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_LRO_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 refilling if we run low on memory. */
217 struct delayed_work refill;
219 /* Work struct for config space updates */
220 struct work_struct config_work;
222 /* Does the affinity hint is set for virtqueues? */
223 bool affinity_hint_set;
225 /* CPU hotplug instances for online & dead */
226 struct hlist_node node;
227 struct hlist_node node_dead;
229 struct control_buf *ctrl;
231 /* Ethtool settings */
235 unsigned long guest_offloads;
236 unsigned long guest_offloads_capable;
238 /* failover when STANDBY feature enabled */
239 struct failover *failover;
242 struct padded_vnet_hdr {
243 struct virtio_net_hdr_mrg_rxbuf hdr;
245 * hdr is in a separate sg buffer, and data sg buffer shares same page
246 * with this header sg. This padding makes next sg 16 byte aligned
252 static bool is_xdp_frame(void *ptr)
254 return (unsigned long)ptr & VIRTIO_XDP_FLAG;
257 static void *xdp_to_ptr(struct xdp_frame *ptr)
259 return (void *)((unsigned long)ptr | VIRTIO_XDP_FLAG);
262 static struct xdp_frame *ptr_to_xdp(void *ptr)
264 return (struct xdp_frame *)((unsigned long)ptr & ~VIRTIO_XDP_FLAG);
267 /* Converting between virtqueue no. and kernel tx/rx queue no.
268 * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
270 static int vq2txq(struct virtqueue *vq)
272 return (vq->index - 1) / 2;
275 static int txq2vq(int txq)
280 static int vq2rxq(struct virtqueue *vq)
282 return vq->index / 2;
285 static int rxq2vq(int rxq)
290 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
292 return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
296 * private is used to chain pages for big packets, put the whole
297 * most recent used list in the beginning for reuse
299 static void give_pages(struct receive_queue *rq, struct page *page)
303 /* Find end of list, sew whole thing into vi->rq.pages. */
304 for (end = page; end->private; end = (struct page *)end->private);
305 end->private = (unsigned long)rq->pages;
309 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
311 struct page *p = rq->pages;
314 rq->pages = (struct page *)p->private;
315 /* clear private here, it is used to chain pages */
318 p = alloc_page(gfp_mask);
322 static void virtqueue_napi_schedule(struct napi_struct *napi,
323 struct virtqueue *vq)
325 if (napi_schedule_prep(napi)) {
326 virtqueue_disable_cb(vq);
327 __napi_schedule(napi);
331 static void virtqueue_napi_complete(struct napi_struct *napi,
332 struct virtqueue *vq, int processed)
336 opaque = virtqueue_enable_cb_prepare(vq);
337 if (napi_complete_done(napi, processed)) {
338 if (unlikely(virtqueue_poll(vq, opaque)))
339 virtqueue_napi_schedule(napi, vq);
341 virtqueue_disable_cb(vq);
345 static void skb_xmit_done(struct virtqueue *vq)
347 struct virtnet_info *vi = vq->vdev->priv;
348 struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi;
350 /* Suppress further interrupts. */
351 virtqueue_disable_cb(vq);
354 virtqueue_napi_schedule(napi, vq);
356 /* We were probably waiting for more output buffers. */
357 netif_wake_subqueue(vi->dev, vq2txq(vq));
360 #define MRG_CTX_HEADER_SHIFT 22
361 static void *mergeable_len_to_ctx(unsigned int truesize,
362 unsigned int headroom)
364 return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) | truesize);
367 static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx)
369 return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT;
372 static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx)
374 return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) - 1);
377 /* Called from bottom half context */
378 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
379 struct receive_queue *rq,
380 struct page *page, unsigned int offset,
381 unsigned int len, unsigned int truesize,
382 bool hdr_valid, unsigned int metasize,
386 struct virtio_net_hdr_mrg_rxbuf *hdr;
387 unsigned int copy, hdr_len, hdr_padded_len;
388 struct page *page_to_free = NULL;
389 int tailroom, shinfo_size;
390 char *p, *hdr_p, *buf;
392 p = page_address(page) + offset;
395 hdr_len = vi->hdr_len;
396 if (vi->mergeable_rx_bufs)
397 hdr_padded_len = sizeof(*hdr);
399 hdr_padded_len = sizeof(struct padded_vnet_hdr);
401 /* If whole_page, there is an offset between the beginning of the
402 * data and the allocated space, otherwise the data and the allocated
405 * Buffers with headroom use PAGE_SIZE as alloc size, see
406 * add_recvbuf_mergeable() + get_mergeable_buf_len()
409 /* Buffers with whole_page use PAGE_SIZE as alloc size,
410 * see add_recvbuf_mergeable() + get_mergeable_buf_len()
412 truesize = PAGE_SIZE;
414 /* page maybe head page, so we should get the buf by p, not the
417 tailroom = truesize - len - offset_in_page(p);
418 buf = (char *)((unsigned long)p & PAGE_MASK);
420 tailroom = truesize - len;
425 offset += hdr_padded_len;
428 shinfo_size = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
430 /* copy small packet so we can reuse these pages */
431 if (!NET_IP_ALIGN && len > GOOD_COPY_LEN && tailroom >= shinfo_size) {
432 skb = build_skb(buf, truesize);
436 skb_reserve(skb, p - buf);
441 /* copy small packet so we can reuse these pages for small data */
442 skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
446 /* Copy all frame if it fits skb->head, otherwise
447 * we let virtio_net_hdr_to_skb() and GRO pull headers as needed.
449 if (len <= skb_tailroom(skb))
452 copy = ETH_HLEN + metasize;
453 skb_put_data(skb, p, copy);
458 if (vi->mergeable_rx_bufs) {
460 skb_add_rx_frag(skb, 0, page, offset, len, truesize);
467 * Verify that we can indeed put this data into a skb.
468 * This is here to handle cases when the device erroneously
469 * tries to receive more than is possible. This is usually
470 * the case of a broken device.
472 if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
473 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
477 BUG_ON(offset >= PAGE_SIZE);
479 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
480 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
481 frag_size, truesize);
483 page = (struct page *)page->private;
488 give_pages(rq, page);
491 /* hdr_valid means no XDP, so we can copy the vnet header */
493 hdr = skb_vnet_hdr(skb);
494 memcpy(hdr, hdr_p, hdr_len);
497 put_page(page_to_free);
500 __skb_pull(skb, metasize);
501 skb_metadata_set(skb, metasize);
507 static int __virtnet_xdp_xmit_one(struct virtnet_info *vi,
508 struct send_queue *sq,
509 struct xdp_frame *xdpf)
511 struct virtio_net_hdr_mrg_rxbuf *hdr;
514 if (unlikely(xdpf->headroom < vi->hdr_len))
517 /* Make room for virtqueue hdr (also change xdpf->headroom?) */
518 xdpf->data -= vi->hdr_len;
519 /* Zero header and leave csum up to XDP layers */
521 memset(hdr, 0, vi->hdr_len);
522 xdpf->len += vi->hdr_len;
524 sg_init_one(sq->sg, xdpf->data, xdpf->len);
526 err = virtqueue_add_outbuf(sq->vq, sq->sg, 1, xdp_to_ptr(xdpf),
529 return -ENOSPC; /* Caller handle free/refcnt */
534 /* when vi->curr_queue_pairs > nr_cpu_ids, the txq/sq is only used for xdp tx on
535 * the current cpu, so it does not need to be locked.
537 * Here we use marco instead of inline functions because we have to deal with
538 * three issues at the same time: 1. the choice of sq. 2. judge and execute the
539 * lock/unlock of txq 3. make sparse happy. It is difficult for two inline
540 * functions to perfectly solve these three problems at the same time.
542 #define virtnet_xdp_get_sq(vi) ({ \
543 struct netdev_queue *txq; \
544 typeof(vi) v = (vi); \
547 if (v->curr_queue_pairs > nr_cpu_ids) { \
548 qp = v->curr_queue_pairs - v->xdp_queue_pairs; \
549 qp += smp_processor_id(); \
550 txq = netdev_get_tx_queue(v->dev, qp); \
551 __netif_tx_acquire(txq); \
553 qp = smp_processor_id() % v->curr_queue_pairs; \
554 txq = netdev_get_tx_queue(v->dev, qp); \
555 __netif_tx_lock(txq, raw_smp_processor_id()); \
560 #define virtnet_xdp_put_sq(vi, q) { \
561 struct netdev_queue *txq; \
562 typeof(vi) v = (vi); \
564 txq = netdev_get_tx_queue(v->dev, (q) - v->sq); \
565 if (v->curr_queue_pairs > nr_cpu_ids) \
566 __netif_tx_release(txq); \
568 __netif_tx_unlock(txq); \
571 static int virtnet_xdp_xmit(struct net_device *dev,
572 int n, struct xdp_frame **frames, u32 flags)
574 struct virtnet_info *vi = netdev_priv(dev);
575 struct receive_queue *rq = vi->rq;
576 struct bpf_prog *xdp_prog;
577 struct send_queue *sq;
587 /* Only allow ndo_xdp_xmit if XDP is loaded on dev, as this
588 * indicate XDP resources have been successfully allocated.
590 xdp_prog = rcu_access_pointer(rq->xdp_prog);
594 sq = virtnet_xdp_get_sq(vi);
596 if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK)) {
601 /* Free up any pending old buffers before queueing new ones. */
602 while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
603 if (likely(is_xdp_frame(ptr))) {
604 struct xdp_frame *frame = ptr_to_xdp(ptr);
607 xdp_return_frame(frame);
609 struct sk_buff *skb = ptr;
612 napi_consume_skb(skb, false);
617 for (i = 0; i < n; i++) {
618 struct xdp_frame *xdpf = frames[i];
620 if (__virtnet_xdp_xmit_one(vi, sq, xdpf))
626 if (flags & XDP_XMIT_FLUSH) {
627 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq))
631 u64_stats_update_begin(&sq->stats.syncp);
632 sq->stats.bytes += bytes;
633 sq->stats.packets += packets;
634 sq->stats.xdp_tx += n;
635 sq->stats.xdp_tx_drops += n - nxmit;
636 sq->stats.kicks += kicks;
637 u64_stats_update_end(&sq->stats.syncp);
639 virtnet_xdp_put_sq(vi, sq);
643 static unsigned int virtnet_get_headroom(struct virtnet_info *vi)
645 return vi->xdp_enabled ? VIRTIO_XDP_HEADROOM : 0;
648 /* We copy the packet for XDP in the following cases:
650 * 1) Packet is scattered across multiple rx buffers.
651 * 2) Headroom space is insufficient.
653 * This is inefficient but it's a temporary condition that
654 * we hit right after XDP is enabled and until queue is refilled
655 * with large buffers with sufficient headroom - so it should affect
656 * at most queue size packets.
657 * Afterwards, the conditions to enable
658 * XDP should preclude the underlying device from sending packets
659 * across multiple buffers (num_buf > 1), and we make sure buffers
660 * have enough headroom.
662 static struct page *xdp_linearize_page(struct receive_queue *rq,
669 struct page *page = alloc_page(GFP_ATOMIC);
674 memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
678 int tailroom = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
683 buf = virtqueue_get_buf(rq->vq, &buflen);
687 p = virt_to_head_page(buf);
688 off = buf - page_address(p);
690 /* guard against a misconfigured or uncooperative backend that
691 * is sending packet larger than the MTU.
693 if ((page_off + buflen + tailroom) > PAGE_SIZE) {
698 memcpy(page_address(page) + page_off,
699 page_address(p) + off, buflen);
704 /* Headroom does not contribute to packet length */
705 *len = page_off - VIRTIO_XDP_HEADROOM;
708 __free_pages(page, 0);
712 static struct sk_buff *receive_small(struct net_device *dev,
713 struct virtnet_info *vi,
714 struct receive_queue *rq,
715 void *buf, void *ctx,
717 unsigned int *xdp_xmit,
718 struct virtnet_rq_stats *stats)
721 struct bpf_prog *xdp_prog;
722 unsigned int xdp_headroom = (unsigned long)ctx;
723 unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom;
724 unsigned int headroom = vi->hdr_len + header_offset;
725 unsigned int buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
726 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
727 struct page *page = virt_to_head_page(buf);
728 unsigned int delta = 0;
729 struct page *xdp_page;
731 unsigned int metasize = 0;
736 if (unlikely(len > GOOD_PACKET_LEN)) {
737 pr_debug("%s: rx error: len %u exceeds max size %d\n",
738 dev->name, len, GOOD_PACKET_LEN);
739 dev->stats.rx_length_errors++;
743 xdp_prog = rcu_dereference(rq->xdp_prog);
745 struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset;
746 struct xdp_frame *xdpf;
751 if (unlikely(hdr->hdr.gso_type))
754 if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) {
755 int offset = buf - page_address(page) + header_offset;
756 unsigned int tlen = len + vi->hdr_len;
759 xdp_headroom = virtnet_get_headroom(vi);
760 header_offset = VIRTNET_RX_PAD + xdp_headroom;
761 headroom = vi->hdr_len + header_offset;
762 buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
763 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
764 xdp_page = xdp_linearize_page(rq, &num_buf, page,
765 offset, header_offset,
770 buf = page_address(xdp_page);
775 xdp_init_buff(&xdp, buflen, &rq->xdp_rxq);
776 xdp_prepare_buff(&xdp, buf + VIRTNET_RX_PAD + vi->hdr_len,
777 xdp_headroom, len, true);
778 orig_data = xdp.data;
779 act = bpf_prog_run_xdp(xdp_prog, &xdp);
780 stats->xdp_packets++;
784 /* Recalculate length in case bpf program changed it */
785 delta = orig_data - xdp.data;
786 len = xdp.data_end - xdp.data;
787 metasize = xdp.data - xdp.data_meta;
791 xdpf = xdp_convert_buff_to_frame(&xdp);
794 err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
795 if (unlikely(!err)) {
796 xdp_return_frame_rx_napi(xdpf);
797 } else if (unlikely(err < 0)) {
798 trace_xdp_exception(vi->dev, xdp_prog, act);
801 *xdp_xmit |= VIRTIO_XDP_TX;
805 stats->xdp_redirects++;
806 err = xdp_do_redirect(dev, &xdp, xdp_prog);
809 *xdp_xmit |= VIRTIO_XDP_REDIR;
813 bpf_warn_invalid_xdp_action(act);
816 trace_xdp_exception(vi->dev, xdp_prog, act);
824 skb = build_skb(buf, buflen);
829 skb_reserve(skb, headroom - delta);
832 buf += header_offset;
833 memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
834 } /* keep zeroed vnet hdr since XDP is loaded */
837 skb_metadata_set(skb, metasize);
852 static struct sk_buff *receive_big(struct net_device *dev,
853 struct virtnet_info *vi,
854 struct receive_queue *rq,
857 struct virtnet_rq_stats *stats)
859 struct page *page = buf;
860 struct sk_buff *skb =
861 page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, true, 0, 0);
863 stats->bytes += len - vi->hdr_len;
871 give_pages(rq, page);
875 static struct sk_buff *receive_mergeable(struct net_device *dev,
876 struct virtnet_info *vi,
877 struct receive_queue *rq,
881 unsigned int *xdp_xmit,
882 struct virtnet_rq_stats *stats)
884 struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
885 u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
886 struct page *page = virt_to_head_page(buf);
887 int offset = buf - page_address(page);
888 struct sk_buff *head_skb, *curr_skb;
889 struct bpf_prog *xdp_prog;
890 unsigned int truesize = mergeable_ctx_to_truesize(ctx);
891 unsigned int headroom = mergeable_ctx_to_headroom(ctx);
892 unsigned int metasize = 0;
893 unsigned int frame_sz;
897 stats->bytes += len - vi->hdr_len;
899 if (unlikely(len > truesize)) {
900 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
901 dev->name, len, (unsigned long)ctx);
902 dev->stats.rx_length_errors++;
906 xdp_prog = rcu_dereference(rq->xdp_prog);
908 struct xdp_frame *xdpf;
909 struct page *xdp_page;
914 /* Transient failure which in theory could occur if
915 * in-flight packets from before XDP was enabled reach
916 * the receive path after XDP is loaded.
918 if (unlikely(hdr->hdr.gso_type))
921 /* Buffers with headroom use PAGE_SIZE as alloc size,
922 * see add_recvbuf_mergeable() + get_mergeable_buf_len()
924 frame_sz = headroom ? PAGE_SIZE : truesize;
926 /* This happens when rx buffer size is underestimated
927 * or headroom is not enough because of the buffer
928 * was refilled before XDP is set. This should only
929 * happen for the first several packets, so we don't
930 * care much about its performance.
932 if (unlikely(num_buf > 1 ||
933 headroom < virtnet_get_headroom(vi))) {
934 /* linearize data for XDP */
935 xdp_page = xdp_linearize_page(rq, &num_buf,
939 frame_sz = PAGE_SIZE;
943 offset = VIRTIO_XDP_HEADROOM;
948 /* Allow consuming headroom but reserve enough space to push
949 * the descriptor on if we get an XDP_TX return code.
951 data = page_address(xdp_page) + offset;
952 xdp_init_buff(&xdp, frame_sz - vi->hdr_len, &rq->xdp_rxq);
953 xdp_prepare_buff(&xdp, data - VIRTIO_XDP_HEADROOM + vi->hdr_len,
954 VIRTIO_XDP_HEADROOM, len - vi->hdr_len, true);
956 act = bpf_prog_run_xdp(xdp_prog, &xdp);
957 stats->xdp_packets++;
961 metasize = xdp.data - xdp.data_meta;
963 /* recalculate offset to account for any header
964 * adjustments and minus the metasize to copy the
965 * metadata in page_to_skb(). Note other cases do not
966 * build an skb and avoid using offset
968 offset = xdp.data - page_address(xdp_page) -
969 vi->hdr_len - metasize;
971 /* recalculate len if xdp.data, xdp.data_end or
972 * xdp.data_meta were adjusted
974 len = xdp.data_end - xdp.data + vi->hdr_len + metasize;
975 /* We can only create skb based on xdp_page. */
976 if (unlikely(xdp_page != page)) {
979 head_skb = page_to_skb(vi, rq, xdp_page, offset,
980 len, PAGE_SIZE, false,
987 xdpf = xdp_convert_buff_to_frame(&xdp);
990 err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
991 if (unlikely(!err)) {
992 xdp_return_frame_rx_napi(xdpf);
993 } else if (unlikely(err < 0)) {
994 trace_xdp_exception(vi->dev, xdp_prog, act);
995 if (unlikely(xdp_page != page))
999 *xdp_xmit |= VIRTIO_XDP_TX;
1000 if (unlikely(xdp_page != page))
1005 stats->xdp_redirects++;
1006 err = xdp_do_redirect(dev, &xdp, xdp_prog);
1008 if (unlikely(xdp_page != page))
1012 *xdp_xmit |= VIRTIO_XDP_REDIR;
1013 if (unlikely(xdp_page != page))
1018 bpf_warn_invalid_xdp_action(act);
1021 trace_xdp_exception(vi->dev, xdp_prog, act);
1024 if (unlikely(xdp_page != page))
1025 __free_pages(xdp_page, 0);
1031 head_skb = page_to_skb(vi, rq, page, offset, len, truesize, !xdp_prog,
1032 metasize, !!headroom);
1033 curr_skb = head_skb;
1035 if (unlikely(!curr_skb))
1040 buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx);
1041 if (unlikely(!buf)) {
1042 pr_debug("%s: rx error: %d buffers out of %d missing\n",
1044 virtio16_to_cpu(vi->vdev,
1046 dev->stats.rx_length_errors++;
1050 stats->bytes += len;
1051 page = virt_to_head_page(buf);
1053 truesize = mergeable_ctx_to_truesize(ctx);
1054 if (unlikely(len > truesize)) {
1055 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
1056 dev->name, len, (unsigned long)ctx);
1057 dev->stats.rx_length_errors++;
1061 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
1062 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
1063 struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
1065 if (unlikely(!nskb))
1067 if (curr_skb == head_skb)
1068 skb_shinfo(curr_skb)->frag_list = nskb;
1070 curr_skb->next = nskb;
1072 head_skb->truesize += nskb->truesize;
1075 if (curr_skb != head_skb) {
1076 head_skb->data_len += len;
1077 head_skb->len += len;
1078 head_skb->truesize += truesize;
1080 offset = buf - page_address(page);
1081 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
1083 skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
1086 skb_add_rx_frag(curr_skb, num_skb_frags, page,
1087 offset, len, truesize);
1091 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
1099 while (num_buf-- > 1) {
1100 buf = virtqueue_get_buf(rq->vq, &len);
1101 if (unlikely(!buf)) {
1102 pr_debug("%s: rx error: %d buffers missing\n",
1103 dev->name, num_buf);
1104 dev->stats.rx_length_errors++;
1107 stats->bytes += len;
1108 page = virt_to_head_page(buf);
1113 dev_kfree_skb(head_skb);
1118 static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
1119 void *buf, unsigned int len, void **ctx,
1120 unsigned int *xdp_xmit,
1121 struct virtnet_rq_stats *stats)
1123 struct net_device *dev = vi->dev;
1124 struct sk_buff *skb;
1125 struct virtio_net_hdr_mrg_rxbuf *hdr;
1127 if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
1128 pr_debug("%s: short packet %i\n", dev->name, len);
1129 dev->stats.rx_length_errors++;
1130 if (vi->mergeable_rx_bufs) {
1131 put_page(virt_to_head_page(buf));
1132 } else if (vi->big_packets) {
1133 give_pages(rq, buf);
1135 put_page(virt_to_head_page(buf));
1140 if (vi->mergeable_rx_bufs)
1141 skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit,
1143 else if (vi->big_packets)
1144 skb = receive_big(dev, vi, rq, buf, len, stats);
1146 skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit, stats);
1151 hdr = skb_vnet_hdr(skb);
1153 if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
1154 skb->ip_summed = CHECKSUM_UNNECESSARY;
1156 if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
1157 virtio_is_little_endian(vi->vdev))) {
1158 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
1159 dev->name, hdr->hdr.gso_type,
1164 skb_record_rx_queue(skb, vq2rxq(rq->vq));
1165 skb->protocol = eth_type_trans(skb, dev);
1166 pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
1167 ntohs(skb->protocol), skb->len, skb->pkt_type);
1169 napi_gro_receive(&rq->napi, skb);
1173 dev->stats.rx_frame_errors++;
1177 /* Unlike mergeable buffers, all buffers are allocated to the
1178 * same size, except for the headroom. For this reason we do
1179 * not need to use mergeable_len_to_ctx here - it is enough
1180 * to store the headroom as the context ignoring the truesize.
1182 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
1185 struct page_frag *alloc_frag = &rq->alloc_frag;
1187 unsigned int xdp_headroom = virtnet_get_headroom(vi);
1188 void *ctx = (void *)(unsigned long)xdp_headroom;
1189 int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
1192 len = SKB_DATA_ALIGN(len) +
1193 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1194 if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
1197 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1198 get_page(alloc_frag->page);
1199 alloc_frag->offset += len;
1200 sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom,
1201 vi->hdr_len + GOOD_PACKET_LEN);
1202 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1204 put_page(virt_to_head_page(buf));
1208 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
1211 struct page *first, *list = NULL;
1215 sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
1217 /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
1218 for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
1219 first = get_a_page(rq, gfp);
1222 give_pages(rq, list);
1225 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
1227 /* chain new page in list head to match sg */
1228 first->private = (unsigned long)list;
1232 first = get_a_page(rq, gfp);
1234 give_pages(rq, list);
1237 p = page_address(first);
1239 /* rq->sg[0], rq->sg[1] share the same page */
1240 /* a separated rq->sg[0] for header - required in case !any_header_sg */
1241 sg_set_buf(&rq->sg[0], p, vi->hdr_len);
1243 /* rq->sg[1] for data packet, from offset */
1244 offset = sizeof(struct padded_vnet_hdr);
1245 sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
1247 /* chain first in list head */
1248 first->private = (unsigned long)list;
1249 err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
1252 give_pages(rq, first);
1257 static unsigned int get_mergeable_buf_len(struct receive_queue *rq,
1258 struct ewma_pkt_len *avg_pkt_len,
1261 const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1265 return PAGE_SIZE - room;
1267 len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
1268 rq->min_buf_len, PAGE_SIZE - hdr_len);
1270 return ALIGN(len, L1_CACHE_BYTES);
1273 static int add_recvbuf_mergeable(struct virtnet_info *vi,
1274 struct receive_queue *rq, gfp_t gfp)
1276 struct page_frag *alloc_frag = &rq->alloc_frag;
1277 unsigned int headroom = virtnet_get_headroom(vi);
1278 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1279 unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
1283 unsigned int len, hole;
1285 /* Extra tailroom is needed to satisfy XDP's assumption. This
1286 * means rx frags coalescing won't work, but consider we've
1287 * disabled GSO for XDP, it won't be a big issue.
1289 len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len, room);
1290 if (unlikely(!skb_page_frag_refill(len + room, alloc_frag, gfp)))
1293 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1294 buf += headroom; /* advance address leaving hole at front of pkt */
1295 get_page(alloc_frag->page);
1296 alloc_frag->offset += len + room;
1297 hole = alloc_frag->size - alloc_frag->offset;
1298 if (hole < len + room) {
1299 /* To avoid internal fragmentation, if there is very likely not
1300 * enough space for another buffer, add the remaining space to
1301 * the current buffer.
1304 alloc_frag->offset += hole;
1307 sg_init_one(rq->sg, buf, len);
1308 ctx = mergeable_len_to_ctx(len, headroom);
1309 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1311 put_page(virt_to_head_page(buf));
1317 * Returns false if we couldn't fill entirely (OOM).
1319 * Normally run in the receive path, but can also be run from ndo_open
1320 * before we're receiving packets, or from refill_work which is
1321 * careful to disable receiving (using napi_disable).
1323 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
1330 if (vi->mergeable_rx_bufs)
1331 err = add_recvbuf_mergeable(vi, rq, gfp);
1332 else if (vi->big_packets)
1333 err = add_recvbuf_big(vi, rq, gfp);
1335 err = add_recvbuf_small(vi, rq, gfp);
1337 oom = err == -ENOMEM;
1340 } while (rq->vq->num_free);
1341 if (virtqueue_kick_prepare(rq->vq) && virtqueue_notify(rq->vq)) {
1342 unsigned long flags;
1344 flags = u64_stats_update_begin_irqsave(&rq->stats.syncp);
1346 u64_stats_update_end_irqrestore(&rq->stats.syncp, flags);
1352 static void skb_recv_done(struct virtqueue *rvq)
1354 struct virtnet_info *vi = rvq->vdev->priv;
1355 struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
1357 virtqueue_napi_schedule(&rq->napi, rvq);
1360 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
1364 /* If all buffers were filled by other side before we napi_enabled, we
1365 * won't get another interrupt, so process any outstanding packets now.
1366 * Call local_bh_enable after to trigger softIRQ processing.
1369 virtqueue_napi_schedule(napi, vq);
1373 static void virtnet_napi_tx_enable(struct virtnet_info *vi,
1374 struct virtqueue *vq,
1375 struct napi_struct *napi)
1380 /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
1381 * enable the feature if this is likely affine with the transmit path.
1383 if (!vi->affinity_hint_set) {
1388 return virtnet_napi_enable(vq, napi);
1391 static void virtnet_napi_tx_disable(struct napi_struct *napi)
1397 static void refill_work(struct work_struct *work)
1399 struct virtnet_info *vi =
1400 container_of(work, struct virtnet_info, refill.work);
1404 for (i = 0; i < vi->curr_queue_pairs; i++) {
1405 struct receive_queue *rq = &vi->rq[i];
1407 napi_disable(&rq->napi);
1408 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
1409 virtnet_napi_enable(rq->vq, &rq->napi);
1411 /* In theory, this can happen: if we don't get any buffers in
1412 * we will *never* try to fill again.
1415 schedule_delayed_work(&vi->refill, HZ/2);
1419 static int virtnet_receive(struct receive_queue *rq, int budget,
1420 unsigned int *xdp_xmit)
1422 struct virtnet_info *vi = rq->vq->vdev->priv;
1423 struct virtnet_rq_stats stats = {};
1428 if (!vi->big_packets || vi->mergeable_rx_bufs) {
1431 while (stats.packets < budget &&
1432 (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) {
1433 receive_buf(vi, rq, buf, len, ctx, xdp_xmit, &stats);
1437 while (stats.packets < budget &&
1438 (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
1439 receive_buf(vi, rq, buf, len, NULL, xdp_xmit, &stats);
1444 if (rq->vq->num_free > min((unsigned int)budget, virtqueue_get_vring_size(rq->vq)) / 2) {
1445 if (!try_fill_recv(vi, rq, GFP_ATOMIC))
1446 schedule_delayed_work(&vi->refill, 0);
1449 u64_stats_update_begin(&rq->stats.syncp);
1450 for (i = 0; i < VIRTNET_RQ_STATS_LEN; i++) {
1451 size_t offset = virtnet_rq_stats_desc[i].offset;
1454 item = (u64 *)((u8 *)&rq->stats + offset);
1455 *item += *(u64 *)((u8 *)&stats + offset);
1457 u64_stats_update_end(&rq->stats.syncp);
1459 return stats.packets;
1462 static void free_old_xmit_skbs(struct send_queue *sq, bool in_napi)
1465 unsigned int packets = 0;
1466 unsigned int bytes = 0;
1469 while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
1470 if (likely(!is_xdp_frame(ptr))) {
1471 struct sk_buff *skb = ptr;
1473 pr_debug("Sent skb %p\n", skb);
1476 napi_consume_skb(skb, in_napi);
1478 struct xdp_frame *frame = ptr_to_xdp(ptr);
1480 bytes += frame->len;
1481 xdp_return_frame(frame);
1486 /* Avoid overhead when no packets have been processed
1487 * happens when called speculatively from start_xmit.
1492 u64_stats_update_begin(&sq->stats.syncp);
1493 sq->stats.bytes += bytes;
1494 sq->stats.packets += packets;
1495 u64_stats_update_end(&sq->stats.syncp);
1498 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
1500 if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
1502 else if (q < vi->curr_queue_pairs)
1508 static void virtnet_poll_cleantx(struct receive_queue *rq)
1510 struct virtnet_info *vi = rq->vq->vdev->priv;
1511 unsigned int index = vq2rxq(rq->vq);
1512 struct send_queue *sq = &vi->sq[index];
1513 struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index);
1515 if (!sq->napi.weight || is_xdp_raw_buffer_queue(vi, index))
1518 if (__netif_tx_trylock(txq)) {
1520 virtqueue_disable_cb(sq->vq);
1521 free_old_xmit_skbs(sq, true);
1522 } while (unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
1524 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1525 netif_tx_wake_queue(txq);
1527 __netif_tx_unlock(txq);
1531 static int virtnet_poll(struct napi_struct *napi, int budget)
1533 struct receive_queue *rq =
1534 container_of(napi, struct receive_queue, napi);
1535 struct virtnet_info *vi = rq->vq->vdev->priv;
1536 struct send_queue *sq;
1537 unsigned int received;
1538 unsigned int xdp_xmit = 0;
1540 virtnet_poll_cleantx(rq);
1542 received = virtnet_receive(rq, budget, &xdp_xmit);
1544 /* Out of packets? */
1545 if (received < budget)
1546 virtqueue_napi_complete(napi, rq->vq, received);
1548 if (xdp_xmit & VIRTIO_XDP_REDIR)
1551 if (xdp_xmit & VIRTIO_XDP_TX) {
1552 sq = virtnet_xdp_get_sq(vi);
1553 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1554 u64_stats_update_begin(&sq->stats.syncp);
1556 u64_stats_update_end(&sq->stats.syncp);
1558 virtnet_xdp_put_sq(vi, sq);
1564 static int virtnet_open(struct net_device *dev)
1566 struct virtnet_info *vi = netdev_priv(dev);
1569 for (i = 0; i < vi->max_queue_pairs; i++) {
1570 if (i < vi->curr_queue_pairs)
1571 /* Make sure we have some buffers: if oom use wq. */
1572 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1573 schedule_delayed_work(&vi->refill, 0);
1575 err = xdp_rxq_info_reg(&vi->rq[i].xdp_rxq, dev, i, vi->rq[i].napi.napi_id);
1579 err = xdp_rxq_info_reg_mem_model(&vi->rq[i].xdp_rxq,
1580 MEM_TYPE_PAGE_SHARED, NULL);
1582 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1586 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1587 virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi);
1593 static int virtnet_poll_tx(struct napi_struct *napi, int budget)
1595 struct send_queue *sq = container_of(napi, struct send_queue, napi);
1596 struct virtnet_info *vi = sq->vq->vdev->priv;
1597 unsigned int index = vq2txq(sq->vq);
1598 struct netdev_queue *txq;
1602 if (unlikely(is_xdp_raw_buffer_queue(vi, index))) {
1603 /* We don't need to enable cb for XDP */
1604 napi_complete_done(napi, 0);
1608 txq = netdev_get_tx_queue(vi->dev, index);
1609 __netif_tx_lock(txq, raw_smp_processor_id());
1610 virtqueue_disable_cb(sq->vq);
1611 free_old_xmit_skbs(sq, true);
1613 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1614 netif_tx_wake_queue(txq);
1616 opaque = virtqueue_enable_cb_prepare(sq->vq);
1618 done = napi_complete_done(napi, 0);
1621 virtqueue_disable_cb(sq->vq);
1623 __netif_tx_unlock(txq);
1626 if (unlikely(virtqueue_poll(sq->vq, opaque))) {
1627 if (napi_schedule_prep(napi)) {
1628 __netif_tx_lock(txq, raw_smp_processor_id());
1629 virtqueue_disable_cb(sq->vq);
1630 __netif_tx_unlock(txq);
1631 __napi_schedule(napi);
1639 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
1641 struct virtio_net_hdr_mrg_rxbuf *hdr;
1642 const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
1643 struct virtnet_info *vi = sq->vq->vdev->priv;
1645 unsigned hdr_len = vi->hdr_len;
1648 pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
1650 can_push = vi->any_header_sg &&
1651 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
1652 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
1653 /* Even if we can, don't push here yet as this would skew
1654 * csum_start offset below. */
1656 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1658 hdr = skb_vnet_hdr(skb);
1660 if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1661 virtio_is_little_endian(vi->vdev), false,
1665 if (vi->mergeable_rx_bufs)
1666 hdr->num_buffers = 0;
1668 sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1670 __skb_push(skb, hdr_len);
1671 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1672 if (unlikely(num_sg < 0))
1674 /* Pull header back to avoid skew in tx bytes calculations. */
1675 __skb_pull(skb, hdr_len);
1677 sg_set_buf(sq->sg, hdr, hdr_len);
1678 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
1679 if (unlikely(num_sg < 0))
1683 return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
1686 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
1688 struct virtnet_info *vi = netdev_priv(dev);
1689 int qnum = skb_get_queue_mapping(skb);
1690 struct send_queue *sq = &vi->sq[qnum];
1692 struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1693 bool kick = !netdev_xmit_more();
1694 bool use_napi = sq->napi.weight;
1696 /* Free up any pending old buffers before queueing new ones. */
1699 virtqueue_disable_cb(sq->vq);
1701 free_old_xmit_skbs(sq, false);
1703 } while (use_napi && kick &&
1704 unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
1706 /* timestamp packet in software */
1707 skb_tx_timestamp(skb);
1709 /* Try to transmit */
1710 err = xmit_skb(sq, skb);
1712 /* This should not happen! */
1713 if (unlikely(err)) {
1714 dev->stats.tx_fifo_errors++;
1715 if (net_ratelimit())
1717 "Unexpected TXQ (%d) queue failure: %d\n",
1719 dev->stats.tx_dropped++;
1720 dev_kfree_skb_any(skb);
1721 return NETDEV_TX_OK;
1724 /* Don't wait up for transmitted skbs to be freed. */
1730 /* If running out of space, stop queue to avoid getting packets that we
1731 * are then unable to transmit.
1732 * An alternative would be to force queuing layer to requeue the skb by
1733 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1734 * returned in a normal path of operation: it means that driver is not
1735 * maintaining the TX queue stop/start state properly, and causes
1736 * the stack to do a non-trivial amount of useless work.
1737 * Since most packets only take 1 or 2 ring slots, stopping the queue
1738 * early means 16 slots are typically wasted.
1740 if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
1741 netif_stop_subqueue(dev, qnum);
1743 unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
1744 /* More just got used, free them then recheck. */
1745 free_old_xmit_skbs(sq, false);
1746 if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
1747 netif_start_subqueue(dev, qnum);
1748 virtqueue_disable_cb(sq->vq);
1753 if (kick || netif_xmit_stopped(txq)) {
1754 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1755 u64_stats_update_begin(&sq->stats.syncp);
1757 u64_stats_update_end(&sq->stats.syncp);
1761 return NETDEV_TX_OK;
1765 * Send command via the control virtqueue and check status. Commands
1766 * supported by the hypervisor, as indicated by feature bits, should
1767 * never fail unless improperly formatted.
1769 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
1770 struct scatterlist *out)
1772 struct scatterlist *sgs[4], hdr, stat;
1773 unsigned out_num = 0, tmp;
1776 /* Caller should know better */
1777 BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
1779 vi->ctrl->status = ~0;
1780 vi->ctrl->hdr.class = class;
1781 vi->ctrl->hdr.cmd = cmd;
1783 sg_init_one(&hdr, &vi->ctrl->hdr, sizeof(vi->ctrl->hdr));
1784 sgs[out_num++] = &hdr;
1787 sgs[out_num++] = out;
1789 /* Add return status. */
1790 sg_init_one(&stat, &vi->ctrl->status, sizeof(vi->ctrl->status));
1791 sgs[out_num] = &stat;
1793 BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1794 ret = virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1796 dev_warn(&vi->vdev->dev,
1797 "Failed to add sgs for command vq: %d\n.", ret);
1801 if (unlikely(!virtqueue_kick(vi->cvq)))
1802 return vi->ctrl->status == VIRTIO_NET_OK;
1804 /* Spin for a response, the kick causes an ioport write, trapping
1805 * into the hypervisor, so the request should be handled immediately.
1807 while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1808 !virtqueue_is_broken(vi->cvq))
1811 return vi->ctrl->status == VIRTIO_NET_OK;
1814 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1816 struct virtnet_info *vi = netdev_priv(dev);
1817 struct virtio_device *vdev = vi->vdev;
1819 struct sockaddr *addr;
1820 struct scatterlist sg;
1822 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
1825 addr = kmemdup(p, sizeof(*addr), GFP_KERNEL);
1829 ret = eth_prepare_mac_addr_change(dev, addr);
1833 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1834 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1835 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1836 VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1837 dev_warn(&vdev->dev,
1838 "Failed to set mac address by vq command.\n");
1842 } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1843 !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1846 /* Naturally, this has an atomicity problem. */
1847 for (i = 0; i < dev->addr_len; i++)
1848 virtio_cwrite8(vdev,
1849 offsetof(struct virtio_net_config, mac) +
1850 i, addr->sa_data[i]);
1853 eth_commit_mac_addr_change(dev, p);
1861 static void virtnet_stats(struct net_device *dev,
1862 struct rtnl_link_stats64 *tot)
1864 struct virtnet_info *vi = netdev_priv(dev);
1868 for (i = 0; i < vi->max_queue_pairs; i++) {
1869 u64 tpackets, tbytes, rpackets, rbytes, rdrops;
1870 struct receive_queue *rq = &vi->rq[i];
1871 struct send_queue *sq = &vi->sq[i];
1874 start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
1875 tpackets = sq->stats.packets;
1876 tbytes = sq->stats.bytes;
1877 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
1880 start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
1881 rpackets = rq->stats.packets;
1882 rbytes = rq->stats.bytes;
1883 rdrops = rq->stats.drops;
1884 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
1886 tot->rx_packets += rpackets;
1887 tot->tx_packets += tpackets;
1888 tot->rx_bytes += rbytes;
1889 tot->tx_bytes += tbytes;
1890 tot->rx_dropped += rdrops;
1893 tot->tx_dropped = dev->stats.tx_dropped;
1894 tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1895 tot->rx_length_errors = dev->stats.rx_length_errors;
1896 tot->rx_frame_errors = dev->stats.rx_frame_errors;
1899 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1902 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1903 VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1904 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1908 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1910 struct scatterlist sg;
1911 struct net_device *dev = vi->dev;
1913 if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1916 vi->ctrl->mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1917 sg_init_one(&sg, &vi->ctrl->mq, sizeof(vi->ctrl->mq));
1919 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1920 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1921 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1925 vi->curr_queue_pairs = queue_pairs;
1926 /* virtnet_open() will refill when device is going to up. */
1927 if (dev->flags & IFF_UP)
1928 schedule_delayed_work(&vi->refill, 0);
1934 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1939 err = _virtnet_set_queues(vi, queue_pairs);
1944 static int virtnet_close(struct net_device *dev)
1946 struct virtnet_info *vi = netdev_priv(dev);
1949 /* Make sure refill_work doesn't re-enable napi! */
1950 cancel_delayed_work_sync(&vi->refill);
1952 for (i = 0; i < vi->max_queue_pairs; i++) {
1953 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1954 napi_disable(&vi->rq[i].napi);
1955 virtnet_napi_tx_disable(&vi->sq[i].napi);
1961 static void virtnet_set_rx_mode(struct net_device *dev)
1963 struct virtnet_info *vi = netdev_priv(dev);
1964 struct scatterlist sg[2];
1965 struct virtio_net_ctrl_mac *mac_data;
1966 struct netdev_hw_addr *ha;
1972 /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1973 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1976 vi->ctrl->promisc = ((dev->flags & IFF_PROMISC) != 0);
1977 vi->ctrl->allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1979 sg_init_one(sg, &vi->ctrl->promisc, sizeof(vi->ctrl->promisc));
1981 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1982 VIRTIO_NET_CTRL_RX_PROMISC, sg))
1983 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1984 vi->ctrl->promisc ? "en" : "dis");
1986 sg_init_one(sg, &vi->ctrl->allmulti, sizeof(vi->ctrl->allmulti));
1988 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1989 VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1990 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1991 vi->ctrl->allmulti ? "en" : "dis");
1993 uc_count = netdev_uc_count(dev);
1994 mc_count = netdev_mc_count(dev);
1995 /* MAC filter - use one buffer for both lists */
1996 buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1997 (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
2002 sg_init_table(sg, 2);
2004 /* Store the unicast list and count in the front of the buffer */
2005 mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
2007 netdev_for_each_uc_addr(ha, dev)
2008 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
2010 sg_set_buf(&sg[0], mac_data,
2011 sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
2013 /* multicast list and count fill the end */
2014 mac_data = (void *)&mac_data->macs[uc_count][0];
2016 mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
2018 netdev_for_each_mc_addr(ha, dev)
2019 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
2021 sg_set_buf(&sg[1], mac_data,
2022 sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
2024 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
2025 VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
2026 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
2031 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
2032 __be16 proto, u16 vid)
2034 struct virtnet_info *vi = netdev_priv(dev);
2035 struct scatterlist sg;
2037 vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
2038 sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
2040 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
2041 VIRTIO_NET_CTRL_VLAN_ADD, &sg))
2042 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
2046 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
2047 __be16 proto, u16 vid)
2049 struct virtnet_info *vi = netdev_priv(dev);
2050 struct scatterlist sg;
2052 vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
2053 sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
2055 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
2056 VIRTIO_NET_CTRL_VLAN_DEL, &sg))
2057 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
2061 static void virtnet_clean_affinity(struct virtnet_info *vi)
2065 if (vi->affinity_hint_set) {
2066 for (i = 0; i < vi->max_queue_pairs; i++) {
2067 virtqueue_set_affinity(vi->rq[i].vq, NULL);
2068 virtqueue_set_affinity(vi->sq[i].vq, NULL);
2071 vi->affinity_hint_set = false;
2075 static void virtnet_set_affinity(struct virtnet_info *vi)
2084 if (!zalloc_cpumask_var(&mask, GFP_KERNEL)) {
2085 virtnet_clean_affinity(vi);
2089 num_cpu = num_online_cpus();
2090 stride = max_t(int, num_cpu / vi->curr_queue_pairs, 1);
2091 stragglers = num_cpu >= vi->curr_queue_pairs ?
2092 num_cpu % vi->curr_queue_pairs :
2094 cpu = cpumask_next(-1, cpu_online_mask);
2096 for (i = 0; i < vi->curr_queue_pairs; i++) {
2097 group_size = stride + (i < stragglers ? 1 : 0);
2099 for (j = 0; j < group_size; j++) {
2100 cpumask_set_cpu(cpu, mask);
2101 cpu = cpumask_next_wrap(cpu, cpu_online_mask,
2104 virtqueue_set_affinity(vi->rq[i].vq, mask);
2105 virtqueue_set_affinity(vi->sq[i].vq, mask);
2106 __netif_set_xps_queue(vi->dev, cpumask_bits(mask), i, XPS_CPUS);
2107 cpumask_clear(mask);
2110 vi->affinity_hint_set = true;
2111 free_cpumask_var(mask);
2114 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
2116 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2118 virtnet_set_affinity(vi);
2122 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
2124 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2126 virtnet_set_affinity(vi);
2130 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
2132 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2135 virtnet_clean_affinity(vi);
2139 static enum cpuhp_state virtionet_online;
2141 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
2145 ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
2148 ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2152 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2156 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
2158 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2159 cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2163 static void virtnet_get_ringparam(struct net_device *dev,
2164 struct ethtool_ringparam *ring)
2166 struct virtnet_info *vi = netdev_priv(dev);
2168 ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
2169 ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
2170 ring->rx_pending = ring->rx_max_pending;
2171 ring->tx_pending = ring->tx_max_pending;
2175 static void virtnet_get_drvinfo(struct net_device *dev,
2176 struct ethtool_drvinfo *info)
2178 struct virtnet_info *vi = netdev_priv(dev);
2179 struct virtio_device *vdev = vi->vdev;
2181 strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
2182 strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
2183 strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
2187 /* TODO: Eliminate OOO packets during switching */
2188 static int virtnet_set_channels(struct net_device *dev,
2189 struct ethtool_channels *channels)
2191 struct virtnet_info *vi = netdev_priv(dev);
2192 u16 queue_pairs = channels->combined_count;
2195 /* We don't support separate rx/tx channels.
2196 * We don't allow setting 'other' channels.
2198 if (channels->rx_count || channels->tx_count || channels->other_count)
2201 if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
2204 /* For now we don't support modifying channels while XDP is loaded
2205 * also when XDP is loaded all RX queues have XDP programs so we only
2206 * need to check a single RX queue.
2208 if (vi->rq[0].xdp_prog)
2212 err = _virtnet_set_queues(vi, queue_pairs);
2217 virtnet_set_affinity(vi);
2220 netif_set_real_num_tx_queues(dev, queue_pairs);
2221 netif_set_real_num_rx_queues(dev, queue_pairs);
2226 static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *data)
2228 struct virtnet_info *vi = netdev_priv(dev);
2232 switch (stringset) {
2234 for (i = 0; i < vi->curr_queue_pairs; i++) {
2235 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++)
2236 ethtool_sprintf(&p, "rx_queue_%u_%s", i,
2237 virtnet_rq_stats_desc[j].desc);
2240 for (i = 0; i < vi->curr_queue_pairs; i++) {
2241 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++)
2242 ethtool_sprintf(&p, "tx_queue_%u_%s", i,
2243 virtnet_sq_stats_desc[j].desc);
2249 static int virtnet_get_sset_count(struct net_device *dev, int sset)
2251 struct virtnet_info *vi = netdev_priv(dev);
2255 return vi->curr_queue_pairs * (VIRTNET_RQ_STATS_LEN +
2256 VIRTNET_SQ_STATS_LEN);
2262 static void virtnet_get_ethtool_stats(struct net_device *dev,
2263 struct ethtool_stats *stats, u64 *data)
2265 struct virtnet_info *vi = netdev_priv(dev);
2266 unsigned int idx = 0, start, i, j;
2267 const u8 *stats_base;
2270 for (i = 0; i < vi->curr_queue_pairs; i++) {
2271 struct receive_queue *rq = &vi->rq[i];
2273 stats_base = (u8 *)&rq->stats;
2275 start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
2276 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
2277 offset = virtnet_rq_stats_desc[j].offset;
2278 data[idx + j] = *(u64 *)(stats_base + offset);
2280 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
2281 idx += VIRTNET_RQ_STATS_LEN;
2284 for (i = 0; i < vi->curr_queue_pairs; i++) {
2285 struct send_queue *sq = &vi->sq[i];
2287 stats_base = (u8 *)&sq->stats;
2289 start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
2290 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
2291 offset = virtnet_sq_stats_desc[j].offset;
2292 data[idx + j] = *(u64 *)(stats_base + offset);
2294 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
2295 idx += VIRTNET_SQ_STATS_LEN;
2299 static void virtnet_get_channels(struct net_device *dev,
2300 struct ethtool_channels *channels)
2302 struct virtnet_info *vi = netdev_priv(dev);
2304 channels->combined_count = vi->curr_queue_pairs;
2305 channels->max_combined = vi->max_queue_pairs;
2306 channels->max_other = 0;
2307 channels->rx_count = 0;
2308 channels->tx_count = 0;
2309 channels->other_count = 0;
2312 static int virtnet_set_link_ksettings(struct net_device *dev,
2313 const struct ethtool_link_ksettings *cmd)
2315 struct virtnet_info *vi = netdev_priv(dev);
2317 return ethtool_virtdev_set_link_ksettings(dev, cmd,
2318 &vi->speed, &vi->duplex);
2321 static int virtnet_get_link_ksettings(struct net_device *dev,
2322 struct ethtool_link_ksettings *cmd)
2324 struct virtnet_info *vi = netdev_priv(dev);
2326 cmd->base.speed = vi->speed;
2327 cmd->base.duplex = vi->duplex;
2328 cmd->base.port = PORT_OTHER;
2333 static int virtnet_set_coalesce(struct net_device *dev,
2334 struct ethtool_coalesce *ec)
2336 struct virtnet_info *vi = netdev_priv(dev);
2339 if (ec->tx_max_coalesced_frames > 1 ||
2340 ec->rx_max_coalesced_frames != 1)
2343 napi_weight = ec->tx_max_coalesced_frames ? NAPI_POLL_WEIGHT : 0;
2344 if (napi_weight ^ vi->sq[0].napi.weight) {
2345 if (dev->flags & IFF_UP)
2347 for (i = 0; i < vi->max_queue_pairs; i++)
2348 vi->sq[i].napi.weight = napi_weight;
2354 static int virtnet_get_coalesce(struct net_device *dev,
2355 struct ethtool_coalesce *ec)
2357 struct ethtool_coalesce ec_default = {
2358 .cmd = ETHTOOL_GCOALESCE,
2359 .rx_max_coalesced_frames = 1,
2361 struct virtnet_info *vi = netdev_priv(dev);
2363 memcpy(ec, &ec_default, sizeof(ec_default));
2365 if (vi->sq[0].napi.weight)
2366 ec->tx_max_coalesced_frames = 1;
2371 static void virtnet_init_settings(struct net_device *dev)
2373 struct virtnet_info *vi = netdev_priv(dev);
2375 vi->speed = SPEED_UNKNOWN;
2376 vi->duplex = DUPLEX_UNKNOWN;
2379 static void virtnet_update_settings(struct virtnet_info *vi)
2384 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_SPEED_DUPLEX))
2387 virtio_cread_le(vi->vdev, struct virtio_net_config, speed, &speed);
2389 if (ethtool_validate_speed(speed))
2392 virtio_cread_le(vi->vdev, struct virtio_net_config, duplex, &duplex);
2394 if (ethtool_validate_duplex(duplex))
2395 vi->duplex = duplex;
2398 static const struct ethtool_ops virtnet_ethtool_ops = {
2399 .supported_coalesce_params = ETHTOOL_COALESCE_MAX_FRAMES,
2400 .get_drvinfo = virtnet_get_drvinfo,
2401 .get_link = ethtool_op_get_link,
2402 .get_ringparam = virtnet_get_ringparam,
2403 .get_strings = virtnet_get_strings,
2404 .get_sset_count = virtnet_get_sset_count,
2405 .get_ethtool_stats = virtnet_get_ethtool_stats,
2406 .set_channels = virtnet_set_channels,
2407 .get_channels = virtnet_get_channels,
2408 .get_ts_info = ethtool_op_get_ts_info,
2409 .get_link_ksettings = virtnet_get_link_ksettings,
2410 .set_link_ksettings = virtnet_set_link_ksettings,
2411 .set_coalesce = virtnet_set_coalesce,
2412 .get_coalesce = virtnet_get_coalesce,
2415 static void virtnet_freeze_down(struct virtio_device *vdev)
2417 struct virtnet_info *vi = vdev->priv;
2420 /* Make sure no work handler is accessing the device */
2421 flush_work(&vi->config_work);
2423 netif_tx_lock_bh(vi->dev);
2424 netif_device_detach(vi->dev);
2425 netif_tx_unlock_bh(vi->dev);
2426 cancel_delayed_work_sync(&vi->refill);
2428 if (netif_running(vi->dev)) {
2429 for (i = 0; i < vi->max_queue_pairs; i++) {
2430 napi_disable(&vi->rq[i].napi);
2431 virtnet_napi_tx_disable(&vi->sq[i].napi);
2436 static int init_vqs(struct virtnet_info *vi);
2438 static int virtnet_restore_up(struct virtio_device *vdev)
2440 struct virtnet_info *vi = vdev->priv;
2447 virtio_device_ready(vdev);
2449 if (netif_running(vi->dev)) {
2450 for (i = 0; i < vi->curr_queue_pairs; i++)
2451 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
2452 schedule_delayed_work(&vi->refill, 0);
2454 for (i = 0; i < vi->max_queue_pairs; i++) {
2455 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2456 virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2461 netif_tx_lock_bh(vi->dev);
2462 netif_device_attach(vi->dev);
2463 netif_tx_unlock_bh(vi->dev);
2467 static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads)
2469 struct scatterlist sg;
2470 vi->ctrl->offloads = cpu_to_virtio64(vi->vdev, offloads);
2472 sg_init_one(&sg, &vi->ctrl->offloads, sizeof(vi->ctrl->offloads));
2474 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
2475 VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) {
2476 dev_warn(&vi->dev->dev, "Fail to set guest offload.\n");
2483 static int virtnet_clear_guest_offloads(struct virtnet_info *vi)
2487 if (!vi->guest_offloads)
2490 return virtnet_set_guest_offloads(vi, offloads);
2493 static int virtnet_restore_guest_offloads(struct virtnet_info *vi)
2495 u64 offloads = vi->guest_offloads;
2497 if (!vi->guest_offloads)
2500 return virtnet_set_guest_offloads(vi, offloads);
2503 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
2504 struct netlink_ext_ack *extack)
2506 unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr);
2507 struct virtnet_info *vi = netdev_priv(dev);
2508 struct bpf_prog *old_prog;
2509 u16 xdp_qp = 0, curr_qp;
2512 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)
2513 && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2514 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2515 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
2516 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) ||
2517 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))) {
2518 NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing LRO/CSUM, disable LRO/CSUM first");
2522 if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
2523 NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required");
2527 if (dev->mtu > max_sz) {
2528 NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP");
2529 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz);
2533 curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
2535 xdp_qp = nr_cpu_ids;
2537 /* XDP requires extra queues for XDP_TX */
2538 if (curr_qp + xdp_qp > vi->max_queue_pairs) {
2539 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",
2540 curr_qp + xdp_qp, vi->max_queue_pairs);
2544 old_prog = rtnl_dereference(vi->rq[0].xdp_prog);
2545 if (!prog && !old_prog)
2549 bpf_prog_add(prog, vi->max_queue_pairs - 1);
2551 /* Make sure NAPI is not using any XDP TX queues for RX. */
2552 if (netif_running(dev)) {
2553 for (i = 0; i < vi->max_queue_pairs; i++) {
2554 napi_disable(&vi->rq[i].napi);
2555 virtnet_napi_tx_disable(&vi->sq[i].napi);
2560 for (i = 0; i < vi->max_queue_pairs; i++) {
2561 rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2563 virtnet_restore_guest_offloads(vi);
2568 err = _virtnet_set_queues(vi, curr_qp + xdp_qp);
2571 netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
2572 vi->xdp_queue_pairs = xdp_qp;
2575 vi->xdp_enabled = true;
2576 for (i = 0; i < vi->max_queue_pairs; i++) {
2577 rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2578 if (i == 0 && !old_prog)
2579 virtnet_clear_guest_offloads(vi);
2582 vi->xdp_enabled = false;
2585 for (i = 0; i < vi->max_queue_pairs; i++) {
2587 bpf_prog_put(old_prog);
2588 if (netif_running(dev)) {
2589 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2590 virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2599 virtnet_clear_guest_offloads(vi);
2600 for (i = 0; i < vi->max_queue_pairs; i++)
2601 rcu_assign_pointer(vi->rq[i].xdp_prog, old_prog);
2604 if (netif_running(dev)) {
2605 for (i = 0; i < vi->max_queue_pairs; i++) {
2606 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2607 virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2612 bpf_prog_sub(prog, vi->max_queue_pairs - 1);
2616 static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp)
2618 switch (xdp->command) {
2619 case XDP_SETUP_PROG:
2620 return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
2626 static int virtnet_get_phys_port_name(struct net_device *dev, char *buf,
2629 struct virtnet_info *vi = netdev_priv(dev);
2632 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
2635 ret = snprintf(buf, len, "sby");
2642 static int virtnet_set_features(struct net_device *dev,
2643 netdev_features_t features)
2645 struct virtnet_info *vi = netdev_priv(dev);
2649 if ((dev->features ^ features) & NETIF_F_LRO) {
2650 if (vi->xdp_enabled)
2653 if (features & NETIF_F_LRO)
2654 offloads = vi->guest_offloads_capable;
2656 offloads = vi->guest_offloads_capable &
2657 ~GUEST_OFFLOAD_LRO_MASK;
2659 err = virtnet_set_guest_offloads(vi, offloads);
2662 vi->guest_offloads = offloads;
2668 static const struct net_device_ops virtnet_netdev = {
2669 .ndo_open = virtnet_open,
2670 .ndo_stop = virtnet_close,
2671 .ndo_start_xmit = start_xmit,
2672 .ndo_validate_addr = eth_validate_addr,
2673 .ndo_set_mac_address = virtnet_set_mac_address,
2674 .ndo_set_rx_mode = virtnet_set_rx_mode,
2675 .ndo_get_stats64 = virtnet_stats,
2676 .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
2677 .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
2678 .ndo_bpf = virtnet_xdp,
2679 .ndo_xdp_xmit = virtnet_xdp_xmit,
2680 .ndo_features_check = passthru_features_check,
2681 .ndo_get_phys_port_name = virtnet_get_phys_port_name,
2682 .ndo_set_features = virtnet_set_features,
2685 static void virtnet_config_changed_work(struct work_struct *work)
2687 struct virtnet_info *vi =
2688 container_of(work, struct virtnet_info, config_work);
2691 if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
2692 struct virtio_net_config, status, &v) < 0)
2695 if (v & VIRTIO_NET_S_ANNOUNCE) {
2696 netdev_notify_peers(vi->dev);
2697 virtnet_ack_link_announce(vi);
2700 /* Ignore unknown (future) status bits */
2701 v &= VIRTIO_NET_S_LINK_UP;
2703 if (vi->status == v)
2708 if (vi->status & VIRTIO_NET_S_LINK_UP) {
2709 virtnet_update_settings(vi);
2710 netif_carrier_on(vi->dev);
2711 netif_tx_wake_all_queues(vi->dev);
2713 netif_carrier_off(vi->dev);
2714 netif_tx_stop_all_queues(vi->dev);
2718 static void virtnet_config_changed(struct virtio_device *vdev)
2720 struct virtnet_info *vi = vdev->priv;
2722 schedule_work(&vi->config_work);
2725 static void virtnet_free_queues(struct virtnet_info *vi)
2729 for (i = 0; i < vi->max_queue_pairs; i++) {
2730 __netif_napi_del(&vi->rq[i].napi);
2731 __netif_napi_del(&vi->sq[i].napi);
2734 /* We called __netif_napi_del(),
2735 * we need to respect an RCU grace period before freeing vi->rq
2744 static void _free_receive_bufs(struct virtnet_info *vi)
2746 struct bpf_prog *old_prog;
2749 for (i = 0; i < vi->max_queue_pairs; i++) {
2750 while (vi->rq[i].pages)
2751 __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
2753 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2754 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
2756 bpf_prog_put(old_prog);
2760 static void free_receive_bufs(struct virtnet_info *vi)
2763 _free_receive_bufs(vi);
2767 static void free_receive_page_frags(struct virtnet_info *vi)
2770 for (i = 0; i < vi->max_queue_pairs; i++)
2771 if (vi->rq[i].alloc_frag.page)
2772 put_page(vi->rq[i].alloc_frag.page);
2775 static void free_unused_bufs(struct virtnet_info *vi)
2780 for (i = 0; i < vi->max_queue_pairs; i++) {
2781 struct virtqueue *vq = vi->sq[i].vq;
2782 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2783 if (!is_xdp_frame(buf))
2786 xdp_return_frame(ptr_to_xdp(buf));
2790 for (i = 0; i < vi->max_queue_pairs; i++) {
2791 struct virtqueue *vq = vi->rq[i].vq;
2793 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2794 if (vi->mergeable_rx_bufs) {
2795 put_page(virt_to_head_page(buf));
2796 } else if (vi->big_packets) {
2797 give_pages(&vi->rq[i], buf);
2799 put_page(virt_to_head_page(buf));
2805 static void virtnet_del_vqs(struct virtnet_info *vi)
2807 struct virtio_device *vdev = vi->vdev;
2809 virtnet_clean_affinity(vi);
2811 vdev->config->del_vqs(vdev);
2813 virtnet_free_queues(vi);
2816 /* How large should a single buffer be so a queue full of these can fit at
2817 * least one full packet?
2818 * Logic below assumes the mergeable buffer header is used.
2820 static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq)
2822 const unsigned int hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2823 unsigned int rq_size = virtqueue_get_vring_size(vq);
2824 unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu;
2825 unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len;
2826 unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size);
2828 return max(max(min_buf_len, hdr_len) - hdr_len,
2829 (unsigned int)GOOD_PACKET_LEN);
2832 static int virtnet_find_vqs(struct virtnet_info *vi)
2834 vq_callback_t **callbacks;
2835 struct virtqueue **vqs;
2841 /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
2842 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
2843 * possible control vq.
2845 total_vqs = vi->max_queue_pairs * 2 +
2846 virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
2848 /* Allocate space for find_vqs parameters */
2849 vqs = kcalloc(total_vqs, sizeof(*vqs), GFP_KERNEL);
2852 callbacks = kmalloc_array(total_vqs, sizeof(*callbacks), GFP_KERNEL);
2855 names = kmalloc_array(total_vqs, sizeof(*names), GFP_KERNEL);
2858 if (!vi->big_packets || vi->mergeable_rx_bufs) {
2859 ctx = kcalloc(total_vqs, sizeof(*ctx), GFP_KERNEL);
2866 /* Parameters for control virtqueue, if any */
2868 callbacks[total_vqs - 1] = NULL;
2869 names[total_vqs - 1] = "control";
2872 /* Allocate/initialize parameters for send/receive virtqueues */
2873 for (i = 0; i < vi->max_queue_pairs; i++) {
2874 callbacks[rxq2vq(i)] = skb_recv_done;
2875 callbacks[txq2vq(i)] = skb_xmit_done;
2876 sprintf(vi->rq[i].name, "input.%d", i);
2877 sprintf(vi->sq[i].name, "output.%d", i);
2878 names[rxq2vq(i)] = vi->rq[i].name;
2879 names[txq2vq(i)] = vi->sq[i].name;
2881 ctx[rxq2vq(i)] = true;
2884 ret = virtio_find_vqs_ctx(vi->vdev, total_vqs, vqs, callbacks,
2890 vi->cvq = vqs[total_vqs - 1];
2891 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
2892 vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
2895 for (i = 0; i < vi->max_queue_pairs; i++) {
2896 vi->rq[i].vq = vqs[rxq2vq(i)];
2897 vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq);
2898 vi->sq[i].vq = vqs[txq2vq(i)];
2901 /* run here: ret == 0. */
2916 static int virtnet_alloc_queues(struct virtnet_info *vi)
2921 vi->ctrl = kzalloc(sizeof(*vi->ctrl), GFP_KERNEL);
2927 vi->sq = kcalloc(vi->max_queue_pairs, sizeof(*vi->sq), GFP_KERNEL);
2930 vi->rq = kcalloc(vi->max_queue_pairs, sizeof(*vi->rq), GFP_KERNEL);
2934 INIT_DELAYED_WORK(&vi->refill, refill_work);
2935 for (i = 0; i < vi->max_queue_pairs; i++) {
2936 vi->rq[i].pages = NULL;
2937 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
2939 netif_tx_napi_add(vi->dev, &vi->sq[i].napi, virtnet_poll_tx,
2940 napi_tx ? napi_weight : 0);
2942 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
2943 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
2944 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
2946 u64_stats_init(&vi->rq[i].stats.syncp);
2947 u64_stats_init(&vi->sq[i].stats.syncp);
2960 static int init_vqs(struct virtnet_info *vi)
2964 /* Allocate send & receive queues */
2965 ret = virtnet_alloc_queues(vi);
2969 ret = virtnet_find_vqs(vi);
2974 virtnet_set_affinity(vi);
2980 virtnet_free_queues(vi);
2986 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
2989 struct virtnet_info *vi = netdev_priv(queue->dev);
2990 unsigned int queue_index = get_netdev_rx_queue_index(queue);
2991 unsigned int headroom = virtnet_get_headroom(vi);
2992 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
2993 struct ewma_pkt_len *avg;
2995 BUG_ON(queue_index >= vi->max_queue_pairs);
2996 avg = &vi->rq[queue_index].mrg_avg_pkt_len;
2997 return sprintf(buf, "%u\n",
2998 get_mergeable_buf_len(&vi->rq[queue_index], avg,
2999 SKB_DATA_ALIGN(headroom + tailroom)));
3002 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
3003 __ATTR_RO(mergeable_rx_buffer_size);
3005 static struct attribute *virtio_net_mrg_rx_attrs[] = {
3006 &mergeable_rx_buffer_size_attribute.attr,
3010 static const struct attribute_group virtio_net_mrg_rx_group = {
3011 .name = "virtio_net",
3012 .attrs = virtio_net_mrg_rx_attrs
3016 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
3018 const char *fname, const char *dname)
3020 if (!virtio_has_feature(vdev, fbit))
3023 dev_err(&vdev->dev, "device advertises feature %s but not %s",
3029 #define VIRTNET_FAIL_ON(vdev, fbit, dbit) \
3030 virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
3032 static bool virtnet_validate_features(struct virtio_device *vdev)
3034 if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
3035 (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
3036 "VIRTIO_NET_F_CTRL_VQ") ||
3037 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
3038 "VIRTIO_NET_F_CTRL_VQ") ||
3039 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
3040 "VIRTIO_NET_F_CTRL_VQ") ||
3041 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
3042 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
3043 "VIRTIO_NET_F_CTRL_VQ"))) {
3050 #define MIN_MTU ETH_MIN_MTU
3051 #define MAX_MTU ETH_MAX_MTU
3053 static int virtnet_validate(struct virtio_device *vdev)
3055 if (!vdev->config->get) {
3056 dev_err(&vdev->dev, "%s failure: config access disabled\n",
3061 if (!virtnet_validate_features(vdev))
3064 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
3065 int mtu = virtio_cread16(vdev,
3066 offsetof(struct virtio_net_config,
3069 __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
3075 static int virtnet_probe(struct virtio_device *vdev)
3077 int i, err = -ENOMEM;
3078 struct net_device *dev;
3079 struct virtnet_info *vi;
3080 u16 max_queue_pairs;
3083 /* Find if host supports multiqueue virtio_net device */
3084 err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
3085 struct virtio_net_config,
3086 max_virtqueue_pairs, &max_queue_pairs);
3088 /* We need at least 2 queue's */
3089 if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
3090 max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
3091 !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
3092 max_queue_pairs = 1;
3094 /* Allocate ourselves a network device with room for our info */
3095 dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
3099 /* Set up network device as normal. */
3100 dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE |
3101 IFF_TX_SKB_NO_LINEAR;
3102 dev->netdev_ops = &virtnet_netdev;
3103 dev->features = NETIF_F_HIGHDMA;
3105 dev->ethtool_ops = &virtnet_ethtool_ops;
3106 SET_NETDEV_DEV(dev, &vdev->dev);
3108 /* Do we support "hardware" checksums? */
3109 if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
3110 /* This opens up the world of extra features. */
3111 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
3113 dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
3115 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
3116 dev->hw_features |= NETIF_F_TSO
3117 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
3119 /* Individual feature bits: what can host handle? */
3120 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
3121 dev->hw_features |= NETIF_F_TSO;
3122 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
3123 dev->hw_features |= NETIF_F_TSO6;
3124 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
3125 dev->hw_features |= NETIF_F_TSO_ECN;
3127 dev->features |= NETIF_F_GSO_ROBUST;
3130 dev->features |= dev->hw_features & NETIF_F_ALL_TSO;
3131 /* (!csum && gso) case will be fixed by register_netdev() */
3133 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
3134 dev->features |= NETIF_F_RXCSUM;
3135 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3136 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6))
3137 dev->features |= NETIF_F_LRO;
3138 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS))
3139 dev->hw_features |= NETIF_F_LRO;
3141 dev->vlan_features = dev->features;
3143 /* MTU range: 68 - 65535 */
3144 dev->min_mtu = MIN_MTU;
3145 dev->max_mtu = MAX_MTU;
3147 /* Configuration may specify what MAC to use. Otherwise random. */
3148 if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
3149 virtio_cread_bytes(vdev,
3150 offsetof(struct virtio_net_config, mac),
3151 dev->dev_addr, dev->addr_len);
3153 eth_hw_addr_random(dev);
3155 /* Set up our device-specific information */
3156 vi = netdev_priv(dev);
3161 INIT_WORK(&vi->config_work, virtnet_config_changed_work);
3163 /* If we can receive ANY GSO packets, we must allocate large ones. */
3164 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3165 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
3166 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
3167 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
3168 vi->big_packets = true;
3170 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
3171 vi->mergeable_rx_bufs = true;
3173 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
3174 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
3175 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
3177 vi->hdr_len = sizeof(struct virtio_net_hdr);
3179 if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
3180 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
3181 vi->any_header_sg = true;
3183 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
3186 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
3187 mtu = virtio_cread16(vdev,
3188 offsetof(struct virtio_net_config,
3190 if (mtu < dev->min_mtu) {
3191 /* Should never trigger: MTU was previously validated
3192 * in virtnet_validate.
3195 "device MTU appears to have changed it is now %d < %d",
3204 /* TODO: size buffers correctly in this case. */
3205 if (dev->mtu > ETH_DATA_LEN)
3206 vi->big_packets = true;
3209 if (vi->any_header_sg)
3210 dev->needed_headroom = vi->hdr_len;
3212 /* Enable multiqueue by default */
3213 if (num_online_cpus() >= max_queue_pairs)
3214 vi->curr_queue_pairs = max_queue_pairs;
3216 vi->curr_queue_pairs = num_online_cpus();
3217 vi->max_queue_pairs = max_queue_pairs;
3219 /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
3225 if (vi->mergeable_rx_bufs)
3226 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
3228 netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
3229 netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
3231 virtnet_init_settings(dev);
3233 if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY)) {
3234 vi->failover = net_failover_create(vi->dev);
3235 if (IS_ERR(vi->failover)) {
3236 err = PTR_ERR(vi->failover);
3241 err = register_netdev(dev);
3243 pr_debug("virtio_net: registering device failed\n");
3247 virtio_device_ready(vdev);
3249 err = virtnet_cpu_notif_add(vi);
3251 pr_debug("virtio_net: registering cpu notifier failed\n");
3252 goto free_unregister_netdev;
3255 virtnet_set_queues(vi, vi->curr_queue_pairs);
3257 /* Assume link up if device can't report link status,
3258 otherwise get link status from config. */
3259 netif_carrier_off(dev);
3260 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
3261 schedule_work(&vi->config_work);
3263 vi->status = VIRTIO_NET_S_LINK_UP;
3264 virtnet_update_settings(vi);
3265 netif_carrier_on(dev);
3268 for (i = 0; i < ARRAY_SIZE(guest_offloads); i++)
3269 if (virtio_has_feature(vi->vdev, guest_offloads[i]))
3270 set_bit(guest_offloads[i], &vi->guest_offloads);
3271 vi->guest_offloads_capable = vi->guest_offloads;
3273 pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
3274 dev->name, max_queue_pairs);
3278 free_unregister_netdev:
3279 vi->vdev->config->reset(vdev);
3281 unregister_netdev(dev);
3283 net_failover_destroy(vi->failover);
3285 cancel_delayed_work_sync(&vi->refill);
3286 free_receive_page_frags(vi);
3287 virtnet_del_vqs(vi);
3293 static void remove_vq_common(struct virtnet_info *vi)
3295 vi->vdev->config->reset(vi->vdev);
3297 /* Free unused buffers in both send and recv, if any. */
3298 free_unused_bufs(vi);
3300 free_receive_bufs(vi);
3302 free_receive_page_frags(vi);
3304 virtnet_del_vqs(vi);
3307 static void virtnet_remove(struct virtio_device *vdev)
3309 struct virtnet_info *vi = vdev->priv;
3311 virtnet_cpu_notif_remove(vi);
3313 /* Make sure no work handler is accessing the device. */
3314 flush_work(&vi->config_work);
3316 unregister_netdev(vi->dev);
3318 net_failover_destroy(vi->failover);
3320 remove_vq_common(vi);
3322 free_netdev(vi->dev);
3325 static __maybe_unused int virtnet_freeze(struct virtio_device *vdev)
3327 struct virtnet_info *vi = vdev->priv;
3329 virtnet_cpu_notif_remove(vi);
3330 virtnet_freeze_down(vdev);
3331 remove_vq_common(vi);
3336 static __maybe_unused int virtnet_restore(struct virtio_device *vdev)
3338 struct virtnet_info *vi = vdev->priv;
3341 err = virtnet_restore_up(vdev);
3344 virtnet_set_queues(vi, vi->curr_queue_pairs);
3346 err = virtnet_cpu_notif_add(vi);
3348 virtnet_freeze_down(vdev);
3349 remove_vq_common(vi);
3356 static struct virtio_device_id id_table[] = {
3357 { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
3361 #define VIRTNET_FEATURES \
3362 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
3364 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
3365 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
3366 VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
3367 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
3368 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
3369 VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
3370 VIRTIO_NET_F_CTRL_MAC_ADDR, \
3371 VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \
3372 VIRTIO_NET_F_SPEED_DUPLEX, VIRTIO_NET_F_STANDBY
3374 static unsigned int features[] = {
3378 static unsigned int features_legacy[] = {
3381 VIRTIO_F_ANY_LAYOUT,
3384 static struct virtio_driver virtio_net_driver = {
3385 .feature_table = features,
3386 .feature_table_size = ARRAY_SIZE(features),
3387 .feature_table_legacy = features_legacy,
3388 .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
3389 .driver.name = KBUILD_MODNAME,
3390 .driver.owner = THIS_MODULE,
3391 .id_table = id_table,
3392 .validate = virtnet_validate,
3393 .probe = virtnet_probe,
3394 .remove = virtnet_remove,
3395 .config_changed = virtnet_config_changed,
3396 #ifdef CONFIG_PM_SLEEP
3397 .freeze = virtnet_freeze,
3398 .restore = virtnet_restore,
3402 static __init int virtio_net_driver_init(void)
3406 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
3408 virtnet_cpu_down_prep);
3411 virtionet_online = ret;
3412 ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
3413 NULL, virtnet_cpu_dead);
3417 ret = register_virtio_driver(&virtio_net_driver);
3422 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3424 cpuhp_remove_multi_state(virtionet_online);
3428 module_init(virtio_net_driver_init);
3430 static __exit void virtio_net_driver_exit(void)
3432 unregister_virtio_driver(&virtio_net_driver);
3433 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3434 cpuhp_remove_multi_state(virtionet_online);
3436 module_exit(virtio_net_driver_exit);
3438 MODULE_DEVICE_TABLE(virtio, id_table);
3439 MODULE_DESCRIPTION("Virtio network driver");
3440 MODULE_LICENSE("GPL");