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,
383 unsigned int headroom)
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;
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 headroom is not 0, there is an offset between the beginning of the
402 * data and the allocated space, otherwise the data and the allocated
406 /* The actual allocated space size is PAGE_SIZE. */
407 truesize = PAGE_SIZE;
408 tailroom = truesize - len - offset;
410 tailroom = truesize - len;
414 offset += hdr_padded_len;
417 shinfo_size = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
419 if (!NET_IP_ALIGN && len > GOOD_COPY_LEN && tailroom >= shinfo_size) {
420 skb = build_skb(p, truesize);
428 /* copy small packet so we can reuse these pages for small data */
429 skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
433 /* Copy all frame if it fits skb->head, otherwise
434 * we let virtio_net_hdr_to_skb() and GRO pull headers as needed.
436 if (len <= skb_tailroom(skb))
439 copy = ETH_HLEN + metasize;
440 skb_put_data(skb, p, copy);
445 if (vi->mergeable_rx_bufs) {
447 skb_add_rx_frag(skb, 0, page, offset, len, truesize);
454 * Verify that we can indeed put this data into a skb.
455 * This is here to handle cases when the device erroneously
456 * tries to receive more than is possible. This is usually
457 * the case of a broken device.
459 if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
460 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
464 BUG_ON(offset >= PAGE_SIZE);
466 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
467 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
468 frag_size, truesize);
470 page = (struct page *)page->private;
475 give_pages(rq, page);
478 /* hdr_valid means no XDP, so we can copy the vnet header */
480 hdr = skb_vnet_hdr(skb);
481 memcpy(hdr, hdr_p, hdr_len);
484 put_page(page_to_free);
487 __skb_pull(skb, metasize);
488 skb_metadata_set(skb, metasize);
494 static int __virtnet_xdp_xmit_one(struct virtnet_info *vi,
495 struct send_queue *sq,
496 struct xdp_frame *xdpf)
498 struct virtio_net_hdr_mrg_rxbuf *hdr;
501 if (unlikely(xdpf->headroom < vi->hdr_len))
504 /* Make room for virtqueue hdr (also change xdpf->headroom?) */
505 xdpf->data -= vi->hdr_len;
506 /* Zero header and leave csum up to XDP layers */
508 memset(hdr, 0, vi->hdr_len);
509 xdpf->len += vi->hdr_len;
511 sg_init_one(sq->sg, xdpf->data, xdpf->len);
513 err = virtqueue_add_outbuf(sq->vq, sq->sg, 1, xdp_to_ptr(xdpf),
516 return -ENOSPC; /* Caller handle free/refcnt */
521 /* when vi->curr_queue_pairs > nr_cpu_ids, the txq/sq is only used for xdp tx on
522 * the current cpu, so it does not need to be locked.
524 * Here we use marco instead of inline functions because we have to deal with
525 * three issues at the same time: 1. the choice of sq. 2. judge and execute the
526 * lock/unlock of txq 3. make sparse happy. It is difficult for two inline
527 * functions to perfectly solve these three problems at the same time.
529 #define virtnet_xdp_get_sq(vi) ({ \
530 struct netdev_queue *txq; \
531 typeof(vi) v = (vi); \
534 if (v->curr_queue_pairs > nr_cpu_ids) { \
535 qp = v->curr_queue_pairs - v->xdp_queue_pairs; \
536 qp += smp_processor_id(); \
537 txq = netdev_get_tx_queue(v->dev, qp); \
538 __netif_tx_acquire(txq); \
540 qp = smp_processor_id() % v->curr_queue_pairs; \
541 txq = netdev_get_tx_queue(v->dev, qp); \
542 __netif_tx_lock(txq, raw_smp_processor_id()); \
547 #define virtnet_xdp_put_sq(vi, q) { \
548 struct netdev_queue *txq; \
549 typeof(vi) v = (vi); \
551 txq = netdev_get_tx_queue(v->dev, (q) - v->sq); \
552 if (v->curr_queue_pairs > nr_cpu_ids) \
553 __netif_tx_release(txq); \
555 __netif_tx_unlock(txq); \
558 static int virtnet_xdp_xmit(struct net_device *dev,
559 int n, struct xdp_frame **frames, u32 flags)
561 struct virtnet_info *vi = netdev_priv(dev);
562 struct receive_queue *rq = vi->rq;
563 struct bpf_prog *xdp_prog;
564 struct send_queue *sq;
574 /* Only allow ndo_xdp_xmit if XDP is loaded on dev, as this
575 * indicate XDP resources have been successfully allocated.
577 xdp_prog = rcu_access_pointer(rq->xdp_prog);
581 sq = virtnet_xdp_get_sq(vi);
583 if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK)) {
588 /* Free up any pending old buffers before queueing new ones. */
589 while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
590 if (likely(is_xdp_frame(ptr))) {
591 struct xdp_frame *frame = ptr_to_xdp(ptr);
594 xdp_return_frame(frame);
596 struct sk_buff *skb = ptr;
599 napi_consume_skb(skb, false);
604 for (i = 0; i < n; i++) {
605 struct xdp_frame *xdpf = frames[i];
607 if (__virtnet_xdp_xmit_one(vi, sq, xdpf))
613 if (flags & XDP_XMIT_FLUSH) {
614 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq))
618 u64_stats_update_begin(&sq->stats.syncp);
619 sq->stats.bytes += bytes;
620 sq->stats.packets += packets;
621 sq->stats.xdp_tx += n;
622 sq->stats.xdp_tx_drops += n - nxmit;
623 sq->stats.kicks += kicks;
624 u64_stats_update_end(&sq->stats.syncp);
626 virtnet_xdp_put_sq(vi, sq);
630 static unsigned int virtnet_get_headroom(struct virtnet_info *vi)
632 return vi->xdp_enabled ? VIRTIO_XDP_HEADROOM : 0;
635 /* We copy the packet for XDP in the following cases:
637 * 1) Packet is scattered across multiple rx buffers.
638 * 2) Headroom space is insufficient.
640 * This is inefficient but it's a temporary condition that
641 * we hit right after XDP is enabled and until queue is refilled
642 * with large buffers with sufficient headroom - so it should affect
643 * at most queue size packets.
644 * Afterwards, the conditions to enable
645 * XDP should preclude the underlying device from sending packets
646 * across multiple buffers (num_buf > 1), and we make sure buffers
647 * have enough headroom.
649 static struct page *xdp_linearize_page(struct receive_queue *rq,
656 struct page *page = alloc_page(GFP_ATOMIC);
661 memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
665 int tailroom = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
670 buf = virtqueue_get_buf(rq->vq, &buflen);
674 p = virt_to_head_page(buf);
675 off = buf - page_address(p);
677 /* guard against a misconfigured or uncooperative backend that
678 * is sending packet larger than the MTU.
680 if ((page_off + buflen + tailroom) > PAGE_SIZE) {
685 memcpy(page_address(page) + page_off,
686 page_address(p) + off, buflen);
691 /* Headroom does not contribute to packet length */
692 *len = page_off - VIRTIO_XDP_HEADROOM;
695 __free_pages(page, 0);
699 static struct sk_buff *receive_small(struct net_device *dev,
700 struct virtnet_info *vi,
701 struct receive_queue *rq,
702 void *buf, void *ctx,
704 unsigned int *xdp_xmit,
705 struct virtnet_rq_stats *stats)
708 struct bpf_prog *xdp_prog;
709 unsigned int xdp_headroom = (unsigned long)ctx;
710 unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom;
711 unsigned int headroom = vi->hdr_len + header_offset;
712 unsigned int buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
713 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
714 struct page *page = virt_to_head_page(buf);
715 unsigned int delta = 0;
716 struct page *xdp_page;
718 unsigned int metasize = 0;
724 xdp_prog = rcu_dereference(rq->xdp_prog);
726 struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset;
727 struct xdp_frame *xdpf;
732 if (unlikely(hdr->hdr.gso_type))
735 if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) {
736 int offset = buf - page_address(page) + header_offset;
737 unsigned int tlen = len + vi->hdr_len;
740 xdp_headroom = virtnet_get_headroom(vi);
741 header_offset = VIRTNET_RX_PAD + xdp_headroom;
742 headroom = vi->hdr_len + header_offset;
743 buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
744 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
745 xdp_page = xdp_linearize_page(rq, &num_buf, page,
746 offset, header_offset,
751 buf = page_address(xdp_page);
756 xdp_init_buff(&xdp, buflen, &rq->xdp_rxq);
757 xdp_prepare_buff(&xdp, buf + VIRTNET_RX_PAD + vi->hdr_len,
758 xdp_headroom, len, true);
759 orig_data = xdp.data;
760 act = bpf_prog_run_xdp(xdp_prog, &xdp);
761 stats->xdp_packets++;
765 /* Recalculate length in case bpf program changed it */
766 delta = orig_data - xdp.data;
767 len = xdp.data_end - xdp.data;
768 metasize = xdp.data - xdp.data_meta;
772 xdpf = xdp_convert_buff_to_frame(&xdp);
775 err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
776 if (unlikely(!err)) {
777 xdp_return_frame_rx_napi(xdpf);
778 } else if (unlikely(err < 0)) {
779 trace_xdp_exception(vi->dev, xdp_prog, act);
782 *xdp_xmit |= VIRTIO_XDP_TX;
786 stats->xdp_redirects++;
787 err = xdp_do_redirect(dev, &xdp, xdp_prog);
790 *xdp_xmit |= VIRTIO_XDP_REDIR;
794 bpf_warn_invalid_xdp_action(act);
797 trace_xdp_exception(vi->dev, xdp_prog, act);
805 skb = build_skb(buf, buflen);
810 skb_reserve(skb, headroom - delta);
813 buf += header_offset;
814 memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
815 } /* keep zeroed vnet hdr since XDP is loaded */
818 skb_metadata_set(skb, metasize);
832 static struct sk_buff *receive_big(struct net_device *dev,
833 struct virtnet_info *vi,
834 struct receive_queue *rq,
837 struct virtnet_rq_stats *stats)
839 struct page *page = buf;
840 struct sk_buff *skb =
841 page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, true, 0, 0);
843 stats->bytes += len - vi->hdr_len;
851 give_pages(rq, page);
855 static struct sk_buff *receive_mergeable(struct net_device *dev,
856 struct virtnet_info *vi,
857 struct receive_queue *rq,
861 unsigned int *xdp_xmit,
862 struct virtnet_rq_stats *stats)
864 struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
865 u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
866 struct page *page = virt_to_head_page(buf);
867 int offset = buf - page_address(page);
868 struct sk_buff *head_skb, *curr_skb;
869 struct bpf_prog *xdp_prog;
870 unsigned int truesize = mergeable_ctx_to_truesize(ctx);
871 unsigned int headroom = mergeable_ctx_to_headroom(ctx);
872 unsigned int metasize = 0;
873 unsigned int frame_sz;
877 stats->bytes += len - vi->hdr_len;
880 xdp_prog = rcu_dereference(rq->xdp_prog);
882 struct xdp_frame *xdpf;
883 struct page *xdp_page;
888 /* Transient failure which in theory could occur if
889 * in-flight packets from before XDP was enabled reach
890 * the receive path after XDP is loaded.
892 if (unlikely(hdr->hdr.gso_type))
895 /* Buffers with headroom use PAGE_SIZE as alloc size,
896 * see add_recvbuf_mergeable() + get_mergeable_buf_len()
898 frame_sz = headroom ? PAGE_SIZE : truesize;
900 /* This happens when rx buffer size is underestimated
901 * or headroom is not enough because of the buffer
902 * was refilled before XDP is set. This should only
903 * happen for the first several packets, so we don't
904 * care much about its performance.
906 if (unlikely(num_buf > 1 ||
907 headroom < virtnet_get_headroom(vi))) {
908 /* linearize data for XDP */
909 xdp_page = xdp_linearize_page(rq, &num_buf,
913 frame_sz = PAGE_SIZE;
917 offset = VIRTIO_XDP_HEADROOM;
922 /* Allow consuming headroom but reserve enough space to push
923 * the descriptor on if we get an XDP_TX return code.
925 data = page_address(xdp_page) + offset;
926 xdp_init_buff(&xdp, frame_sz - vi->hdr_len, &rq->xdp_rxq);
927 xdp_prepare_buff(&xdp, data - VIRTIO_XDP_HEADROOM + vi->hdr_len,
928 VIRTIO_XDP_HEADROOM, len - vi->hdr_len, true);
930 act = bpf_prog_run_xdp(xdp_prog, &xdp);
931 stats->xdp_packets++;
935 metasize = xdp.data - xdp.data_meta;
937 /* recalculate offset to account for any header
938 * adjustments and minus the metasize to copy the
939 * metadata in page_to_skb(). Note other cases do not
940 * build an skb and avoid using offset
942 offset = xdp.data - page_address(xdp_page) -
943 vi->hdr_len - metasize;
945 /* recalculate len if xdp.data, xdp.data_end or
946 * xdp.data_meta were adjusted
948 len = xdp.data_end - xdp.data + vi->hdr_len + metasize;
949 /* We can only create skb based on xdp_page. */
950 if (unlikely(xdp_page != page)) {
953 head_skb = page_to_skb(vi, rq, xdp_page, offset,
954 len, PAGE_SIZE, false,
961 xdpf = xdp_convert_buff_to_frame(&xdp);
964 err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
965 if (unlikely(!err)) {
966 xdp_return_frame_rx_napi(xdpf);
967 } else if (unlikely(err < 0)) {
968 trace_xdp_exception(vi->dev, xdp_prog, act);
969 if (unlikely(xdp_page != page))
973 *xdp_xmit |= VIRTIO_XDP_TX;
974 if (unlikely(xdp_page != page))
979 stats->xdp_redirects++;
980 err = xdp_do_redirect(dev, &xdp, xdp_prog);
982 if (unlikely(xdp_page != page))
986 *xdp_xmit |= VIRTIO_XDP_REDIR;
987 if (unlikely(xdp_page != page))
992 bpf_warn_invalid_xdp_action(act);
995 trace_xdp_exception(vi->dev, xdp_prog, act);
998 if (unlikely(xdp_page != page))
999 __free_pages(xdp_page, 0);
1005 if (unlikely(len > truesize)) {
1006 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
1007 dev->name, len, (unsigned long)ctx);
1008 dev->stats.rx_length_errors++;
1012 head_skb = page_to_skb(vi, rq, page, offset, len, truesize, !xdp_prog,
1013 metasize, headroom);
1014 curr_skb = head_skb;
1016 if (unlikely(!curr_skb))
1021 buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx);
1022 if (unlikely(!buf)) {
1023 pr_debug("%s: rx error: %d buffers out of %d missing\n",
1025 virtio16_to_cpu(vi->vdev,
1027 dev->stats.rx_length_errors++;
1031 stats->bytes += len;
1032 page = virt_to_head_page(buf);
1034 truesize = mergeable_ctx_to_truesize(ctx);
1035 if (unlikely(len > truesize)) {
1036 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
1037 dev->name, len, (unsigned long)ctx);
1038 dev->stats.rx_length_errors++;
1042 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
1043 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
1044 struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
1046 if (unlikely(!nskb))
1048 if (curr_skb == head_skb)
1049 skb_shinfo(curr_skb)->frag_list = nskb;
1051 curr_skb->next = nskb;
1053 head_skb->truesize += nskb->truesize;
1056 if (curr_skb != head_skb) {
1057 head_skb->data_len += len;
1058 head_skb->len += len;
1059 head_skb->truesize += truesize;
1061 offset = buf - page_address(page);
1062 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
1064 skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
1067 skb_add_rx_frag(curr_skb, num_skb_frags, page,
1068 offset, len, truesize);
1072 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
1080 while (num_buf-- > 1) {
1081 buf = virtqueue_get_buf(rq->vq, &len);
1082 if (unlikely(!buf)) {
1083 pr_debug("%s: rx error: %d buffers missing\n",
1084 dev->name, num_buf);
1085 dev->stats.rx_length_errors++;
1088 stats->bytes += len;
1089 page = virt_to_head_page(buf);
1094 dev_kfree_skb(head_skb);
1099 static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
1100 void *buf, unsigned int len, void **ctx,
1101 unsigned int *xdp_xmit,
1102 struct virtnet_rq_stats *stats)
1104 struct net_device *dev = vi->dev;
1105 struct sk_buff *skb;
1106 struct virtio_net_hdr_mrg_rxbuf *hdr;
1108 if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
1109 pr_debug("%s: short packet %i\n", dev->name, len);
1110 dev->stats.rx_length_errors++;
1111 if (vi->mergeable_rx_bufs) {
1112 put_page(virt_to_head_page(buf));
1113 } else if (vi->big_packets) {
1114 give_pages(rq, buf);
1116 put_page(virt_to_head_page(buf));
1121 if (vi->mergeable_rx_bufs)
1122 skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit,
1124 else if (vi->big_packets)
1125 skb = receive_big(dev, vi, rq, buf, len, stats);
1127 skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit, stats);
1132 hdr = skb_vnet_hdr(skb);
1134 if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
1135 skb->ip_summed = CHECKSUM_UNNECESSARY;
1137 if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
1138 virtio_is_little_endian(vi->vdev))) {
1139 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
1140 dev->name, hdr->hdr.gso_type,
1145 skb_record_rx_queue(skb, vq2rxq(rq->vq));
1146 skb->protocol = eth_type_trans(skb, dev);
1147 pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
1148 ntohs(skb->protocol), skb->len, skb->pkt_type);
1150 napi_gro_receive(&rq->napi, skb);
1154 dev->stats.rx_frame_errors++;
1158 /* Unlike mergeable buffers, all buffers are allocated to the
1159 * same size, except for the headroom. For this reason we do
1160 * not need to use mergeable_len_to_ctx here - it is enough
1161 * to store the headroom as the context ignoring the truesize.
1163 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
1166 struct page_frag *alloc_frag = &rq->alloc_frag;
1168 unsigned int xdp_headroom = virtnet_get_headroom(vi);
1169 void *ctx = (void *)(unsigned long)xdp_headroom;
1170 int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
1173 len = SKB_DATA_ALIGN(len) +
1174 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1175 if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
1178 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1179 get_page(alloc_frag->page);
1180 alloc_frag->offset += len;
1181 sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom,
1182 vi->hdr_len + GOOD_PACKET_LEN);
1183 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1185 put_page(virt_to_head_page(buf));
1189 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
1192 struct page *first, *list = NULL;
1196 sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
1198 /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
1199 for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
1200 first = get_a_page(rq, gfp);
1203 give_pages(rq, list);
1206 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
1208 /* chain new page in list head to match sg */
1209 first->private = (unsigned long)list;
1213 first = get_a_page(rq, gfp);
1215 give_pages(rq, list);
1218 p = page_address(first);
1220 /* rq->sg[0], rq->sg[1] share the same page */
1221 /* a separated rq->sg[0] for header - required in case !any_header_sg */
1222 sg_set_buf(&rq->sg[0], p, vi->hdr_len);
1224 /* rq->sg[1] for data packet, from offset */
1225 offset = sizeof(struct padded_vnet_hdr);
1226 sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
1228 /* chain first in list head */
1229 first->private = (unsigned long)list;
1230 err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
1233 give_pages(rq, first);
1238 static unsigned int get_mergeable_buf_len(struct receive_queue *rq,
1239 struct ewma_pkt_len *avg_pkt_len,
1242 const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1246 return PAGE_SIZE - room;
1248 len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
1249 rq->min_buf_len, PAGE_SIZE - hdr_len);
1251 return ALIGN(len, L1_CACHE_BYTES);
1254 static int add_recvbuf_mergeable(struct virtnet_info *vi,
1255 struct receive_queue *rq, gfp_t gfp)
1257 struct page_frag *alloc_frag = &rq->alloc_frag;
1258 unsigned int headroom = virtnet_get_headroom(vi);
1259 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1260 unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
1264 unsigned int len, hole;
1266 /* Extra tailroom is needed to satisfy XDP's assumption. This
1267 * means rx frags coalescing won't work, but consider we've
1268 * disabled GSO for XDP, it won't be a big issue.
1270 len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len, room);
1271 if (unlikely(!skb_page_frag_refill(len + room, alloc_frag, gfp)))
1274 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1275 buf += headroom; /* advance address leaving hole at front of pkt */
1276 get_page(alloc_frag->page);
1277 alloc_frag->offset += len + room;
1278 hole = alloc_frag->size - alloc_frag->offset;
1279 if (hole < len + room) {
1280 /* To avoid internal fragmentation, if there is very likely not
1281 * enough space for another buffer, add the remaining space to
1282 * the current buffer.
1285 alloc_frag->offset += hole;
1288 sg_init_one(rq->sg, buf, len);
1289 ctx = mergeable_len_to_ctx(len, headroom);
1290 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1292 put_page(virt_to_head_page(buf));
1298 * Returns false if we couldn't fill entirely (OOM).
1300 * Normally run in the receive path, but can also be run from ndo_open
1301 * before we're receiving packets, or from refill_work which is
1302 * careful to disable receiving (using napi_disable).
1304 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
1311 if (vi->mergeable_rx_bufs)
1312 err = add_recvbuf_mergeable(vi, rq, gfp);
1313 else if (vi->big_packets)
1314 err = add_recvbuf_big(vi, rq, gfp);
1316 err = add_recvbuf_small(vi, rq, gfp);
1318 oom = err == -ENOMEM;
1321 } while (rq->vq->num_free);
1322 if (virtqueue_kick_prepare(rq->vq) && virtqueue_notify(rq->vq)) {
1323 unsigned long flags;
1325 flags = u64_stats_update_begin_irqsave(&rq->stats.syncp);
1327 u64_stats_update_end_irqrestore(&rq->stats.syncp, flags);
1333 static void skb_recv_done(struct virtqueue *rvq)
1335 struct virtnet_info *vi = rvq->vdev->priv;
1336 struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
1338 virtqueue_napi_schedule(&rq->napi, rvq);
1341 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
1345 /* If all buffers were filled by other side before we napi_enabled, we
1346 * won't get another interrupt, so process any outstanding packets now.
1347 * Call local_bh_enable after to trigger softIRQ processing.
1350 virtqueue_napi_schedule(napi, vq);
1354 static void virtnet_napi_tx_enable(struct virtnet_info *vi,
1355 struct virtqueue *vq,
1356 struct napi_struct *napi)
1361 /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
1362 * enable the feature if this is likely affine with the transmit path.
1364 if (!vi->affinity_hint_set) {
1369 return virtnet_napi_enable(vq, napi);
1372 static void virtnet_napi_tx_disable(struct napi_struct *napi)
1378 static void refill_work(struct work_struct *work)
1380 struct virtnet_info *vi =
1381 container_of(work, struct virtnet_info, refill.work);
1385 for (i = 0; i < vi->curr_queue_pairs; i++) {
1386 struct receive_queue *rq = &vi->rq[i];
1388 napi_disable(&rq->napi);
1389 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
1390 virtnet_napi_enable(rq->vq, &rq->napi);
1392 /* In theory, this can happen: if we don't get any buffers in
1393 * we will *never* try to fill again.
1396 schedule_delayed_work(&vi->refill, HZ/2);
1400 static int virtnet_receive(struct receive_queue *rq, int budget,
1401 unsigned int *xdp_xmit)
1403 struct virtnet_info *vi = rq->vq->vdev->priv;
1404 struct virtnet_rq_stats stats = {};
1409 if (!vi->big_packets || vi->mergeable_rx_bufs) {
1412 while (stats.packets < budget &&
1413 (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) {
1414 receive_buf(vi, rq, buf, len, ctx, xdp_xmit, &stats);
1418 while (stats.packets < budget &&
1419 (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
1420 receive_buf(vi, rq, buf, len, NULL, xdp_xmit, &stats);
1425 if (rq->vq->num_free > min((unsigned int)budget, virtqueue_get_vring_size(rq->vq)) / 2) {
1426 if (!try_fill_recv(vi, rq, GFP_ATOMIC))
1427 schedule_delayed_work(&vi->refill, 0);
1430 u64_stats_update_begin(&rq->stats.syncp);
1431 for (i = 0; i < VIRTNET_RQ_STATS_LEN; i++) {
1432 size_t offset = virtnet_rq_stats_desc[i].offset;
1435 item = (u64 *)((u8 *)&rq->stats + offset);
1436 *item += *(u64 *)((u8 *)&stats + offset);
1438 u64_stats_update_end(&rq->stats.syncp);
1440 return stats.packets;
1443 static void free_old_xmit_skbs(struct send_queue *sq, bool in_napi)
1446 unsigned int packets = 0;
1447 unsigned int bytes = 0;
1450 while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
1451 if (likely(!is_xdp_frame(ptr))) {
1452 struct sk_buff *skb = ptr;
1454 pr_debug("Sent skb %p\n", skb);
1457 napi_consume_skb(skb, in_napi);
1459 struct xdp_frame *frame = ptr_to_xdp(ptr);
1461 bytes += frame->len;
1462 xdp_return_frame(frame);
1467 /* Avoid overhead when no packets have been processed
1468 * happens when called speculatively from start_xmit.
1473 u64_stats_update_begin(&sq->stats.syncp);
1474 sq->stats.bytes += bytes;
1475 sq->stats.packets += packets;
1476 u64_stats_update_end(&sq->stats.syncp);
1479 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
1481 if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
1483 else if (q < vi->curr_queue_pairs)
1489 static void virtnet_poll_cleantx(struct receive_queue *rq)
1491 struct virtnet_info *vi = rq->vq->vdev->priv;
1492 unsigned int index = vq2rxq(rq->vq);
1493 struct send_queue *sq = &vi->sq[index];
1494 struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index);
1496 if (!sq->napi.weight || is_xdp_raw_buffer_queue(vi, index))
1499 if (__netif_tx_trylock(txq)) {
1500 free_old_xmit_skbs(sq, true);
1501 __netif_tx_unlock(txq);
1504 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1505 netif_tx_wake_queue(txq);
1508 static int virtnet_poll(struct napi_struct *napi, int budget)
1510 struct receive_queue *rq =
1511 container_of(napi, struct receive_queue, napi);
1512 struct virtnet_info *vi = rq->vq->vdev->priv;
1513 struct send_queue *sq;
1514 unsigned int received;
1515 unsigned int xdp_xmit = 0;
1517 virtnet_poll_cleantx(rq);
1519 received = virtnet_receive(rq, budget, &xdp_xmit);
1521 /* Out of packets? */
1522 if (received < budget)
1523 virtqueue_napi_complete(napi, rq->vq, received);
1525 if (xdp_xmit & VIRTIO_XDP_REDIR)
1528 if (xdp_xmit & VIRTIO_XDP_TX) {
1529 sq = virtnet_xdp_get_sq(vi);
1530 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1531 u64_stats_update_begin(&sq->stats.syncp);
1533 u64_stats_update_end(&sq->stats.syncp);
1535 virtnet_xdp_put_sq(vi, sq);
1541 static int virtnet_open(struct net_device *dev)
1543 struct virtnet_info *vi = netdev_priv(dev);
1546 for (i = 0; i < vi->max_queue_pairs; i++) {
1547 if (i < vi->curr_queue_pairs)
1548 /* Make sure we have some buffers: if oom use wq. */
1549 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1550 schedule_delayed_work(&vi->refill, 0);
1552 err = xdp_rxq_info_reg(&vi->rq[i].xdp_rxq, dev, i, vi->rq[i].napi.napi_id);
1556 err = xdp_rxq_info_reg_mem_model(&vi->rq[i].xdp_rxq,
1557 MEM_TYPE_PAGE_SHARED, NULL);
1559 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1563 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1564 virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi);
1570 static int virtnet_poll_tx(struct napi_struct *napi, int budget)
1572 struct send_queue *sq = container_of(napi, struct send_queue, napi);
1573 struct virtnet_info *vi = sq->vq->vdev->priv;
1574 unsigned int index = vq2txq(sq->vq);
1575 struct netdev_queue *txq;
1577 if (unlikely(is_xdp_raw_buffer_queue(vi, index))) {
1578 /* We don't need to enable cb for XDP */
1579 napi_complete_done(napi, 0);
1583 txq = netdev_get_tx_queue(vi->dev, index);
1584 __netif_tx_lock(txq, raw_smp_processor_id());
1585 free_old_xmit_skbs(sq, true);
1586 __netif_tx_unlock(txq);
1588 virtqueue_napi_complete(napi, sq->vq, 0);
1590 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1591 netif_tx_wake_queue(txq);
1596 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
1598 struct virtio_net_hdr_mrg_rxbuf *hdr;
1599 const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
1600 struct virtnet_info *vi = sq->vq->vdev->priv;
1602 unsigned hdr_len = vi->hdr_len;
1605 pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
1607 can_push = vi->any_header_sg &&
1608 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
1609 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
1610 /* Even if we can, don't push here yet as this would skew
1611 * csum_start offset below. */
1613 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1615 hdr = skb_vnet_hdr(skb);
1617 if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1618 virtio_is_little_endian(vi->vdev), false,
1622 if (vi->mergeable_rx_bufs)
1623 hdr->num_buffers = 0;
1625 sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1627 __skb_push(skb, hdr_len);
1628 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1629 if (unlikely(num_sg < 0))
1631 /* Pull header back to avoid skew in tx bytes calculations. */
1632 __skb_pull(skb, hdr_len);
1634 sg_set_buf(sq->sg, hdr, hdr_len);
1635 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
1636 if (unlikely(num_sg < 0))
1640 return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
1643 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
1645 struct virtnet_info *vi = netdev_priv(dev);
1646 int qnum = skb_get_queue_mapping(skb);
1647 struct send_queue *sq = &vi->sq[qnum];
1649 struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1650 bool kick = !netdev_xmit_more();
1651 bool use_napi = sq->napi.weight;
1653 /* Free up any pending old buffers before queueing new ones. */
1654 free_old_xmit_skbs(sq, false);
1656 if (use_napi && kick)
1657 virtqueue_enable_cb_delayed(sq->vq);
1659 /* timestamp packet in software */
1660 skb_tx_timestamp(skb);
1662 /* Try to transmit */
1663 err = xmit_skb(sq, skb);
1665 /* This should not happen! */
1666 if (unlikely(err)) {
1667 dev->stats.tx_fifo_errors++;
1668 if (net_ratelimit())
1670 "Unexpected TXQ (%d) queue failure: %d\n",
1672 dev->stats.tx_dropped++;
1673 dev_kfree_skb_any(skb);
1674 return NETDEV_TX_OK;
1677 /* Don't wait up for transmitted skbs to be freed. */
1683 /* If running out of space, stop queue to avoid getting packets that we
1684 * are then unable to transmit.
1685 * An alternative would be to force queuing layer to requeue the skb by
1686 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1687 * returned in a normal path of operation: it means that driver is not
1688 * maintaining the TX queue stop/start state properly, and causes
1689 * the stack to do a non-trivial amount of useless work.
1690 * Since most packets only take 1 or 2 ring slots, stopping the queue
1691 * early means 16 slots are typically wasted.
1693 if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
1694 netif_stop_subqueue(dev, qnum);
1696 unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
1697 /* More just got used, free them then recheck. */
1698 free_old_xmit_skbs(sq, false);
1699 if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
1700 netif_start_subqueue(dev, qnum);
1701 virtqueue_disable_cb(sq->vq);
1706 if (kick || netif_xmit_stopped(txq)) {
1707 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1708 u64_stats_update_begin(&sq->stats.syncp);
1710 u64_stats_update_end(&sq->stats.syncp);
1714 return NETDEV_TX_OK;
1718 * Send command via the control virtqueue and check status. Commands
1719 * supported by the hypervisor, as indicated by feature bits, should
1720 * never fail unless improperly formatted.
1722 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
1723 struct scatterlist *out)
1725 struct scatterlist *sgs[4], hdr, stat;
1726 unsigned out_num = 0, tmp;
1728 /* Caller should know better */
1729 BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
1731 vi->ctrl->status = ~0;
1732 vi->ctrl->hdr.class = class;
1733 vi->ctrl->hdr.cmd = cmd;
1735 sg_init_one(&hdr, &vi->ctrl->hdr, sizeof(vi->ctrl->hdr));
1736 sgs[out_num++] = &hdr;
1739 sgs[out_num++] = out;
1741 /* Add return status. */
1742 sg_init_one(&stat, &vi->ctrl->status, sizeof(vi->ctrl->status));
1743 sgs[out_num] = &stat;
1745 BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1746 virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1748 if (unlikely(!virtqueue_kick(vi->cvq)))
1749 return vi->ctrl->status == VIRTIO_NET_OK;
1751 /* Spin for a response, the kick causes an ioport write, trapping
1752 * into the hypervisor, so the request should be handled immediately.
1754 while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1755 !virtqueue_is_broken(vi->cvq))
1758 return vi->ctrl->status == VIRTIO_NET_OK;
1761 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1763 struct virtnet_info *vi = netdev_priv(dev);
1764 struct virtio_device *vdev = vi->vdev;
1766 struct sockaddr *addr;
1767 struct scatterlist sg;
1769 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
1772 addr = kmemdup(p, sizeof(*addr), GFP_KERNEL);
1776 ret = eth_prepare_mac_addr_change(dev, addr);
1780 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1781 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1782 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1783 VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1784 dev_warn(&vdev->dev,
1785 "Failed to set mac address by vq command.\n");
1789 } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1790 !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1793 /* Naturally, this has an atomicity problem. */
1794 for (i = 0; i < dev->addr_len; i++)
1795 virtio_cwrite8(vdev,
1796 offsetof(struct virtio_net_config, mac) +
1797 i, addr->sa_data[i]);
1800 eth_commit_mac_addr_change(dev, p);
1808 static void virtnet_stats(struct net_device *dev,
1809 struct rtnl_link_stats64 *tot)
1811 struct virtnet_info *vi = netdev_priv(dev);
1815 for (i = 0; i < vi->max_queue_pairs; i++) {
1816 u64 tpackets, tbytes, rpackets, rbytes, rdrops;
1817 struct receive_queue *rq = &vi->rq[i];
1818 struct send_queue *sq = &vi->sq[i];
1821 start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
1822 tpackets = sq->stats.packets;
1823 tbytes = sq->stats.bytes;
1824 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
1827 start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
1828 rpackets = rq->stats.packets;
1829 rbytes = rq->stats.bytes;
1830 rdrops = rq->stats.drops;
1831 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
1833 tot->rx_packets += rpackets;
1834 tot->tx_packets += tpackets;
1835 tot->rx_bytes += rbytes;
1836 tot->tx_bytes += tbytes;
1837 tot->rx_dropped += rdrops;
1840 tot->tx_dropped = dev->stats.tx_dropped;
1841 tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1842 tot->rx_length_errors = dev->stats.rx_length_errors;
1843 tot->rx_frame_errors = dev->stats.rx_frame_errors;
1846 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1849 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1850 VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1851 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1855 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1857 struct scatterlist sg;
1858 struct net_device *dev = vi->dev;
1860 if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1863 vi->ctrl->mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1864 sg_init_one(&sg, &vi->ctrl->mq, sizeof(vi->ctrl->mq));
1866 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1867 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1868 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1872 vi->curr_queue_pairs = queue_pairs;
1873 /* virtnet_open() will refill when device is going to up. */
1874 if (dev->flags & IFF_UP)
1875 schedule_delayed_work(&vi->refill, 0);
1881 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1886 err = _virtnet_set_queues(vi, queue_pairs);
1891 static int virtnet_close(struct net_device *dev)
1893 struct virtnet_info *vi = netdev_priv(dev);
1896 /* Make sure refill_work doesn't re-enable napi! */
1897 cancel_delayed_work_sync(&vi->refill);
1899 for (i = 0; i < vi->max_queue_pairs; i++) {
1900 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1901 napi_disable(&vi->rq[i].napi);
1902 virtnet_napi_tx_disable(&vi->sq[i].napi);
1908 static void virtnet_set_rx_mode(struct net_device *dev)
1910 struct virtnet_info *vi = netdev_priv(dev);
1911 struct scatterlist sg[2];
1912 struct virtio_net_ctrl_mac *mac_data;
1913 struct netdev_hw_addr *ha;
1919 /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1920 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1923 vi->ctrl->promisc = ((dev->flags & IFF_PROMISC) != 0);
1924 vi->ctrl->allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1926 sg_init_one(sg, &vi->ctrl->promisc, sizeof(vi->ctrl->promisc));
1928 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1929 VIRTIO_NET_CTRL_RX_PROMISC, sg))
1930 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1931 vi->ctrl->promisc ? "en" : "dis");
1933 sg_init_one(sg, &vi->ctrl->allmulti, sizeof(vi->ctrl->allmulti));
1935 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1936 VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1937 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1938 vi->ctrl->allmulti ? "en" : "dis");
1940 uc_count = netdev_uc_count(dev);
1941 mc_count = netdev_mc_count(dev);
1942 /* MAC filter - use one buffer for both lists */
1943 buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1944 (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1949 sg_init_table(sg, 2);
1951 /* Store the unicast list and count in the front of the buffer */
1952 mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1954 netdev_for_each_uc_addr(ha, dev)
1955 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1957 sg_set_buf(&sg[0], mac_data,
1958 sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1960 /* multicast list and count fill the end */
1961 mac_data = (void *)&mac_data->macs[uc_count][0];
1963 mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1965 netdev_for_each_mc_addr(ha, dev)
1966 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1968 sg_set_buf(&sg[1], mac_data,
1969 sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1971 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1972 VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1973 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1978 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1979 __be16 proto, u16 vid)
1981 struct virtnet_info *vi = netdev_priv(dev);
1982 struct scatterlist sg;
1984 vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
1985 sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
1987 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1988 VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1989 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1993 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1994 __be16 proto, u16 vid)
1996 struct virtnet_info *vi = netdev_priv(dev);
1997 struct scatterlist sg;
1999 vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
2000 sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
2002 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
2003 VIRTIO_NET_CTRL_VLAN_DEL, &sg))
2004 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
2008 static void virtnet_clean_affinity(struct virtnet_info *vi)
2012 if (vi->affinity_hint_set) {
2013 for (i = 0; i < vi->max_queue_pairs; i++) {
2014 virtqueue_set_affinity(vi->rq[i].vq, NULL);
2015 virtqueue_set_affinity(vi->sq[i].vq, NULL);
2018 vi->affinity_hint_set = false;
2022 static void virtnet_set_affinity(struct virtnet_info *vi)
2031 if (!zalloc_cpumask_var(&mask, GFP_KERNEL)) {
2032 virtnet_clean_affinity(vi);
2036 num_cpu = num_online_cpus();
2037 stride = max_t(int, num_cpu / vi->curr_queue_pairs, 1);
2038 stragglers = num_cpu >= vi->curr_queue_pairs ?
2039 num_cpu % vi->curr_queue_pairs :
2041 cpu = cpumask_next(-1, cpu_online_mask);
2043 for (i = 0; i < vi->curr_queue_pairs; i++) {
2044 group_size = stride + (i < stragglers ? 1 : 0);
2046 for (j = 0; j < group_size; j++) {
2047 cpumask_set_cpu(cpu, mask);
2048 cpu = cpumask_next_wrap(cpu, cpu_online_mask,
2051 virtqueue_set_affinity(vi->rq[i].vq, mask);
2052 virtqueue_set_affinity(vi->sq[i].vq, mask);
2053 __netif_set_xps_queue(vi->dev, cpumask_bits(mask), i, XPS_CPUS);
2054 cpumask_clear(mask);
2057 vi->affinity_hint_set = true;
2058 free_cpumask_var(mask);
2061 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
2063 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2065 virtnet_set_affinity(vi);
2069 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
2071 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2073 virtnet_set_affinity(vi);
2077 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
2079 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2082 virtnet_clean_affinity(vi);
2086 static enum cpuhp_state virtionet_online;
2088 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
2092 ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
2095 ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2099 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2103 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
2105 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2106 cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2110 static void virtnet_get_ringparam(struct net_device *dev,
2111 struct ethtool_ringparam *ring)
2113 struct virtnet_info *vi = netdev_priv(dev);
2115 ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
2116 ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
2117 ring->rx_pending = ring->rx_max_pending;
2118 ring->tx_pending = ring->tx_max_pending;
2122 static void virtnet_get_drvinfo(struct net_device *dev,
2123 struct ethtool_drvinfo *info)
2125 struct virtnet_info *vi = netdev_priv(dev);
2126 struct virtio_device *vdev = vi->vdev;
2128 strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
2129 strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
2130 strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
2134 /* TODO: Eliminate OOO packets during switching */
2135 static int virtnet_set_channels(struct net_device *dev,
2136 struct ethtool_channels *channels)
2138 struct virtnet_info *vi = netdev_priv(dev);
2139 u16 queue_pairs = channels->combined_count;
2142 /* We don't support separate rx/tx channels.
2143 * We don't allow setting 'other' channels.
2145 if (channels->rx_count || channels->tx_count || channels->other_count)
2148 if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
2151 /* For now we don't support modifying channels while XDP is loaded
2152 * also when XDP is loaded all RX queues have XDP programs so we only
2153 * need to check a single RX queue.
2155 if (vi->rq[0].xdp_prog)
2159 err = _virtnet_set_queues(vi, queue_pairs);
2164 virtnet_set_affinity(vi);
2167 netif_set_real_num_tx_queues(dev, queue_pairs);
2168 netif_set_real_num_rx_queues(dev, queue_pairs);
2173 static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *data)
2175 struct virtnet_info *vi = netdev_priv(dev);
2179 switch (stringset) {
2181 for (i = 0; i < vi->curr_queue_pairs; i++) {
2182 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++)
2183 ethtool_sprintf(&p, "rx_queue_%u_%s", i,
2184 virtnet_rq_stats_desc[j].desc);
2187 for (i = 0; i < vi->curr_queue_pairs; i++) {
2188 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++)
2189 ethtool_sprintf(&p, "tx_queue_%u_%s", i,
2190 virtnet_sq_stats_desc[j].desc);
2196 static int virtnet_get_sset_count(struct net_device *dev, int sset)
2198 struct virtnet_info *vi = netdev_priv(dev);
2202 return vi->curr_queue_pairs * (VIRTNET_RQ_STATS_LEN +
2203 VIRTNET_SQ_STATS_LEN);
2209 static void virtnet_get_ethtool_stats(struct net_device *dev,
2210 struct ethtool_stats *stats, u64 *data)
2212 struct virtnet_info *vi = netdev_priv(dev);
2213 unsigned int idx = 0, start, i, j;
2214 const u8 *stats_base;
2217 for (i = 0; i < vi->curr_queue_pairs; i++) {
2218 struct receive_queue *rq = &vi->rq[i];
2220 stats_base = (u8 *)&rq->stats;
2222 start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
2223 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
2224 offset = virtnet_rq_stats_desc[j].offset;
2225 data[idx + j] = *(u64 *)(stats_base + offset);
2227 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
2228 idx += VIRTNET_RQ_STATS_LEN;
2231 for (i = 0; i < vi->curr_queue_pairs; i++) {
2232 struct send_queue *sq = &vi->sq[i];
2234 stats_base = (u8 *)&sq->stats;
2236 start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
2237 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
2238 offset = virtnet_sq_stats_desc[j].offset;
2239 data[idx + j] = *(u64 *)(stats_base + offset);
2241 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
2242 idx += VIRTNET_SQ_STATS_LEN;
2246 static void virtnet_get_channels(struct net_device *dev,
2247 struct ethtool_channels *channels)
2249 struct virtnet_info *vi = netdev_priv(dev);
2251 channels->combined_count = vi->curr_queue_pairs;
2252 channels->max_combined = vi->max_queue_pairs;
2253 channels->max_other = 0;
2254 channels->rx_count = 0;
2255 channels->tx_count = 0;
2256 channels->other_count = 0;
2259 static int virtnet_set_link_ksettings(struct net_device *dev,
2260 const struct ethtool_link_ksettings *cmd)
2262 struct virtnet_info *vi = netdev_priv(dev);
2264 return ethtool_virtdev_set_link_ksettings(dev, cmd,
2265 &vi->speed, &vi->duplex);
2268 static int virtnet_get_link_ksettings(struct net_device *dev,
2269 struct ethtool_link_ksettings *cmd)
2271 struct virtnet_info *vi = netdev_priv(dev);
2273 cmd->base.speed = vi->speed;
2274 cmd->base.duplex = vi->duplex;
2275 cmd->base.port = PORT_OTHER;
2280 static int virtnet_set_coalesce(struct net_device *dev,
2281 struct ethtool_coalesce *ec)
2283 struct virtnet_info *vi = netdev_priv(dev);
2286 if (ec->tx_max_coalesced_frames > 1 ||
2287 ec->rx_max_coalesced_frames != 1)
2290 napi_weight = ec->tx_max_coalesced_frames ? NAPI_POLL_WEIGHT : 0;
2291 if (napi_weight ^ vi->sq[0].napi.weight) {
2292 if (dev->flags & IFF_UP)
2294 for (i = 0; i < vi->max_queue_pairs; i++)
2295 vi->sq[i].napi.weight = napi_weight;
2301 static int virtnet_get_coalesce(struct net_device *dev,
2302 struct ethtool_coalesce *ec)
2304 struct ethtool_coalesce ec_default = {
2305 .cmd = ETHTOOL_GCOALESCE,
2306 .rx_max_coalesced_frames = 1,
2308 struct virtnet_info *vi = netdev_priv(dev);
2310 memcpy(ec, &ec_default, sizeof(ec_default));
2312 if (vi->sq[0].napi.weight)
2313 ec->tx_max_coalesced_frames = 1;
2318 static void virtnet_init_settings(struct net_device *dev)
2320 struct virtnet_info *vi = netdev_priv(dev);
2322 vi->speed = SPEED_UNKNOWN;
2323 vi->duplex = DUPLEX_UNKNOWN;
2326 static void virtnet_update_settings(struct virtnet_info *vi)
2331 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_SPEED_DUPLEX))
2334 virtio_cread_le(vi->vdev, struct virtio_net_config, speed, &speed);
2336 if (ethtool_validate_speed(speed))
2339 virtio_cread_le(vi->vdev, struct virtio_net_config, duplex, &duplex);
2341 if (ethtool_validate_duplex(duplex))
2342 vi->duplex = duplex;
2345 static const struct ethtool_ops virtnet_ethtool_ops = {
2346 .supported_coalesce_params = ETHTOOL_COALESCE_MAX_FRAMES,
2347 .get_drvinfo = virtnet_get_drvinfo,
2348 .get_link = ethtool_op_get_link,
2349 .get_ringparam = virtnet_get_ringparam,
2350 .get_strings = virtnet_get_strings,
2351 .get_sset_count = virtnet_get_sset_count,
2352 .get_ethtool_stats = virtnet_get_ethtool_stats,
2353 .set_channels = virtnet_set_channels,
2354 .get_channels = virtnet_get_channels,
2355 .get_ts_info = ethtool_op_get_ts_info,
2356 .get_link_ksettings = virtnet_get_link_ksettings,
2357 .set_link_ksettings = virtnet_set_link_ksettings,
2358 .set_coalesce = virtnet_set_coalesce,
2359 .get_coalesce = virtnet_get_coalesce,
2362 static void virtnet_freeze_down(struct virtio_device *vdev)
2364 struct virtnet_info *vi = vdev->priv;
2367 /* Make sure no work handler is accessing the device */
2368 flush_work(&vi->config_work);
2370 netif_tx_lock_bh(vi->dev);
2371 netif_device_detach(vi->dev);
2372 netif_tx_unlock_bh(vi->dev);
2373 cancel_delayed_work_sync(&vi->refill);
2375 if (netif_running(vi->dev)) {
2376 for (i = 0; i < vi->max_queue_pairs; i++) {
2377 napi_disable(&vi->rq[i].napi);
2378 virtnet_napi_tx_disable(&vi->sq[i].napi);
2383 static int init_vqs(struct virtnet_info *vi);
2385 static int virtnet_restore_up(struct virtio_device *vdev)
2387 struct virtnet_info *vi = vdev->priv;
2394 virtio_device_ready(vdev);
2396 if (netif_running(vi->dev)) {
2397 for (i = 0; i < vi->curr_queue_pairs; i++)
2398 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
2399 schedule_delayed_work(&vi->refill, 0);
2401 for (i = 0; i < vi->max_queue_pairs; i++) {
2402 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2403 virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2408 netif_tx_lock_bh(vi->dev);
2409 netif_device_attach(vi->dev);
2410 netif_tx_unlock_bh(vi->dev);
2414 static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads)
2416 struct scatterlist sg;
2417 vi->ctrl->offloads = cpu_to_virtio64(vi->vdev, offloads);
2419 sg_init_one(&sg, &vi->ctrl->offloads, sizeof(vi->ctrl->offloads));
2421 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
2422 VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) {
2423 dev_warn(&vi->dev->dev, "Fail to set guest offload.\n");
2430 static int virtnet_clear_guest_offloads(struct virtnet_info *vi)
2434 if (!vi->guest_offloads)
2437 return virtnet_set_guest_offloads(vi, offloads);
2440 static int virtnet_restore_guest_offloads(struct virtnet_info *vi)
2442 u64 offloads = vi->guest_offloads;
2444 if (!vi->guest_offloads)
2447 return virtnet_set_guest_offloads(vi, offloads);
2450 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
2451 struct netlink_ext_ack *extack)
2453 unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr);
2454 struct virtnet_info *vi = netdev_priv(dev);
2455 struct bpf_prog *old_prog;
2456 u16 xdp_qp = 0, curr_qp;
2459 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)
2460 && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2461 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2462 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
2463 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) ||
2464 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))) {
2465 NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing LRO/CSUM, disable LRO/CSUM first");
2469 if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
2470 NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required");
2474 if (dev->mtu > max_sz) {
2475 NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP");
2476 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz);
2480 curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
2482 xdp_qp = nr_cpu_ids;
2484 /* XDP requires extra queues for XDP_TX */
2485 if (curr_qp + xdp_qp > vi->max_queue_pairs) {
2486 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",
2487 curr_qp + xdp_qp, vi->max_queue_pairs);
2491 old_prog = rtnl_dereference(vi->rq[0].xdp_prog);
2492 if (!prog && !old_prog)
2496 bpf_prog_add(prog, vi->max_queue_pairs - 1);
2498 /* Make sure NAPI is not using any XDP TX queues for RX. */
2499 if (netif_running(dev)) {
2500 for (i = 0; i < vi->max_queue_pairs; i++) {
2501 napi_disable(&vi->rq[i].napi);
2502 virtnet_napi_tx_disable(&vi->sq[i].napi);
2507 for (i = 0; i < vi->max_queue_pairs; i++) {
2508 rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2510 virtnet_restore_guest_offloads(vi);
2515 err = _virtnet_set_queues(vi, curr_qp + xdp_qp);
2518 netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
2519 vi->xdp_queue_pairs = xdp_qp;
2522 vi->xdp_enabled = true;
2523 for (i = 0; i < vi->max_queue_pairs; i++) {
2524 rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2525 if (i == 0 && !old_prog)
2526 virtnet_clear_guest_offloads(vi);
2529 vi->xdp_enabled = false;
2532 for (i = 0; i < vi->max_queue_pairs; i++) {
2534 bpf_prog_put(old_prog);
2535 if (netif_running(dev)) {
2536 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2537 virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2546 virtnet_clear_guest_offloads(vi);
2547 for (i = 0; i < vi->max_queue_pairs; i++)
2548 rcu_assign_pointer(vi->rq[i].xdp_prog, old_prog);
2551 if (netif_running(dev)) {
2552 for (i = 0; i < vi->max_queue_pairs; i++) {
2553 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2554 virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2559 bpf_prog_sub(prog, vi->max_queue_pairs - 1);
2563 static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp)
2565 switch (xdp->command) {
2566 case XDP_SETUP_PROG:
2567 return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
2573 static int virtnet_get_phys_port_name(struct net_device *dev, char *buf,
2576 struct virtnet_info *vi = netdev_priv(dev);
2579 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
2582 ret = snprintf(buf, len, "sby");
2589 static int virtnet_set_features(struct net_device *dev,
2590 netdev_features_t features)
2592 struct virtnet_info *vi = netdev_priv(dev);
2596 if ((dev->features ^ features) & NETIF_F_LRO) {
2597 if (vi->xdp_enabled)
2600 if (features & NETIF_F_LRO)
2601 offloads = vi->guest_offloads_capable;
2603 offloads = vi->guest_offloads_capable &
2604 ~GUEST_OFFLOAD_LRO_MASK;
2606 err = virtnet_set_guest_offloads(vi, offloads);
2609 vi->guest_offloads = offloads;
2615 static const struct net_device_ops virtnet_netdev = {
2616 .ndo_open = virtnet_open,
2617 .ndo_stop = virtnet_close,
2618 .ndo_start_xmit = start_xmit,
2619 .ndo_validate_addr = eth_validate_addr,
2620 .ndo_set_mac_address = virtnet_set_mac_address,
2621 .ndo_set_rx_mode = virtnet_set_rx_mode,
2622 .ndo_get_stats64 = virtnet_stats,
2623 .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
2624 .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
2625 .ndo_bpf = virtnet_xdp,
2626 .ndo_xdp_xmit = virtnet_xdp_xmit,
2627 .ndo_features_check = passthru_features_check,
2628 .ndo_get_phys_port_name = virtnet_get_phys_port_name,
2629 .ndo_set_features = virtnet_set_features,
2632 static void virtnet_config_changed_work(struct work_struct *work)
2634 struct virtnet_info *vi =
2635 container_of(work, struct virtnet_info, config_work);
2638 if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
2639 struct virtio_net_config, status, &v) < 0)
2642 if (v & VIRTIO_NET_S_ANNOUNCE) {
2643 netdev_notify_peers(vi->dev);
2644 virtnet_ack_link_announce(vi);
2647 /* Ignore unknown (future) status bits */
2648 v &= VIRTIO_NET_S_LINK_UP;
2650 if (vi->status == v)
2655 if (vi->status & VIRTIO_NET_S_LINK_UP) {
2656 virtnet_update_settings(vi);
2657 netif_carrier_on(vi->dev);
2658 netif_tx_wake_all_queues(vi->dev);
2660 netif_carrier_off(vi->dev);
2661 netif_tx_stop_all_queues(vi->dev);
2665 static void virtnet_config_changed(struct virtio_device *vdev)
2667 struct virtnet_info *vi = vdev->priv;
2669 schedule_work(&vi->config_work);
2672 static void virtnet_free_queues(struct virtnet_info *vi)
2676 for (i = 0; i < vi->max_queue_pairs; i++) {
2677 __netif_napi_del(&vi->rq[i].napi);
2678 __netif_napi_del(&vi->sq[i].napi);
2681 /* We called __netif_napi_del(),
2682 * we need to respect an RCU grace period before freeing vi->rq
2691 static void _free_receive_bufs(struct virtnet_info *vi)
2693 struct bpf_prog *old_prog;
2696 for (i = 0; i < vi->max_queue_pairs; i++) {
2697 while (vi->rq[i].pages)
2698 __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
2700 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2701 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
2703 bpf_prog_put(old_prog);
2707 static void free_receive_bufs(struct virtnet_info *vi)
2710 _free_receive_bufs(vi);
2714 static void free_receive_page_frags(struct virtnet_info *vi)
2717 for (i = 0; i < vi->max_queue_pairs; i++)
2718 if (vi->rq[i].alloc_frag.page)
2719 put_page(vi->rq[i].alloc_frag.page);
2722 static void free_unused_bufs(struct virtnet_info *vi)
2727 for (i = 0; i < vi->max_queue_pairs; i++) {
2728 struct virtqueue *vq = vi->sq[i].vq;
2729 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2730 if (!is_xdp_frame(buf))
2733 xdp_return_frame(ptr_to_xdp(buf));
2737 for (i = 0; i < vi->max_queue_pairs; i++) {
2738 struct virtqueue *vq = vi->rq[i].vq;
2740 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2741 if (vi->mergeable_rx_bufs) {
2742 put_page(virt_to_head_page(buf));
2743 } else if (vi->big_packets) {
2744 give_pages(&vi->rq[i], buf);
2746 put_page(virt_to_head_page(buf));
2752 static void virtnet_del_vqs(struct virtnet_info *vi)
2754 struct virtio_device *vdev = vi->vdev;
2756 virtnet_clean_affinity(vi);
2758 vdev->config->del_vqs(vdev);
2760 virtnet_free_queues(vi);
2763 /* How large should a single buffer be so a queue full of these can fit at
2764 * least one full packet?
2765 * Logic below assumes the mergeable buffer header is used.
2767 static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq)
2769 const unsigned int hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2770 unsigned int rq_size = virtqueue_get_vring_size(vq);
2771 unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu;
2772 unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len;
2773 unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size);
2775 return max(max(min_buf_len, hdr_len) - hdr_len,
2776 (unsigned int)GOOD_PACKET_LEN);
2779 static int virtnet_find_vqs(struct virtnet_info *vi)
2781 vq_callback_t **callbacks;
2782 struct virtqueue **vqs;
2788 /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
2789 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
2790 * possible control vq.
2792 total_vqs = vi->max_queue_pairs * 2 +
2793 virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
2795 /* Allocate space for find_vqs parameters */
2796 vqs = kcalloc(total_vqs, sizeof(*vqs), GFP_KERNEL);
2799 callbacks = kmalloc_array(total_vqs, sizeof(*callbacks), GFP_KERNEL);
2802 names = kmalloc_array(total_vqs, sizeof(*names), GFP_KERNEL);
2805 if (!vi->big_packets || vi->mergeable_rx_bufs) {
2806 ctx = kcalloc(total_vqs, sizeof(*ctx), GFP_KERNEL);
2813 /* Parameters for control virtqueue, if any */
2815 callbacks[total_vqs - 1] = NULL;
2816 names[total_vqs - 1] = "control";
2819 /* Allocate/initialize parameters for send/receive virtqueues */
2820 for (i = 0; i < vi->max_queue_pairs; i++) {
2821 callbacks[rxq2vq(i)] = skb_recv_done;
2822 callbacks[txq2vq(i)] = skb_xmit_done;
2823 sprintf(vi->rq[i].name, "input.%d", i);
2824 sprintf(vi->sq[i].name, "output.%d", i);
2825 names[rxq2vq(i)] = vi->rq[i].name;
2826 names[txq2vq(i)] = vi->sq[i].name;
2828 ctx[rxq2vq(i)] = true;
2831 ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
2837 vi->cvq = vqs[total_vqs - 1];
2838 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
2839 vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
2842 for (i = 0; i < vi->max_queue_pairs; i++) {
2843 vi->rq[i].vq = vqs[rxq2vq(i)];
2844 vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq);
2845 vi->sq[i].vq = vqs[txq2vq(i)];
2848 /* run here: ret == 0. */
2863 static int virtnet_alloc_queues(struct virtnet_info *vi)
2867 vi->ctrl = kzalloc(sizeof(*vi->ctrl), GFP_KERNEL);
2870 vi->sq = kcalloc(vi->max_queue_pairs, sizeof(*vi->sq), GFP_KERNEL);
2873 vi->rq = kcalloc(vi->max_queue_pairs, sizeof(*vi->rq), GFP_KERNEL);
2877 INIT_DELAYED_WORK(&vi->refill, refill_work);
2878 for (i = 0; i < vi->max_queue_pairs; i++) {
2879 vi->rq[i].pages = NULL;
2880 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
2882 netif_tx_napi_add(vi->dev, &vi->sq[i].napi, virtnet_poll_tx,
2883 napi_tx ? napi_weight : 0);
2885 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
2886 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
2887 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
2889 u64_stats_init(&vi->rq[i].stats.syncp);
2890 u64_stats_init(&vi->sq[i].stats.syncp);
2903 static int init_vqs(struct virtnet_info *vi)
2907 /* Allocate send & receive queues */
2908 ret = virtnet_alloc_queues(vi);
2912 ret = virtnet_find_vqs(vi);
2917 virtnet_set_affinity(vi);
2923 virtnet_free_queues(vi);
2929 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
2932 struct virtnet_info *vi = netdev_priv(queue->dev);
2933 unsigned int queue_index = get_netdev_rx_queue_index(queue);
2934 unsigned int headroom = virtnet_get_headroom(vi);
2935 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
2936 struct ewma_pkt_len *avg;
2938 BUG_ON(queue_index >= vi->max_queue_pairs);
2939 avg = &vi->rq[queue_index].mrg_avg_pkt_len;
2940 return sprintf(buf, "%u\n",
2941 get_mergeable_buf_len(&vi->rq[queue_index], avg,
2942 SKB_DATA_ALIGN(headroom + tailroom)));
2945 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
2946 __ATTR_RO(mergeable_rx_buffer_size);
2948 static struct attribute *virtio_net_mrg_rx_attrs[] = {
2949 &mergeable_rx_buffer_size_attribute.attr,
2953 static const struct attribute_group virtio_net_mrg_rx_group = {
2954 .name = "virtio_net",
2955 .attrs = virtio_net_mrg_rx_attrs
2959 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
2961 const char *fname, const char *dname)
2963 if (!virtio_has_feature(vdev, fbit))
2966 dev_err(&vdev->dev, "device advertises feature %s but not %s",
2972 #define VIRTNET_FAIL_ON(vdev, fbit, dbit) \
2973 virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
2975 static bool virtnet_validate_features(struct virtio_device *vdev)
2977 if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
2978 (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
2979 "VIRTIO_NET_F_CTRL_VQ") ||
2980 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
2981 "VIRTIO_NET_F_CTRL_VQ") ||
2982 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
2983 "VIRTIO_NET_F_CTRL_VQ") ||
2984 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
2985 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
2986 "VIRTIO_NET_F_CTRL_VQ"))) {
2993 #define MIN_MTU ETH_MIN_MTU
2994 #define MAX_MTU ETH_MAX_MTU
2996 static int virtnet_validate(struct virtio_device *vdev)
2998 if (!vdev->config->get) {
2999 dev_err(&vdev->dev, "%s failure: config access disabled\n",
3004 if (!virtnet_validate_features(vdev))
3007 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
3008 int mtu = virtio_cread16(vdev,
3009 offsetof(struct virtio_net_config,
3012 __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
3018 static int virtnet_probe(struct virtio_device *vdev)
3020 int i, err = -ENOMEM;
3021 struct net_device *dev;
3022 struct virtnet_info *vi;
3023 u16 max_queue_pairs;
3026 /* Find if host supports multiqueue virtio_net device */
3027 err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
3028 struct virtio_net_config,
3029 max_virtqueue_pairs, &max_queue_pairs);
3031 /* We need at least 2 queue's */
3032 if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
3033 max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
3034 !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
3035 max_queue_pairs = 1;
3037 /* Allocate ourselves a network device with room for our info */
3038 dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
3042 /* Set up network device as normal. */
3043 dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE |
3044 IFF_TX_SKB_NO_LINEAR;
3045 dev->netdev_ops = &virtnet_netdev;
3046 dev->features = NETIF_F_HIGHDMA;
3048 dev->ethtool_ops = &virtnet_ethtool_ops;
3049 SET_NETDEV_DEV(dev, &vdev->dev);
3051 /* Do we support "hardware" checksums? */
3052 if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
3053 /* This opens up the world of extra features. */
3054 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
3056 dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
3058 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
3059 dev->hw_features |= NETIF_F_TSO
3060 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
3062 /* Individual feature bits: what can host handle? */
3063 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
3064 dev->hw_features |= NETIF_F_TSO;
3065 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
3066 dev->hw_features |= NETIF_F_TSO6;
3067 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
3068 dev->hw_features |= NETIF_F_TSO_ECN;
3070 dev->features |= NETIF_F_GSO_ROBUST;
3073 dev->features |= dev->hw_features & NETIF_F_ALL_TSO;
3074 /* (!csum && gso) case will be fixed by register_netdev() */
3076 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
3077 dev->features |= NETIF_F_RXCSUM;
3078 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3079 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6))
3080 dev->features |= NETIF_F_LRO;
3081 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS))
3082 dev->hw_features |= NETIF_F_LRO;
3084 dev->vlan_features = dev->features;
3086 /* MTU range: 68 - 65535 */
3087 dev->min_mtu = MIN_MTU;
3088 dev->max_mtu = MAX_MTU;
3090 /* Configuration may specify what MAC to use. Otherwise random. */
3091 if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
3092 virtio_cread_bytes(vdev,
3093 offsetof(struct virtio_net_config, mac),
3094 dev->dev_addr, dev->addr_len);
3096 eth_hw_addr_random(dev);
3098 /* Set up our device-specific information */
3099 vi = netdev_priv(dev);
3104 INIT_WORK(&vi->config_work, virtnet_config_changed_work);
3106 /* If we can receive ANY GSO packets, we must allocate large ones. */
3107 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3108 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
3109 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
3110 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
3111 vi->big_packets = true;
3113 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
3114 vi->mergeable_rx_bufs = true;
3116 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
3117 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
3118 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
3120 vi->hdr_len = sizeof(struct virtio_net_hdr);
3122 if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
3123 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
3124 vi->any_header_sg = true;
3126 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
3129 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
3130 mtu = virtio_cread16(vdev,
3131 offsetof(struct virtio_net_config,
3133 if (mtu < dev->min_mtu) {
3134 /* Should never trigger: MTU was previously validated
3135 * in virtnet_validate.
3138 "device MTU appears to have changed it is now %d < %d",
3147 /* TODO: size buffers correctly in this case. */
3148 if (dev->mtu > ETH_DATA_LEN)
3149 vi->big_packets = true;
3152 if (vi->any_header_sg)
3153 dev->needed_headroom = vi->hdr_len;
3155 /* Enable multiqueue by default */
3156 if (num_online_cpus() >= max_queue_pairs)
3157 vi->curr_queue_pairs = max_queue_pairs;
3159 vi->curr_queue_pairs = num_online_cpus();
3160 vi->max_queue_pairs = max_queue_pairs;
3162 /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
3168 if (vi->mergeable_rx_bufs)
3169 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
3171 netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
3172 netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
3174 virtnet_init_settings(dev);
3176 if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY)) {
3177 vi->failover = net_failover_create(vi->dev);
3178 if (IS_ERR(vi->failover)) {
3179 err = PTR_ERR(vi->failover);
3184 err = register_netdev(dev);
3186 pr_debug("virtio_net: registering device failed\n");
3190 virtio_device_ready(vdev);
3192 err = virtnet_cpu_notif_add(vi);
3194 pr_debug("virtio_net: registering cpu notifier failed\n");
3195 goto free_unregister_netdev;
3198 virtnet_set_queues(vi, vi->curr_queue_pairs);
3200 /* Assume link up if device can't report link status,
3201 otherwise get link status from config. */
3202 netif_carrier_off(dev);
3203 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
3204 schedule_work(&vi->config_work);
3206 vi->status = VIRTIO_NET_S_LINK_UP;
3207 virtnet_update_settings(vi);
3208 netif_carrier_on(dev);
3211 for (i = 0; i < ARRAY_SIZE(guest_offloads); i++)
3212 if (virtio_has_feature(vi->vdev, guest_offloads[i]))
3213 set_bit(guest_offloads[i], &vi->guest_offloads);
3214 vi->guest_offloads_capable = vi->guest_offloads;
3216 pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
3217 dev->name, max_queue_pairs);
3221 free_unregister_netdev:
3222 vi->vdev->config->reset(vdev);
3224 unregister_netdev(dev);
3226 net_failover_destroy(vi->failover);
3228 cancel_delayed_work_sync(&vi->refill);
3229 free_receive_page_frags(vi);
3230 virtnet_del_vqs(vi);
3236 static void remove_vq_common(struct virtnet_info *vi)
3238 vi->vdev->config->reset(vi->vdev);
3240 /* Free unused buffers in both send and recv, if any. */
3241 free_unused_bufs(vi);
3243 free_receive_bufs(vi);
3245 free_receive_page_frags(vi);
3247 virtnet_del_vqs(vi);
3250 static void virtnet_remove(struct virtio_device *vdev)
3252 struct virtnet_info *vi = vdev->priv;
3254 virtnet_cpu_notif_remove(vi);
3256 /* Make sure no work handler is accessing the device. */
3257 flush_work(&vi->config_work);
3259 unregister_netdev(vi->dev);
3261 net_failover_destroy(vi->failover);
3263 remove_vq_common(vi);
3265 free_netdev(vi->dev);
3268 static __maybe_unused int virtnet_freeze(struct virtio_device *vdev)
3270 struct virtnet_info *vi = vdev->priv;
3272 virtnet_cpu_notif_remove(vi);
3273 virtnet_freeze_down(vdev);
3274 remove_vq_common(vi);
3279 static __maybe_unused int virtnet_restore(struct virtio_device *vdev)
3281 struct virtnet_info *vi = vdev->priv;
3284 err = virtnet_restore_up(vdev);
3287 virtnet_set_queues(vi, vi->curr_queue_pairs);
3289 err = virtnet_cpu_notif_add(vi);
3296 static struct virtio_device_id id_table[] = {
3297 { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
3301 #define VIRTNET_FEATURES \
3302 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
3304 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
3305 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
3306 VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
3307 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
3308 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
3309 VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
3310 VIRTIO_NET_F_CTRL_MAC_ADDR, \
3311 VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \
3312 VIRTIO_NET_F_SPEED_DUPLEX, VIRTIO_NET_F_STANDBY
3314 static unsigned int features[] = {
3318 static unsigned int features_legacy[] = {
3321 VIRTIO_F_ANY_LAYOUT,
3324 static struct virtio_driver virtio_net_driver = {
3325 .feature_table = features,
3326 .feature_table_size = ARRAY_SIZE(features),
3327 .feature_table_legacy = features_legacy,
3328 .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
3329 .driver.name = KBUILD_MODNAME,
3330 .driver.owner = THIS_MODULE,
3331 .id_table = id_table,
3332 .validate = virtnet_validate,
3333 .probe = virtnet_probe,
3334 .remove = virtnet_remove,
3335 .config_changed = virtnet_config_changed,
3336 #ifdef CONFIG_PM_SLEEP
3337 .freeze = virtnet_freeze,
3338 .restore = virtnet_restore,
3342 static __init int virtio_net_driver_init(void)
3346 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
3348 virtnet_cpu_down_prep);
3351 virtionet_online = ret;
3352 ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
3353 NULL, virtnet_cpu_dead);
3357 ret = register_virtio_driver(&virtio_net_driver);
3362 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3364 cpuhp_remove_multi_state(virtionet_online);
3368 module_init(virtio_net_driver_init);
3370 static __exit void virtio_net_driver_exit(void)
3372 unregister_virtio_driver(&virtio_net_driver);
3373 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3374 cpuhp_remove_multi_state(virtionet_online);
3376 module_exit(virtio_net_driver_exit);
3378 MODULE_DEVICE_TABLE(virtio, id_table);
3379 MODULE_DESCRIPTION("Virtio network driver");
3380 MODULE_LICENSE("GPL");