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;
242 /* failover when STANDBY feature enabled */
243 struct failover *failover;
246 struct padded_vnet_hdr {
247 struct virtio_net_hdr_mrg_rxbuf hdr;
249 * hdr is in a separate sg buffer, and data sg buffer shares same page
250 * with this header sg. This padding makes next sg 16 byte aligned
256 static bool is_xdp_frame(void *ptr)
258 return (unsigned long)ptr & VIRTIO_XDP_FLAG;
261 static void *xdp_to_ptr(struct xdp_frame *ptr)
263 return (void *)((unsigned long)ptr | VIRTIO_XDP_FLAG);
266 static struct xdp_frame *ptr_to_xdp(void *ptr)
268 return (struct xdp_frame *)((unsigned long)ptr & ~VIRTIO_XDP_FLAG);
271 /* Converting between virtqueue no. and kernel tx/rx queue no.
272 * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
274 static int vq2txq(struct virtqueue *vq)
276 return (vq->index - 1) / 2;
279 static int txq2vq(int txq)
284 static int vq2rxq(struct virtqueue *vq)
286 return vq->index / 2;
289 static int rxq2vq(int rxq)
294 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
296 return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
300 * private is used to chain pages for big packets, put the whole
301 * most recent used list in the beginning for reuse
303 static void give_pages(struct receive_queue *rq, struct page *page)
307 /* Find end of list, sew whole thing into vi->rq.pages. */
308 for (end = page; end->private; end = (struct page *)end->private);
309 end->private = (unsigned long)rq->pages;
313 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
315 struct page *p = rq->pages;
318 rq->pages = (struct page *)p->private;
319 /* clear private here, it is used to chain pages */
322 p = alloc_page(gfp_mask);
326 static void virtqueue_napi_schedule(struct napi_struct *napi,
327 struct virtqueue *vq)
329 if (napi_schedule_prep(napi)) {
330 virtqueue_disable_cb(vq);
331 __napi_schedule(napi);
335 static void virtqueue_napi_complete(struct napi_struct *napi,
336 struct virtqueue *vq, int processed)
340 opaque = virtqueue_enable_cb_prepare(vq);
341 if (napi_complete_done(napi, processed)) {
342 if (unlikely(virtqueue_poll(vq, opaque)))
343 virtqueue_napi_schedule(napi, vq);
345 virtqueue_disable_cb(vq);
349 static void skb_xmit_done(struct virtqueue *vq)
351 struct virtnet_info *vi = vq->vdev->priv;
352 struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi;
354 /* Suppress further interrupts. */
355 virtqueue_disable_cb(vq);
358 virtqueue_napi_schedule(napi, vq);
360 /* We were probably waiting for more output buffers. */
361 netif_wake_subqueue(vi->dev, vq2txq(vq));
364 #define MRG_CTX_HEADER_SHIFT 22
365 static void *mergeable_len_to_ctx(unsigned int truesize,
366 unsigned int headroom)
368 return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) | truesize);
371 static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx)
373 return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT;
376 static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx)
378 return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) - 1);
381 /* Called from bottom half context */
382 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
383 struct receive_queue *rq,
384 struct page *page, unsigned int offset,
385 unsigned int len, unsigned int truesize,
386 bool hdr_valid, unsigned int metasize)
389 struct virtio_net_hdr_mrg_rxbuf *hdr;
390 unsigned int copy, hdr_len, hdr_padded_len;
393 p = page_address(page) + offset;
395 /* copy small packet so we can reuse these pages for small data */
396 skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
400 hdr = skb_vnet_hdr(skb);
402 hdr_len = vi->hdr_len;
403 if (vi->mergeable_rx_bufs)
404 hdr_padded_len = sizeof(*hdr);
406 hdr_padded_len = sizeof(struct padded_vnet_hdr);
408 /* hdr_valid means no XDP, so we can copy the vnet header */
410 memcpy(hdr, p, hdr_len);
413 offset += hdr_padded_len;
416 /* Copy all frame if it fits skb->head, otherwise
417 * we let virtio_net_hdr_to_skb() and GRO pull headers as needed.
419 if (len <= skb_tailroom(skb))
422 copy = ETH_HLEN + metasize;
423 skb_put_data(skb, p, copy);
426 __skb_pull(skb, metasize);
427 skb_metadata_set(skb, metasize);
433 if (vi->mergeable_rx_bufs) {
435 skb_add_rx_frag(skb, 0, page, offset, len, truesize);
442 * Verify that we can indeed put this data into a skb.
443 * This is here to handle cases when the device erroneously
444 * tries to receive more than is possible. This is usually
445 * the case of a broken device.
447 if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
448 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
452 BUG_ON(offset >= PAGE_SIZE);
454 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
455 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
456 frag_size, truesize);
458 page = (struct page *)page->private;
463 give_pages(rq, page);
468 static int __virtnet_xdp_xmit_one(struct virtnet_info *vi,
469 struct send_queue *sq,
470 struct xdp_frame *xdpf)
472 struct virtio_net_hdr_mrg_rxbuf *hdr;
475 if (unlikely(xdpf->headroom < vi->hdr_len))
478 /* Make room for virtqueue hdr (also change xdpf->headroom?) */
479 xdpf->data -= vi->hdr_len;
480 /* Zero header and leave csum up to XDP layers */
482 memset(hdr, 0, vi->hdr_len);
483 xdpf->len += vi->hdr_len;
485 sg_init_one(sq->sg, xdpf->data, xdpf->len);
487 err = virtqueue_add_outbuf(sq->vq, sq->sg, 1, xdp_to_ptr(xdpf),
490 return -ENOSPC; /* Caller handle free/refcnt */
495 static struct send_queue *virtnet_xdp_sq(struct virtnet_info *vi)
499 qp = vi->curr_queue_pairs - vi->xdp_queue_pairs + smp_processor_id();
503 static int virtnet_xdp_xmit(struct net_device *dev,
504 int n, struct xdp_frame **frames, u32 flags)
506 struct virtnet_info *vi = netdev_priv(dev);
507 struct receive_queue *rq = vi->rq;
508 struct bpf_prog *xdp_prog;
509 struct send_queue *sq;
519 /* Only allow ndo_xdp_xmit if XDP is loaded on dev, as this
520 * indicate XDP resources have been successfully allocated.
522 xdp_prog = rcu_dereference(rq->xdp_prog);
526 sq = virtnet_xdp_sq(vi);
528 if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK)) {
534 /* Free up any pending old buffers before queueing new ones. */
535 while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
536 if (likely(is_xdp_frame(ptr))) {
537 struct xdp_frame *frame = ptr_to_xdp(ptr);
540 xdp_return_frame(frame);
542 struct sk_buff *skb = ptr;
545 napi_consume_skb(skb, false);
550 for (i = 0; i < n; i++) {
551 struct xdp_frame *xdpf = frames[i];
553 err = __virtnet_xdp_xmit_one(vi, sq, xdpf);
555 xdp_return_frame_rx_napi(xdpf);
561 if (flags & XDP_XMIT_FLUSH) {
562 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq))
566 u64_stats_update_begin(&sq->stats.syncp);
567 sq->stats.bytes += bytes;
568 sq->stats.packets += packets;
569 sq->stats.xdp_tx += n;
570 sq->stats.xdp_tx_drops += drops;
571 sq->stats.kicks += kicks;
572 u64_stats_update_end(&sq->stats.syncp);
577 static unsigned int virtnet_get_headroom(struct virtnet_info *vi)
579 return vi->xdp_queue_pairs ? VIRTIO_XDP_HEADROOM : 0;
582 /* We copy the packet for XDP in the following cases:
584 * 1) Packet is scattered across multiple rx buffers.
585 * 2) Headroom space is insufficient.
587 * This is inefficient but it's a temporary condition that
588 * we hit right after XDP is enabled and until queue is refilled
589 * with large buffers with sufficient headroom - so it should affect
590 * at most queue size packets.
591 * Afterwards, the conditions to enable
592 * XDP should preclude the underlying device from sending packets
593 * across multiple buffers (num_buf > 1), and we make sure buffers
594 * have enough headroom.
596 static struct page *xdp_linearize_page(struct receive_queue *rq,
603 int tailroom = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
606 if (page_off + *len + tailroom > PAGE_SIZE)
609 page = alloc_page(GFP_ATOMIC);
613 memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
621 buf = virtqueue_get_buf(rq->vq, &buflen);
625 p = virt_to_head_page(buf);
626 off = buf - page_address(p);
628 /* guard against a misconfigured or uncooperative backend that
629 * is sending packet larger than the MTU.
631 if ((page_off + buflen + tailroom) > PAGE_SIZE) {
636 memcpy(page_address(page) + page_off,
637 page_address(p) + off, buflen);
642 /* Headroom does not contribute to packet length */
643 *len = page_off - VIRTIO_XDP_HEADROOM;
646 __free_pages(page, 0);
650 static struct sk_buff *receive_small(struct net_device *dev,
651 struct virtnet_info *vi,
652 struct receive_queue *rq,
653 void *buf, void *ctx,
655 unsigned int *xdp_xmit,
656 struct virtnet_rq_stats *stats)
659 struct bpf_prog *xdp_prog;
660 unsigned int xdp_headroom = (unsigned long)ctx;
661 unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom;
662 unsigned int headroom = vi->hdr_len + header_offset;
663 unsigned int buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
664 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
665 struct page *page = virt_to_head_page(buf);
666 unsigned int delta = 0;
667 struct page *xdp_page;
669 unsigned int metasize = 0;
675 xdp_prog = rcu_dereference(rq->xdp_prog);
677 struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset;
678 struct xdp_frame *xdpf;
683 if (unlikely(hdr->hdr.gso_type))
686 if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) {
687 int offset = buf - page_address(page) + header_offset;
688 unsigned int tlen = len + vi->hdr_len;
691 xdp_headroom = virtnet_get_headroom(vi);
692 header_offset = VIRTNET_RX_PAD + xdp_headroom;
693 headroom = vi->hdr_len + header_offset;
694 buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
695 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
696 xdp_page = xdp_linearize_page(rq, &num_buf, page,
697 offset, header_offset,
702 buf = page_address(xdp_page);
707 xdp.data_hard_start = buf + VIRTNET_RX_PAD + vi->hdr_len;
708 xdp.data = xdp.data_hard_start + xdp_headroom;
709 xdp.data_end = xdp.data + len;
710 xdp.data_meta = xdp.data;
711 xdp.rxq = &rq->xdp_rxq;
712 orig_data = xdp.data;
713 act = bpf_prog_run_xdp(xdp_prog, &xdp);
714 stats->xdp_packets++;
718 /* Recalculate length in case bpf program changed it */
719 delta = orig_data - xdp.data;
720 len = xdp.data_end - xdp.data;
721 metasize = xdp.data - xdp.data_meta;
725 xdpf = convert_to_xdp_frame(&xdp);
728 err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
729 if (unlikely(err < 0)) {
730 trace_xdp_exception(vi->dev, xdp_prog, act);
733 *xdp_xmit |= VIRTIO_XDP_TX;
737 stats->xdp_redirects++;
738 err = xdp_do_redirect(dev, &xdp, xdp_prog);
741 *xdp_xmit |= VIRTIO_XDP_REDIR;
745 bpf_warn_invalid_xdp_action(act);
748 trace_xdp_exception(vi->dev, xdp_prog, act);
755 skb = build_skb(buf, buflen);
760 skb_reserve(skb, headroom - delta);
763 buf += header_offset;
764 memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
765 } /* keep zeroed vnet hdr since packet was changed by bpf */
768 skb_metadata_set(skb, metasize);
782 static struct sk_buff *receive_big(struct net_device *dev,
783 struct virtnet_info *vi,
784 struct receive_queue *rq,
787 struct virtnet_rq_stats *stats)
789 struct page *page = buf;
790 struct sk_buff *skb =
791 page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, true, 0);
793 stats->bytes += len - vi->hdr_len;
801 give_pages(rq, page);
805 static struct sk_buff *receive_mergeable(struct net_device *dev,
806 struct virtnet_info *vi,
807 struct receive_queue *rq,
811 unsigned int *xdp_xmit,
812 struct virtnet_rq_stats *stats)
814 struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
815 u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
816 struct page *page = virt_to_head_page(buf);
817 int offset = buf - page_address(page);
818 struct sk_buff *head_skb, *curr_skb;
819 struct bpf_prog *xdp_prog;
820 unsigned int truesize;
821 unsigned int headroom = mergeable_ctx_to_headroom(ctx);
823 unsigned int metasize = 0;
826 stats->bytes += len - vi->hdr_len;
829 xdp_prog = rcu_dereference(rq->xdp_prog);
831 struct xdp_frame *xdpf;
832 struct page *xdp_page;
837 /* Transient failure which in theory could occur if
838 * in-flight packets from before XDP was enabled reach
839 * the receive path after XDP is loaded.
841 if (unlikely(hdr->hdr.gso_type))
844 /* This happens when rx buffer size is underestimated
845 * or headroom is not enough because of the buffer
846 * was refilled before XDP is set. This should only
847 * happen for the first several packets, so we don't
848 * care much about its performance.
850 if (unlikely(num_buf > 1 ||
851 headroom < virtnet_get_headroom(vi))) {
852 /* linearize data for XDP */
853 xdp_page = xdp_linearize_page(rq, &num_buf,
859 offset = VIRTIO_XDP_HEADROOM;
864 /* Allow consuming headroom but reserve enough space to push
865 * the descriptor on if we get an XDP_TX return code.
867 data = page_address(xdp_page) + offset;
868 xdp.data_hard_start = data - VIRTIO_XDP_HEADROOM + vi->hdr_len;
869 xdp.data = data + vi->hdr_len;
870 xdp.data_end = xdp.data + (len - vi->hdr_len);
871 xdp.data_meta = xdp.data;
872 xdp.rxq = &rq->xdp_rxq;
874 act = bpf_prog_run_xdp(xdp_prog, &xdp);
875 stats->xdp_packets++;
879 metasize = xdp.data - xdp.data_meta;
881 /* recalculate offset to account for any header
882 * adjustments and minus the metasize to copy the
883 * metadata in page_to_skb(). Note other cases do not
884 * build an skb and avoid using offset
886 offset = xdp.data - page_address(xdp_page) -
887 vi->hdr_len - metasize;
889 /* recalculate len if xdp.data, xdp.data_end or
890 * xdp.data_meta were adjusted
892 len = xdp.data_end - xdp.data + vi->hdr_len + metasize;
893 /* We can only create skb based on xdp_page. */
894 if (unlikely(xdp_page != page)) {
897 head_skb = page_to_skb(vi, rq, xdp_page, offset,
898 len, PAGE_SIZE, false,
905 xdpf = convert_to_xdp_frame(&xdp);
908 err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
909 if (unlikely(err < 0)) {
910 trace_xdp_exception(vi->dev, xdp_prog, act);
911 if (unlikely(xdp_page != page))
915 *xdp_xmit |= VIRTIO_XDP_TX;
916 if (unlikely(xdp_page != page))
921 stats->xdp_redirects++;
922 err = xdp_do_redirect(dev, &xdp, xdp_prog);
924 if (unlikely(xdp_page != page))
928 *xdp_xmit |= VIRTIO_XDP_REDIR;
929 if (unlikely(xdp_page != page))
934 bpf_warn_invalid_xdp_action(act);
937 trace_xdp_exception(vi->dev, xdp_prog, act);
940 if (unlikely(xdp_page != page))
941 __free_pages(xdp_page, 0);
947 truesize = mergeable_ctx_to_truesize(ctx);
948 if (unlikely(len > truesize)) {
949 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
950 dev->name, len, (unsigned long)ctx);
951 dev->stats.rx_length_errors++;
955 head_skb = page_to_skb(vi, rq, page, offset, len, truesize, !xdp_prog,
959 if (unlikely(!curr_skb))
964 buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx);
965 if (unlikely(!buf)) {
966 pr_debug("%s: rx error: %d buffers out of %d missing\n",
968 virtio16_to_cpu(vi->vdev,
970 dev->stats.rx_length_errors++;
975 page = virt_to_head_page(buf);
977 truesize = mergeable_ctx_to_truesize(ctx);
978 if (unlikely(len > truesize)) {
979 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
980 dev->name, len, (unsigned long)ctx);
981 dev->stats.rx_length_errors++;
985 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
986 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
987 struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
991 if (curr_skb == head_skb)
992 skb_shinfo(curr_skb)->frag_list = nskb;
994 curr_skb->next = nskb;
996 head_skb->truesize += nskb->truesize;
999 if (curr_skb != head_skb) {
1000 head_skb->data_len += len;
1001 head_skb->len += len;
1002 head_skb->truesize += truesize;
1004 offset = buf - page_address(page);
1005 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
1007 skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
1010 skb_add_rx_frag(curr_skb, num_skb_frags, page,
1011 offset, len, truesize);
1015 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
1023 while (num_buf-- > 1) {
1024 buf = virtqueue_get_buf(rq->vq, &len);
1025 if (unlikely(!buf)) {
1026 pr_debug("%s: rx error: %d buffers missing\n",
1027 dev->name, num_buf);
1028 dev->stats.rx_length_errors++;
1031 stats->bytes += len;
1032 page = virt_to_head_page(buf);
1037 dev_kfree_skb(head_skb);
1042 static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
1043 void *buf, unsigned int len, void **ctx,
1044 unsigned int *xdp_xmit,
1045 struct virtnet_rq_stats *stats)
1047 struct net_device *dev = vi->dev;
1048 struct sk_buff *skb;
1049 struct virtio_net_hdr_mrg_rxbuf *hdr;
1051 if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
1052 pr_debug("%s: short packet %i\n", dev->name, len);
1053 dev->stats.rx_length_errors++;
1054 if (vi->mergeable_rx_bufs) {
1055 put_page(virt_to_head_page(buf));
1056 } else if (vi->big_packets) {
1057 give_pages(rq, buf);
1059 put_page(virt_to_head_page(buf));
1064 if (vi->mergeable_rx_bufs)
1065 skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit,
1067 else if (vi->big_packets)
1068 skb = receive_big(dev, vi, rq, buf, len, stats);
1070 skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit, stats);
1075 hdr = skb_vnet_hdr(skb);
1077 if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
1078 skb->ip_summed = CHECKSUM_UNNECESSARY;
1080 if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
1081 virtio_is_little_endian(vi->vdev))) {
1082 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
1083 dev->name, hdr->hdr.gso_type,
1088 skb->protocol = eth_type_trans(skb, dev);
1089 pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
1090 ntohs(skb->protocol), skb->len, skb->pkt_type);
1092 napi_gro_receive(&rq->napi, skb);
1096 dev->stats.rx_frame_errors++;
1100 /* Unlike mergeable buffers, all buffers are allocated to the
1101 * same size, except for the headroom. For this reason we do
1102 * not need to use mergeable_len_to_ctx here - it is enough
1103 * to store the headroom as the context ignoring the truesize.
1105 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
1108 struct page_frag *alloc_frag = &rq->alloc_frag;
1110 unsigned int xdp_headroom = virtnet_get_headroom(vi);
1111 void *ctx = (void *)(unsigned long)xdp_headroom;
1112 int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
1115 len = SKB_DATA_ALIGN(len) +
1116 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1117 if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
1120 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1121 get_page(alloc_frag->page);
1122 alloc_frag->offset += len;
1123 sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom,
1124 vi->hdr_len + GOOD_PACKET_LEN);
1125 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1127 put_page(virt_to_head_page(buf));
1131 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
1134 struct page *first, *list = NULL;
1138 sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
1140 /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
1141 for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
1142 first = get_a_page(rq, gfp);
1145 give_pages(rq, list);
1148 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
1150 /* chain new page in list head to match sg */
1151 first->private = (unsigned long)list;
1155 first = get_a_page(rq, gfp);
1157 give_pages(rq, list);
1160 p = page_address(first);
1162 /* rq->sg[0], rq->sg[1] share the same page */
1163 /* a separated rq->sg[0] for header - required in case !any_header_sg */
1164 sg_set_buf(&rq->sg[0], p, vi->hdr_len);
1166 /* rq->sg[1] for data packet, from offset */
1167 offset = sizeof(struct padded_vnet_hdr);
1168 sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
1170 /* chain first in list head */
1171 first->private = (unsigned long)list;
1172 err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
1175 give_pages(rq, first);
1180 static unsigned int get_mergeable_buf_len(struct receive_queue *rq,
1181 struct ewma_pkt_len *avg_pkt_len,
1184 const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
1188 return PAGE_SIZE - room;
1190 len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
1191 rq->min_buf_len, PAGE_SIZE - hdr_len);
1193 return ALIGN(len, L1_CACHE_BYTES);
1196 static int add_recvbuf_mergeable(struct virtnet_info *vi,
1197 struct receive_queue *rq, gfp_t gfp)
1199 struct page_frag *alloc_frag = &rq->alloc_frag;
1200 unsigned int headroom = virtnet_get_headroom(vi);
1201 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1202 unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
1206 unsigned int len, hole;
1208 /* Extra tailroom is needed to satisfy XDP's assumption. This
1209 * means rx frags coalescing won't work, but consider we've
1210 * disabled GSO for XDP, it won't be a big issue.
1212 len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len, room);
1213 if (unlikely(!skb_page_frag_refill(len + room, alloc_frag, gfp)))
1216 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1217 buf += headroom; /* advance address leaving hole at front of pkt */
1218 get_page(alloc_frag->page);
1219 alloc_frag->offset += len + room;
1220 hole = alloc_frag->size - alloc_frag->offset;
1221 if (hole < len + room) {
1222 /* To avoid internal fragmentation, if there is very likely not
1223 * enough space for another buffer, add the remaining space to
1224 * the current buffer.
1227 alloc_frag->offset += hole;
1230 sg_init_one(rq->sg, buf, len);
1231 ctx = mergeable_len_to_ctx(len, headroom);
1232 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1234 put_page(virt_to_head_page(buf));
1240 * Returns false if we couldn't fill entirely (OOM).
1242 * Normally run in the receive path, but can also be run from ndo_open
1243 * before we're receiving packets, or from refill_work which is
1244 * careful to disable receiving (using napi_disable).
1246 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
1253 if (vi->mergeable_rx_bufs)
1254 err = add_recvbuf_mergeable(vi, rq, gfp);
1255 else if (vi->big_packets)
1256 err = add_recvbuf_big(vi, rq, gfp);
1258 err = add_recvbuf_small(vi, rq, gfp);
1260 oom = err == -ENOMEM;
1263 } while (rq->vq->num_free);
1264 if (virtqueue_kick_prepare(rq->vq) && virtqueue_notify(rq->vq)) {
1265 unsigned long flags;
1267 flags = u64_stats_update_begin_irqsave(&rq->stats.syncp);
1269 u64_stats_update_end_irqrestore(&rq->stats.syncp, flags);
1275 static void skb_recv_done(struct virtqueue *rvq)
1277 struct virtnet_info *vi = rvq->vdev->priv;
1278 struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
1280 virtqueue_napi_schedule(&rq->napi, rvq);
1283 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
1287 /* If all buffers were filled by other side before we napi_enabled, we
1288 * won't get another interrupt, so process any outstanding packets now.
1289 * Call local_bh_enable after to trigger softIRQ processing.
1292 virtqueue_napi_schedule(napi, vq);
1296 static void virtnet_napi_tx_enable(struct virtnet_info *vi,
1297 struct virtqueue *vq,
1298 struct napi_struct *napi)
1303 /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
1304 * enable the feature if this is likely affine with the transmit path.
1306 if (!vi->affinity_hint_set) {
1311 return virtnet_napi_enable(vq, napi);
1314 static void virtnet_napi_tx_disable(struct napi_struct *napi)
1320 static void refill_work(struct work_struct *work)
1322 struct virtnet_info *vi =
1323 container_of(work, struct virtnet_info, refill.work);
1327 for (i = 0; i < vi->curr_queue_pairs; i++) {
1328 struct receive_queue *rq = &vi->rq[i];
1330 napi_disable(&rq->napi);
1331 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
1332 virtnet_napi_enable(rq->vq, &rq->napi);
1334 /* In theory, this can happen: if we don't get any buffers in
1335 * we will *never* try to fill again.
1338 schedule_delayed_work(&vi->refill, HZ/2);
1342 static int virtnet_receive(struct receive_queue *rq, int budget,
1343 unsigned int *xdp_xmit)
1345 struct virtnet_info *vi = rq->vq->vdev->priv;
1346 struct virtnet_rq_stats stats = {};
1351 if (!vi->big_packets || vi->mergeable_rx_bufs) {
1354 while (stats.packets < budget &&
1355 (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) {
1356 receive_buf(vi, rq, buf, len, ctx, xdp_xmit, &stats);
1360 while (stats.packets < budget &&
1361 (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
1362 receive_buf(vi, rq, buf, len, NULL, xdp_xmit, &stats);
1367 if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) {
1368 if (!try_fill_recv(vi, rq, GFP_ATOMIC))
1369 schedule_delayed_work(&vi->refill, 0);
1372 u64_stats_update_begin(&rq->stats.syncp);
1373 for (i = 0; i < VIRTNET_RQ_STATS_LEN; i++) {
1374 size_t offset = virtnet_rq_stats_desc[i].offset;
1377 item = (u64 *)((u8 *)&rq->stats + offset);
1378 *item += *(u64 *)((u8 *)&stats + offset);
1380 u64_stats_update_end(&rq->stats.syncp);
1382 return stats.packets;
1385 static void free_old_xmit_skbs(struct send_queue *sq, bool in_napi)
1388 unsigned int packets = 0;
1389 unsigned int bytes = 0;
1392 while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
1393 if (likely(!is_xdp_frame(ptr))) {
1394 struct sk_buff *skb = ptr;
1396 pr_debug("Sent skb %p\n", skb);
1399 napi_consume_skb(skb, in_napi);
1401 struct xdp_frame *frame = ptr_to_xdp(ptr);
1403 bytes += frame->len;
1404 xdp_return_frame(frame);
1409 /* Avoid overhead when no packets have been processed
1410 * happens when called speculatively from start_xmit.
1415 u64_stats_update_begin(&sq->stats.syncp);
1416 sq->stats.bytes += bytes;
1417 sq->stats.packets += packets;
1418 u64_stats_update_end(&sq->stats.syncp);
1421 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
1423 if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
1425 else if (q < vi->curr_queue_pairs)
1431 static void virtnet_poll_cleantx(struct receive_queue *rq)
1433 struct virtnet_info *vi = rq->vq->vdev->priv;
1434 unsigned int index = vq2rxq(rq->vq);
1435 struct send_queue *sq = &vi->sq[index];
1436 struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index);
1438 if (!sq->napi.weight || is_xdp_raw_buffer_queue(vi, index))
1441 if (__netif_tx_trylock(txq)) {
1442 free_old_xmit_skbs(sq, true);
1443 __netif_tx_unlock(txq);
1446 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1447 netif_tx_wake_queue(txq);
1450 static int virtnet_poll(struct napi_struct *napi, int budget)
1452 struct receive_queue *rq =
1453 container_of(napi, struct receive_queue, napi);
1454 struct virtnet_info *vi = rq->vq->vdev->priv;
1455 struct send_queue *sq;
1456 unsigned int received;
1457 unsigned int xdp_xmit = 0;
1459 virtnet_poll_cleantx(rq);
1461 received = virtnet_receive(rq, budget, &xdp_xmit);
1463 /* Out of packets? */
1464 if (received < budget)
1465 virtqueue_napi_complete(napi, rq->vq, received);
1467 if (xdp_xmit & VIRTIO_XDP_REDIR)
1470 if (xdp_xmit & VIRTIO_XDP_TX) {
1471 sq = virtnet_xdp_sq(vi);
1472 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1473 u64_stats_update_begin(&sq->stats.syncp);
1475 u64_stats_update_end(&sq->stats.syncp);
1482 static int virtnet_open(struct net_device *dev)
1484 struct virtnet_info *vi = netdev_priv(dev);
1487 for (i = 0; i < vi->max_queue_pairs; i++) {
1488 if (i < vi->curr_queue_pairs)
1489 /* Make sure we have some buffers: if oom use wq. */
1490 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1491 schedule_delayed_work(&vi->refill, 0);
1493 err = xdp_rxq_info_reg(&vi->rq[i].xdp_rxq, dev, i);
1497 err = xdp_rxq_info_reg_mem_model(&vi->rq[i].xdp_rxq,
1498 MEM_TYPE_PAGE_SHARED, NULL);
1500 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1504 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1505 virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi);
1511 static int virtnet_poll_tx(struct napi_struct *napi, int budget)
1513 struct send_queue *sq = container_of(napi, struct send_queue, napi);
1514 struct virtnet_info *vi = sq->vq->vdev->priv;
1515 unsigned int index = vq2txq(sq->vq);
1516 struct netdev_queue *txq;
1520 if (unlikely(is_xdp_raw_buffer_queue(vi, index))) {
1521 /* We don't need to enable cb for XDP */
1522 napi_complete_done(napi, 0);
1526 txq = netdev_get_tx_queue(vi->dev, index);
1527 __netif_tx_lock(txq, raw_smp_processor_id());
1528 virtqueue_disable_cb(sq->vq);
1529 free_old_xmit_skbs(sq, true);
1531 opaque = virtqueue_enable_cb_prepare(sq->vq);
1533 done = napi_complete_done(napi, 0);
1536 virtqueue_disable_cb(sq->vq);
1538 __netif_tx_unlock(txq);
1541 if (unlikely(virtqueue_poll(sq->vq, opaque))) {
1542 if (napi_schedule_prep(napi)) {
1543 __netif_tx_lock(txq, raw_smp_processor_id());
1544 virtqueue_disable_cb(sq->vq);
1545 __netif_tx_unlock(txq);
1546 __napi_schedule(napi);
1551 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1552 netif_tx_wake_queue(txq);
1557 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
1559 struct virtio_net_hdr_mrg_rxbuf *hdr;
1560 const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
1561 struct virtnet_info *vi = sq->vq->vdev->priv;
1563 unsigned hdr_len = vi->hdr_len;
1566 pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
1568 can_push = vi->any_header_sg &&
1569 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
1570 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
1571 /* Even if we can, don't push here yet as this would skew
1572 * csum_start offset below. */
1574 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1576 hdr = skb_vnet_hdr(skb);
1578 if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1579 virtio_is_little_endian(vi->vdev), false,
1583 if (vi->mergeable_rx_bufs)
1584 hdr->num_buffers = 0;
1586 sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1588 __skb_push(skb, hdr_len);
1589 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1590 if (unlikely(num_sg < 0))
1592 /* Pull header back to avoid skew in tx bytes calculations. */
1593 __skb_pull(skb, hdr_len);
1595 sg_set_buf(sq->sg, hdr, hdr_len);
1596 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
1597 if (unlikely(num_sg < 0))
1601 return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
1604 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
1606 struct virtnet_info *vi = netdev_priv(dev);
1607 int qnum = skb_get_queue_mapping(skb);
1608 struct send_queue *sq = &vi->sq[qnum];
1610 struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1611 bool kick = !skb->xmit_more;
1612 bool use_napi = sq->napi.weight;
1614 /* Free up any pending old buffers before queueing new ones. */
1615 free_old_xmit_skbs(sq, false);
1617 if (use_napi && kick)
1618 virtqueue_enable_cb_delayed(sq->vq);
1620 /* timestamp packet in software */
1621 skb_tx_timestamp(skb);
1623 /* Try to transmit */
1624 err = xmit_skb(sq, skb);
1626 /* This should not happen! */
1627 if (unlikely(err)) {
1628 dev->stats.tx_fifo_errors++;
1629 if (net_ratelimit())
1631 "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
1632 dev->stats.tx_dropped++;
1633 dev_kfree_skb_any(skb);
1634 return NETDEV_TX_OK;
1637 /* Don't wait up for transmitted skbs to be freed. */
1643 /* If running out of space, stop queue to avoid getting packets that we
1644 * are then unable to transmit.
1645 * An alternative would be to force queuing layer to requeue the skb by
1646 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1647 * returned in a normal path of operation: it means that driver is not
1648 * maintaining the TX queue stop/start state properly, and causes
1649 * the stack to do a non-trivial amount of useless work.
1650 * Since most packets only take 1 or 2 ring slots, stopping the queue
1651 * early means 16 slots are typically wasted.
1653 if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
1654 netif_stop_subqueue(dev, qnum);
1656 unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
1657 /* More just got used, free them then recheck. */
1658 free_old_xmit_skbs(sq, false);
1659 if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
1660 netif_start_subqueue(dev, qnum);
1661 virtqueue_disable_cb(sq->vq);
1666 if (kick || netif_xmit_stopped(txq)) {
1667 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1668 u64_stats_update_begin(&sq->stats.syncp);
1670 u64_stats_update_end(&sq->stats.syncp);
1674 return NETDEV_TX_OK;
1678 * Send command via the control virtqueue and check status. Commands
1679 * supported by the hypervisor, as indicated by feature bits, should
1680 * never fail unless improperly formatted.
1682 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
1683 struct scatterlist *out)
1685 struct scatterlist *sgs[4], hdr, stat;
1686 unsigned out_num = 0, tmp;
1688 /* Caller should know better */
1689 BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
1691 vi->ctrl->status = ~0;
1692 vi->ctrl->hdr.class = class;
1693 vi->ctrl->hdr.cmd = cmd;
1695 sg_init_one(&hdr, &vi->ctrl->hdr, sizeof(vi->ctrl->hdr));
1696 sgs[out_num++] = &hdr;
1699 sgs[out_num++] = out;
1701 /* Add return status. */
1702 sg_init_one(&stat, &vi->ctrl->status, sizeof(vi->ctrl->status));
1703 sgs[out_num] = &stat;
1705 BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1706 virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1708 if (unlikely(!virtqueue_kick(vi->cvq)))
1709 return vi->ctrl->status == VIRTIO_NET_OK;
1711 /* Spin for a response, the kick causes an ioport write, trapping
1712 * into the hypervisor, so the request should be handled immediately.
1714 while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1715 !virtqueue_is_broken(vi->cvq))
1718 return vi->ctrl->status == VIRTIO_NET_OK;
1721 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1723 struct virtnet_info *vi = netdev_priv(dev);
1724 struct virtio_device *vdev = vi->vdev;
1726 struct sockaddr *addr;
1727 struct scatterlist sg;
1729 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
1732 addr = kmemdup(p, sizeof(*addr), GFP_KERNEL);
1736 ret = eth_prepare_mac_addr_change(dev, addr);
1740 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1741 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1742 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1743 VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1744 dev_warn(&vdev->dev,
1745 "Failed to set mac address by vq command.\n");
1749 } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1750 !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1753 /* Naturally, this has an atomicity problem. */
1754 for (i = 0; i < dev->addr_len; i++)
1755 virtio_cwrite8(vdev,
1756 offsetof(struct virtio_net_config, mac) +
1757 i, addr->sa_data[i]);
1760 eth_commit_mac_addr_change(dev, p);
1768 static void virtnet_stats(struct net_device *dev,
1769 struct rtnl_link_stats64 *tot)
1771 struct virtnet_info *vi = netdev_priv(dev);
1775 for (i = 0; i < vi->max_queue_pairs; i++) {
1776 u64 tpackets, tbytes, rpackets, rbytes, rdrops;
1777 struct receive_queue *rq = &vi->rq[i];
1778 struct send_queue *sq = &vi->sq[i];
1781 start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
1782 tpackets = sq->stats.packets;
1783 tbytes = sq->stats.bytes;
1784 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
1787 start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
1788 rpackets = rq->stats.packets;
1789 rbytes = rq->stats.bytes;
1790 rdrops = rq->stats.drops;
1791 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
1793 tot->rx_packets += rpackets;
1794 tot->tx_packets += tpackets;
1795 tot->rx_bytes += rbytes;
1796 tot->tx_bytes += tbytes;
1797 tot->rx_dropped += rdrops;
1800 tot->tx_dropped = dev->stats.tx_dropped;
1801 tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1802 tot->rx_length_errors = dev->stats.rx_length_errors;
1803 tot->rx_frame_errors = dev->stats.rx_frame_errors;
1806 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1809 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1810 VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1811 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1815 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1817 struct scatterlist sg;
1818 struct net_device *dev = vi->dev;
1820 if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1823 vi->ctrl->mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1824 sg_init_one(&sg, &vi->ctrl->mq, sizeof(vi->ctrl->mq));
1826 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1827 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1828 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1832 vi->curr_queue_pairs = queue_pairs;
1833 /* virtnet_open() will refill when device is going to up. */
1834 if (dev->flags & IFF_UP)
1835 schedule_delayed_work(&vi->refill, 0);
1841 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1846 err = _virtnet_set_queues(vi, queue_pairs);
1851 static int virtnet_close(struct net_device *dev)
1853 struct virtnet_info *vi = netdev_priv(dev);
1856 /* Make sure refill_work doesn't re-enable napi! */
1857 cancel_delayed_work_sync(&vi->refill);
1859 for (i = 0; i < vi->max_queue_pairs; i++) {
1860 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1861 napi_disable(&vi->rq[i].napi);
1862 virtnet_napi_tx_disable(&vi->sq[i].napi);
1868 static void virtnet_set_rx_mode(struct net_device *dev)
1870 struct virtnet_info *vi = netdev_priv(dev);
1871 struct scatterlist sg[2];
1872 struct virtio_net_ctrl_mac *mac_data;
1873 struct netdev_hw_addr *ha;
1879 /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1880 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1883 vi->ctrl->promisc = ((dev->flags & IFF_PROMISC) != 0);
1884 vi->ctrl->allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1886 sg_init_one(sg, &vi->ctrl->promisc, sizeof(vi->ctrl->promisc));
1888 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1889 VIRTIO_NET_CTRL_RX_PROMISC, sg))
1890 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1891 vi->ctrl->promisc ? "en" : "dis");
1893 sg_init_one(sg, &vi->ctrl->allmulti, sizeof(vi->ctrl->allmulti));
1895 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1896 VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1897 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1898 vi->ctrl->allmulti ? "en" : "dis");
1900 uc_count = netdev_uc_count(dev);
1901 mc_count = netdev_mc_count(dev);
1902 /* MAC filter - use one buffer for both lists */
1903 buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1904 (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1909 sg_init_table(sg, 2);
1911 /* Store the unicast list and count in the front of the buffer */
1912 mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1914 netdev_for_each_uc_addr(ha, dev)
1915 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1917 sg_set_buf(&sg[0], mac_data,
1918 sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1920 /* multicast list and count fill the end */
1921 mac_data = (void *)&mac_data->macs[uc_count][0];
1923 mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1925 netdev_for_each_mc_addr(ha, dev)
1926 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1928 sg_set_buf(&sg[1], mac_data,
1929 sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1931 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1932 VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1933 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1938 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1939 __be16 proto, u16 vid)
1941 struct virtnet_info *vi = netdev_priv(dev);
1942 struct scatterlist sg;
1944 vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
1945 sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
1947 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1948 VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1949 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1953 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1954 __be16 proto, u16 vid)
1956 struct virtnet_info *vi = netdev_priv(dev);
1957 struct scatterlist sg;
1959 vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
1960 sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
1962 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1963 VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1964 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1968 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1972 if (vi->affinity_hint_set) {
1973 for (i = 0; i < vi->max_queue_pairs; i++) {
1974 virtqueue_set_affinity(vi->rq[i].vq, NULL);
1975 virtqueue_set_affinity(vi->sq[i].vq, NULL);
1978 vi->affinity_hint_set = false;
1982 static void virtnet_set_affinity(struct virtnet_info *vi)
1991 if (!zalloc_cpumask_var(&mask, GFP_KERNEL)) {
1992 virtnet_clean_affinity(vi, -1);
1996 num_cpu = num_online_cpus();
1997 stride = max_t(int, num_cpu / vi->curr_queue_pairs, 1);
1998 stragglers = num_cpu >= vi->curr_queue_pairs ?
1999 num_cpu % vi->curr_queue_pairs :
2001 cpu = cpumask_next(-1, cpu_online_mask);
2003 for (i = 0; i < vi->curr_queue_pairs; i++) {
2004 group_size = stride + (i < stragglers ? 1 : 0);
2006 for (j = 0; j < group_size; j++) {
2007 cpumask_set_cpu(cpu, mask);
2008 cpu = cpumask_next_wrap(cpu, cpu_online_mask,
2011 virtqueue_set_affinity(vi->rq[i].vq, mask);
2012 virtqueue_set_affinity(vi->sq[i].vq, mask);
2013 __netif_set_xps_queue(vi->dev, cpumask_bits(mask), i, false);
2014 cpumask_clear(mask);
2017 vi->affinity_hint_set = true;
2018 free_cpumask_var(mask);
2021 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
2023 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2025 virtnet_set_affinity(vi);
2029 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
2031 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2033 virtnet_set_affinity(vi);
2037 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
2039 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2042 virtnet_clean_affinity(vi, cpu);
2046 static enum cpuhp_state virtionet_online;
2048 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
2052 ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
2055 ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2059 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2063 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
2065 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2066 cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2070 static void virtnet_get_ringparam(struct net_device *dev,
2071 struct ethtool_ringparam *ring)
2073 struct virtnet_info *vi = netdev_priv(dev);
2075 ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
2076 ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
2077 ring->rx_pending = ring->rx_max_pending;
2078 ring->tx_pending = ring->tx_max_pending;
2082 static void virtnet_get_drvinfo(struct net_device *dev,
2083 struct ethtool_drvinfo *info)
2085 struct virtnet_info *vi = netdev_priv(dev);
2086 struct virtio_device *vdev = vi->vdev;
2088 strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
2089 strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
2090 strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
2094 /* TODO: Eliminate OOO packets during switching */
2095 static int virtnet_set_channels(struct net_device *dev,
2096 struct ethtool_channels *channels)
2098 struct virtnet_info *vi = netdev_priv(dev);
2099 u16 queue_pairs = channels->combined_count;
2102 /* We don't support separate rx/tx channels.
2103 * We don't allow setting 'other' channels.
2105 if (channels->rx_count || channels->tx_count || channels->other_count)
2108 if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
2111 /* For now we don't support modifying channels while XDP is loaded
2112 * also when XDP is loaded all RX queues have XDP programs so we only
2113 * need to check a single RX queue.
2115 if (vi->rq[0].xdp_prog)
2119 err = _virtnet_set_queues(vi, queue_pairs);
2124 virtnet_set_affinity(vi);
2127 netif_set_real_num_tx_queues(dev, queue_pairs);
2128 netif_set_real_num_rx_queues(dev, queue_pairs);
2133 static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *data)
2135 struct virtnet_info *vi = netdev_priv(dev);
2136 char *p = (char *)data;
2139 switch (stringset) {
2141 for (i = 0; i < vi->curr_queue_pairs; i++) {
2142 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
2143 snprintf(p, ETH_GSTRING_LEN, "rx_queue_%u_%s",
2144 i, virtnet_rq_stats_desc[j].desc);
2145 p += ETH_GSTRING_LEN;
2149 for (i = 0; i < vi->curr_queue_pairs; i++) {
2150 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
2151 snprintf(p, ETH_GSTRING_LEN, "tx_queue_%u_%s",
2152 i, virtnet_sq_stats_desc[j].desc);
2153 p += ETH_GSTRING_LEN;
2160 static int virtnet_get_sset_count(struct net_device *dev, int sset)
2162 struct virtnet_info *vi = netdev_priv(dev);
2166 return vi->curr_queue_pairs * (VIRTNET_RQ_STATS_LEN +
2167 VIRTNET_SQ_STATS_LEN);
2173 static void virtnet_get_ethtool_stats(struct net_device *dev,
2174 struct ethtool_stats *stats, u64 *data)
2176 struct virtnet_info *vi = netdev_priv(dev);
2177 unsigned int idx = 0, start, i, j;
2178 const u8 *stats_base;
2181 for (i = 0; i < vi->curr_queue_pairs; i++) {
2182 struct receive_queue *rq = &vi->rq[i];
2184 stats_base = (u8 *)&rq->stats;
2186 start = u64_stats_fetch_begin_irq(&rq->stats.syncp);
2187 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
2188 offset = virtnet_rq_stats_desc[j].offset;
2189 data[idx + j] = *(u64 *)(stats_base + offset);
2191 } while (u64_stats_fetch_retry_irq(&rq->stats.syncp, start));
2192 idx += VIRTNET_RQ_STATS_LEN;
2195 for (i = 0; i < vi->curr_queue_pairs; i++) {
2196 struct send_queue *sq = &vi->sq[i];
2198 stats_base = (u8 *)&sq->stats;
2200 start = u64_stats_fetch_begin_irq(&sq->stats.syncp);
2201 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
2202 offset = virtnet_sq_stats_desc[j].offset;
2203 data[idx + j] = *(u64 *)(stats_base + offset);
2205 } while (u64_stats_fetch_retry_irq(&sq->stats.syncp, start));
2206 idx += VIRTNET_SQ_STATS_LEN;
2210 static void virtnet_get_channels(struct net_device *dev,
2211 struct ethtool_channels *channels)
2213 struct virtnet_info *vi = netdev_priv(dev);
2215 channels->combined_count = vi->curr_queue_pairs;
2216 channels->max_combined = vi->max_queue_pairs;
2217 channels->max_other = 0;
2218 channels->rx_count = 0;
2219 channels->tx_count = 0;
2220 channels->other_count = 0;
2223 /* Check if the user is trying to change anything besides speed/duplex */
2225 virtnet_validate_ethtool_cmd(const struct ethtool_link_ksettings *cmd)
2227 struct ethtool_link_ksettings diff1 = *cmd;
2228 struct ethtool_link_ksettings diff2 = {};
2230 /* cmd is always set so we need to clear it, validate the port type
2231 * and also without autonegotiation we can ignore advertising
2233 diff1.base.speed = 0;
2234 diff2.base.port = PORT_OTHER;
2235 ethtool_link_ksettings_zero_link_mode(&diff1, advertising);
2236 diff1.base.duplex = 0;
2238 diff1.base.link_mode_masks_nwords = 0;
2240 return !memcmp(&diff1.base, &diff2.base, sizeof(diff1.base)) &&
2241 bitmap_empty(diff1.link_modes.supported,
2242 __ETHTOOL_LINK_MODE_MASK_NBITS) &&
2243 bitmap_empty(diff1.link_modes.advertising,
2244 __ETHTOOL_LINK_MODE_MASK_NBITS) &&
2245 bitmap_empty(diff1.link_modes.lp_advertising,
2246 __ETHTOOL_LINK_MODE_MASK_NBITS);
2249 static int virtnet_set_link_ksettings(struct net_device *dev,
2250 const struct ethtool_link_ksettings *cmd)
2252 struct virtnet_info *vi = netdev_priv(dev);
2255 speed = cmd->base.speed;
2256 /* don't allow custom speed and duplex */
2257 if (!ethtool_validate_speed(speed) ||
2258 !ethtool_validate_duplex(cmd->base.duplex) ||
2259 !virtnet_validate_ethtool_cmd(cmd))
2262 vi->duplex = cmd->base.duplex;
2267 static int virtnet_get_link_ksettings(struct net_device *dev,
2268 struct ethtool_link_ksettings *cmd)
2270 struct virtnet_info *vi = netdev_priv(dev);
2272 cmd->base.speed = vi->speed;
2273 cmd->base.duplex = vi->duplex;
2274 cmd->base.port = PORT_OTHER;
2279 static void virtnet_init_settings(struct net_device *dev)
2281 struct virtnet_info *vi = netdev_priv(dev);
2283 vi->speed = SPEED_UNKNOWN;
2284 vi->duplex = DUPLEX_UNKNOWN;
2287 static void virtnet_update_settings(struct virtnet_info *vi)
2292 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_SPEED_DUPLEX))
2295 speed = virtio_cread32(vi->vdev, offsetof(struct virtio_net_config,
2297 if (ethtool_validate_speed(speed))
2299 duplex = virtio_cread8(vi->vdev, offsetof(struct virtio_net_config,
2301 if (ethtool_validate_duplex(duplex))
2302 vi->duplex = duplex;
2305 static const struct ethtool_ops virtnet_ethtool_ops = {
2306 .get_drvinfo = virtnet_get_drvinfo,
2307 .get_link = ethtool_op_get_link,
2308 .get_ringparam = virtnet_get_ringparam,
2309 .get_strings = virtnet_get_strings,
2310 .get_sset_count = virtnet_get_sset_count,
2311 .get_ethtool_stats = virtnet_get_ethtool_stats,
2312 .set_channels = virtnet_set_channels,
2313 .get_channels = virtnet_get_channels,
2314 .get_ts_info = ethtool_op_get_ts_info,
2315 .get_link_ksettings = virtnet_get_link_ksettings,
2316 .set_link_ksettings = virtnet_set_link_ksettings,
2319 static void virtnet_freeze_down(struct virtio_device *vdev)
2321 struct virtnet_info *vi = vdev->priv;
2323 /* Make sure no work handler is accessing the device */
2324 flush_work(&vi->config_work);
2326 netif_tx_lock_bh(vi->dev);
2327 netif_device_detach(vi->dev);
2328 netif_tx_unlock_bh(vi->dev);
2329 if (netif_running(vi->dev))
2330 virtnet_close(vi->dev);
2333 static int init_vqs(struct virtnet_info *vi);
2335 static int virtnet_restore_up(struct virtio_device *vdev)
2337 struct virtnet_info *vi = vdev->priv;
2344 virtio_device_ready(vdev);
2346 if (netif_running(vi->dev)) {
2347 err = virtnet_open(vi->dev);
2352 netif_tx_lock_bh(vi->dev);
2353 netif_device_attach(vi->dev);
2354 netif_tx_unlock_bh(vi->dev);
2358 static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads)
2360 struct scatterlist sg;
2361 vi->ctrl->offloads = cpu_to_virtio64(vi->vdev, offloads);
2363 sg_init_one(&sg, &vi->ctrl->offloads, sizeof(vi->ctrl->offloads));
2365 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
2366 VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) {
2367 dev_warn(&vi->dev->dev, "Fail to set guest offload. \n");
2374 static int virtnet_clear_guest_offloads(struct virtnet_info *vi)
2378 if (!vi->guest_offloads)
2381 return virtnet_set_guest_offloads(vi, offloads);
2384 static int virtnet_restore_guest_offloads(struct virtnet_info *vi)
2386 u64 offloads = vi->guest_offloads;
2388 if (!vi->guest_offloads)
2391 return virtnet_set_guest_offloads(vi, offloads);
2394 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
2395 struct netlink_ext_ack *extack)
2397 unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr);
2398 struct virtnet_info *vi = netdev_priv(dev);
2399 struct bpf_prog *old_prog;
2400 u16 xdp_qp = 0, curr_qp;
2403 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)
2404 && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2405 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2406 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
2407 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) ||
2408 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))) {
2409 NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing LRO/CSUM, disable LRO/CSUM first");
2413 if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
2414 NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required");
2418 if (dev->mtu > max_sz) {
2419 NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP");
2420 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz);
2424 curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
2426 xdp_qp = nr_cpu_ids;
2428 /* XDP requires extra queues for XDP_TX */
2429 if (curr_qp + xdp_qp > vi->max_queue_pairs) {
2430 NL_SET_ERR_MSG_MOD(extack, "Too few free TX rings available");
2431 netdev_warn(dev, "request %i queues but max is %i\n",
2432 curr_qp + xdp_qp, vi->max_queue_pairs);
2436 old_prog = rtnl_dereference(vi->rq[0].xdp_prog);
2437 if (!prog && !old_prog)
2441 prog = bpf_prog_add(prog, vi->max_queue_pairs - 1);
2443 return PTR_ERR(prog);
2446 /* Make sure NAPI is not using any XDP TX queues for RX. */
2447 if (netif_running(dev)) {
2448 for (i = 0; i < vi->max_queue_pairs; i++) {
2449 napi_disable(&vi->rq[i].napi);
2450 virtnet_napi_tx_disable(&vi->sq[i].napi);
2455 for (i = 0; i < vi->max_queue_pairs; i++) {
2456 rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2458 virtnet_restore_guest_offloads(vi);
2463 err = _virtnet_set_queues(vi, curr_qp + xdp_qp);
2466 netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
2467 vi->xdp_queue_pairs = xdp_qp;
2470 for (i = 0; i < vi->max_queue_pairs; i++) {
2471 rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2472 if (i == 0 && !old_prog)
2473 virtnet_clear_guest_offloads(vi);
2477 for (i = 0; i < vi->max_queue_pairs; i++) {
2479 bpf_prog_put(old_prog);
2480 if (netif_running(dev)) {
2481 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2482 virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2491 virtnet_clear_guest_offloads(vi);
2492 for (i = 0; i < vi->max_queue_pairs; i++)
2493 rcu_assign_pointer(vi->rq[i].xdp_prog, old_prog);
2496 if (netif_running(dev)) {
2497 for (i = 0; i < vi->max_queue_pairs; i++) {
2498 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2499 virtnet_napi_tx_enable(vi, vi->sq[i].vq,
2504 bpf_prog_sub(prog, vi->max_queue_pairs - 1);
2508 static u32 virtnet_xdp_query(struct net_device *dev)
2510 struct virtnet_info *vi = netdev_priv(dev);
2511 const struct bpf_prog *xdp_prog;
2514 for (i = 0; i < vi->max_queue_pairs; i++) {
2515 xdp_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2517 return xdp_prog->aux->id;
2522 static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp)
2524 switch (xdp->command) {
2525 case XDP_SETUP_PROG:
2526 return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
2527 case XDP_QUERY_PROG:
2528 xdp->prog_id = virtnet_xdp_query(dev);
2535 static int virtnet_get_phys_port_name(struct net_device *dev, char *buf,
2538 struct virtnet_info *vi = netdev_priv(dev);
2541 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
2544 ret = snprintf(buf, len, "sby");
2551 static const struct net_device_ops virtnet_netdev = {
2552 .ndo_open = virtnet_open,
2553 .ndo_stop = virtnet_close,
2554 .ndo_start_xmit = start_xmit,
2555 .ndo_validate_addr = eth_validate_addr,
2556 .ndo_set_mac_address = virtnet_set_mac_address,
2557 .ndo_set_rx_mode = virtnet_set_rx_mode,
2558 .ndo_get_stats64 = virtnet_stats,
2559 .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
2560 .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
2561 .ndo_bpf = virtnet_xdp,
2562 .ndo_xdp_xmit = virtnet_xdp_xmit,
2563 .ndo_features_check = passthru_features_check,
2564 .ndo_get_phys_port_name = virtnet_get_phys_port_name,
2567 static void virtnet_config_changed_work(struct work_struct *work)
2569 struct virtnet_info *vi =
2570 container_of(work, struct virtnet_info, config_work);
2573 if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
2574 struct virtio_net_config, status, &v) < 0)
2577 if (v & VIRTIO_NET_S_ANNOUNCE) {
2578 netdev_notify_peers(vi->dev);
2579 virtnet_ack_link_announce(vi);
2582 /* Ignore unknown (future) status bits */
2583 v &= VIRTIO_NET_S_LINK_UP;
2585 if (vi->status == v)
2590 if (vi->status & VIRTIO_NET_S_LINK_UP) {
2591 virtnet_update_settings(vi);
2592 netif_carrier_on(vi->dev);
2593 netif_tx_wake_all_queues(vi->dev);
2595 netif_carrier_off(vi->dev);
2596 netif_tx_stop_all_queues(vi->dev);
2600 static void virtnet_config_changed(struct virtio_device *vdev)
2602 struct virtnet_info *vi = vdev->priv;
2604 schedule_work(&vi->config_work);
2607 static void virtnet_free_queues(struct virtnet_info *vi)
2611 for (i = 0; i < vi->max_queue_pairs; i++) {
2612 napi_hash_del(&vi->rq[i].napi);
2613 netif_napi_del(&vi->rq[i].napi);
2614 netif_napi_del(&vi->sq[i].napi);
2617 /* We called napi_hash_del() before netif_napi_del(),
2618 * we need to respect an RCU grace period before freeing vi->rq
2627 static void _free_receive_bufs(struct virtnet_info *vi)
2629 struct bpf_prog *old_prog;
2632 for (i = 0; i < vi->max_queue_pairs; i++) {
2633 while (vi->rq[i].pages)
2634 __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
2636 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2637 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
2639 bpf_prog_put(old_prog);
2643 static void free_receive_bufs(struct virtnet_info *vi)
2646 _free_receive_bufs(vi);
2650 static void free_receive_page_frags(struct virtnet_info *vi)
2653 for (i = 0; i < vi->max_queue_pairs; i++)
2654 if (vi->rq[i].alloc_frag.page)
2655 put_page(vi->rq[i].alloc_frag.page);
2658 static void virtnet_sq_free_unused_buf(struct virtqueue *vq, void *buf)
2660 if (!is_xdp_frame(buf))
2663 xdp_return_frame(ptr_to_xdp(buf));
2666 static void virtnet_rq_free_unused_buf(struct virtqueue *vq, void *buf)
2668 struct virtnet_info *vi = vq->vdev->priv;
2671 if (vi->mergeable_rx_bufs)
2672 put_page(virt_to_head_page(buf));
2673 else if (vi->big_packets)
2674 give_pages(&vi->rq[i], buf);
2676 put_page(virt_to_head_page(buf));
2679 static void free_unused_bufs(struct virtnet_info *vi)
2684 for (i = 0; i < vi->max_queue_pairs; i++) {
2685 struct virtqueue *vq = vi->sq[i].vq;
2686 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
2687 virtnet_sq_free_unused_buf(vq, buf);
2691 for (i = 0; i < vi->max_queue_pairs; i++) {
2692 struct virtqueue *vq = vi->rq[i].vq;
2693 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
2694 virtnet_rq_free_unused_buf(vq, buf);
2699 static void virtnet_del_vqs(struct virtnet_info *vi)
2701 struct virtio_device *vdev = vi->vdev;
2703 virtnet_clean_affinity(vi, -1);
2705 vdev->config->del_vqs(vdev);
2707 virtnet_free_queues(vi);
2710 /* How large should a single buffer be so a queue full of these can fit at
2711 * least one full packet?
2712 * Logic below assumes the mergeable buffer header is used.
2714 static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq)
2716 const unsigned int hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2717 unsigned int rq_size = virtqueue_get_vring_size(vq);
2718 unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu;
2719 unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len;
2720 unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size);
2722 return max(max(min_buf_len, hdr_len) - hdr_len,
2723 (unsigned int)GOOD_PACKET_LEN);
2726 static int virtnet_find_vqs(struct virtnet_info *vi)
2728 vq_callback_t **callbacks;
2729 struct virtqueue **vqs;
2735 /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
2736 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
2737 * possible control vq.
2739 total_vqs = vi->max_queue_pairs * 2 +
2740 virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
2742 /* Allocate space for find_vqs parameters */
2743 vqs = kcalloc(total_vqs, sizeof(*vqs), GFP_KERNEL);
2746 callbacks = kmalloc_array(total_vqs, sizeof(*callbacks), GFP_KERNEL);
2749 names = kmalloc_array(total_vqs, sizeof(*names), GFP_KERNEL);
2752 if (!vi->big_packets || vi->mergeable_rx_bufs) {
2753 ctx = kcalloc(total_vqs, sizeof(*ctx), GFP_KERNEL);
2760 /* Parameters for control virtqueue, if any */
2762 callbacks[total_vqs - 1] = NULL;
2763 names[total_vqs - 1] = "control";
2766 /* Allocate/initialize parameters for send/receive virtqueues */
2767 for (i = 0; i < vi->max_queue_pairs; i++) {
2768 callbacks[rxq2vq(i)] = skb_recv_done;
2769 callbacks[txq2vq(i)] = skb_xmit_done;
2770 sprintf(vi->rq[i].name, "input.%d", i);
2771 sprintf(vi->sq[i].name, "output.%d", i);
2772 names[rxq2vq(i)] = vi->rq[i].name;
2773 names[txq2vq(i)] = vi->sq[i].name;
2775 ctx[rxq2vq(i)] = true;
2778 ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
2784 vi->cvq = vqs[total_vqs - 1];
2785 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
2786 vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
2789 for (i = 0; i < vi->max_queue_pairs; i++) {
2790 vi->rq[i].vq = vqs[rxq2vq(i)];
2791 vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq);
2792 vi->sq[i].vq = vqs[txq2vq(i)];
2795 /* run here: ret == 0. */
2810 static int virtnet_alloc_queues(struct virtnet_info *vi)
2814 vi->ctrl = kzalloc(sizeof(*vi->ctrl), GFP_KERNEL);
2817 vi->sq = kcalloc(vi->max_queue_pairs, sizeof(*vi->sq), GFP_KERNEL);
2820 vi->rq = kcalloc(vi->max_queue_pairs, sizeof(*vi->rq), GFP_KERNEL);
2824 INIT_DELAYED_WORK(&vi->refill, refill_work);
2825 for (i = 0; i < vi->max_queue_pairs; i++) {
2826 vi->rq[i].pages = NULL;
2827 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
2829 netif_tx_napi_add(vi->dev, &vi->sq[i].napi, virtnet_poll_tx,
2830 napi_tx ? napi_weight : 0);
2832 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
2833 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
2834 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
2836 u64_stats_init(&vi->rq[i].stats.syncp);
2837 u64_stats_init(&vi->sq[i].stats.syncp);
2850 static int init_vqs(struct virtnet_info *vi)
2854 /* Allocate send & receive queues */
2855 ret = virtnet_alloc_queues(vi);
2859 ret = virtnet_find_vqs(vi);
2864 virtnet_set_affinity(vi);
2870 virtnet_free_queues(vi);
2876 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
2879 struct virtnet_info *vi = netdev_priv(queue->dev);
2880 unsigned int queue_index = get_netdev_rx_queue_index(queue);
2881 unsigned int headroom = virtnet_get_headroom(vi);
2882 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
2883 struct ewma_pkt_len *avg;
2885 BUG_ON(queue_index >= vi->max_queue_pairs);
2886 avg = &vi->rq[queue_index].mrg_avg_pkt_len;
2887 return sprintf(buf, "%u\n",
2888 get_mergeable_buf_len(&vi->rq[queue_index], avg,
2889 SKB_DATA_ALIGN(headroom + tailroom)));
2892 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
2893 __ATTR_RO(mergeable_rx_buffer_size);
2895 static struct attribute *virtio_net_mrg_rx_attrs[] = {
2896 &mergeable_rx_buffer_size_attribute.attr,
2900 static const struct attribute_group virtio_net_mrg_rx_group = {
2901 .name = "virtio_net",
2902 .attrs = virtio_net_mrg_rx_attrs
2906 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
2908 const char *fname, const char *dname)
2910 if (!virtio_has_feature(vdev, fbit))
2913 dev_err(&vdev->dev, "device advertises feature %s but not %s",
2919 #define VIRTNET_FAIL_ON(vdev, fbit, dbit) \
2920 virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
2922 static bool virtnet_validate_features(struct virtio_device *vdev)
2924 if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
2925 (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
2926 "VIRTIO_NET_F_CTRL_VQ") ||
2927 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
2928 "VIRTIO_NET_F_CTRL_VQ") ||
2929 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
2930 "VIRTIO_NET_F_CTRL_VQ") ||
2931 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
2932 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
2933 "VIRTIO_NET_F_CTRL_VQ"))) {
2940 #define MIN_MTU ETH_MIN_MTU
2941 #define MAX_MTU ETH_MAX_MTU
2943 static int virtnet_validate(struct virtio_device *vdev)
2945 if (!vdev->config->get) {
2946 dev_err(&vdev->dev, "%s failure: config access disabled\n",
2951 if (!virtnet_validate_features(vdev))
2954 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2955 int mtu = virtio_cread16(vdev,
2956 offsetof(struct virtio_net_config,
2959 __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
2965 static int virtnet_probe(struct virtio_device *vdev)
2967 int i, err = -ENOMEM;
2968 struct net_device *dev;
2969 struct virtnet_info *vi;
2970 u16 max_queue_pairs;
2973 /* Find if host supports multiqueue virtio_net device */
2974 err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
2975 struct virtio_net_config,
2976 max_virtqueue_pairs, &max_queue_pairs);
2978 /* We need at least 2 queue's */
2979 if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
2980 max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
2981 !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2982 max_queue_pairs = 1;
2984 /* Allocate ourselves a network device with room for our info */
2985 dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
2989 /* Set up network device as normal. */
2990 dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
2991 dev->netdev_ops = &virtnet_netdev;
2992 dev->features = NETIF_F_HIGHDMA;
2994 dev->ethtool_ops = &virtnet_ethtool_ops;
2995 SET_NETDEV_DEV(dev, &vdev->dev);
2997 /* Do we support "hardware" checksums? */
2998 if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
2999 /* This opens up the world of extra features. */
3000 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
3002 dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
3004 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
3005 dev->hw_features |= NETIF_F_TSO
3006 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
3008 /* Individual feature bits: what can host handle? */
3009 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
3010 dev->hw_features |= NETIF_F_TSO;
3011 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
3012 dev->hw_features |= NETIF_F_TSO6;
3013 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
3014 dev->hw_features |= NETIF_F_TSO_ECN;
3016 dev->features |= NETIF_F_GSO_ROBUST;
3019 dev->features |= dev->hw_features & NETIF_F_ALL_TSO;
3020 /* (!csum && gso) case will be fixed by register_netdev() */
3022 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
3023 dev->features |= NETIF_F_RXCSUM;
3025 dev->vlan_features = dev->features;
3027 /* MTU range: 68 - 65535 */
3028 dev->min_mtu = MIN_MTU;
3029 dev->max_mtu = MAX_MTU;
3031 /* Configuration may specify what MAC to use. Otherwise random. */
3032 if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
3033 virtio_cread_bytes(vdev,
3034 offsetof(struct virtio_net_config, mac),
3035 dev->dev_addr, dev->addr_len);
3037 eth_hw_addr_random(dev);
3039 /* Set up our device-specific information */
3040 vi = netdev_priv(dev);
3045 INIT_WORK(&vi->config_work, virtnet_config_changed_work);
3047 /* If we can receive ANY GSO packets, we must allocate large ones. */
3048 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3049 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
3050 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
3051 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
3052 vi->big_packets = true;
3054 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
3055 vi->mergeable_rx_bufs = true;
3057 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
3058 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
3059 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
3061 vi->hdr_len = sizeof(struct virtio_net_hdr);
3063 if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
3064 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
3065 vi->any_header_sg = true;
3067 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
3070 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
3071 mtu = virtio_cread16(vdev,
3072 offsetof(struct virtio_net_config,
3074 if (mtu < dev->min_mtu) {
3075 /* Should never trigger: MTU was previously validated
3076 * in virtnet_validate.
3078 dev_err(&vdev->dev, "device MTU appears to have changed "
3079 "it is now %d < %d", mtu, dev->min_mtu);
3086 /* TODO: size buffers correctly in this case. */
3087 if (dev->mtu > ETH_DATA_LEN)
3088 vi->big_packets = true;
3091 if (vi->any_header_sg)
3092 dev->needed_headroom = vi->hdr_len;
3094 /* Enable multiqueue by default */
3095 if (num_online_cpus() >= max_queue_pairs)
3096 vi->curr_queue_pairs = max_queue_pairs;
3098 vi->curr_queue_pairs = num_online_cpus();
3099 vi->max_queue_pairs = max_queue_pairs;
3101 /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
3107 if (vi->mergeable_rx_bufs)
3108 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
3110 netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
3111 netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
3113 virtnet_init_settings(dev);
3115 if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY)) {
3116 vi->failover = net_failover_create(vi->dev);
3117 if (IS_ERR(vi->failover)) {
3118 err = PTR_ERR(vi->failover);
3123 /* serialize netdev register + virtio_device_ready() with ndo_open() */
3126 err = register_netdevice(dev);
3128 pr_debug("virtio_net: registering device failed\n");
3133 virtio_device_ready(vdev);
3135 _virtnet_set_queues(vi, vi->curr_queue_pairs);
3139 err = virtnet_cpu_notif_add(vi);
3141 pr_debug("virtio_net: registering cpu notifier failed\n");
3142 goto free_unregister_netdev;
3145 /* Assume link up if device can't report link status,
3146 otherwise get link status from config. */
3147 netif_carrier_off(dev);
3148 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
3149 schedule_work(&vi->config_work);
3151 vi->status = VIRTIO_NET_S_LINK_UP;
3152 virtnet_update_settings(vi);
3153 netif_carrier_on(dev);
3156 for (i = 0; i < ARRAY_SIZE(guest_offloads); i++)
3157 if (virtio_has_feature(vi->vdev, guest_offloads[i]))
3158 set_bit(guest_offloads[i], &vi->guest_offloads);
3160 pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
3161 dev->name, max_queue_pairs);
3165 free_unregister_netdev:
3166 vi->vdev->config->reset(vdev);
3168 unregister_netdev(dev);
3170 net_failover_destroy(vi->failover);
3172 cancel_delayed_work_sync(&vi->refill);
3173 free_receive_page_frags(vi);
3174 virtnet_del_vqs(vi);
3180 static void remove_vq_common(struct virtnet_info *vi)
3182 vi->vdev->config->reset(vi->vdev);
3184 /* Free unused buffers in both send and recv, if any. */
3185 free_unused_bufs(vi);
3187 free_receive_bufs(vi);
3189 free_receive_page_frags(vi);
3191 virtnet_del_vqs(vi);
3194 static void virtnet_remove(struct virtio_device *vdev)
3196 struct virtnet_info *vi = vdev->priv;
3198 virtnet_cpu_notif_remove(vi);
3200 /* Make sure no work handler is accessing the device. */
3201 flush_work(&vi->config_work);
3203 unregister_netdev(vi->dev);
3205 net_failover_destroy(vi->failover);
3207 remove_vq_common(vi);
3209 free_netdev(vi->dev);
3212 static __maybe_unused int virtnet_freeze(struct virtio_device *vdev)
3214 struct virtnet_info *vi = vdev->priv;
3216 virtnet_cpu_notif_remove(vi);
3217 virtnet_freeze_down(vdev);
3218 remove_vq_common(vi);
3223 static __maybe_unused int virtnet_restore(struct virtio_device *vdev)
3225 struct virtnet_info *vi = vdev->priv;
3228 err = virtnet_restore_up(vdev);
3231 virtnet_set_queues(vi, vi->curr_queue_pairs);
3233 err = virtnet_cpu_notif_add(vi);
3235 virtnet_freeze_down(vdev);
3236 remove_vq_common(vi);
3243 static struct virtio_device_id id_table[] = {
3244 { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
3248 #define VIRTNET_FEATURES \
3249 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
3251 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
3252 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
3253 VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
3254 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
3255 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
3256 VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
3257 VIRTIO_NET_F_CTRL_MAC_ADDR, \
3258 VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \
3259 VIRTIO_NET_F_SPEED_DUPLEX, VIRTIO_NET_F_STANDBY
3261 static unsigned int features[] = {
3265 static unsigned int features_legacy[] = {
3268 VIRTIO_F_ANY_LAYOUT,
3271 static struct virtio_driver virtio_net_driver = {
3272 .feature_table = features,
3273 .feature_table_size = ARRAY_SIZE(features),
3274 .feature_table_legacy = features_legacy,
3275 .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
3276 .driver.name = KBUILD_MODNAME,
3277 .driver.owner = THIS_MODULE,
3278 .id_table = id_table,
3279 .validate = virtnet_validate,
3280 .probe = virtnet_probe,
3281 .remove = virtnet_remove,
3282 .config_changed = virtnet_config_changed,
3283 #ifdef CONFIG_PM_SLEEP
3284 .freeze = virtnet_freeze,
3285 .restore = virtnet_restore,
3289 static __init int virtio_net_driver_init(void)
3293 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
3295 virtnet_cpu_down_prep);
3298 virtionet_online = ret;
3299 ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
3300 NULL, virtnet_cpu_dead);
3304 ret = register_virtio_driver(&virtio_net_driver);
3309 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3311 cpuhp_remove_multi_state(virtionet_online);
3315 module_init(virtio_net_driver_init);
3317 static __exit void virtio_net_driver_exit(void)
3319 unregister_virtio_driver(&virtio_net_driver);
3320 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
3321 cpuhp_remove_multi_state(virtionet_online);
3323 module_exit(virtio_net_driver_exit);
3325 MODULE_DEVICE_TABLE(virtio, id_table);
3326 MODULE_DESCRIPTION("Virtio network driver");
3327 MODULE_LICENSE("GPL");