1 /* A network driver using virtio.
3 * Copyright 2007 Rusty Russell <rusty@rustcorp.com.au> IBM Corporation
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation; either version 2 of the License, or
8 * (at your option) any later version.
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, see <http://www.gnu.org/licenses/>.
19 #include <linux/netdevice.h>
20 #include <linux/etherdevice.h>
21 #include <linux/ethtool.h>
22 #include <linux/module.h>
23 #include <linux/virtio.h>
24 #include <linux/virtio_net.h>
25 #include <linux/bpf.h>
26 #include <linux/bpf_trace.h>
27 #include <linux/scatterlist.h>
28 #include <linux/if_vlan.h>
29 #include <linux/slab.h>
30 #include <linux/cpu.h>
31 #include <linux/average.h>
32 #include <linux/filter.h>
33 #include <linux/kernel.h>
34 #include <linux/pci.h>
35 #include <net/route.h>
37 #include <net/net_failover.h>
39 static int napi_weight = NAPI_POLL_WEIGHT;
40 module_param(napi_weight, int, 0444);
42 static bool csum = true, gso = true, napi_tx;
43 module_param(csum, bool, 0444);
44 module_param(gso, bool, 0444);
45 module_param(napi_tx, bool, 0644);
47 /* FIXME: MTU in config. */
48 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
49 #define GOOD_COPY_LEN 128
51 #define VIRTNET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
53 /* Amount of XDP headroom to prepend to packets for use by xdp_adjust_head */
54 #define VIRTIO_XDP_HEADROOM 256
56 /* Separating two types of XDP xmit */
57 #define VIRTIO_XDP_TX BIT(0)
58 #define VIRTIO_XDP_REDIR BIT(1)
60 #define VIRTIO_XDP_FLAG BIT(0)
62 /* RX packet size EWMA. The average packet size is used to determine the packet
63 * buffer size when refilling RX rings. As the entire RX ring may be refilled
64 * at once, the weight is chosen so that the EWMA will be insensitive to short-
65 * term, transient changes in packet size.
67 DECLARE_EWMA(pkt_len, 0, 64)
69 #define VIRTNET_DRIVER_VERSION "1.0.0"
71 static const unsigned long guest_offloads[] = {
72 VIRTIO_NET_F_GUEST_TSO4,
73 VIRTIO_NET_F_GUEST_TSO6,
74 VIRTIO_NET_F_GUEST_ECN,
75 VIRTIO_NET_F_GUEST_UFO,
76 VIRTIO_NET_F_GUEST_CSUM
79 struct virtnet_stat_desc {
80 char desc[ETH_GSTRING_LEN];
84 struct virtnet_sq_stats {
85 struct u64_stats_sync syncp;
93 struct virtnet_rq_stats {
94 struct u64_stats_sync syncp;
105 #define VIRTNET_SQ_STAT(m) offsetof(struct virtnet_sq_stats, m)
106 #define VIRTNET_RQ_STAT(m) offsetof(struct virtnet_rq_stats, m)
108 static const struct virtnet_stat_desc virtnet_sq_stats_desc[] = {
109 { "packets", VIRTNET_SQ_STAT(packets) },
110 { "bytes", VIRTNET_SQ_STAT(bytes) },
111 { "xdp_tx", VIRTNET_SQ_STAT(xdp_tx) },
112 { "xdp_tx_drops", VIRTNET_SQ_STAT(xdp_tx_drops) },
113 { "kicks", VIRTNET_SQ_STAT(kicks) },
116 static const struct virtnet_stat_desc virtnet_rq_stats_desc[] = {
117 { "packets", VIRTNET_RQ_STAT(packets) },
118 { "bytes", VIRTNET_RQ_STAT(bytes) },
119 { "drops", VIRTNET_RQ_STAT(drops) },
120 { "xdp_packets", VIRTNET_RQ_STAT(xdp_packets) },
121 { "xdp_tx", VIRTNET_RQ_STAT(xdp_tx) },
122 { "xdp_redirects", VIRTNET_RQ_STAT(xdp_redirects) },
123 { "xdp_drops", VIRTNET_RQ_STAT(xdp_drops) },
124 { "kicks", VIRTNET_RQ_STAT(kicks) },
127 #define VIRTNET_SQ_STATS_LEN ARRAY_SIZE(virtnet_sq_stats_desc)
128 #define VIRTNET_RQ_STATS_LEN ARRAY_SIZE(virtnet_rq_stats_desc)
130 /* Internal representation of a send virtqueue */
132 /* Virtqueue associated with this send _queue */
133 struct virtqueue *vq;
135 /* TX: fragments + linear part + virtio header */
136 struct scatterlist sg[MAX_SKB_FRAGS + 2];
138 /* Name of the send queue: output.$index */
141 struct virtnet_sq_stats stats;
143 struct napi_struct napi;
146 /* Internal representation of a receive virtqueue */
147 struct receive_queue {
148 /* Virtqueue associated with this receive_queue */
149 struct virtqueue *vq;
151 struct napi_struct napi;
153 struct bpf_prog __rcu *xdp_prog;
155 struct virtnet_rq_stats stats;
157 /* Chain pages by the private ptr. */
160 /* Average packet length for mergeable receive buffers. */
161 struct ewma_pkt_len mrg_avg_pkt_len;
163 /* Page frag for packet buffer allocation. */
164 struct page_frag alloc_frag;
166 /* RX: fragments + linear part + virtio header */
167 struct scatterlist sg[MAX_SKB_FRAGS + 2];
169 /* Min single buffer size for mergeable buffers case. */
170 unsigned int min_buf_len;
172 /* Name of this receive queue: input.$index */
175 struct xdp_rxq_info xdp_rxq;
178 /* Control VQ buffers: protected by the rtnl lock */
180 struct virtio_net_ctrl_hdr hdr;
181 virtio_net_ctrl_ack status;
182 struct virtio_net_ctrl_mq mq;
189 struct virtnet_info {
190 struct virtio_device *vdev;
191 struct virtqueue *cvq;
192 struct net_device *dev;
193 struct send_queue *sq;
194 struct receive_queue *rq;
197 /* Max # of queue pairs supported by the device */
200 /* # of queue pairs currently used by the driver */
201 u16 curr_queue_pairs;
203 /* # of XDP queue pairs currently used by the driver */
206 /* I like... big packets and I cannot lie! */
209 /* Host will merge rx buffers for big packets (shake it! shake it!) */
210 bool mergeable_rx_bufs;
212 /* Has control virtqueue */
215 /* Host can handle any s/g split between our header and packet data */
218 /* Packet virtio header size */
221 /* Work struct for refilling if we run low on memory. */
222 struct delayed_work refill;
224 /* Work struct for config space updates */
225 struct work_struct config_work;
227 /* Does the affinity hint is set for virtqueues? */
228 bool affinity_hint_set;
230 /* CPU hotplug instances for online & dead */
231 struct hlist_node node;
232 struct hlist_node node_dead;
234 struct control_buf *ctrl;
236 /* Ethtool settings */
240 unsigned long guest_offloads;
241 unsigned long guest_offloads_capable;
243 /* failover when STANDBY feature enabled */
244 struct failover *failover;
247 struct padded_vnet_hdr {
248 struct virtio_net_hdr_mrg_rxbuf hdr;
250 * hdr is in a separate sg buffer, and data sg buffer shares same page
251 * with this header sg. This padding makes next sg 16 byte aligned
257 static bool is_xdp_frame(void *ptr)
259 return (unsigned long)ptr & VIRTIO_XDP_FLAG;
262 static void *xdp_to_ptr(struct xdp_frame *ptr)
264 return (void *)((unsigned long)ptr | VIRTIO_XDP_FLAG);
267 static struct xdp_frame *ptr_to_xdp(void *ptr)
269 return (struct xdp_frame *)((unsigned long)ptr & ~VIRTIO_XDP_FLAG);
272 /* Converting between virtqueue no. and kernel tx/rx queue no.
273 * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
275 static int vq2txq(struct virtqueue *vq)
277 return (vq->index - 1) / 2;
280 static int txq2vq(int txq)
285 static int vq2rxq(struct virtqueue *vq)
287 return vq->index / 2;
290 static int rxq2vq(int rxq)
295 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
297 return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
301 * private is used to chain pages for big packets, put the whole
302 * most recent used list in the beginning for reuse
304 static void give_pages(struct receive_queue *rq, struct page *page)
308 /* Find end of list, sew whole thing into vi->rq.pages. */
309 for (end = page; end->private; end = (struct page *)end->private);
310 end->private = (unsigned long)rq->pages;
314 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
316 struct page *p = rq->pages;
319 rq->pages = (struct page *)p->private;
320 /* clear private here, it is used to chain pages */
323 p = alloc_page(gfp_mask);
327 static void virtqueue_napi_schedule(struct napi_struct *napi,
328 struct virtqueue *vq)
330 if (napi_schedule_prep(napi)) {
331 virtqueue_disable_cb(vq);
332 __napi_schedule(napi);
336 static void virtqueue_napi_complete(struct napi_struct *napi,
337 struct virtqueue *vq, int processed)
341 opaque = virtqueue_enable_cb_prepare(vq);
342 if (napi_complete_done(napi, processed)) {
343 if (unlikely(virtqueue_poll(vq, opaque)))
344 virtqueue_napi_schedule(napi, vq);
346 virtqueue_disable_cb(vq);
350 static void skb_xmit_done(struct virtqueue *vq)
352 struct virtnet_info *vi = vq->vdev->priv;
353 struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi;
355 /* Suppress further interrupts. */
356 virtqueue_disable_cb(vq);
359 virtqueue_napi_schedule(napi, vq);
361 /* We were probably waiting for more output buffers. */
362 netif_wake_subqueue(vi->dev, vq2txq(vq));
365 #define MRG_CTX_HEADER_SHIFT 22
366 static void *mergeable_len_to_ctx(unsigned int truesize,
367 unsigned int headroom)
369 return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) | truesize);
372 static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx)
374 return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT;
377 static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx)
379 return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) - 1);
382 /* Called from bottom half context */
383 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
384 struct receive_queue *rq,
385 struct page *page, unsigned int offset,
386 unsigned int len, unsigned int truesize,
390 struct virtio_net_hdr_mrg_rxbuf *hdr;
391 unsigned int copy, hdr_len, hdr_padded_len;
394 p = page_address(page) + offset;
396 /* copy small packet so we can reuse these pages for small data */
397 skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
401 hdr = skb_vnet_hdr(skb);
403 hdr_len = vi->hdr_len;
404 if (vi->mergeable_rx_bufs)
405 hdr_padded_len = sizeof(*hdr);
407 hdr_padded_len = sizeof(struct padded_vnet_hdr);
410 memcpy(hdr, p, hdr_len);
413 offset += hdr_padded_len;
417 if (copy > skb_tailroom(skb))
418 copy = skb_tailroom(skb);
419 skb_put_data(skb, p, copy);
424 if (vi->mergeable_rx_bufs) {
426 skb_add_rx_frag(skb, 0, page, offset, len, truesize);
433 * Verify that we can indeed put this data into a skb.
434 * This is here to handle cases when the device erroneously
435 * tries to receive more than is possible. This is usually
436 * the case of a broken device.
438 if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
439 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
443 BUG_ON(offset >= PAGE_SIZE);
445 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
446 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
447 frag_size, truesize);
449 page = (struct page *)page->private;
454 give_pages(rq, page);
459 static int __virtnet_xdp_xmit_one(struct virtnet_info *vi,
460 struct send_queue *sq,
461 struct xdp_frame *xdpf)
463 struct virtio_net_hdr_mrg_rxbuf *hdr;
466 /* virtqueue want to use data area in-front of packet */
467 if (unlikely(xdpf->metasize > 0))
470 if (unlikely(xdpf->headroom < vi->hdr_len))
473 /* Make room for virtqueue hdr (also change xdpf->headroom?) */
474 xdpf->data -= vi->hdr_len;
475 /* Zero header and leave csum up to XDP layers */
477 memset(hdr, 0, vi->hdr_len);
478 xdpf->len += vi->hdr_len;
480 sg_init_one(sq->sg, xdpf->data, xdpf->len);
482 err = virtqueue_add_outbuf(sq->vq, sq->sg, 1, xdp_to_ptr(xdpf),
485 return -ENOSPC; /* Caller handle free/refcnt */
490 static struct send_queue *virtnet_xdp_sq(struct virtnet_info *vi)
494 qp = vi->curr_queue_pairs - vi->xdp_queue_pairs + smp_processor_id();
498 static int virtnet_xdp_xmit(struct net_device *dev,
499 int n, struct xdp_frame **frames, u32 flags)
501 struct virtnet_info *vi = netdev_priv(dev);
502 struct receive_queue *rq = vi->rq;
503 struct bpf_prog *xdp_prog;
504 struct send_queue *sq;
514 /* Only allow ndo_xdp_xmit if XDP is loaded on dev, as this
515 * indicate XDP resources have been successfully allocated.
517 xdp_prog = rcu_dereference(rq->xdp_prog);
521 sq = virtnet_xdp_sq(vi);
523 if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK)) {
529 /* Free up any pending old buffers before queueing new ones. */
530 while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
531 if (likely(is_xdp_frame(ptr))) {
532 struct xdp_frame *frame = ptr_to_xdp(ptr);
535 xdp_return_frame(frame);
537 struct sk_buff *skb = ptr;
540 napi_consume_skb(skb, false);
545 for (i = 0; i < n; i++) {
546 struct xdp_frame *xdpf = frames[i];
548 err = __virtnet_xdp_xmit_one(vi, sq, xdpf);
550 xdp_return_frame_rx_napi(xdpf);
556 if (flags & XDP_XMIT_FLUSH) {
557 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq))
561 u64_stats_update_begin(&sq->stats.syncp);
562 sq->stats.bytes += bytes;
563 sq->stats.packets += packets;
564 sq->stats.xdp_tx += n;
565 sq->stats.xdp_tx_drops += drops;
566 sq->stats.kicks += kicks;
567 u64_stats_update_end(&sq->stats.syncp);
572 static unsigned int virtnet_get_headroom(struct virtnet_info *vi)
574 return vi->xdp_queue_pairs ? VIRTIO_XDP_HEADROOM : 0;
577 /* We copy the packet for XDP in the following cases:
579 * 1) Packet is scattered across multiple rx buffers.
580 * 2) Headroom space is insufficient.
582 * This is inefficient but it's a temporary condition that
583 * we hit right after XDP is enabled and until queue is refilled
584 * with large buffers with sufficient headroom - so it should affect
585 * at most queue size packets.
586 * Afterwards, the conditions to enable
587 * XDP should preclude the underlying device from sending packets
588 * across multiple buffers (num_buf > 1), and we make sure buffers
589 * have enough headroom.
591 static struct page *xdp_linearize_page(struct receive_queue *rq,
598 struct page *page = alloc_page(GFP_ATOMIC);
603 memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
607 int tailroom = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
612 buf = virtqueue_get_buf(rq->vq, &buflen);
616 p = virt_to_head_page(buf);
617 off = buf - page_address(p);
619 /* guard against a misconfigured or uncooperative backend that
620 * is sending packet larger than the MTU.
622 if ((page_off + buflen + tailroom) > PAGE_SIZE) {
627 memcpy(page_address(page) + page_off,
628 page_address(p) + off, buflen);
633 /* Headroom does not contribute to packet length */
634 *len = page_off - VIRTIO_XDP_HEADROOM;
637 __free_pages(page, 0);
641 static struct sk_buff *receive_small(struct net_device *dev,
642 struct virtnet_info *vi,
643 struct receive_queue *rq,
644 void *buf, void *ctx,
646 unsigned int *xdp_xmit,
647 struct virtnet_rq_stats *stats)
650 struct bpf_prog *xdp_prog;
651 unsigned int xdp_headroom = (unsigned long)ctx;
652 unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom;
653 unsigned int headroom = vi->hdr_len + header_offset;
654 unsigned int buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
655 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
656 struct page *page = virt_to_head_page(buf);
657 unsigned int delta = 0;
658 struct page *xdp_page;
665 xdp_prog = rcu_dereference(rq->xdp_prog);
667 struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset;
668 struct xdp_frame *xdpf;
673 if (unlikely(hdr->hdr.gso_type))
676 if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) {
677 int offset = buf - page_address(page) + header_offset;
678 unsigned int tlen = len + vi->hdr_len;
681 xdp_headroom = virtnet_get_headroom(vi);
682 header_offset = VIRTNET_RX_PAD + xdp_headroom;
683 headroom = vi->hdr_len + header_offset;
684 buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
685 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
686 xdp_page = xdp_linearize_page(rq, &num_buf, page,
687 offset, header_offset,
692 buf = page_address(xdp_page);
697 xdp.data_hard_start = buf + VIRTNET_RX_PAD + vi->hdr_len;
698 xdp.data = xdp.data_hard_start + xdp_headroom;
699 xdp_set_data_meta_invalid(&xdp);
700 xdp.data_end = xdp.data + len;
701 xdp.rxq = &rq->xdp_rxq;
702 orig_data = xdp.data;
703 act = bpf_prog_run_xdp(xdp_prog, &xdp);
704 stats->xdp_packets++;
708 /* Recalculate length in case bpf program changed it */
709 delta = orig_data - xdp.data;
710 len = xdp.data_end - xdp.data;
714 xdpf = convert_to_xdp_frame(&xdp);
717 err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
718 if (unlikely(err < 0)) {
719 trace_xdp_exception(vi->dev, xdp_prog, act);
722 *xdp_xmit |= VIRTIO_XDP_TX;
726 stats->xdp_redirects++;
727 err = xdp_do_redirect(dev, &xdp, xdp_prog);
730 *xdp_xmit |= VIRTIO_XDP_REDIR;
734 bpf_warn_invalid_xdp_action(act);
737 trace_xdp_exception(vi->dev, xdp_prog, act);
744 skb = build_skb(buf, buflen);
749 skb_reserve(skb, headroom - delta);
752 buf += header_offset;
753 memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
754 } /* keep zeroed vnet hdr since packet was changed by bpf */
768 static struct sk_buff *receive_big(struct net_device *dev,
769 struct virtnet_info *vi,
770 struct receive_queue *rq,
773 struct virtnet_rq_stats *stats)
775 struct page *page = buf;
776 struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len,
779 stats->bytes += len - vi->hdr_len;
787 give_pages(rq, page);
791 static struct sk_buff *receive_mergeable(struct net_device *dev,
792 struct virtnet_info *vi,
793 struct receive_queue *rq,
797 unsigned int *xdp_xmit,
798 struct virtnet_rq_stats *stats)
800 struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
801 u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
802 struct page *page = virt_to_head_page(buf);
803 int offset = buf - page_address(page);
804 struct sk_buff *head_skb, *curr_skb;
805 struct bpf_prog *xdp_prog;
806 unsigned int truesize;
807 unsigned int headroom = mergeable_ctx_to_headroom(ctx);
811 stats->bytes += len - vi->hdr_len;
814 xdp_prog = rcu_dereference(rq->xdp_prog);
816 struct xdp_frame *xdpf;
817 struct page *xdp_page;
822 /* Transient failure which in theory could occur if
823 * in-flight packets from before XDP was enabled reach
824 * the receive path after XDP is loaded.
826 if (unlikely(hdr->hdr.gso_type))
829 /* This happens when rx buffer size is underestimated
830 * or headroom is not enough because of the buffer
831 * was refilled before XDP is set. This should only
832 * happen for the first several packets, so we don't
833 * care much about its performance.
835 if (unlikely(num_buf > 1 ||
836 headroom < virtnet_get_headroom(vi))) {
837 /* linearize data for XDP */
838 xdp_page = xdp_linearize_page(rq, &num_buf,
844 offset = VIRTIO_XDP_HEADROOM;
849 /* Allow consuming headroom but reserve enough space to push
850 * the descriptor on if we get an XDP_TX return code.
852 data = page_address(xdp_page) + offset;
853 xdp.data_hard_start = data - VIRTIO_XDP_HEADROOM + vi->hdr_len;
854 xdp.data = data + vi->hdr_len;
855 xdp_set_data_meta_invalid(&xdp);
856 xdp.data_end = xdp.data + (len - vi->hdr_len);
857 xdp.rxq = &rq->xdp_rxq;
859 act = bpf_prog_run_xdp(xdp_prog, &xdp);
860 stats->xdp_packets++;
864 /* recalculate offset to account for any header
865 * adjustments. Note other cases do not build an
866 * skb and avoid using offset
869 page_address(xdp_page) - vi->hdr_len;
871 /* recalculate len if xdp.data or xdp.data_end were
874 len = xdp.data_end - xdp.data + vi->hdr_len;
875 /* We can only create skb based on xdp_page. */
876 if (unlikely(xdp_page != page)) {
879 head_skb = page_to_skb(vi, rq, xdp_page,
887 xdpf = convert_to_xdp_frame(&xdp);
890 err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
891 if (unlikely(err < 0)) {
892 trace_xdp_exception(vi->dev, xdp_prog, act);
893 if (unlikely(xdp_page != page))
897 *xdp_xmit |= VIRTIO_XDP_TX;
898 if (unlikely(xdp_page != page))
903 stats->xdp_redirects++;
904 err = xdp_do_redirect(dev, &xdp, xdp_prog);
906 if (unlikely(xdp_page != page))
910 *xdp_xmit |= VIRTIO_XDP_REDIR;
911 if (unlikely(xdp_page != page))
916 bpf_warn_invalid_xdp_action(act);
919 trace_xdp_exception(vi->dev, xdp_prog, act);
922 if (unlikely(xdp_page != page))
923 __free_pages(xdp_page, 0);
929 truesize = mergeable_ctx_to_truesize(ctx);
930 if (unlikely(len > truesize)) {
931 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
932 dev->name, len, (unsigned long)ctx);
933 dev->stats.rx_length_errors++;
937 head_skb = page_to_skb(vi, rq, page, offset, len, truesize, !xdp_prog);
940 if (unlikely(!curr_skb))
945 buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx);
946 if (unlikely(!buf)) {
947 pr_debug("%s: rx error: %d buffers out of %d missing\n",
949 virtio16_to_cpu(vi->vdev,
951 dev->stats.rx_length_errors++;
956 page = virt_to_head_page(buf);
958 truesize = mergeable_ctx_to_truesize(ctx);
959 if (unlikely(len > truesize)) {
960 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
961 dev->name, len, (unsigned long)ctx);
962 dev->stats.rx_length_errors++;
966 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
967 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
968 struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
972 if (curr_skb == head_skb)
973 skb_shinfo(curr_skb)->frag_list = nskb;
975 curr_skb->next = nskb;
977 head_skb->truesize += nskb->truesize;
980 if (curr_skb != head_skb) {
981 head_skb->data_len += len;
982 head_skb->len += len;
983 head_skb->truesize += truesize;
985 offset = buf - page_address(page);
986 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
988 skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
991 skb_add_rx_frag(curr_skb, num_skb_frags, page,
992 offset, len, truesize);
996 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
1004 while (num_buf-- > 1) {
1005 buf = virtqueue_get_buf(rq->vq, &len);
1006 if (unlikely(!buf)) {
1007 pr_debug("%s: rx error: %d buffers missing\n",
1008 dev->name, num_buf);
1009 dev->stats.rx_length_errors++;
1012 stats->bytes += len;
1013 page = virt_to_head_page(buf);
1018 dev_kfree_skb(head_skb);
1023 static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
1024 void *buf, unsigned int len, void **ctx,
1025 unsigned int *xdp_xmit,
1026 struct virtnet_rq_stats *stats)
1028 struct net_device *dev = vi->dev;
1029 struct sk_buff *skb;
1030 struct virtio_net_hdr_mrg_rxbuf *hdr;
1032 if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
1033 pr_debug("%s: short packet %i\n", dev->name, len);
1034 dev->stats.rx_length_errors++;
1035 if (vi->mergeable_rx_bufs) {
1036 put_page(virt_to_head_page(buf));
1037 } else if (vi->big_packets) {
1038 give_pages(rq, buf);
1040 put_page(virt_to_head_page(buf));
1045 if (vi->mergeable_rx_bufs)
1046 skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit,
1048 else if (vi->big_packets)
1049 skb = receive_big(dev, vi, rq, buf, len, stats);
1051 skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit, stats);
1056 hdr = skb_vnet_hdr(skb);
1058 if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
1059 skb->ip_summed = CHECKSUM_UNNECESSARY;
1061 if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
1062 virtio_is_little_endian(vi->vdev))) {
1063 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
1064 dev->name, hdr->hdr.gso_type,
1069 skb->protocol = eth_type_trans(skb, dev);
1070 pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
1071 ntohs(skb->protocol), skb->len, skb->pkt_type);
1073 napi_gro_receive(&rq->napi, skb);
1077 dev->stats.rx_frame_errors++;
1081 /* Unlike mergeable buffers, all buffers are allocated to the
1082 * same size, except for the headroom. For this reason we do
1083 * not need to use mergeable_len_to_ctx here - it is enough
1084 * to store the headroom as the context ignoring the truesize.
1086 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
1089 struct page_frag *alloc_frag = &rq->alloc_frag;
1091 unsigned int xdp_headroom = virtnet_get_headroom(vi);
1092 void *ctx = (void *)(unsigned long)xdp_headroom;
1093 int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
1096 len = SKB_DATA_ALIGN(len) +
1097 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1098 if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
1101 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1102 get_page(alloc_frag->page);
1103 alloc_frag->offset += len;
1104 sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom,
1105 vi->hdr_len + GOOD_PACKET_LEN);
1106 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1108 put_page(virt_to_head_page(buf));
1112 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
1115 struct page *first, *list = NULL;
1119 sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
1121 /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
1122 for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
1123 first = get_a_page(rq, gfp);
1126 give_pages(rq, list);
1129 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
1131 /* chain new page in list head to match sg */
1132 first->private = (unsigned long)list;
1136 first = get_a_page(rq, gfp);
1138 give_pages(rq, list);
1141 p = page_address(first);
1143 /* rq->sg[0], rq->sg[1] share the same page */
1144 /* a separated rq->sg[0] for header - required in case !any_header_sg */
1145 sg_set_buf(&rq->sg[0], p, vi->hdr_len);
1147 /* rq->sg[1] for data packet, from offset */
1148 offset = sizeof(struct padded_vnet_hdr);
1149 sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
1151 /* chain first in list head */
1152 first->private = (unsigned long)list;
1153 err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
1156 give_pages(rq, first);
1161 static unsigned int get_mergeable_buf_len(struct receive_queue *rq,
1162 struct ewma_pkt_len *avg_pkt_len,
1165 const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1169 return PAGE_SIZE - room;
1171 len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
1172 rq->min_buf_len, PAGE_SIZE - hdr_len);
1174 return ALIGN(len, L1_CACHE_BYTES);
1177 static int add_recvbuf_mergeable(struct virtnet_info *vi,
1178 struct receive_queue *rq, gfp_t gfp)
1180 struct page_frag *alloc_frag = &rq->alloc_frag;
1181 unsigned int headroom = virtnet_get_headroom(vi);
1182 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1183 unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
1187 unsigned int len, hole;
1189 /* Extra tailroom is needed to satisfy XDP's assumption. This
1190 * means rx frags coalescing won't work, but consider we've
1191 * disabled GSO for XDP, it won't be a big issue.
1193 len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len, room);
1194 if (unlikely(!skb_page_frag_refill(len + room, alloc_frag, gfp)))
1197 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1198 buf += headroom; /* advance address leaving hole at front of pkt */
1199 get_page(alloc_frag->page);
1200 alloc_frag->offset += len + room;
1201 hole = alloc_frag->size - alloc_frag->offset;
1202 if (hole < len + room) {
1203 /* To avoid internal fragmentation, if there is very likely not
1204 * enough space for another buffer, add the remaining space to
1205 * the current buffer.
1208 alloc_frag->offset += hole;
1211 sg_init_one(rq->sg, buf, len);
1212 ctx = mergeable_len_to_ctx(len, headroom);
1213 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1215 put_page(virt_to_head_page(buf));
1221 * Returns false if we couldn't fill entirely (OOM).
1223 * Normally run in the receive path, but can also be run from ndo_open
1224 * before we're receiving packets, or from refill_work which is
1225 * careful to disable receiving (using napi_disable).
1227 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
1234 if (vi->mergeable_rx_bufs)
1235 err = add_recvbuf_mergeable(vi, rq, gfp);
1236 else if (vi->big_packets)
1237 err = add_recvbuf_big(vi, rq, gfp);
1239 err = add_recvbuf_small(vi, rq, gfp);
1241 oom = err == -ENOMEM;
1244 } while (rq->vq->num_free);
1245 if (virtqueue_kick_prepare(rq->vq) && virtqueue_notify(rq->vq)) {
1246 u64_stats_update_begin(&rq->stats.syncp);
1248 u64_stats_update_end(&rq->stats.syncp);
1254 static void skb_recv_done(struct virtqueue *rvq)
1256 struct virtnet_info *vi = rvq->vdev->priv;
1257 struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
1259 virtqueue_napi_schedule(&rq->napi, rvq);
1262 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
1266 /* If all buffers were filled by other side before we napi_enabled, we
1267 * won't get another interrupt, so process any outstanding packets now.
1268 * Call local_bh_enable after to trigger softIRQ processing.
1271 virtqueue_napi_schedule(napi, vq);
1275 static void virtnet_napi_tx_enable(struct virtnet_info *vi,
1276 struct virtqueue *vq,
1277 struct napi_struct *napi)
1282 /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
1283 * enable the feature if this is likely affine with the transmit path.
1285 if (!vi->affinity_hint_set) {
1290 return virtnet_napi_enable(vq, napi);
1293 static void virtnet_napi_tx_disable(struct napi_struct *napi)
1299 static void refill_work(struct work_struct *work)
1301 struct virtnet_info *vi =
1302 container_of(work, struct virtnet_info, refill.work);
1306 for (i = 0; i < vi->curr_queue_pairs; i++) {
1307 struct receive_queue *rq = &vi->rq[i];
1309 napi_disable(&rq->napi);
1310 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
1311 virtnet_napi_enable(rq->vq, &rq->napi);
1313 /* In theory, this can happen: if we don't get any buffers in
1314 * we will *never* try to fill again.
1317 schedule_delayed_work(&vi->refill, HZ/2);
1321 static int virtnet_receive(struct receive_queue *rq, int budget,
1322 unsigned int *xdp_xmit)
1324 struct virtnet_info *vi = rq->vq->vdev->priv;
1325 struct virtnet_rq_stats stats = {};
1330 if (!vi->big_packets || vi->mergeable_rx_bufs) {
1333 while (stats.packets < budget &&
1334 (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) {
1335 receive_buf(vi, rq, buf, len, ctx, xdp_xmit, &stats);
1339 while (stats.packets < budget &&
1340 (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
1341 receive_buf(vi, rq, buf, len, NULL, xdp_xmit, &stats);
1346 if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) {
1347 if (!try_fill_recv(vi, rq, GFP_ATOMIC))
1348 schedule_delayed_work(&vi->refill, 0);
1351 u64_stats_update_begin(&rq->stats.syncp);
1352 for (i = 0; i < VIRTNET_RQ_STATS_LEN; i++) {
1353 size_t offset = virtnet_rq_stats_desc[i].offset;
1356 item = (u64 *)((u8 *)&rq->stats + offset);
1357 *item += *(u64 *)((u8 *)&stats + offset);
1359 u64_stats_update_end(&rq->stats.syncp);
1361 return stats.packets;
1364 static void free_old_xmit_skbs(struct send_queue *sq, bool in_napi)
1367 unsigned int packets = 0;
1368 unsigned int bytes = 0;
1371 while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
1372 if (likely(!is_xdp_frame(ptr))) {
1373 struct sk_buff *skb = ptr;
1375 pr_debug("Sent skb %p\n", skb);
1378 napi_consume_skb(skb, in_napi);
1380 struct xdp_frame *frame = ptr_to_xdp(ptr);
1382 bytes += frame->len;
1383 xdp_return_frame(frame);
1388 /* Avoid overhead when no packets have been processed
1389 * happens when called speculatively from start_xmit.
1394 u64_stats_update_begin(&sq->stats.syncp);
1395 sq->stats.bytes += bytes;
1396 sq->stats.packets += packets;
1397 u64_stats_update_end(&sq->stats.syncp);
1400 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
1402 if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
1404 else if (q < vi->curr_queue_pairs)
1410 static void virtnet_poll_cleantx(struct receive_queue *rq)
1412 struct virtnet_info *vi = rq->vq->vdev->priv;
1413 unsigned int index = vq2rxq(rq->vq);
1414 struct send_queue *sq = &vi->sq[index];
1415 struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index);
1417 if (!sq->napi.weight || is_xdp_raw_buffer_queue(vi, index))
1420 if (__netif_tx_trylock(txq)) {
1421 free_old_xmit_skbs(sq, true);
1422 __netif_tx_unlock(txq);
1425 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1426 netif_tx_wake_queue(txq);
1429 static int virtnet_poll(struct napi_struct *napi, int budget)
1431 struct receive_queue *rq =
1432 container_of(napi, struct receive_queue, napi);
1433 struct virtnet_info *vi = rq->vq->vdev->priv;
1434 struct send_queue *sq;
1435 unsigned int received;
1436 unsigned int xdp_xmit = 0;
1438 virtnet_poll_cleantx(rq);
1440 received = virtnet_receive(rq, budget, &xdp_xmit);
1442 /* Out of packets? */
1443 if (received < budget)
1444 virtqueue_napi_complete(napi, rq->vq, received);
1446 if (xdp_xmit & VIRTIO_XDP_REDIR)
1449 if (xdp_xmit & VIRTIO_XDP_TX) {
1450 sq = virtnet_xdp_sq(vi);
1451 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1452 u64_stats_update_begin(&sq->stats.syncp);
1454 u64_stats_update_end(&sq->stats.syncp);
1461 static int virtnet_open(struct net_device *dev)
1463 struct virtnet_info *vi = netdev_priv(dev);
1466 for (i = 0; i < vi->max_queue_pairs; i++) {
1467 if (i < vi->curr_queue_pairs)
1468 /* Make sure we have some buffers: if oom use wq. */
1469 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1470 schedule_delayed_work(&vi->refill, 0);
1472 err = xdp_rxq_info_reg(&vi->rq[i].xdp_rxq, dev, i);
1476 err = xdp_rxq_info_reg_mem_model(&vi->rq[i].xdp_rxq,
1477 MEM_TYPE_PAGE_SHARED, NULL);
1479 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1483 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1484 virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi);
1490 static int virtnet_poll_tx(struct napi_struct *napi, int budget)
1492 struct send_queue *sq = container_of(napi, struct send_queue, napi);
1493 struct virtnet_info *vi = sq->vq->vdev->priv;
1494 unsigned int index = vq2txq(sq->vq);
1495 struct netdev_queue *txq;
1497 if (unlikely(is_xdp_raw_buffer_queue(vi, index))) {
1498 /* We don't need to enable cb for XDP */
1499 napi_complete_done(napi, 0);
1503 txq = netdev_get_tx_queue(vi->dev, index);
1504 __netif_tx_lock(txq, raw_smp_processor_id());
1505 free_old_xmit_skbs(sq, true);
1506 __netif_tx_unlock(txq);
1508 virtqueue_napi_complete(napi, sq->vq, 0);
1510 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1511 netif_tx_wake_queue(txq);
1516 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
1518 struct virtio_net_hdr_mrg_rxbuf *hdr;
1519 const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
1520 struct virtnet_info *vi = sq->vq->vdev->priv;
1522 unsigned hdr_len = vi->hdr_len;
1525 pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
1527 can_push = vi->any_header_sg &&
1528 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
1529 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
1530 /* Even if we can, don't push here yet as this would skew
1531 * csum_start offset below. */
1533 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1535 hdr = skb_vnet_hdr(skb);
1537 if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1538 virtio_is_little_endian(vi->vdev), false,
1542 if (vi->mergeable_rx_bufs)
1543 hdr->num_buffers = 0;
1545 sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1547 __skb_push(skb, hdr_len);
1548 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1549 if (unlikely(num_sg < 0))
1551 /* Pull header back to avoid skew in tx bytes calculations. */
1552 __skb_pull(skb, hdr_len);
1554 sg_set_buf(sq->sg, hdr, hdr_len);
1555 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
1556 if (unlikely(num_sg < 0))
1560 return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
1563 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
1565 struct virtnet_info *vi = netdev_priv(dev);
1566 int qnum = skb_get_queue_mapping(skb);
1567 struct send_queue *sq = &vi->sq[qnum];
1569 struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1570 bool kick = !skb->xmit_more;
1571 bool use_napi = sq->napi.weight;
1573 /* Free up any pending old buffers before queueing new ones. */
1574 free_old_xmit_skbs(sq, false);
1576 if (use_napi && kick)
1577 virtqueue_enable_cb_delayed(sq->vq);
1579 /* timestamp packet in software */
1580 skb_tx_timestamp(skb);
1582 /* Try to transmit */
1583 err = xmit_skb(sq, skb);
1585 /* This should not happen! */
1586 if (unlikely(err)) {
1587 dev->stats.tx_fifo_errors++;
1588 if (net_ratelimit())
1590 "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
1591 dev->stats.tx_dropped++;
1592 dev_kfree_skb_any(skb);
1593 return NETDEV_TX_OK;
1596 /* Don't wait up for transmitted skbs to be freed. */
1602 /* If running out of space, stop queue to avoid getting packets that we
1603 * are then unable to transmit.
1604 * An alternative would be to force queuing layer to requeue the skb by
1605 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1606 * returned in a normal path of operation: it means that driver is not
1607 * maintaining the TX queue stop/start state properly, and causes
1608 * the stack to do a non-trivial amount of useless work.
1609 * Since most packets only take 1 or 2 ring slots, stopping the queue
1610 * early means 16 slots are typically wasted.
1612 if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
1613 netif_stop_subqueue(dev, qnum);
1615 unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
1616 /* More just got used, free them then recheck. */
1617 free_old_xmit_skbs(sq, false);
1618 if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
1619 netif_start_subqueue(dev, qnum);
1620 virtqueue_disable_cb(sq->vq);
1625 if (kick || netif_xmit_stopped(txq)) {
1626 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1627 u64_stats_update_begin(&sq->stats.syncp);
1629 u64_stats_update_end(&sq->stats.syncp);
1633 return NETDEV_TX_OK;
1637 * Send command via the control virtqueue and check status. Commands
1638 * supported by the hypervisor, as indicated by feature bits, should
1639 * never fail unless improperly formatted.
1641 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
1642 struct scatterlist *out)
1644 struct scatterlist *sgs[4], hdr, stat;
1645 unsigned out_num = 0, tmp;
1647 /* Caller should know better */
1648 BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
1650 vi->ctrl->status = ~0;
1651 vi->ctrl->hdr.class = class;
1652 vi->ctrl->hdr.cmd = cmd;
1654 sg_init_one(&hdr, &vi->ctrl->hdr, sizeof(vi->ctrl->hdr));
1655 sgs[out_num++] = &hdr;
1658 sgs[out_num++] = out;
1660 /* Add return status. */
1661 sg_init_one(&stat, &vi->ctrl->status, sizeof(vi->ctrl->status));
1662 sgs[out_num] = &stat;
1664 BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1665 virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1667 if (unlikely(!virtqueue_kick(vi->cvq)))
1668 return vi->ctrl->status == VIRTIO_NET_OK;
1670 /* Spin for a response, the kick causes an ioport write, trapping
1671 * into the hypervisor, so the request should be handled immediately.
1673 while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1674 !virtqueue_is_broken(vi->cvq))
1677 return vi->ctrl->status == VIRTIO_NET_OK;
1680 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1682 struct virtnet_info *vi = netdev_priv(dev);
1683 struct virtio_device *vdev = vi->vdev;
1685 struct sockaddr *addr;
1686 struct scatterlist sg;
1688 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
1691 addr = kmemdup(p, sizeof(*addr), GFP_KERNEL);
1695 ret = eth_prepare_mac_addr_change(dev, addr);
1699 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1700 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1701 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1702 VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1703 dev_warn(&vdev->dev,
1704 "Failed to set mac address by vq command.\n");
1708 } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1709 !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1712 /* Naturally, this has an atomicity problem. */
1713 for (i = 0; i < dev->addr_len; i++)
1714 virtio_cwrite8(vdev,
1715 offsetof(struct virtio_net_config, mac) +
1716 i, addr->sa_data[i]);
1719 eth_commit_mac_addr_change(dev, p);
1727 static void virtnet_stats(struct net_device *dev,
1728 struct rtnl_link_stats64 *tot)
1730 struct virtnet_info *vi = netdev_priv(dev);
1734 for (i = 0; i < vi->max_queue_pairs; i++) {
1735 u64 tpackets, tbytes, rpackets, rbytes, rdrops;
1736 struct receive_queue *rq = &vi->rq[i];
1737 struct send_queue *sq = &vi->sq[i];
1740 start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
1741 tpackets = sq->stats.packets;
1742 tbytes = sq->stats.bytes;
1743 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
1746 start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
1747 rpackets = rq->stats.packets;
1748 rbytes = rq->stats.bytes;
1749 rdrops = rq->stats.drops;
1750 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
1752 tot->rx_packets += rpackets;
1753 tot->tx_packets += tpackets;
1754 tot->rx_bytes += rbytes;
1755 tot->tx_bytes += tbytes;
1756 tot->rx_dropped += rdrops;
1759 tot->tx_dropped = dev->stats.tx_dropped;
1760 tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1761 tot->rx_length_errors = dev->stats.rx_length_errors;
1762 tot->rx_frame_errors = dev->stats.rx_frame_errors;
1765 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1768 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1769 VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1770 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1774 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1776 struct scatterlist sg;
1777 struct net_device *dev = vi->dev;
1779 if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1782 vi->ctrl->mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1783 sg_init_one(&sg, &vi->ctrl->mq, sizeof(vi->ctrl->mq));
1785 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1786 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1787 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1791 vi->curr_queue_pairs = queue_pairs;
1792 /* virtnet_open() will refill when device is going to up. */
1793 if (dev->flags & IFF_UP)
1794 schedule_delayed_work(&vi->refill, 0);
1800 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1805 err = _virtnet_set_queues(vi, queue_pairs);
1810 static int virtnet_close(struct net_device *dev)
1812 struct virtnet_info *vi = netdev_priv(dev);
1815 /* Make sure refill_work doesn't re-enable napi! */
1816 cancel_delayed_work_sync(&vi->refill);
1818 for (i = 0; i < vi->max_queue_pairs; i++) {
1819 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1820 napi_disable(&vi->rq[i].napi);
1821 virtnet_napi_tx_disable(&vi->sq[i].napi);
1827 static void virtnet_set_rx_mode(struct net_device *dev)
1829 struct virtnet_info *vi = netdev_priv(dev);
1830 struct scatterlist sg[2];
1831 struct virtio_net_ctrl_mac *mac_data;
1832 struct netdev_hw_addr *ha;
1838 /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1839 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1842 vi->ctrl->promisc = ((dev->flags & IFF_PROMISC) != 0);
1843 vi->ctrl->allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1845 sg_init_one(sg, &vi->ctrl->promisc, sizeof(vi->ctrl->promisc));
1847 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1848 VIRTIO_NET_CTRL_RX_PROMISC, sg))
1849 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1850 vi->ctrl->promisc ? "en" : "dis");
1852 sg_init_one(sg, &vi->ctrl->allmulti, sizeof(vi->ctrl->allmulti));
1854 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1855 VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1856 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1857 vi->ctrl->allmulti ? "en" : "dis");
1859 uc_count = netdev_uc_count(dev);
1860 mc_count = netdev_mc_count(dev);
1861 /* MAC filter - use one buffer for both lists */
1862 buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1863 (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1868 sg_init_table(sg, 2);
1870 /* Store the unicast list and count in the front of the buffer */
1871 mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1873 netdev_for_each_uc_addr(ha, dev)
1874 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1876 sg_set_buf(&sg[0], mac_data,
1877 sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1879 /* multicast list and count fill the end */
1880 mac_data = (void *)&mac_data->macs[uc_count][0];
1882 mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1884 netdev_for_each_mc_addr(ha, dev)
1885 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1887 sg_set_buf(&sg[1], mac_data,
1888 sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1890 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1891 VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1892 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1897 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1898 __be16 proto, u16 vid)
1900 struct virtnet_info *vi = netdev_priv(dev);
1901 struct scatterlist sg;
1903 vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
1904 sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
1906 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1907 VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1908 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1912 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1913 __be16 proto, u16 vid)
1915 struct virtnet_info *vi = netdev_priv(dev);
1916 struct scatterlist sg;
1918 vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
1919 sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
1921 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1922 VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1923 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1927 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1931 if (vi->affinity_hint_set) {
1932 for (i = 0; i < vi->max_queue_pairs; i++) {
1933 virtqueue_set_affinity(vi->rq[i].vq, NULL);
1934 virtqueue_set_affinity(vi->sq[i].vq, NULL);
1937 vi->affinity_hint_set = false;
1941 static void virtnet_set_affinity(struct virtnet_info *vi)
1950 if (!zalloc_cpumask_var(&mask, GFP_KERNEL)) {
1951 virtnet_clean_affinity(vi, -1);
1955 num_cpu = num_online_cpus();
1956 stride = max_t(int, num_cpu / vi->curr_queue_pairs, 1);
1957 stragglers = num_cpu >= vi->curr_queue_pairs ?
1958 num_cpu % vi->curr_queue_pairs :
1960 cpu = cpumask_next(-1, cpu_online_mask);
1962 for (i = 0; i < vi->curr_queue_pairs; i++) {
1963 group_size = stride + (i < stragglers ? 1 : 0);
1965 for (j = 0; j < group_size; j++) {
1966 cpumask_set_cpu(cpu, mask);
1967 cpu = cpumask_next_wrap(cpu, cpu_online_mask,
1970 virtqueue_set_affinity(vi->rq[i].vq, mask);
1971 virtqueue_set_affinity(vi->sq[i].vq, mask);
1972 __netif_set_xps_queue(vi->dev, cpumask_bits(mask), i, false);
1973 cpumask_clear(mask);
1976 vi->affinity_hint_set = true;
1977 free_cpumask_var(mask);
1980 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
1982 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1984 virtnet_set_affinity(vi);
1988 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
1990 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1992 virtnet_set_affinity(vi);
1996 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
1998 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2001 virtnet_clean_affinity(vi, cpu);
2005 static enum cpuhp_state virtionet_online;
2007 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
2011 ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
2014 ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2018 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2022 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
2024 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2025 cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2029 static void virtnet_get_ringparam(struct net_device *dev,
2030 struct ethtool_ringparam *ring)
2032 struct virtnet_info *vi = netdev_priv(dev);
2034 ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
2035 ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
2036 ring->rx_pending = ring->rx_max_pending;
2037 ring->tx_pending = ring->tx_max_pending;
2041 static void virtnet_get_drvinfo(struct net_device *dev,
2042 struct ethtool_drvinfo *info)
2044 struct virtnet_info *vi = netdev_priv(dev);
2045 struct virtio_device *vdev = vi->vdev;
2047 strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
2048 strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
2049 strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
2053 /* TODO: Eliminate OOO packets during switching */
2054 static int virtnet_set_channels(struct net_device *dev,
2055 struct ethtool_channels *channels)
2057 struct virtnet_info *vi = netdev_priv(dev);
2058 u16 queue_pairs = channels->combined_count;
2061 /* We don't support separate rx/tx channels.
2062 * We don't allow setting 'other' channels.
2064 if (channels->rx_count || channels->tx_count || channels->other_count)
2067 if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
2070 /* For now we don't support modifying channels while XDP is loaded
2071 * also when XDP is loaded all RX queues have XDP programs so we only
2072 * need to check a single RX queue.
2074 if (vi->rq[0].xdp_prog)
2078 err = _virtnet_set_queues(vi, queue_pairs);
2080 netif_set_real_num_tx_queues(dev, queue_pairs);
2081 netif_set_real_num_rx_queues(dev, queue_pairs);
2083 virtnet_set_affinity(vi);
2090 static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *data)
2092 struct virtnet_info *vi = netdev_priv(dev);
2093 char *p = (char *)data;
2096 switch (stringset) {
2098 for (i = 0; i < vi->curr_queue_pairs; i++) {
2099 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
2100 snprintf(p, ETH_GSTRING_LEN, "rx_queue_%u_%s",
2101 i, virtnet_rq_stats_desc[j].desc);
2102 p += ETH_GSTRING_LEN;
2106 for (i = 0; i < vi->curr_queue_pairs; i++) {
2107 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
2108 snprintf(p, ETH_GSTRING_LEN, "tx_queue_%u_%s",
2109 i, virtnet_sq_stats_desc[j].desc);
2110 p += ETH_GSTRING_LEN;
2117 static int virtnet_get_sset_count(struct net_device *dev, int sset)
2119 struct virtnet_info *vi = netdev_priv(dev);
2123 return vi->curr_queue_pairs * (VIRTNET_RQ_STATS_LEN +
2124 VIRTNET_SQ_STATS_LEN);
2130 static void virtnet_get_ethtool_stats(struct net_device *dev,
2131 struct ethtool_stats *stats, u64 *data)
2133 struct virtnet_info *vi = netdev_priv(dev);
2134 unsigned int idx = 0, start, i, j;
2135 const u8 *stats_base;
2138 for (i = 0; i < vi->curr_queue_pairs; i++) {
2139 struct receive_queue *rq = &vi->rq[i];
2141 stats_base = (u8 *)&rq->stats;
2143 start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
2144 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
2145 offset = virtnet_rq_stats_desc[j].offset;
2146 data[idx + j] = *(u64 *)(stats_base + offset);
2148 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
2149 idx += VIRTNET_RQ_STATS_LEN;
2152 for (i = 0; i < vi->curr_queue_pairs; i++) {
2153 struct send_queue *sq = &vi->sq[i];
2155 stats_base = (u8 *)&sq->stats;
2157 start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
2158 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
2159 offset = virtnet_sq_stats_desc[j].offset;
2160 data[idx + j] = *(u64 *)(stats_base + offset);
2162 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
2163 idx += VIRTNET_SQ_STATS_LEN;
2167 static void virtnet_get_channels(struct net_device *dev,
2168 struct ethtool_channels *channels)
2170 struct virtnet_info *vi = netdev_priv(dev);
2172 channels->combined_count = vi->curr_queue_pairs;
2173 channels->max_combined = vi->max_queue_pairs;
2174 channels->max_other = 0;
2175 channels->rx_count = 0;
2176 channels->tx_count = 0;
2177 channels->other_count = 0;
2180 /* Check if the user is trying to change anything besides speed/duplex */
2182 virtnet_validate_ethtool_cmd(const struct ethtool_link_ksettings *cmd)
2184 struct ethtool_link_ksettings diff1 = *cmd;
2185 struct ethtool_link_ksettings diff2 = {};
2187 /* cmd is always set so we need to clear it, validate the port type
2188 * and also without autonegotiation we can ignore advertising
2190 diff1.base.speed = 0;
2191 diff2.base.port = PORT_OTHER;
2192 ethtool_link_ksettings_zero_link_mode(&diff1, advertising);
2193 diff1.base.duplex = 0;
2195 diff1.base.link_mode_masks_nwords = 0;
2197 return !memcmp(&diff1.base, &diff2.base, sizeof(diff1.base)) &&
2198 bitmap_empty(diff1.link_modes.supported,
2199 __ETHTOOL_LINK_MODE_MASK_NBITS) &&
2200 bitmap_empty(diff1.link_modes.advertising,
2201 __ETHTOOL_LINK_MODE_MASK_NBITS) &&
2202 bitmap_empty(diff1.link_modes.lp_advertising,
2203 __ETHTOOL_LINK_MODE_MASK_NBITS);
2206 static int virtnet_set_link_ksettings(struct net_device *dev,
2207 const struct ethtool_link_ksettings *cmd)
2209 struct virtnet_info *vi = netdev_priv(dev);
2212 speed = cmd->base.speed;
2213 /* don't allow custom speed and duplex */
2214 if (!ethtool_validate_speed(speed) ||
2215 !ethtool_validate_duplex(cmd->base.duplex) ||
2216 !virtnet_validate_ethtool_cmd(cmd))
2219 vi->duplex = cmd->base.duplex;
2224 static int virtnet_get_link_ksettings(struct net_device *dev,
2225 struct ethtool_link_ksettings *cmd)
2227 struct virtnet_info *vi = netdev_priv(dev);
2229 cmd->base.speed = vi->speed;
2230 cmd->base.duplex = vi->duplex;
2231 cmd->base.port = PORT_OTHER;
2236 static int virtnet_set_coalesce(struct net_device *dev,
2237 struct ethtool_coalesce *ec)
2239 struct ethtool_coalesce ec_default = {
2240 .cmd = ETHTOOL_SCOALESCE,
2241 .rx_max_coalesced_frames = 1,
2243 struct virtnet_info *vi = netdev_priv(dev);
2246 if (ec->tx_max_coalesced_frames > 1)
2249 ec_default.tx_max_coalesced_frames = ec->tx_max_coalesced_frames;
2250 napi_weight = ec->tx_max_coalesced_frames ? NAPI_POLL_WEIGHT : 0;
2252 /* disallow changes to fields not explicitly tested above */
2253 if (memcmp(ec, &ec_default, sizeof(ec_default)))
2256 if (napi_weight ^ vi->sq[0].napi.weight) {
2257 if (dev->flags & IFF_UP)
2259 for (i = 0; i < vi->max_queue_pairs; i++)
2260 vi->sq[i].napi.weight = napi_weight;
2266 static int virtnet_get_coalesce(struct net_device *dev,
2267 struct ethtool_coalesce *ec)
2269 struct ethtool_coalesce ec_default = {
2270 .cmd = ETHTOOL_GCOALESCE,
2271 .rx_max_coalesced_frames = 1,
2273 struct virtnet_info *vi = netdev_priv(dev);
2275 memcpy(ec, &ec_default, sizeof(ec_default));
2277 if (vi->sq[0].napi.weight)
2278 ec->tx_max_coalesced_frames = 1;
2283 static void virtnet_init_settings(struct net_device *dev)
2285 struct virtnet_info *vi = netdev_priv(dev);
2287 vi->speed = SPEED_UNKNOWN;
2288 vi->duplex = DUPLEX_UNKNOWN;
2291 static void virtnet_update_settings(struct virtnet_info *vi)
2296 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_SPEED_DUPLEX))
2299 speed = virtio_cread32(vi->vdev, offsetof(struct virtio_net_config,
2301 if (ethtool_validate_speed(speed))
2303 duplex = virtio_cread8(vi->vdev, offsetof(struct virtio_net_config,
2305 if (ethtool_validate_duplex(duplex))
2306 vi->duplex = duplex;
2309 static const struct ethtool_ops virtnet_ethtool_ops = {
2310 .get_drvinfo = virtnet_get_drvinfo,
2311 .get_link = ethtool_op_get_link,
2312 .get_ringparam = virtnet_get_ringparam,
2313 .get_strings = virtnet_get_strings,
2314 .get_sset_count = virtnet_get_sset_count,
2315 .get_ethtool_stats = virtnet_get_ethtool_stats,
2316 .set_channels = virtnet_set_channels,
2317 .get_channels = virtnet_get_channels,
2318 .get_ts_info = ethtool_op_get_ts_info,
2319 .get_link_ksettings = virtnet_get_link_ksettings,
2320 .set_link_ksettings = virtnet_set_link_ksettings,
2321 .set_coalesce = virtnet_set_coalesce,
2322 .get_coalesce = virtnet_get_coalesce,
2325 static void virtnet_freeze_down(struct virtio_device *vdev)
2327 struct virtnet_info *vi = vdev->priv;
2330 /* Make sure no work handler is accessing the device */
2331 flush_work(&vi->config_work);
2333 netif_tx_lock_bh(vi->dev);
2334 netif_device_detach(vi->dev);
2335 netif_tx_unlock_bh(vi->dev);
2336 cancel_delayed_work_sync(&vi->refill);
2338 if (netif_running(vi->dev)) {
2339 for (i = 0; i < vi->max_queue_pairs; i++) {
2340 napi_disable(&vi->rq[i].napi);
2341 virtnet_napi_tx_disable(&vi->sq[i].napi);
2346 static int init_vqs(struct virtnet_info *vi);
2348 static int virtnet_restore_up(struct virtio_device *vdev)
2350 struct virtnet_info *vi = vdev->priv;
2357 virtio_device_ready(vdev);
2359 if (netif_running(vi->dev)) {
2360 for (i = 0; i < vi->curr_queue_pairs; i++)
2361 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
2362 schedule_delayed_work(&vi->refill, 0);
2364 for (i = 0; i < vi->max_queue_pairs; i++) {
2365 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2366 virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2371 netif_tx_lock_bh(vi->dev);
2372 netif_device_attach(vi->dev);
2373 netif_tx_unlock_bh(vi->dev);
2377 static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads)
2379 struct scatterlist sg;
2380 vi->ctrl->offloads = cpu_to_virtio64(vi->vdev, offloads);
2382 sg_init_one(&sg, &vi->ctrl->offloads, sizeof(vi->ctrl->offloads));
2384 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
2385 VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) {
2386 dev_warn(&vi->dev->dev, "Fail to set guest offload. \n");
2393 static int virtnet_clear_guest_offloads(struct virtnet_info *vi)
2397 if (!vi->guest_offloads)
2400 return virtnet_set_guest_offloads(vi, offloads);
2403 static int virtnet_restore_guest_offloads(struct virtnet_info *vi)
2405 u64 offloads = vi->guest_offloads;
2407 if (!vi->guest_offloads)
2410 return virtnet_set_guest_offloads(vi, offloads);
2413 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
2414 struct netlink_ext_ack *extack)
2416 unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr);
2417 struct virtnet_info *vi = netdev_priv(dev);
2418 struct bpf_prog *old_prog;
2419 u16 xdp_qp = 0, curr_qp;
2422 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)
2423 && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2424 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2425 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
2426 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) ||
2427 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))) {
2428 NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing LRO/CSUM, disable LRO/CSUM first");
2432 if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
2433 NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required");
2437 if (dev->mtu > max_sz) {
2438 NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP");
2439 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz);
2443 curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
2445 xdp_qp = nr_cpu_ids;
2447 /* XDP requires extra queues for XDP_TX */
2448 if (curr_qp + xdp_qp > vi->max_queue_pairs) {
2449 NL_SET_ERR_MSG_MOD(extack, "Too few free TX rings available");
2450 netdev_warn(dev, "request %i queues but max is %i\n",
2451 curr_qp + xdp_qp, vi->max_queue_pairs);
2455 old_prog = rtnl_dereference(vi->rq[0].xdp_prog);
2456 if (!prog && !old_prog)
2460 prog = bpf_prog_add(prog, vi->max_queue_pairs - 1);
2462 return PTR_ERR(prog);
2465 /* Make sure NAPI is not using any XDP TX queues for RX. */
2466 if (netif_running(dev)) {
2467 for (i = 0; i < vi->max_queue_pairs; i++) {
2468 napi_disable(&vi->rq[i].napi);
2469 virtnet_napi_tx_disable(&vi->sq[i].napi);
2474 for (i = 0; i < vi->max_queue_pairs; i++) {
2475 rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2477 virtnet_restore_guest_offloads(vi);
2482 err = _virtnet_set_queues(vi, curr_qp + xdp_qp);
2485 netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
2486 vi->xdp_queue_pairs = xdp_qp;
2489 for (i = 0; i < vi->max_queue_pairs; i++) {
2490 rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2491 if (i == 0 && !old_prog)
2492 virtnet_clear_guest_offloads(vi);
2496 for (i = 0; i < vi->max_queue_pairs; i++) {
2498 bpf_prog_put(old_prog);
2499 if (netif_running(dev)) {
2500 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2501 virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2510 virtnet_clear_guest_offloads(vi);
2511 for (i = 0; i < vi->max_queue_pairs; i++)
2512 rcu_assign_pointer(vi->rq[i].xdp_prog, old_prog);
2515 if (netif_running(dev)) {
2516 for (i = 0; i < vi->max_queue_pairs; i++) {
2517 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2518 virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2523 bpf_prog_sub(prog, vi->max_queue_pairs - 1);
2527 static u32 virtnet_xdp_query(struct net_device *dev)
2529 struct virtnet_info *vi = netdev_priv(dev);
2530 const struct bpf_prog *xdp_prog;
2533 for (i = 0; i < vi->max_queue_pairs; i++) {
2534 xdp_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2536 return xdp_prog->aux->id;
2541 static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp)
2543 switch (xdp->command) {
2544 case XDP_SETUP_PROG:
2545 return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
2546 case XDP_QUERY_PROG:
2547 xdp->prog_id = virtnet_xdp_query(dev);
2554 static int virtnet_get_phys_port_name(struct net_device *dev, char *buf,
2557 struct virtnet_info *vi = netdev_priv(dev);
2560 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
2563 ret = snprintf(buf, len, "sby");
2570 static int virtnet_set_features(struct net_device *dev,
2571 netdev_features_t features)
2573 struct virtnet_info *vi = netdev_priv(dev);
2577 if ((dev->features ^ features) & NETIF_F_LRO) {
2578 if (vi->xdp_queue_pairs)
2581 if (features & NETIF_F_LRO)
2582 offloads = vi->guest_offloads_capable;
2586 err = virtnet_set_guest_offloads(vi, offloads);
2589 vi->guest_offloads = offloads;
2595 static const struct net_device_ops virtnet_netdev = {
2596 .ndo_open = virtnet_open,
2597 .ndo_stop = virtnet_close,
2598 .ndo_start_xmit = start_xmit,
2599 .ndo_validate_addr = eth_validate_addr,
2600 .ndo_set_mac_address = virtnet_set_mac_address,
2601 .ndo_set_rx_mode = virtnet_set_rx_mode,
2602 .ndo_get_stats64 = virtnet_stats,
2603 .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
2604 .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
2605 .ndo_bpf = virtnet_xdp,
2606 .ndo_xdp_xmit = virtnet_xdp_xmit,
2607 .ndo_features_check = passthru_features_check,
2608 .ndo_get_phys_port_name = virtnet_get_phys_port_name,
2609 .ndo_set_features = virtnet_set_features,
2612 static void virtnet_config_changed_work(struct work_struct *work)
2614 struct virtnet_info *vi =
2615 container_of(work, struct virtnet_info, config_work);
2618 if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
2619 struct virtio_net_config, status, &v) < 0)
2622 if (v & VIRTIO_NET_S_ANNOUNCE) {
2623 netdev_notify_peers(vi->dev);
2624 virtnet_ack_link_announce(vi);
2627 /* Ignore unknown (future) status bits */
2628 v &= VIRTIO_NET_S_LINK_UP;
2630 if (vi->status == v)
2635 if (vi->status & VIRTIO_NET_S_LINK_UP) {
2636 virtnet_update_settings(vi);
2637 netif_carrier_on(vi->dev);
2638 netif_tx_wake_all_queues(vi->dev);
2640 netif_carrier_off(vi->dev);
2641 netif_tx_stop_all_queues(vi->dev);
2645 static void virtnet_config_changed(struct virtio_device *vdev)
2647 struct virtnet_info *vi = vdev->priv;
2649 schedule_work(&vi->config_work);
2652 static void virtnet_free_queues(struct virtnet_info *vi)
2656 for (i = 0; i < vi->max_queue_pairs; i++) {
2657 napi_hash_del(&vi->rq[i].napi);
2658 netif_napi_del(&vi->rq[i].napi);
2659 netif_napi_del(&vi->sq[i].napi);
2662 /* We called napi_hash_del() before netif_napi_del(),
2663 * we need to respect an RCU grace period before freeing vi->rq
2672 static void _free_receive_bufs(struct virtnet_info *vi)
2674 struct bpf_prog *old_prog;
2677 for (i = 0; i < vi->max_queue_pairs; i++) {
2678 while (vi->rq[i].pages)
2679 __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
2681 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2682 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
2684 bpf_prog_put(old_prog);
2688 static void free_receive_bufs(struct virtnet_info *vi)
2691 _free_receive_bufs(vi);
2695 static void free_receive_page_frags(struct virtnet_info *vi)
2698 for (i = 0; i < vi->max_queue_pairs; i++)
2699 if (vi->rq[i].alloc_frag.page)
2700 put_page(vi->rq[i].alloc_frag.page);
2703 static void free_unused_bufs(struct virtnet_info *vi)
2708 for (i = 0; i < vi->max_queue_pairs; i++) {
2709 struct virtqueue *vq = vi->sq[i].vq;
2710 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2711 if (!is_xdp_frame(buf))
2714 xdp_return_frame(ptr_to_xdp(buf));
2718 for (i = 0; i < vi->max_queue_pairs; i++) {
2719 struct virtqueue *vq = vi->rq[i].vq;
2721 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2722 if (vi->mergeable_rx_bufs) {
2723 put_page(virt_to_head_page(buf));
2724 } else if (vi->big_packets) {
2725 give_pages(&vi->rq[i], buf);
2727 put_page(virt_to_head_page(buf));
2733 static void virtnet_del_vqs(struct virtnet_info *vi)
2735 struct virtio_device *vdev = vi->vdev;
2737 virtnet_clean_affinity(vi, -1);
2739 vdev->config->del_vqs(vdev);
2741 virtnet_free_queues(vi);
2744 /* How large should a single buffer be so a queue full of these can fit at
2745 * least one full packet?
2746 * Logic below assumes the mergeable buffer header is used.
2748 static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq)
2750 const unsigned int hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2751 unsigned int rq_size = virtqueue_get_vring_size(vq);
2752 unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu;
2753 unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len;
2754 unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size);
2756 return max(max(min_buf_len, hdr_len) - hdr_len,
2757 (unsigned int)GOOD_PACKET_LEN);
2760 static int virtnet_find_vqs(struct virtnet_info *vi)
2762 vq_callback_t **callbacks;
2763 struct virtqueue **vqs;
2769 /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
2770 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
2771 * possible control vq.
2773 total_vqs = vi->max_queue_pairs * 2 +
2774 virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
2776 /* Allocate space for find_vqs parameters */
2777 vqs = kcalloc(total_vqs, sizeof(*vqs), GFP_KERNEL);
2780 callbacks = kmalloc_array(total_vqs, sizeof(*callbacks), GFP_KERNEL);
2783 names = kmalloc_array(total_vqs, sizeof(*names), GFP_KERNEL);
2786 if (!vi->big_packets || vi->mergeable_rx_bufs) {
2787 ctx = kcalloc(total_vqs, sizeof(*ctx), GFP_KERNEL);
2794 /* Parameters for control virtqueue, if any */
2796 callbacks[total_vqs - 1] = NULL;
2797 names[total_vqs - 1] = "control";
2800 /* Allocate/initialize parameters for send/receive virtqueues */
2801 for (i = 0; i < vi->max_queue_pairs; i++) {
2802 callbacks[rxq2vq(i)] = skb_recv_done;
2803 callbacks[txq2vq(i)] = skb_xmit_done;
2804 sprintf(vi->rq[i].name, "input.%d", i);
2805 sprintf(vi->sq[i].name, "output.%d", i);
2806 names[rxq2vq(i)] = vi->rq[i].name;
2807 names[txq2vq(i)] = vi->sq[i].name;
2809 ctx[rxq2vq(i)] = true;
2812 ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
2818 vi->cvq = vqs[total_vqs - 1];
2819 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
2820 vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
2823 for (i = 0; i < vi->max_queue_pairs; i++) {
2824 vi->rq[i].vq = vqs[rxq2vq(i)];
2825 vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq);
2826 vi->sq[i].vq = vqs[txq2vq(i)];
2829 /* run here: ret == 0. */
2844 static int virtnet_alloc_queues(struct virtnet_info *vi)
2848 vi->ctrl = kzalloc(sizeof(*vi->ctrl), GFP_KERNEL);
2851 vi->sq = kcalloc(vi->max_queue_pairs, sizeof(*vi->sq), GFP_KERNEL);
2854 vi->rq = kcalloc(vi->max_queue_pairs, sizeof(*vi->rq), GFP_KERNEL);
2858 INIT_DELAYED_WORK(&vi->refill, refill_work);
2859 for (i = 0; i < vi->max_queue_pairs; i++) {
2860 vi->rq[i].pages = NULL;
2861 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
2863 netif_tx_napi_add(vi->dev, &vi->sq[i].napi, virtnet_poll_tx,
2864 napi_tx ? napi_weight : 0);
2866 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
2867 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
2868 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
2870 u64_stats_init(&vi->rq[i].stats.syncp);
2871 u64_stats_init(&vi->sq[i].stats.syncp);
2884 static int init_vqs(struct virtnet_info *vi)
2888 /* Allocate send & receive queues */
2889 ret = virtnet_alloc_queues(vi);
2893 ret = virtnet_find_vqs(vi);
2898 virtnet_set_affinity(vi);
2904 virtnet_free_queues(vi);
2910 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
2913 struct virtnet_info *vi = netdev_priv(queue->dev);
2914 unsigned int queue_index = get_netdev_rx_queue_index(queue);
2915 unsigned int headroom = virtnet_get_headroom(vi);
2916 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
2917 struct ewma_pkt_len *avg;
2919 BUG_ON(queue_index >= vi->max_queue_pairs);
2920 avg = &vi->rq[queue_index].mrg_avg_pkt_len;
2921 return sprintf(buf, "%u\n",
2922 get_mergeable_buf_len(&vi->rq[queue_index], avg,
2923 SKB_DATA_ALIGN(headroom + tailroom)));
2926 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
2927 __ATTR_RO(mergeable_rx_buffer_size);
2929 static struct attribute *virtio_net_mrg_rx_attrs[] = {
2930 &mergeable_rx_buffer_size_attribute.attr,
2934 static const struct attribute_group virtio_net_mrg_rx_group = {
2935 .name = "virtio_net",
2936 .attrs = virtio_net_mrg_rx_attrs
2940 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
2942 const char *fname, const char *dname)
2944 if (!virtio_has_feature(vdev, fbit))
2947 dev_err(&vdev->dev, "device advertises feature %s but not %s",
2953 #define VIRTNET_FAIL_ON(vdev, fbit, dbit) \
2954 virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
2956 static bool virtnet_validate_features(struct virtio_device *vdev)
2958 if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
2959 (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
2960 "VIRTIO_NET_F_CTRL_VQ") ||
2961 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
2962 "VIRTIO_NET_F_CTRL_VQ") ||
2963 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
2964 "VIRTIO_NET_F_CTRL_VQ") ||
2965 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
2966 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
2967 "VIRTIO_NET_F_CTRL_VQ"))) {
2974 #define MIN_MTU ETH_MIN_MTU
2975 #define MAX_MTU ETH_MAX_MTU
2977 static int virtnet_validate(struct virtio_device *vdev)
2979 if (!vdev->config->get) {
2980 dev_err(&vdev->dev, "%s failure: config access disabled\n",
2985 if (!virtnet_validate_features(vdev))
2988 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2989 int mtu = virtio_cread16(vdev,
2990 offsetof(struct virtio_net_config,
2993 __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
2999 static int virtnet_probe(struct virtio_device *vdev)
3001 int i, err = -ENOMEM;
3002 struct net_device *dev;
3003 struct virtnet_info *vi;
3004 u16 max_queue_pairs;
3007 /* Find if host supports multiqueue virtio_net device */
3008 err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
3009 struct virtio_net_config,
3010 max_virtqueue_pairs, &max_queue_pairs);
3012 /* We need at least 2 queue's */
3013 if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
3014 max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
3015 !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
3016 max_queue_pairs = 1;
3018 /* Allocate ourselves a network device with room for our info */
3019 dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
3023 /* Set up network device as normal. */
3024 dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
3025 dev->netdev_ops = &virtnet_netdev;
3026 dev->features = NETIF_F_HIGHDMA;
3028 dev->ethtool_ops = &virtnet_ethtool_ops;
3029 SET_NETDEV_DEV(dev, &vdev->dev);
3031 /* Do we support "hardware" checksums? */
3032 if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
3033 /* This opens up the world of extra features. */
3034 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
3036 dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
3038 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
3039 dev->hw_features |= NETIF_F_TSO
3040 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
3042 /* Individual feature bits: what can host handle? */
3043 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
3044 dev->hw_features |= NETIF_F_TSO;
3045 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
3046 dev->hw_features |= NETIF_F_TSO6;
3047 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
3048 dev->hw_features |= NETIF_F_TSO_ECN;
3050 dev->features |= NETIF_F_GSO_ROBUST;
3053 dev->features |= dev->hw_features & NETIF_F_ALL_TSO;
3054 /* (!csum && gso) case will be fixed by register_netdev() */
3056 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
3057 dev->features |= NETIF_F_RXCSUM;
3058 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3059 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6))
3060 dev->features |= NETIF_F_LRO;
3061 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS))
3062 dev->hw_features |= NETIF_F_LRO;
3064 dev->vlan_features = dev->features;
3066 /* MTU range: 68 - 65535 */
3067 dev->min_mtu = MIN_MTU;
3068 dev->max_mtu = MAX_MTU;
3070 /* Configuration may specify what MAC to use. Otherwise random. */
3071 if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
3072 virtio_cread_bytes(vdev,
3073 offsetof(struct virtio_net_config, mac),
3074 dev->dev_addr, dev->addr_len);
3076 eth_hw_addr_random(dev);
3078 /* Set up our device-specific information */
3079 vi = netdev_priv(dev);
3084 INIT_WORK(&vi->config_work, virtnet_config_changed_work);
3086 /* If we can receive ANY GSO packets, we must allocate large ones. */
3087 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3088 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
3089 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
3090 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
3091 vi->big_packets = true;
3093 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
3094 vi->mergeable_rx_bufs = true;
3096 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
3097 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
3098 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
3100 vi->hdr_len = sizeof(struct virtio_net_hdr);
3102 if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
3103 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
3104 vi->any_header_sg = true;
3106 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
3109 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
3110 mtu = virtio_cread16(vdev,
3111 offsetof(struct virtio_net_config,
3113 if (mtu < dev->min_mtu) {
3114 /* Should never trigger: MTU was previously validated
3115 * in virtnet_validate.
3117 dev_err(&vdev->dev, "device MTU appears to have changed "
3118 "it is now %d < %d", mtu, dev->min_mtu);
3125 /* TODO: size buffers correctly in this case. */
3126 if (dev->mtu > ETH_DATA_LEN)
3127 vi->big_packets = true;
3130 if (vi->any_header_sg)
3131 dev->needed_headroom = vi->hdr_len;
3133 /* Enable multiqueue by default */
3134 if (num_online_cpus() >= max_queue_pairs)
3135 vi->curr_queue_pairs = max_queue_pairs;
3137 vi->curr_queue_pairs = num_online_cpus();
3138 vi->max_queue_pairs = max_queue_pairs;
3140 /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
3146 if (vi->mergeable_rx_bufs)
3147 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
3149 netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
3150 netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
3152 virtnet_init_settings(dev);
3154 if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY)) {
3155 vi->failover = net_failover_create(vi->dev);
3156 if (IS_ERR(vi->failover)) {
3157 err = PTR_ERR(vi->failover);
3162 err = register_netdev(dev);
3164 pr_debug("virtio_net: registering device failed\n");
3168 virtio_device_ready(vdev);
3170 err = virtnet_cpu_notif_add(vi);
3172 pr_debug("virtio_net: registering cpu notifier failed\n");
3173 goto free_unregister_netdev;
3176 virtnet_set_queues(vi, vi->curr_queue_pairs);
3178 /* Assume link up if device can't report link status,
3179 otherwise get link status from config. */
3180 netif_carrier_off(dev);
3181 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
3182 schedule_work(&vi->config_work);
3184 vi->status = VIRTIO_NET_S_LINK_UP;
3185 virtnet_update_settings(vi);
3186 netif_carrier_on(dev);
3189 for (i = 0; i < ARRAY_SIZE(guest_offloads); i++)
3190 if (virtio_has_feature(vi->vdev, guest_offloads[i]))
3191 set_bit(guest_offloads[i], &vi->guest_offloads);
3192 vi->guest_offloads_capable = vi->guest_offloads;
3194 pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
3195 dev->name, max_queue_pairs);
3199 free_unregister_netdev:
3200 vi->vdev->config->reset(vdev);
3202 unregister_netdev(dev);
3204 net_failover_destroy(vi->failover);
3206 cancel_delayed_work_sync(&vi->refill);
3207 free_receive_page_frags(vi);
3208 virtnet_del_vqs(vi);
3214 static void remove_vq_common(struct virtnet_info *vi)
3216 vi->vdev->config->reset(vi->vdev);
3218 /* Free unused buffers in both send and recv, if any. */
3219 free_unused_bufs(vi);
3221 free_receive_bufs(vi);
3223 free_receive_page_frags(vi);
3225 virtnet_del_vqs(vi);
3228 static void virtnet_remove(struct virtio_device *vdev)
3230 struct virtnet_info *vi = vdev->priv;
3232 virtnet_cpu_notif_remove(vi);
3234 /* Make sure no work handler is accessing the device. */
3235 flush_work(&vi->config_work);
3237 unregister_netdev(vi->dev);
3239 net_failover_destroy(vi->failover);
3241 remove_vq_common(vi);
3243 free_netdev(vi->dev);
3246 static __maybe_unused int virtnet_freeze(struct virtio_device *vdev)
3248 struct virtnet_info *vi = vdev->priv;
3250 virtnet_cpu_notif_remove(vi);
3251 virtnet_freeze_down(vdev);
3252 remove_vq_common(vi);
3257 static __maybe_unused int virtnet_restore(struct virtio_device *vdev)
3259 struct virtnet_info *vi = vdev->priv;
3262 err = virtnet_restore_up(vdev);
3265 virtnet_set_queues(vi, vi->curr_queue_pairs);
3267 err = virtnet_cpu_notif_add(vi);
3274 static struct virtio_device_id id_table[] = {
3275 { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
3279 #define VIRTNET_FEATURES \
3280 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
3282 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
3283 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
3284 VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
3285 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
3286 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
3287 VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
3288 VIRTIO_NET_F_CTRL_MAC_ADDR, \
3289 VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \
3290 VIRTIO_NET_F_SPEED_DUPLEX, VIRTIO_NET_F_STANDBY
3292 static unsigned int features[] = {
3296 static unsigned int features_legacy[] = {
3299 VIRTIO_F_ANY_LAYOUT,
3302 static struct virtio_driver virtio_net_driver = {
3303 .feature_table = features,
3304 .feature_table_size = ARRAY_SIZE(features),
3305 .feature_table_legacy = features_legacy,
3306 .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
3307 .driver.name = KBUILD_MODNAME,
3308 .driver.owner = THIS_MODULE,
3309 .id_table = id_table,
3310 .validate = virtnet_validate,
3311 .probe = virtnet_probe,
3312 .remove = virtnet_remove,
3313 .config_changed = virtnet_config_changed,
3314 #ifdef CONFIG_PM_SLEEP
3315 .freeze = virtnet_freeze,
3316 .restore = virtnet_restore,
3320 static __init int virtio_net_driver_init(void)
3324 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
3326 virtnet_cpu_down_prep);
3329 virtionet_online = ret;
3330 ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
3331 NULL, virtnet_cpu_dead);
3335 ret = register_virtio_driver(&virtio_net_driver);
3340 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3342 cpuhp_remove_multi_state(virtionet_online);
3346 module_init(virtio_net_driver_init);
3348 static __exit void virtio_net_driver_exit(void)
3350 unregister_virtio_driver(&virtio_net_driver);
3351 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3352 cpuhp_remove_multi_state(virtionet_online);
3354 module_exit(virtio_net_driver_exit);
3356 MODULE_DEVICE_TABLE(virtio, id_table);
3357 MODULE_DESCRIPTION("Virtio network driver");
3358 MODULE_LICENSE("GPL");