest_offloads(vi, offloads);
2539 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
2540 struct netlink_ext_ack *extack)
2542 unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr);
2543 struct virtnet_info *vi = netdev_priv(dev);
2544 struct bpf_prog *old_prog;
2545 u16 xdp_qp = 0, curr_qp;
2548 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)
2549 && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2550 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2551 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
2552 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) ||
2553 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))) {
2554 NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing GRO_HW/CSUM, disable GRO_HW/CSUM first");
2558 if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
2559 NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required");
2563 if (dev->mtu > max_sz) {
2564 NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP");
2565 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz);
2569 curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
2571 xdp_qp = nr_cpu_ids;
2573 /* XDP requires extra queues for XDP_TX */
2574 if (curr_qp + xdp_qp > vi->max_queue_pairs) {
2575 netdev_warn(dev, "XDP request %i queues but max is %i. XDP_TX and XDP_REDIRECT will operate in a slower locked tx mode.\n",
2576 curr_qp + xdp_qp, vi->max_queue_pairs);
2580 old_prog = rtnl_dereference(vi->rq[0].xdp_prog);
2581 if (!prog && !old_prog)
2585 bpf_prog_add(prog, vi->max_queue_pairs - 1);
2587 /* Make sure NAPI is not using any XDP TX queues for RX. */
2588 if (netif_running(dev)) {
2589 for (i = 0; i < vi->max_queue_pairs; i++) {
2590 napi_disable(&vi->rq[i].napi);
2591 virtnet_napi_tx_disable(&vi->sq[i].napi);
2596 for (i = 0; i < vi->max_queue_pairs; i++) {
2597 rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2599 virtnet_restore_guest_offloads(vi);
2604 err = _virtnet_set_queues(vi, curr_qp + xdp_qp);
2607 netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
2608 vi->xdp_queue_pairs = xdp_qp;
2611 vi->xdp_enabled = true;
2612 for (i = 0; i < vi->max_queue_pairs; i++) {
2613 rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2614 if (i == 0 && !old_prog)
2615 virtnet_clear_guest_offloads(vi);
2618 vi->xdp_enabled = false;
2621 for (i = 0; i < vi->max_queue_pairs; i++) {
2623 bpf_prog_put(old_prog);
2624 if (netif_running(dev)) {
2625 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2626 virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2635 virtnet_clear_guest_offloads(vi);
2636 for (i = 0; i < vi->max_queue_pairs; i++)
2637 rcu_assign_pointer(vi->rq[i].xdp_prog, old_prog);
2640 if (netif_running(dev)) {
2641 for (i = 0; i < vi->max_queue_pairs; i++) {
2642 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2643 virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2648 bpf_prog_sub(prog, vi->max_queue_pairs - 1);
2652 static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp)
2654 switch (xdp->command) {
2655 case XDP_SETUP_PROG:
2656 return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
2662 static int virtnet_get_phys_port_name(struct net_device *dev, char *buf,
2665 struct virtnet_info *vi = netdev_priv(dev);
2668 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
2671 ret = snprintf(buf, len, "sby");
2678 static int virtnet_set_features(struct net_device *dev,
2679 netdev_features_t features)
2681 struct virtnet_info *vi = netdev_priv(dev);
2685 if ((dev->features ^ features) & NETIF_F_GRO_HW) {
2686 if (vi->xdp_enabled)
2689 if (features & NETIF_F_GRO_HW)
2690 offloads = vi->guest_offloads_capable;
2692 offloads = vi->guest_offloads_capable &
2693 ~GUEST_OFFLOAD_GRO_HW_MASK;
2695 err = virtnet_set_guest_offloads(vi, offloads);
2698 vi->guest_offloads = offloads;
2704 static const struct net_device_ops virtnet_netdev = {
2705 .ndo_open = virtnet_open,
2706 .ndo_stop = virtnet_close,
2707 .ndo_start_xmit = start_xmit,
2708 .ndo_validate_addr = eth_validate_addr,
2709 .ndo_set_mac_address = virtnet_set_mac_address,
2710 .ndo_set_rx_mode = virtnet_set_rx_mode,
2711 .ndo_get_stats64 = virtnet_stats,
2712 .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
2713 .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
2714 .ndo_bpf = virtnet_xdp,
2715 .ndo_xdp_xmit = virtnet_xdp_xmit,
2716 .ndo_features_check = passthru_features_check,
2717 .ndo_get_phys_port_name = virtnet_get_phys_port_name,
2718 .ndo_set_features = virtnet_set_features,
2721 static void virtnet_config_changed_work(struct work_struct *work)
2723 struct virtnet_info *vi =
2724 container_of(work, struct virtnet_info, config_work);
2727 if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
2728 struct virtio_net_config, status, &v) < 0)
2731 if (v & VIRTIO_NET_S_ANNOUNCE) {
2732 netdev_notify_peers(vi->dev);
2733 virtnet_ack_link_announce(vi);
2736 /* Ignore unknown (future) status bits */
2737 v &= VIRTIO_NET_S_LINK_UP;
2739 if (vi->status == v)
2744 if (vi->status & VIRTIO_NET_S_LINK_UP) {
2745 virtnet_update_settings(vi);
2746 netif_carrier_on(vi->dev);
2747 netif_tx_wake_all_queues(vi->dev);
2749 netif_carrier_off(vi->dev);
2750 netif_tx_stop_all_queues(vi->dev);
2754 static void virtnet_config_changed(struct virtio_device *vdev)
2756 struct virtnet_info *vi = vdev->priv;
2758 schedule_work(&vi->config_work);
2761 static void virtnet_free_queues(struct virtnet_info *vi)
2765 for (i = 0; i < vi->max_queue_pairs; i++) {
2766 __netif_napi_del(&vi->rq[i].napi);
2767 __netif_napi_del(&vi->sq[i].napi);
2770 /* We called __netif_napi_del(),
2771 * we need to respect an RCU grace period before freeing vi->rq
2780 static void _free_receive_bufs(struct virtnet_info *vi)
2782 struct bpf_prog *old_prog;
2785 for (i = 0; i < vi->max_queue_pairs; i++) {
2786 while (vi->rq[i].pages)
2787 __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
2789 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2790 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
2792 bpf_prog_put(old_prog);
2796 static void free_receive_bufs(struct virtnet_info *vi)
2799 _free_receive_bufs(vi);
2803 static void free_receive_page_frags(struct virtnet_info *vi)
2806 for (i = 0; i < vi->max_queue_pairs; i++)
2807 if (vi->rq[i].alloc_frag.page)
2808 put_page(vi->rq[i].alloc_frag.page);
2811 static void free_unused_bufs(struct virtnet_info *vi)
2816 for (i = 0; i < vi->max_queue_pairs; i++) {
2817 struct virtqueue *vq = vi->sq[i].vq;
2818 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2819 if (!is_xdp_frame(buf))
2822 xdp_return_frame(ptr_to_xdp(buf));
2826 for (i = 0; i < vi->max_queue_pairs; i++) {
2827 struct virtqueue *vq = vi->rq[i].vq;
2829 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2830 if (vi->mergeable_rx_bufs) {
2831 put_page(virt_to_head_page(buf));
2832 } else if (vi->big_packets) {
2833 give_pages(&vi->rq[i], buf);
2835 put_page(virt_to_head_page(buf));
2841 static void virtnet_del_vqs(struct virtnet_info *vi)
2843 struct virtio_device *vdev = vi->vdev;
2845 virtnet_clean_affinity(vi);
2847 vdev->config->del_vqs(vdev);
2849 virtnet_free_queues(vi);
2852 /* How large should a single buffer be so a queue full of these can fit at
2853 * least one full packet?
2854 * Logic below assumes the mergeable buffer header is used.
2856 static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq)
2858 const unsigned int hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2859 unsigned int rq_size = virtqueue_get_vring_size(vq);
2860 unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu;
2861 unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len;
2862 unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size);
2864 return max(max(min_buf_len, hdr_len) - hdr_len,
2865 (unsigned int)GOOD_PACKET_LEN);
2868 static int virtnet_find_vqs(struct virtnet_info *vi)
2870 vq_callback_t **callbacks;
2871 struct virtqueue **vqs;
2877 /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
2878 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
2879 * possible control vq.
2881 total_vqs = vi->max_queue_pairs * 2 +
2882 virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
2884 /* Allocate space for find_vqs parameters */
2885 vqs = kcalloc(total_vqs, sizeof(*vqs), GFP_KERNEL);
2888 callbacks = kmalloc_array(total_vqs, sizeof(*callbacks), GFP_KERNEL);
2891 names = kmalloc_array(total_vqs, sizeof(*names), GFP_KERNEL);
2894 if (!vi->big_packets || vi->mergeable_rx_bufs) {
2895 ctx = kcalloc(total_vqs, sizeof(*ctx), GFP_KERNEL);
2902 /* Parameters for control virtqueue, if any */
2904 callbacks[total_vqs - 1] = NULL;
2905 names[total_vqs - 1] = "control";
2908 /* Allocate/initialize parameters for send/receive virtqueues */
2909 for (i = 0; i < vi->max_queue_pairs; i++) {
2910 callbacks[rxq2vq(i)] = skb_recv_done;
2911 callbacks[txq2vq(i)] = skb_xmit_done;
2912 sprintf(vi->rq[i].name, "input.%d", i);
2913 sprintf(vi->sq[i].name, "output.%d", i);
2914 names[rxq2vq(i)] = vi->rq[i].name;
2915 names[txq2vq(i)] = vi->sq[i].name;
2917 ctx[rxq2vq(i)] = true;
2920 ret = virtio_find_vqs_ctx(vi->vdev, total_vqs, vqs, callbacks,
2926 vi->cvq = vqs[total_vqs - 1];
2927 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
2928 vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
2931 for (i = 0; i < vi->max_queue_pairs; i++) {
2932 vi->rq[i].vq = vqs[rxq2vq(i)];
2933 vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq);
2934 vi->sq[i].vq = vqs[txq2vq(i)];
2937 /* run here: ret == 0. */
2952 static int virtnet_alloc_queues(struct virtnet_info *vi)
2957 vi->ctrl = kzalloc(sizeof(*vi->ctrl), GFP_KERNEL);
2963 vi->sq = kcalloc(vi->max_queue_pairs, sizeof(*vi->sq), GFP_KERNEL);
2966 vi->rq = kcalloc(vi->max_queue_pairs, sizeof(*vi->rq), GFP_KERNEL);
2970 INIT_DELAYED_WORK(&vi->refill, refill_work);
2971 for (i = 0; i < vi->max_queue_pairs; i++) {
2972 vi->rq[i].pages = NULL;
2973 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
2975 netif_tx_napi_add(vi->dev, &vi->sq[i].napi, virtnet_poll_tx,
2976 napi_tx ? napi_weight : 0);
2978 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
2979 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
2980 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
2982 u64_stats_init(&vi->rq[i].stats.syncp);
2983 u64_stats_init(&vi->sq[i].stats.syncp);
2996 static int init_vqs(struct virtnet_info *vi)
3000 /* Allocate send & receive queues */
3001 ret = virtnet_alloc_queues(vi);
3005 ret = virtnet_find_vqs(vi);
3010 virtnet_set_affinity(vi);
3016 virtnet_free_queues(vi);
3022 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
3025 struct virtnet_info *vi = netdev_priv(queue->dev);
3026 unsigned int queue_index = get_netdev_rx_queue_index(queue);
3027 unsigned int headroom = virtnet_get_headroom(vi);
3028 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
3029 struct ewma_pkt_len *avg;
3031 BUG_ON(queue_index >= vi->max_queue_pairs);
3032 avg = &vi->rq[queue_index].mrg_avg_pkt_len;
3033 return sprintf(buf, "%u\n",
3034 get_mergeable_buf_len(&vi->rq[queue_index], avg,
3035 SKB_DATA_ALIGN(headroom + tailroom)));
3038 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
3039 __ATTR_RO(mergeable_rx_buffer_size);
3041 static struct attribute *virtio_net_mrg_rx_attrs[] = {
3042 &mergeable_rx_buffer_size_attribute.attr,
3046 static const struct attribute_group virtio_net_mrg_rx_group = {
3047 .name = "virtio_net",
3048 .attrs = virtio_net_mrg_rx_attrs
3052 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
3054 const char *fname, const char *dname)
3056 if (!virtio_has_feature(vdev, fbit))
3059 dev_err(&vdev->dev, "device advertises feature %s but not %s",
3065 #define VIRTNET_FAIL_ON(vdev, fbit, dbit) \
3066 virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
3068 static bool virtnet_validate_features(struct virtio_device *vdev)
3070 if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
3071 (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
3072 "VIRTIO_NET_F_CTRL_VQ") ||
3073 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
3074 "VIRTIO_NET_F_CTRL_VQ") ||
3075 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
3076 "VIRTIO_NET_F_CTRL_VQ") ||
3077 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
3078 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
3079 "VIRTIO_NET_F_CTRL_VQ"))) {
3086 #define MIN_MTU ETH_MIN_MTU
3087 #define MAX_MTU ETH_MAX_MTU
3089 static int virtnet_validate(struct virtio_device *vdev)
3091 if (!vdev->config->get) {
3092 dev_err(&vdev->dev, "%s failure: config access disabled\n",
3097 if (!virtnet_validate_features(vdev))
3100 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
3101 int mtu = virtio_cread16(vdev,
3102 offsetof(struct virtio_net_config,
3105 __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
3111 static int virtnet_probe(struct virtio_device *vdev)
3113 int i, err = -ENOMEM;
3114 struct net_device *dev;
3115 struct virtnet_info *vi;
3116 u16 max_queue_pairs;
3119 /* Find if host supports multiqueue virtio_net device */
3120 err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
3121 struct virtio_net_config,
3122 max_virtqueue_pairs, &max_queue_pairs);
3124 /* We need at least 2 queue's */
3125 if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
3126 max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
3127 !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
3128 max_queue_pairs = 1;
3130 /* Allocate ourselves a network device with room for our info */
3131 dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
3135 /* Set up network device as normal. */
3136 dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE |
3137 IFF_TX_SKB_NO_LINEAR;
3138 dev->netdev_ops = &virtnet_netdev;
3139 dev->features = NETIF_F_HIGHDMA;
3141 dev->ethtool_ops = &virtnet_ethtool_ops;
3142 SET_NETDEV_DEV(dev, &vdev->dev);
3144 /* Do we support "hardware" checksums? */
3145 if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
3146 /* This opens up the world of extra features. */
3147 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
3149 dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
3151 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
3152 dev->hw_features |= NETIF_F_TSO
3153 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
3155 /* Individual feature bits: what can host handle? */
3156 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
3157 dev->hw_features |= NETIF_F_TSO;
3158 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
3159 dev->hw_features |= NETIF_F_TSO6;
3160 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
3161 dev->hw_features |= NETIF_F_TSO_ECN;
3163 dev->features |= NETIF_F_GSO_ROBUST;
3166 dev->features |= dev->hw_features & NETIF_F_ALL_TSO;
3167 /* (!csum && gso) case will be fixed by register_netdev() */
3169 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
3170 dev->features |= NETIF_F_RXCSUM;
3171 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3172 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6))
3173 dev->features |= NETIF_F_GRO_HW;
3174 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS))
3175 dev->hw_features |= NETIF_F_GRO_HW;
3177 dev->vlan_features = dev->features;
3179 /* MTU range: 68 - 65535 */
3180 dev->min_mtu = MIN_MTU;
3181 dev->max_mtu = MAX_MTU;
3183 /* Configuration may specify what MAC to use. Otherwise random. */
3184 if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
3185 virtio_cread_bytes(vdev,
3186 offsetof(struct virtio_net_config, mac),
3187 dev->dev_addr, dev->addr_len);
3189 eth_hw_addr_random(dev);
3191 /* Set up our device-specific information */
3192 vi = netdev_priv(dev);
3197 INIT_WORK(&vi->config_work, virtnet_config_changed_work);
3198 spin_lock_init(&vi->refill_lock);
3200 /* If we can receive ANY GSO packets, we must allocate large ones. */
3201 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3202 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
3203 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
3204 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
3205 vi->big_packets = true;
3207 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
3208 vi->mergeable_rx_bufs = true;
3210 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
3211 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
3212 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
3214 vi->hdr_len = sizeof(struct virtio_net_hdr);
3216 if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
3217 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
3218 vi->any_header_sg = true;
3220 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
3223 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
3224 mtu = virtio_cread16(vdev,
3225 offsetof(struct virtio_net_config,
3227 if (mtu < dev->min_mtu) {
3228 /* Should never trigger: MTU was previously validated
3229 * in virtnet_validate.
3232 "device MTU appears to have changed it is now %d < %d",
3241 /* TODO: size buffers correctly in this case. */
3242 if (dev->mtu > ETH_DATA_LEN)
3243 vi->big_packets = true;
3246 if (vi->any_header_sg)
3247 dev->needed_headroom = vi->hdr_len;
3249 /* Enable multiqueue by default */
3250 if (num_online_cpus() >= max_queue_pairs)
3251 vi->curr_queue_pairs = max_queue_pairs;
3253 vi->curr_queue_pairs = num_online_cpus();
3254 vi->max_queue_pairs = max_queue_pairs;
3256 /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
3262 if (vi->mergeable_rx_bufs)
3263 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
3265 netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
3266 netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
3268 virtnet_init_settings(dev);
3270 if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY)) {
3271 vi->failover = net_failover_create(vi->dev);
3272 if (IS_ERR(vi->failover)) {
3273 err = PTR_ERR(vi->failover);
3278 /* serialize netdev register + virtio_device_ready() with ndo_open() */
3281 err = register_netdevice(dev);
3283 pr_debug("virtio_net: registering device failed\n");
3288 virtio_device_ready(vdev);
3292 err = virtnet_cpu_notif_add(vi);
3294 pr_debug("virtio_net: registering cpu notifier failed\n");
3295 goto free_unregister_netdev;
3298 virtnet_set_queues(vi, vi->curr_queue_pairs);
3300 /* Assume link up if device can't report link status,
3301 otherwise get link status from config. */
3302 netif_carrier_off(dev);
3303 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
3304 schedule_work(&vi->config_work);
3306 vi->status = VIRTIO_NET_S_LINK_UP;
3307 virtnet_update_settings(vi);
3308 netif_carrier_on(dev);
3311 for (i = 0; i < ARRAY_SIZE(guest_offloads); i++)
3312 if (virtio_has_feature(vi->vdev, guest_offloads[i]))
3313 set_bit(guest_offloads[i], &vi->guest_offloads);
3314 vi->guest_offloads_capable = vi->guest_offloads;
3316 pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
3317 dev->name, max_queue_pairs);
3321 free_unregister_netdev:
3322 vi->vdev->config->reset(vdev);
3324 unregister_netdev(dev);
3326 net_failover_destroy(vi->failover);
3328 cancel_delayed_work_sync(&vi->refill);
3329 free_receive_page_frags(vi);
3330 virtnet_del_vqs(vi);
3336 static void remove_vq_common(struct virtnet_info *vi)
3338 vi->vdev->config->reset(vi->vdev);
3340 /* Free unused buffers in both send and recv, if any. */
3341 free_unused_bufs(vi);
3343 free_receive_bufs(vi);
3345 free_receive_page_frags(vi);
3347 virtnet_del_vqs(vi);
3350 static void virtnet_remove(struct virtio_device *vdev)
3352 struct virtnet_info *vi = vdev->priv;
3354 virtnet_cpu_notif_remove(vi);
3356 /* Make sure no work handler is accessing the device. */
3357 flush_work(&vi->config_work);
3359 unregister_netdev(vi->dev);
3361 net_failover_destroy(vi->failover);
3363 remove_vq_common(vi);
3365 free_netdev(vi->dev);
3368 static __maybe_unused int virtnet_freeze(struct virtio_device *vdev)
3370 struct virtnet_info *vi = vdev->priv;
3372 virtnet_cpu_notif_remove(vi);
3373 virtnet_freeze_down(vdev);
3374 remove_vq_common(vi);
3379 static __maybe_unused int virtnet_restore(struct virtio_device *vdev)
3381 struct virtnet_info *vi = vdev->priv;
3384 err = virtnet_restore_up(vdev);
3387 virtnet_set_queues(vi, vi->curr_queue_pairs);
3389 err = virtnet_cpu_notif_add(vi);
3391 virtnet_freeze_down(vdev);
3392 remove_vq_common(vi);
3399 static struct virtio_device_id id_table[] = {
3400 { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
3404 #define VIRTNET_FEATURES \
3405 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
3407 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
3408 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
3409 VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
3410 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
3411 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
3412 VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
3413 VIRTIO_NET_F_CTRL_MAC_ADDR, \
3414 VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \
3415 VIRTIO_NET_F_SPEED_DUPLEX, VIRTIO_NET_F_STANDBY
3417 static unsigned int features[] = {
3421 static unsigned int features_legacy[] = {
3424 VIRTIO_F_ANY_LAYOUT,
3427 static struct virtio_driver virtio_net_driver = {
3428 .feature_table = features,
3429 .feature_table_size = ARRAY_SIZE(features),
3430 .feature_table_legacy = features_legacy,
3431 .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
3432 .driver.name = KBUILD_MODNAME,
3433 .driver.owner = THIS_MODULE,
3434 .id_table = id_table,
3435 .validate = virtnet_validate,
3436 .probe = virtnet_probe,
3437 .remove = virtnet_remove,
3438 .config_changed = virtnet_config_changed,
3439 #ifdef CONFIG_PM_SLEEP
3440 .freeze = virtnet_freeze,
3441 .restore = virtnet_restore,
3445 static __init int virtio_net_driver_init(void)
3449 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
3451 virtnet_cpu_down_prep);
3454 virtionet_online = ret;
3455 ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
3456 NULL, virtnet_cpu_dead);
3460 ret = register_virtio_driver(&virtio_net_driver);
3465 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3467 cpuhp_remove_multi_state(virtionet_online);
3471 module_init(virtio_net_driver_init);
3473 static __exit void virtio_net_driver_exit(void)
3475 unregister_virtio_driver(&virtio_net_driver);
3476 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3477 cpuhp_remove_multi_state(virtionet_online);
3479 module_exit(virtio_net_driver_exit);
3481 MODULE_DEVICE_TABLE(virtio, id_table);
3482 MODULE_DESCRIPTION("Virtio network driver");
3483 MODULE_LICENSE("GPL");