1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* A network driver using virtio.
4 * Copyright 2007 Rusty Russell <rusty@rustcorp.com.au> IBM Corporation
7 #include <linux/netdevice.h>
8 #include <linux/etherdevice.h>
9 #include <linux/ethtool.h>
10 #include <linux/module.h>
11 #include <linux/virtio.h>
12 #include <linux/virtio_net.h>
13 #include <linux/bpf.h>
14 #include <linux/bpf_trace.h>
15 #include <linux/scatterlist.h>
16 #include <linux/if_vlan.h>
17 #include <linux/slab.h>
18 #include <linux/cpu.h>
19 #include <linux/average.h>
20 #include <linux/filter.h>
21 #include <linux/kernel.h>
22 #include <net/route.h>
24 #include <net/net_failover.h>
26 static int napi_weight = NAPI_POLL_WEIGHT;
27 module_param(napi_weight, int, 0444);
29 static bool csum = true, gso = true, napi_tx = true;
30 module_param(csum, bool, 0444);
31 module_param(gso, bool, 0444);
32 module_param(napi_tx, bool, 0644);
34 /* FIXME: MTU in config. */
35 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
36 #define GOOD_COPY_LEN 128
38 #define VIRTNET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
40 /* Amount of XDP headroom to prepend to packets for use by xdp_adjust_head */
41 #define VIRTIO_XDP_HEADROOM 256
43 /* Separating two types of XDP xmit */
44 #define VIRTIO_XDP_TX BIT(0)
45 #define VIRTIO_XDP_REDIR BIT(1)
47 #define VIRTIO_XDP_FLAG BIT(0)
49 /* RX packet size EWMA. The average packet size is used to determine the packet
50 * buffer size when refilling RX rings. As the entire RX ring may be refilled
51 * at once, the weight is chosen so that the EWMA will be insensitive to short-
52 * term, transient changes in packet size.
54 DECLARE_EWMA(pkt_len, 0, 64)
56 #define VIRTNET_DRIVER_VERSION "1.0.0"
58 static const unsigned long guest_offloads[] = {
59 VIRTIO_NET_F_GUEST_TSO4,
60 VIRTIO_NET_F_GUEST_TSO6,
61 VIRTIO_NET_F_GUEST_ECN,
62 VIRTIO_NET_F_GUEST_UFO,
63 VIRTIO_NET_F_GUEST_CSUM,
64 VIRTIO_NET_F_GUEST_USO4,
65 VIRTIO_NET_F_GUEST_USO6,
66 VIRTIO_NET_F_GUEST_HDRLEN
69 #define GUEST_OFFLOAD_GRO_HW_MASK ((1ULL << VIRTIO_NET_F_GUEST_TSO4) | \
70 (1ULL << VIRTIO_NET_F_GUEST_TSO6) | \
71 (1ULL << VIRTIO_NET_F_GUEST_ECN) | \
72 (1ULL << VIRTIO_NET_F_GUEST_UFO) | \
73 (1ULL << VIRTIO_NET_F_GUEST_USO4) | \
74 (1ULL << VIRTIO_NET_F_GUEST_USO6))
76 struct virtnet_stat_desc {
77 char desc[ETH_GSTRING_LEN];
81 struct virtnet_sq_stats {
82 struct u64_stats_sync syncp;
91 struct virtnet_rq_stats {
92 struct u64_stats_sync syncp;
103 #define VIRTNET_SQ_STAT(m) offsetof(struct virtnet_sq_stats, m)
104 #define VIRTNET_RQ_STAT(m) offsetof(struct virtnet_rq_stats, m)
106 static const struct virtnet_stat_desc virtnet_sq_stats_desc[] = {
107 { "packets", VIRTNET_SQ_STAT(packets) },
108 { "bytes", VIRTNET_SQ_STAT(bytes) },
109 { "xdp_tx", VIRTNET_SQ_STAT(xdp_tx) },
110 { "xdp_tx_drops", VIRTNET_SQ_STAT(xdp_tx_drops) },
111 { "kicks", VIRTNET_SQ_STAT(kicks) },
112 { "tx_timeouts", VIRTNET_SQ_STAT(tx_timeouts) },
115 static const struct virtnet_stat_desc virtnet_rq_stats_desc[] = {
116 { "packets", VIRTNET_RQ_STAT(packets) },
117 { "bytes", VIRTNET_RQ_STAT(bytes) },
118 { "drops", VIRTNET_RQ_STAT(drops) },
119 { "xdp_packets", VIRTNET_RQ_STAT(xdp_packets) },
120 { "xdp_tx", VIRTNET_RQ_STAT(xdp_tx) },
121 { "xdp_redirects", VIRTNET_RQ_STAT(xdp_redirects) },
122 { "xdp_drops", VIRTNET_RQ_STAT(xdp_drops) },
123 { "kicks", VIRTNET_RQ_STAT(kicks) },
126 #define VIRTNET_SQ_STATS_LEN ARRAY_SIZE(virtnet_sq_stats_desc)
127 #define VIRTNET_RQ_STATS_LEN ARRAY_SIZE(virtnet_rq_stats_desc)
129 /* Internal representation of a send virtqueue */
131 /* Virtqueue associated with this send _queue */
132 struct virtqueue *vq;
134 /* TX: fragments + linear part + virtio header */
135 struct scatterlist sg[MAX_SKB_FRAGS + 2];
137 /* Name of the send queue: output.$index */
140 struct virtnet_sq_stats stats;
142 struct napi_struct napi;
144 /* Record whether sq is in reset state. */
148 /* Internal representation of a receive virtqueue */
149 struct receive_queue {
150 /* Virtqueue associated with this receive_queue */
151 struct virtqueue *vq;
153 struct napi_struct napi;
155 struct bpf_prog __rcu *xdp_prog;
157 struct virtnet_rq_stats stats;
159 /* Chain pages by the private ptr. */
162 /* Average packet length for mergeable receive buffers. */
163 struct ewma_pkt_len mrg_avg_pkt_len;
165 /* Page frag for packet buffer allocation. */
166 struct page_frag alloc_frag;
168 /* RX: fragments + linear part + virtio header */
169 struct scatterlist sg[MAX_SKB_FRAGS + 2];
171 /* Min single buffer size for mergeable buffers case. */
172 unsigned int min_buf_len;
174 /* Name of this receive queue: input.$index */
177 struct xdp_rxq_info xdp_rxq;
180 /* This structure can contain rss message with maximum settings for indirection table and keysize
181 * Note, that default structure that describes RSS configuration virtio_net_rss_config
182 * contains same info but can't handle table values.
183 * In any case, structure would be passed to virtio hw through sg_buf split by parts
184 * because table sizes may be differ according to the device configuration.
186 #define VIRTIO_NET_RSS_MAX_KEY_SIZE 40
187 #define VIRTIO_NET_RSS_MAX_TABLE_LEN 128
188 struct virtio_net_ctrl_rss {
190 u16 indirection_table_mask;
191 u16 unclassified_queue;
192 u16 indirection_table[VIRTIO_NET_RSS_MAX_TABLE_LEN];
195 u8 key[VIRTIO_NET_RSS_MAX_KEY_SIZE];
198 /* Control VQ buffers: protected by the rtnl lock */
200 struct virtio_net_ctrl_hdr hdr;
201 virtio_net_ctrl_ack status;
202 struct virtio_net_ctrl_mq mq;
207 struct virtio_net_ctrl_rss rss;
210 struct virtnet_info {
211 struct virtio_device *vdev;
212 struct virtqueue *cvq;
213 struct net_device *dev;
214 struct send_queue *sq;
215 struct receive_queue *rq;
218 /* Max # of queue pairs supported by the device */
221 /* # of queue pairs currently used by the driver */
222 u16 curr_queue_pairs;
224 /* # of XDP queue pairs currently used by the driver */
227 /* xdp_queue_pairs may be 0, when xdp is already loaded. So add this. */
230 /* I like... big packets and I cannot lie! */
233 /* number of sg entries allocated for big packets */
234 unsigned int big_packets_num_skbfrags;
236 /* Host will merge rx buffers for big packets (shake it! shake it!) */
237 bool mergeable_rx_bufs;
239 /* Host supports rss and/or hash report */
241 bool has_rss_hash_report;
243 u16 rss_indir_table_size;
244 u32 rss_hash_types_supported;
245 u32 rss_hash_types_saved;
247 /* Has control virtqueue */
250 /* Host can handle any s/g split between our header and packet data */
253 /* Packet virtio header size */
256 /* Work struct for delayed refilling if we run low on memory. */
257 struct delayed_work refill;
259 /* Is delayed refill enabled? */
262 /* The lock to synchronize the access to refill_enabled */
263 spinlock_t refill_lock;
265 /* Work struct for config space updates */
266 struct work_struct config_work;
268 /* Does the affinity hint is set for virtqueues? */
269 bool affinity_hint_set;
271 /* CPU hotplug instances for online & dead */
272 struct hlist_node node;
273 struct hlist_node node_dead;
275 struct control_buf *ctrl;
277 /* Ethtool settings */
281 /* Interrupt coalescing settings */
287 unsigned long guest_offloads;
288 unsigned long guest_offloads_capable;
290 /* failover when STANDBY feature enabled */
291 struct failover *failover;
294 struct padded_vnet_hdr {
295 struct virtio_net_hdr_v1_hash hdr;
297 * hdr is in a separate sg buffer, and data sg buffer shares same page
298 * with this header sg. This padding makes next sg 16 byte aligned
304 static void virtnet_rq_free_unused_buf(struct virtqueue *vq, void *buf);
305 static void virtnet_sq_free_unused_buf(struct virtqueue *vq, void *buf);
307 static bool is_xdp_frame(void *ptr)
309 return (unsigned long)ptr & VIRTIO_XDP_FLAG;
312 static void *xdp_to_ptr(struct xdp_frame *ptr)
314 return (void *)((unsigned long)ptr | VIRTIO_XDP_FLAG);
317 static struct xdp_frame *ptr_to_xdp(void *ptr)
319 return (struct xdp_frame *)((unsigned long)ptr & ~VIRTIO_XDP_FLAG);
322 /* Converting between virtqueue no. and kernel tx/rx queue no.
323 * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
325 static int vq2txq(struct virtqueue *vq)
327 return (vq->index - 1) / 2;
330 static int txq2vq(int txq)
335 static int vq2rxq(struct virtqueue *vq)
337 return vq->index / 2;
340 static int rxq2vq(int rxq)
345 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
347 return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
351 * private is used to chain pages for big packets, put the whole
352 * most recent used list in the beginning for reuse
354 static void give_pages(struct receive_queue *rq, struct page *page)
358 /* Find end of list, sew whole thing into vi->rq.pages. */
359 for (end = page; end->private; end = (struct page *)end->private);
360 end->private = (unsigned long)rq->pages;
364 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
366 struct page *p = rq->pages;
369 rq->pages = (struct page *)p->private;
370 /* clear private here, it is used to chain pages */
373 p = alloc_page(gfp_mask);
377 static void enable_delayed_refill(struct virtnet_info *vi)
379 spin_lock_bh(&vi->refill_lock);
380 vi->refill_enabled = true;
381 spin_unlock_bh(&vi->refill_lock);
384 static void disable_delayed_refill(struct virtnet_info *vi)
386 spin_lock_bh(&vi->refill_lock);
387 vi->refill_enabled = false;
388 spin_unlock_bh(&vi->refill_lock);
391 static void virtqueue_napi_schedule(struct napi_struct *napi,
392 struct virtqueue *vq)
394 if (napi_schedule_prep(napi)) {
395 virtqueue_disable_cb(vq);
396 __napi_schedule(napi);
400 static void virtqueue_napi_complete(struct napi_struct *napi,
401 struct virtqueue *vq, int processed)
405 opaque = virtqueue_enable_cb_prepare(vq);
406 if (napi_complete_done(napi, processed)) {
407 if (unlikely(virtqueue_poll(vq, opaque)))
408 virtqueue_napi_schedule(napi, vq);
410 virtqueue_disable_cb(vq);
414 static void skb_xmit_done(struct virtqueue *vq)
416 struct virtnet_info *vi = vq->vdev->priv;
417 struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi;
419 /* Suppress further interrupts. */
420 virtqueue_disable_cb(vq);
423 virtqueue_napi_schedule(napi, vq);
425 /* We were probably waiting for more output buffers. */
426 netif_wake_subqueue(vi->dev, vq2txq(vq));
429 #define MRG_CTX_HEADER_SHIFT 22
430 static void *mergeable_len_to_ctx(unsigned int truesize,
431 unsigned int headroom)
433 return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) | truesize);
436 static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx)
438 return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT;
441 static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx)
443 return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) - 1);
446 /* Called from bottom half context */
447 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
448 struct receive_queue *rq,
449 struct page *page, unsigned int offset,
450 unsigned int len, unsigned int truesize,
451 unsigned int headroom)
454 struct virtio_net_hdr_mrg_rxbuf *hdr;
455 unsigned int copy, hdr_len, hdr_padded_len;
456 struct page *page_to_free = NULL;
457 int tailroom, shinfo_size;
458 char *p, *hdr_p, *buf;
460 p = page_address(page) + offset;
463 hdr_len = vi->hdr_len;
464 if (vi->mergeable_rx_bufs)
465 hdr_padded_len = hdr_len;
467 hdr_padded_len = sizeof(struct padded_vnet_hdr);
471 offset += hdr_padded_len;
473 tailroom = truesize - headroom - hdr_padded_len - len;
475 shinfo_size = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
477 /* copy small packet so we can reuse these pages */
478 if (!NET_IP_ALIGN && len > GOOD_COPY_LEN && tailroom >= shinfo_size) {
479 skb = build_skb(buf, truesize);
483 skb_reserve(skb, p - buf);
486 page = (struct page *)page->private;
488 give_pages(rq, page);
492 /* copy small packet so we can reuse these pages for small data */
493 skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
497 /* Copy all frame if it fits skb->head, otherwise
498 * we let virtio_net_hdr_to_skb() and GRO pull headers as needed.
500 if (len <= skb_tailroom(skb))
504 skb_put_data(skb, p, copy);
509 if (vi->mergeable_rx_bufs) {
511 skb_add_rx_frag(skb, 0, page, offset, len, truesize);
518 * Verify that we can indeed put this data into a skb.
519 * This is here to handle cases when the device erroneously
520 * tries to receive more than is possible. This is usually
521 * the case of a broken device.
523 if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
524 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
528 BUG_ON(offset >= PAGE_SIZE);
530 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
531 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
532 frag_size, truesize);
534 page = (struct page *)page->private;
539 give_pages(rq, page);
542 hdr = skb_vnet_hdr(skb);
543 memcpy(hdr, hdr_p, hdr_len);
545 put_page(page_to_free);
550 static void free_old_xmit_skbs(struct send_queue *sq, bool in_napi)
553 unsigned int packets = 0;
554 unsigned int bytes = 0;
557 while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
558 if (likely(!is_xdp_frame(ptr))) {
559 struct sk_buff *skb = ptr;
561 pr_debug("Sent skb %p\n", skb);
564 napi_consume_skb(skb, in_napi);
566 struct xdp_frame *frame = ptr_to_xdp(ptr);
568 bytes += xdp_get_frame_len(frame);
569 xdp_return_frame(frame);
574 /* Avoid overhead when no packets have been processed
575 * happens when called speculatively from start_xmit.
580 u64_stats_update_begin(&sq->stats.syncp);
581 sq->stats.bytes += bytes;
582 sq->stats.packets += packets;
583 u64_stats_update_end(&sq->stats.syncp);
586 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
588 if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
590 else if (q < vi->curr_queue_pairs)
596 static void check_sq_full_and_disable(struct virtnet_info *vi,
597 struct net_device *dev,
598 struct send_queue *sq)
600 bool use_napi = sq->napi.weight;
605 /* If running out of space, stop queue to avoid getting packets that we
606 * are then unable to transmit.
607 * An alternative would be to force queuing layer to requeue the skb by
608 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
609 * returned in a normal path of operation: it means that driver is not
610 * maintaining the TX queue stop/start state properly, and causes
611 * the stack to do a non-trivial amount of useless work.
612 * Since most packets only take 1 or 2 ring slots, stopping the queue
613 * early means 16 slots are typically wasted.
615 if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
616 netif_stop_subqueue(dev, qnum);
618 if (unlikely(!virtqueue_enable_cb_delayed(sq->vq)))
619 virtqueue_napi_schedule(&sq->napi, sq->vq);
620 } else if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
621 /* More just got used, free them then recheck. */
622 free_old_xmit_skbs(sq, false);
623 if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
624 netif_start_subqueue(dev, qnum);
625 virtqueue_disable_cb(sq->vq);
631 static int __virtnet_xdp_xmit_one(struct virtnet_info *vi,
632 struct send_queue *sq,
633 struct xdp_frame *xdpf)
635 struct virtio_net_hdr_mrg_rxbuf *hdr;
636 struct skb_shared_info *shinfo;
640 if (unlikely(xdpf->headroom < vi->hdr_len))
643 if (unlikely(xdp_frame_has_frags(xdpf))) {
644 shinfo = xdp_get_shared_info_from_frame(xdpf);
645 nr_frags = shinfo->nr_frags;
648 /* In wrapping function virtnet_xdp_xmit(), we need to free
649 * up the pending old buffers, where we need to calculate the
650 * position of skb_shared_info in xdp_get_frame_len() and
651 * xdp_return_frame(), which will involve to xdpf->data and
652 * xdpf->headroom. Therefore, we need to update the value of
653 * headroom synchronously here.
655 xdpf->headroom -= vi->hdr_len;
656 xdpf->data -= vi->hdr_len;
657 /* Zero header and leave csum up to XDP layers */
659 memset(hdr, 0, vi->hdr_len);
660 xdpf->len += vi->hdr_len;
662 sg_init_table(sq->sg, nr_frags + 1);
663 sg_set_buf(sq->sg, xdpf->data, xdpf->len);
664 for (i = 0; i < nr_frags; i++) {
665 skb_frag_t *frag = &shinfo->frags[i];
667 sg_set_page(&sq->sg[i + 1], skb_frag_page(frag),
668 skb_frag_size(frag), skb_frag_off(frag));
671 err = virtqueue_add_outbuf(sq->vq, sq->sg, nr_frags + 1,
672 xdp_to_ptr(xdpf), GFP_ATOMIC);
674 return -ENOSPC; /* Caller handle free/refcnt */
679 /* when vi->curr_queue_pairs > nr_cpu_ids, the txq/sq is only used for xdp tx on
680 * the current cpu, so it does not need to be locked.
682 * Here we use marco instead of inline functions because we have to deal with
683 * three issues at the same time: 1. the choice of sq. 2. judge and execute the
684 * lock/unlock of txq 3. make sparse happy. It is difficult for two inline
685 * functions to perfectly solve these three problems at the same time.
687 #define virtnet_xdp_get_sq(vi) ({ \
688 int cpu = smp_processor_id(); \
689 struct netdev_queue *txq; \
690 typeof(vi) v = (vi); \
693 if (v->curr_queue_pairs > nr_cpu_ids) { \
694 qp = v->curr_queue_pairs - v->xdp_queue_pairs; \
696 txq = netdev_get_tx_queue(v->dev, qp); \
697 __netif_tx_acquire(txq); \
699 qp = cpu % v->curr_queue_pairs; \
700 txq = netdev_get_tx_queue(v->dev, qp); \
701 __netif_tx_lock(txq, cpu); \
706 #define virtnet_xdp_put_sq(vi, q) { \
707 struct netdev_queue *txq; \
708 typeof(vi) v = (vi); \
710 txq = netdev_get_tx_queue(v->dev, (q) - v->sq); \
711 if (v->curr_queue_pairs > nr_cpu_ids) \
712 __netif_tx_release(txq); \
714 __netif_tx_unlock(txq); \
717 static int virtnet_xdp_xmit(struct net_device *dev,
718 int n, struct xdp_frame **frames, u32 flags)
720 struct virtnet_info *vi = netdev_priv(dev);
721 struct receive_queue *rq = vi->rq;
722 struct bpf_prog *xdp_prog;
723 struct send_queue *sq;
733 /* Only allow ndo_xdp_xmit if XDP is loaded on dev, as this
734 * indicate XDP resources have been successfully allocated.
736 xdp_prog = rcu_access_pointer(rq->xdp_prog);
740 sq = virtnet_xdp_get_sq(vi);
742 if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK)) {
747 /* Free up any pending old buffers before queueing new ones. */
748 while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
749 if (likely(is_xdp_frame(ptr))) {
750 struct xdp_frame *frame = ptr_to_xdp(ptr);
752 bytes += xdp_get_frame_len(frame);
753 xdp_return_frame(frame);
755 struct sk_buff *skb = ptr;
758 napi_consume_skb(skb, false);
763 for (i = 0; i < n; i++) {
764 struct xdp_frame *xdpf = frames[i];
766 if (__virtnet_xdp_xmit_one(vi, sq, xdpf))
772 if (!is_xdp_raw_buffer_queue(vi, sq - vi->sq))
773 check_sq_full_and_disable(vi, dev, sq);
775 if (flags & XDP_XMIT_FLUSH) {
776 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq))
780 u64_stats_update_begin(&sq->stats.syncp);
781 sq->stats.bytes += bytes;
782 sq->stats.packets += packets;
783 sq->stats.xdp_tx += n;
784 sq->stats.xdp_tx_drops += n - nxmit;
785 sq->stats.kicks += kicks;
786 u64_stats_update_end(&sq->stats.syncp);
788 virtnet_xdp_put_sq(vi, sq);
792 static unsigned int virtnet_get_headroom(struct virtnet_info *vi)
794 return vi->xdp_enabled ? VIRTIO_XDP_HEADROOM : 0;
797 /* We copy the packet for XDP in the following cases:
799 * 1) Packet is scattered across multiple rx buffers.
800 * 2) Headroom space is insufficient.
802 * This is inefficient but it's a temporary condition that
803 * we hit right after XDP is enabled and until queue is refilled
804 * with large buffers with sufficient headroom - so it should affect
805 * at most queue size packets.
806 * Afterwards, the conditions to enable
807 * XDP should preclude the underlying device from sending packets
808 * across multiple buffers (num_buf > 1), and we make sure buffers
809 * have enough headroom.
811 static struct page *xdp_linearize_page(struct receive_queue *rq,
818 struct page *page = alloc_page(GFP_ATOMIC);
823 memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
827 int tailroom = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
832 buf = virtqueue_get_buf(rq->vq, &buflen);
836 p = virt_to_head_page(buf);
837 off = buf - page_address(p);
839 /* guard against a misconfigured or uncooperative backend that
840 * is sending packet larger than the MTU.
842 if ((page_off + buflen + tailroom) > PAGE_SIZE) {
847 memcpy(page_address(page) + page_off,
848 page_address(p) + off, buflen);
853 /* Headroom does not contribute to packet length */
854 *len = page_off - VIRTIO_XDP_HEADROOM;
857 __free_pages(page, 0);
861 static struct sk_buff *receive_small(struct net_device *dev,
862 struct virtnet_info *vi,
863 struct receive_queue *rq,
864 void *buf, void *ctx,
866 unsigned int *xdp_xmit,
867 struct virtnet_rq_stats *stats)
870 struct bpf_prog *xdp_prog;
871 unsigned int xdp_headroom = (unsigned long)ctx;
872 unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom;
873 unsigned int headroom = vi->hdr_len + header_offset;
874 unsigned int buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
875 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
876 struct page *page = virt_to_head_page(buf);
877 unsigned int delta = 0;
878 struct page *xdp_page;
880 unsigned int metasize = 0;
885 if (unlikely(len > GOOD_PACKET_LEN)) {
886 pr_debug("%s: rx error: len %u exceeds max size %d\n",
887 dev->name, len, GOOD_PACKET_LEN);
888 dev->stats.rx_length_errors++;
892 if (likely(!vi->xdp_enabled)) {
898 xdp_prog = rcu_dereference(rq->xdp_prog);
900 struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset;
901 struct xdp_frame *xdpf;
906 if (unlikely(hdr->hdr.gso_type))
909 if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) {
910 int offset = buf - page_address(page) + header_offset;
911 unsigned int tlen = len + vi->hdr_len;
914 xdp_headroom = virtnet_get_headroom(vi);
915 header_offset = VIRTNET_RX_PAD + xdp_headroom;
916 headroom = vi->hdr_len + header_offset;
917 buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
918 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
919 xdp_page = xdp_linearize_page(rq, &num_buf, page,
920 offset, header_offset,
925 buf = page_address(xdp_page);
930 xdp_init_buff(&xdp, buflen, &rq->xdp_rxq);
931 xdp_prepare_buff(&xdp, buf + VIRTNET_RX_PAD + vi->hdr_len,
932 xdp_headroom, len, true);
933 orig_data = xdp.data;
934 act = bpf_prog_run_xdp(xdp_prog, &xdp);
935 stats->xdp_packets++;
939 /* Recalculate length in case bpf program changed it */
940 delta = orig_data - xdp.data;
941 len = xdp.data_end - xdp.data;
942 metasize = xdp.data - xdp.data_meta;
946 xdpf = xdp_convert_buff_to_frame(&xdp);
949 err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
950 if (unlikely(!err)) {
951 xdp_return_frame_rx_napi(xdpf);
952 } else if (unlikely(err < 0)) {
953 trace_xdp_exception(vi->dev, xdp_prog, act);
956 *xdp_xmit |= VIRTIO_XDP_TX;
960 stats->xdp_redirects++;
961 err = xdp_do_redirect(dev, &xdp, xdp_prog);
964 *xdp_xmit |= VIRTIO_XDP_REDIR;
968 bpf_warn_invalid_xdp_action(vi->dev, xdp_prog, act);
971 trace_xdp_exception(vi->dev, xdp_prog, act);
980 skb = build_skb(buf, buflen);
983 skb_reserve(skb, headroom - delta);
986 buf += header_offset;
987 memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
988 } /* keep zeroed vnet hdr since XDP is loaded */
991 skb_metadata_set(skb, metasize);
1005 static struct sk_buff *receive_big(struct net_device *dev,
1006 struct virtnet_info *vi,
1007 struct receive_queue *rq,
1010 struct virtnet_rq_stats *stats)
1012 struct page *page = buf;
1013 struct sk_buff *skb =
1014 page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, 0);
1016 stats->bytes += len - vi->hdr_len;
1024 give_pages(rq, page);
1028 /* Why not use xdp_build_skb_from_frame() ?
1029 * XDP core assumes that xdp frags are PAGE_SIZE in length, while in
1030 * virtio-net there are 2 points that do not match its requirements:
1031 * 1. The size of the prefilled buffer is not fixed before xdp is set.
1032 * 2. xdp_build_skb_from_frame() does more checks that we don't need,
1033 * like eth_type_trans() (which virtio-net does in receive_buf()).
1035 static struct sk_buff *build_skb_from_xdp_buff(struct net_device *dev,
1036 struct virtnet_info *vi,
1037 struct xdp_buff *xdp,
1038 unsigned int xdp_frags_truesz)
1040 struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(xdp);
1041 unsigned int headroom, data_len;
1042 struct sk_buff *skb;
1046 if (unlikely(xdp->data_end > xdp_data_hard_end(xdp))) {
1047 pr_debug("Error building skb as missing reserved tailroom for xdp");
1051 if (unlikely(xdp_buff_has_frags(xdp)))
1052 nr_frags = sinfo->nr_frags;
1054 skb = build_skb(xdp->data_hard_start, xdp->frame_sz);
1058 headroom = xdp->data - xdp->data_hard_start;
1059 data_len = xdp->data_end - xdp->data;
1060 skb_reserve(skb, headroom);
1061 __skb_put(skb, data_len);
1063 metasize = xdp->data - xdp->data_meta;
1064 metasize = metasize > 0 ? metasize : 0;
1066 skb_metadata_set(skb, metasize);
1068 if (unlikely(xdp_buff_has_frags(xdp)))
1069 xdp_update_skb_shared_info(skb, nr_frags,
1070 sinfo->xdp_frags_size,
1072 xdp_buff_is_frag_pfmemalloc(xdp));
1077 /* TODO: build xdp in big mode */
1078 static int virtnet_build_xdp_buff_mrg(struct net_device *dev,
1079 struct virtnet_info *vi,
1080 struct receive_queue *rq,
1081 struct xdp_buff *xdp,
1084 unsigned int frame_sz,
1086 unsigned int *xdp_frags_truesize,
1087 struct virtnet_rq_stats *stats)
1089 struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
1090 unsigned int headroom, tailroom, room;
1091 unsigned int truesize, cur_frag_size;
1092 struct skb_shared_info *shinfo;
1093 unsigned int xdp_frags_truesz = 0;
1099 xdp_init_buff(xdp, frame_sz, &rq->xdp_rxq);
1100 xdp_prepare_buff(xdp, buf - VIRTIO_XDP_HEADROOM,
1101 VIRTIO_XDP_HEADROOM + vi->hdr_len, len - vi->hdr_len, true);
1107 /* If we want to build multi-buffer xdp, we need
1108 * to specify that the flags of xdp_buff have the
1109 * XDP_FLAGS_HAS_FRAG bit.
1111 if (!xdp_buff_has_frags(xdp))
1112 xdp_buff_set_frags_flag(xdp);
1114 shinfo = xdp_get_shared_info_from_buff(xdp);
1115 shinfo->nr_frags = 0;
1116 shinfo->xdp_frags_size = 0;
1119 if (*num_buf > MAX_SKB_FRAGS + 1)
1122 while (--*num_buf > 0) {
1123 buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx);
1124 if (unlikely(!buf)) {
1125 pr_debug("%s: rx error: %d buffers out of %d missing\n",
1126 dev->name, *num_buf,
1127 virtio16_to_cpu(vi->vdev, hdr->num_buffers));
1128 dev->stats.rx_length_errors++;
1132 stats->bytes += len;
1133 page = virt_to_head_page(buf);
1134 offset = buf - page_address(page);
1136 truesize = mergeable_ctx_to_truesize(ctx);
1137 headroom = mergeable_ctx_to_headroom(ctx);
1138 tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1139 room = SKB_DATA_ALIGN(headroom + tailroom);
1141 cur_frag_size = truesize;
1142 xdp_frags_truesz += cur_frag_size;
1143 if (unlikely(len > truesize - room || cur_frag_size > PAGE_SIZE)) {
1145 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
1146 dev->name, len, (unsigned long)(truesize - room));
1147 dev->stats.rx_length_errors++;
1151 frag = &shinfo->frags[shinfo->nr_frags++];
1152 __skb_frag_set_page(frag, page);
1153 skb_frag_off_set(frag, offset);
1154 skb_frag_size_set(frag, len);
1155 if (page_is_pfmemalloc(page))
1156 xdp_buff_set_frag_pfmemalloc(xdp);
1158 shinfo->xdp_frags_size += len;
1161 *xdp_frags_truesize = xdp_frags_truesz;
1165 static struct sk_buff *receive_mergeable(struct net_device *dev,
1166 struct virtnet_info *vi,
1167 struct receive_queue *rq,
1171 unsigned int *xdp_xmit,
1172 struct virtnet_rq_stats *stats)
1174 struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
1175 int num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
1176 struct page *page = virt_to_head_page(buf);
1177 int offset = buf - page_address(page);
1178 struct sk_buff *head_skb, *curr_skb;
1179 struct bpf_prog *xdp_prog;
1180 unsigned int truesize = mergeable_ctx_to_truesize(ctx);
1181 unsigned int headroom = mergeable_ctx_to_headroom(ctx);
1182 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1183 unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
1184 unsigned int frame_sz, xdp_room;
1188 stats->bytes += len - vi->hdr_len;
1190 if (unlikely(len > truesize - room)) {
1191 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
1192 dev->name, len, (unsigned long)(truesize - room));
1193 dev->stats.rx_length_errors++;
1197 if (likely(!vi->xdp_enabled)) {
1203 xdp_prog = rcu_dereference(rq->xdp_prog);
1205 unsigned int xdp_frags_truesz = 0;
1206 struct skb_shared_info *shinfo;
1207 struct xdp_frame *xdpf;
1208 struct page *xdp_page;
1209 struct xdp_buff xdp;
1214 /* Transient failure which in theory could occur if
1215 * in-flight packets from before XDP was enabled reach
1216 * the receive path after XDP is loaded.
1218 if (unlikely(hdr->hdr.gso_type))
1221 /* Now XDP core assumes frag size is PAGE_SIZE, but buffers
1222 * with headroom may add hole in truesize, which
1223 * make their length exceed PAGE_SIZE. So we disabled the
1224 * hole mechanism for xdp. See add_recvbuf_mergeable().
1226 frame_sz = truesize;
1228 /* This happens when headroom is not enough because
1229 * of the buffer was prefilled before XDP is set.
1230 * This should only happen for the first several packets.
1231 * In fact, vq reset can be used here to help us clean up
1232 * the prefilled buffers, but many existing devices do not
1233 * support it, and we don't want to bother users who are
1234 * using xdp normally.
1236 if (!xdp_prog->aux->xdp_has_frags &&
1237 (num_buf > 1 || headroom < virtnet_get_headroom(vi))) {
1238 /* linearize data for XDP */
1239 xdp_page = xdp_linearize_page(rq, &num_buf,
1241 VIRTIO_XDP_HEADROOM,
1243 frame_sz = PAGE_SIZE;
1247 offset = VIRTIO_XDP_HEADROOM;
1248 } else if (unlikely(headroom < virtnet_get_headroom(vi))) {
1249 xdp_room = SKB_DATA_ALIGN(VIRTIO_XDP_HEADROOM +
1250 sizeof(struct skb_shared_info));
1251 if (len + xdp_room > PAGE_SIZE)
1254 xdp_page = alloc_page(GFP_ATOMIC);
1258 memcpy(page_address(xdp_page) + VIRTIO_XDP_HEADROOM,
1259 page_address(page) + offset, len);
1260 frame_sz = PAGE_SIZE;
1261 offset = VIRTIO_XDP_HEADROOM;
1266 data = page_address(xdp_page) + offset;
1267 err = virtnet_build_xdp_buff_mrg(dev, vi, rq, &xdp, data, len, frame_sz,
1268 &num_buf, &xdp_frags_truesz, stats);
1272 act = bpf_prog_run_xdp(xdp_prog, &xdp);
1273 stats->xdp_packets++;
1277 head_skb = build_skb_from_xdp_buff(dev, vi, &xdp, xdp_frags_truesz);
1278 if (unlikely(!head_skb))
1281 if (unlikely(xdp_page != page))
1287 xdpf = xdp_convert_buff_to_frame(&xdp);
1288 if (unlikely(!xdpf)) {
1289 netdev_dbg(dev, "convert buff to frame failed for xdp\n");
1292 err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
1293 if (unlikely(!err)) {
1294 xdp_return_frame_rx_napi(xdpf);
1295 } else if (unlikely(err < 0)) {
1296 trace_xdp_exception(vi->dev, xdp_prog, act);
1299 *xdp_xmit |= VIRTIO_XDP_TX;
1300 if (unlikely(xdp_page != page))
1305 stats->xdp_redirects++;
1306 err = xdp_do_redirect(dev, &xdp, xdp_prog);
1309 *xdp_xmit |= VIRTIO_XDP_REDIR;
1310 if (unlikely(xdp_page != page))
1315 bpf_warn_invalid_xdp_action(vi->dev, xdp_prog, act);
1318 trace_xdp_exception(vi->dev, xdp_prog, act);
1324 if (unlikely(xdp_page != page))
1325 __free_pages(xdp_page, 0);
1327 if (xdp_buff_has_frags(&xdp)) {
1328 shinfo = xdp_get_shared_info_from_buff(&xdp);
1329 for (i = 0; i < shinfo->nr_frags; i++) {
1330 xdp_page = skb_frag_page(&shinfo->frags[i]);
1340 head_skb = page_to_skb(vi, rq, page, offset, len, truesize, headroom);
1341 curr_skb = head_skb;
1343 if (unlikely(!curr_skb))
1348 buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx);
1349 if (unlikely(!buf)) {
1350 pr_debug("%s: rx error: %d buffers out of %d missing\n",
1352 virtio16_to_cpu(vi->vdev,
1354 dev->stats.rx_length_errors++;
1358 stats->bytes += len;
1359 page = virt_to_head_page(buf);
1361 truesize = mergeable_ctx_to_truesize(ctx);
1362 headroom = mergeable_ctx_to_headroom(ctx);
1363 tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1364 room = SKB_DATA_ALIGN(headroom + tailroom);
1365 if (unlikely(len > truesize - room)) {
1366 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
1367 dev->name, len, (unsigned long)(truesize - room));
1368 dev->stats.rx_length_errors++;
1372 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
1373 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
1374 struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
1376 if (unlikely(!nskb))
1378 if (curr_skb == head_skb)
1379 skb_shinfo(curr_skb)->frag_list = nskb;
1381 curr_skb->next = nskb;
1383 head_skb->truesize += nskb->truesize;
1386 if (curr_skb != head_skb) {
1387 head_skb->data_len += len;
1388 head_skb->len += len;
1389 head_skb->truesize += truesize;
1391 offset = buf - page_address(page);
1392 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
1394 skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
1397 skb_add_rx_frag(curr_skb, num_skb_frags, page,
1398 offset, len, truesize);
1402 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
1410 while (num_buf-- > 1) {
1411 buf = virtqueue_get_buf(rq->vq, &len);
1412 if (unlikely(!buf)) {
1413 pr_debug("%s: rx error: %d buffers missing\n",
1414 dev->name, num_buf);
1415 dev->stats.rx_length_errors++;
1418 stats->bytes += len;
1419 page = virt_to_head_page(buf);
1424 dev_kfree_skb(head_skb);
1429 static void virtio_skb_set_hash(const struct virtio_net_hdr_v1_hash *hdr_hash,
1430 struct sk_buff *skb)
1432 enum pkt_hash_types rss_hash_type;
1434 if (!hdr_hash || !skb)
1437 switch (__le16_to_cpu(hdr_hash->hash_report)) {
1438 case VIRTIO_NET_HASH_REPORT_TCPv4:
1439 case VIRTIO_NET_HASH_REPORT_UDPv4:
1440 case VIRTIO_NET_HASH_REPORT_TCPv6:
1441 case VIRTIO_NET_HASH_REPORT_UDPv6:
1442 case VIRTIO_NET_HASH_REPORT_TCPv6_EX:
1443 case VIRTIO_NET_HASH_REPORT_UDPv6_EX:
1444 rss_hash_type = PKT_HASH_TYPE_L4;
1446 case VIRTIO_NET_HASH_REPORT_IPv4:
1447 case VIRTIO_NET_HASH_REPORT_IPv6:
1448 case VIRTIO_NET_HASH_REPORT_IPv6_EX:
1449 rss_hash_type = PKT_HASH_TYPE_L3;
1451 case VIRTIO_NET_HASH_REPORT_NONE:
1453 rss_hash_type = PKT_HASH_TYPE_NONE;
1455 skb_set_hash(skb, __le32_to_cpu(hdr_hash->hash_value), rss_hash_type);
1458 static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
1459 void *buf, unsigned int len, void **ctx,
1460 unsigned int *xdp_xmit,
1461 struct virtnet_rq_stats *stats)
1463 struct net_device *dev = vi->dev;
1464 struct sk_buff *skb;
1465 struct virtio_net_hdr_mrg_rxbuf *hdr;
1467 if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
1468 pr_debug("%s: short packet %i\n", dev->name, len);
1469 dev->stats.rx_length_errors++;
1470 virtnet_rq_free_unused_buf(rq->vq, buf);
1474 if (vi->mergeable_rx_bufs)
1475 skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit,
1477 else if (vi->big_packets)
1478 skb = receive_big(dev, vi, rq, buf, len, stats);
1480 skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit, stats);
1485 hdr = skb_vnet_hdr(skb);
1486 if (dev->features & NETIF_F_RXHASH && vi->has_rss_hash_report)
1487 virtio_skb_set_hash((const struct virtio_net_hdr_v1_hash *)hdr, skb);
1489 if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
1490 skb->ip_summed = CHECKSUM_UNNECESSARY;
1492 if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
1493 virtio_is_little_endian(vi->vdev))) {
1494 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
1495 dev->name, hdr->hdr.gso_type,
1500 skb_record_rx_queue(skb, vq2rxq(rq->vq));
1501 skb->protocol = eth_type_trans(skb, dev);
1502 pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
1503 ntohs(skb->protocol), skb->len, skb->pkt_type);
1505 napi_gro_receive(&rq->napi, skb);
1509 dev->stats.rx_frame_errors++;
1513 /* Unlike mergeable buffers, all buffers are allocated to the
1514 * same size, except for the headroom. For this reason we do
1515 * not need to use mergeable_len_to_ctx here - it is enough
1516 * to store the headroom as the context ignoring the truesize.
1518 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
1521 struct page_frag *alloc_frag = &rq->alloc_frag;
1523 unsigned int xdp_headroom = virtnet_get_headroom(vi);
1524 void *ctx = (void *)(unsigned long)xdp_headroom;
1525 int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
1528 len = SKB_DATA_ALIGN(len) +
1529 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1530 if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
1533 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1534 get_page(alloc_frag->page);
1535 alloc_frag->offset += len;
1536 sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom,
1537 vi->hdr_len + GOOD_PACKET_LEN);
1538 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1540 put_page(virt_to_head_page(buf));
1544 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
1547 struct page *first, *list = NULL;
1551 sg_init_table(rq->sg, vi->big_packets_num_skbfrags + 2);
1553 /* page in rq->sg[vi->big_packets_num_skbfrags + 1] is list tail */
1554 for (i = vi->big_packets_num_skbfrags + 1; i > 1; --i) {
1555 first = get_a_page(rq, gfp);
1558 give_pages(rq, list);
1561 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
1563 /* chain new page in list head to match sg */
1564 first->private = (unsigned long)list;
1568 first = get_a_page(rq, gfp);
1570 give_pages(rq, list);
1573 p = page_address(first);
1575 /* rq->sg[0], rq->sg[1] share the same page */
1576 /* a separated rq->sg[0] for header - required in case !any_header_sg */
1577 sg_set_buf(&rq->sg[0], p, vi->hdr_len);
1579 /* rq->sg[1] for data packet, from offset */
1580 offset = sizeof(struct padded_vnet_hdr);
1581 sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
1583 /* chain first in list head */
1584 first->private = (unsigned long)list;
1585 err = virtqueue_add_inbuf(rq->vq, rq->sg, vi->big_packets_num_skbfrags + 2,
1588 give_pages(rq, first);
1593 static unsigned int get_mergeable_buf_len(struct receive_queue *rq,
1594 struct ewma_pkt_len *avg_pkt_len,
1597 struct virtnet_info *vi = rq->vq->vdev->priv;
1598 const size_t hdr_len = vi->hdr_len;
1602 return PAGE_SIZE - room;
1604 len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
1605 rq->min_buf_len, PAGE_SIZE - hdr_len);
1607 return ALIGN(len, L1_CACHE_BYTES);
1610 static int add_recvbuf_mergeable(struct virtnet_info *vi,
1611 struct receive_queue *rq, gfp_t gfp)
1613 struct page_frag *alloc_frag = &rq->alloc_frag;
1614 unsigned int headroom = virtnet_get_headroom(vi);
1615 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1616 unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
1620 unsigned int len, hole;
1622 /* Extra tailroom is needed to satisfy XDP's assumption. This
1623 * means rx frags coalescing won't work, but consider we've
1624 * disabled GSO for XDP, it won't be a big issue.
1626 len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len, room);
1627 if (unlikely(!skb_page_frag_refill(len + room, alloc_frag, gfp)))
1630 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
1631 buf += headroom; /* advance address leaving hole at front of pkt */
1632 get_page(alloc_frag->page);
1633 alloc_frag->offset += len + room;
1634 hole = alloc_frag->size - alloc_frag->offset;
1635 if (hole < len + room) {
1636 /* To avoid internal fragmentation, if there is very likely not
1637 * enough space for another buffer, add the remaining space to
1638 * the current buffer.
1639 * XDP core assumes that frame_size of xdp_buff and the length
1640 * of the frag are PAGE_SIZE, so we disable the hole mechanism.
1644 alloc_frag->offset += hole;
1647 sg_init_one(rq->sg, buf, len);
1648 ctx = mergeable_len_to_ctx(len + room, headroom);
1649 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1651 put_page(virt_to_head_page(buf));
1657 * Returns false if we couldn't fill entirely (OOM).
1659 * Normally run in the receive path, but can also be run from ndo_open
1660 * before we're receiving packets, or from refill_work which is
1661 * careful to disable receiving (using napi_disable).
1663 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
1670 if (vi->mergeable_rx_bufs)
1671 err = add_recvbuf_mergeable(vi, rq, gfp);
1672 else if (vi->big_packets)
1673 err = add_recvbuf_big(vi, rq, gfp);
1675 err = add_recvbuf_small(vi, rq, gfp);
1677 oom = err == -ENOMEM;
1680 } while (rq->vq->num_free);
1681 if (virtqueue_kick_prepare(rq->vq) && virtqueue_notify(rq->vq)) {
1682 unsigned long flags;
1684 flags = u64_stats_update_begin_irqsave(&rq->stats.syncp);
1686 u64_stats_update_end_irqrestore(&rq->stats.syncp, flags);
1692 static void skb_recv_done(struct virtqueue *rvq)
1694 struct virtnet_info *vi = rvq->vdev->priv;
1695 struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
1697 virtqueue_napi_schedule(&rq->napi, rvq);
1700 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
1704 /* If all buffers were filled by other side before we napi_enabled, we
1705 * won't get another interrupt, so process any outstanding packets now.
1706 * Call local_bh_enable after to trigger softIRQ processing.
1709 virtqueue_napi_schedule(napi, vq);
1713 static void virtnet_napi_tx_enable(struct virtnet_info *vi,
1714 struct virtqueue *vq,
1715 struct napi_struct *napi)
1720 /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
1721 * enable the feature if this is likely affine with the transmit path.
1723 if (!vi->affinity_hint_set) {
1728 return virtnet_napi_enable(vq, napi);
1731 static void virtnet_napi_tx_disable(struct napi_struct *napi)
1737 static void refill_work(struct work_struct *work)
1739 struct virtnet_info *vi =
1740 container_of(work, struct virtnet_info, refill.work);
1744 for (i = 0; i < vi->curr_queue_pairs; i++) {
1745 struct receive_queue *rq = &vi->rq[i];
1747 napi_disable(&rq->napi);
1748 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
1749 virtnet_napi_enable(rq->vq, &rq->napi);
1751 /* In theory, this can happen: if we don't get any buffers in
1752 * we will *never* try to fill again.
1755 schedule_delayed_work(&vi->refill, HZ/2);
1759 static int virtnet_receive(struct receive_queue *rq, int budget,
1760 unsigned int *xdp_xmit)
1762 struct virtnet_info *vi = rq->vq->vdev->priv;
1763 struct virtnet_rq_stats stats = {};
1768 if (!vi->big_packets || vi->mergeable_rx_bufs) {
1771 while (stats.packets < budget &&
1772 (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) {
1773 receive_buf(vi, rq, buf, len, ctx, xdp_xmit, &stats);
1777 while (stats.packets < budget &&
1778 (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
1779 receive_buf(vi, rq, buf, len, NULL, xdp_xmit, &stats);
1784 if (rq->vq->num_free > min((unsigned int)budget, virtqueue_get_vring_size(rq->vq)) / 2) {
1785 if (!try_fill_recv(vi, rq, GFP_ATOMIC)) {
1786 spin_lock(&vi->refill_lock);
1787 if (vi->refill_enabled)
1788 schedule_delayed_work(&vi->refill, 0);
1789 spin_unlock(&vi->refill_lock);
1793 u64_stats_update_begin(&rq->stats.syncp);
1794 for (i = 0; i < VIRTNET_RQ_STATS_LEN; i++) {
1795 size_t offset = virtnet_rq_stats_desc[i].offset;
1798 item = (u64 *)((u8 *)&rq->stats + offset);
1799 *item += *(u64 *)((u8 *)&stats + offset);
1801 u64_stats_update_end(&rq->stats.syncp);
1803 return stats.packets;
1806 static void virtnet_poll_cleantx(struct receive_queue *rq)
1808 struct virtnet_info *vi = rq->vq->vdev->priv;
1809 unsigned int index = vq2rxq(rq->vq);
1810 struct send_queue *sq = &vi->sq[index];
1811 struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index);
1813 if (!sq->napi.weight || is_xdp_raw_buffer_queue(vi, index))
1816 if (__netif_tx_trylock(txq)) {
1818 __netif_tx_unlock(txq);
1823 virtqueue_disable_cb(sq->vq);
1824 free_old_xmit_skbs(sq, true);
1825 } while (unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
1827 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1828 netif_tx_wake_queue(txq);
1830 __netif_tx_unlock(txq);
1834 static int virtnet_poll(struct napi_struct *napi, int budget)
1836 struct receive_queue *rq =
1837 container_of(napi, struct receive_queue, napi);
1838 struct virtnet_info *vi = rq->vq->vdev->priv;
1839 struct send_queue *sq;
1840 unsigned int received;
1841 unsigned int xdp_xmit = 0;
1843 virtnet_poll_cleantx(rq);
1845 received = virtnet_receive(rq, budget, &xdp_xmit);
1847 if (xdp_xmit & VIRTIO_XDP_REDIR)
1850 /* Out of packets? */
1851 if (received < budget)
1852 virtqueue_napi_complete(napi, rq->vq, received);
1854 if (xdp_xmit & VIRTIO_XDP_TX) {
1855 sq = virtnet_xdp_get_sq(vi);
1856 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
1857 u64_stats_update_begin(&sq->stats.syncp);
1859 u64_stats_update_end(&sq->stats.syncp);
1861 virtnet_xdp_put_sq(vi, sq);
1867 static int virtnet_open(struct net_device *dev)
1869 struct virtnet_info *vi = netdev_priv(dev);
1872 enable_delayed_refill(vi);
1874 for (i = 0; i < vi->max_queue_pairs; i++) {
1875 if (i < vi->curr_queue_pairs)
1876 /* Make sure we have some buffers: if oom use wq. */
1877 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1878 schedule_delayed_work(&vi->refill, 0);
1880 err = xdp_rxq_info_reg(&vi->rq[i].xdp_rxq, dev, i, vi->rq[i].napi.napi_id);
1884 err = xdp_rxq_info_reg_mem_model(&vi->rq[i].xdp_rxq,
1885 MEM_TYPE_PAGE_SHARED, NULL);
1887 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
1891 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1892 virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi);
1898 static int virtnet_poll_tx(struct napi_struct *napi, int budget)
1900 struct send_queue *sq = container_of(napi, struct send_queue, napi);
1901 struct virtnet_info *vi = sq->vq->vdev->priv;
1902 unsigned int index = vq2txq(sq->vq);
1903 struct netdev_queue *txq;
1907 if (unlikely(is_xdp_raw_buffer_queue(vi, index))) {
1908 /* We don't need to enable cb for XDP */
1909 napi_complete_done(napi, 0);
1913 txq = netdev_get_tx_queue(vi->dev, index);
1914 __netif_tx_lock(txq, raw_smp_processor_id());
1915 virtqueue_disable_cb(sq->vq);
1916 free_old_xmit_skbs(sq, true);
1918 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1919 netif_tx_wake_queue(txq);
1921 opaque = virtqueue_enable_cb_prepare(sq->vq);
1923 done = napi_complete_done(napi, 0);
1926 virtqueue_disable_cb(sq->vq);
1928 __netif_tx_unlock(txq);
1931 if (unlikely(virtqueue_poll(sq->vq, opaque))) {
1932 if (napi_schedule_prep(napi)) {
1933 __netif_tx_lock(txq, raw_smp_processor_id());
1934 virtqueue_disable_cb(sq->vq);
1935 __netif_tx_unlock(txq);
1936 __napi_schedule(napi);
1944 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
1946 struct virtio_net_hdr_mrg_rxbuf *hdr;
1947 const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
1948 struct virtnet_info *vi = sq->vq->vdev->priv;
1950 unsigned hdr_len = vi->hdr_len;
1953 pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
1955 can_push = vi->any_header_sg &&
1956 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
1957 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
1958 /* Even if we can, don't push here yet as this would skew
1959 * csum_start offset below. */
1961 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1963 hdr = skb_vnet_hdr(skb);
1965 if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1966 virtio_is_little_endian(vi->vdev), false,
1970 if (vi->mergeable_rx_bufs)
1971 hdr->num_buffers = 0;
1973 sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1975 __skb_push(skb, hdr_len);
1976 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1977 if (unlikely(num_sg < 0))
1979 /* Pull header back to avoid skew in tx bytes calculations. */
1980 __skb_pull(skb, hdr_len);
1982 sg_set_buf(sq->sg, hdr, hdr_len);
1983 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
1984 if (unlikely(num_sg < 0))
1988 return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
1991 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
1993 struct virtnet_info *vi = netdev_priv(dev);
1994 int qnum = skb_get_queue_mapping(skb);
1995 struct send_queue *sq = &vi->sq[qnum];
1997 struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1998 bool kick = !netdev_xmit_more();
1999 bool use_napi = sq->napi.weight;
2001 /* Free up any pending old buffers before queueing new ones. */
2004 virtqueue_disable_cb(sq->vq);
2006 free_old_xmit_skbs(sq, false);
2008 } while (use_napi && kick &&
2009 unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
2011 /* timestamp packet in software */
2012 skb_tx_timestamp(skb);
2014 /* Try to transmit */
2015 err = xmit_skb(sq, skb);
2017 /* This should not happen! */
2018 if (unlikely(err)) {
2019 dev->stats.tx_fifo_errors++;
2020 if (net_ratelimit())
2022 "Unexpected TXQ (%d) queue failure: %d\n",
2024 dev->stats.tx_dropped++;
2025 dev_kfree_skb_any(skb);
2026 return NETDEV_TX_OK;
2029 /* Don't wait up for transmitted skbs to be freed. */
2035 check_sq_full_and_disable(vi, dev, sq);
2037 if (kick || netif_xmit_stopped(txq)) {
2038 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
2039 u64_stats_update_begin(&sq->stats.syncp);
2041 u64_stats_update_end(&sq->stats.syncp);
2045 return NETDEV_TX_OK;
2048 static int virtnet_rx_resize(struct virtnet_info *vi,
2049 struct receive_queue *rq, u32 ring_num)
2051 bool running = netif_running(vi->dev);
2054 qindex = rq - vi->rq;
2057 napi_disable(&rq->napi);
2059 err = virtqueue_resize(rq->vq, ring_num, virtnet_rq_free_unused_buf);
2061 netdev_err(vi->dev, "resize rx fail: rx queue index: %d err: %d\n", qindex, err);
2063 if (!try_fill_recv(vi, rq, GFP_KERNEL))
2064 schedule_delayed_work(&vi->refill, 0);
2067 virtnet_napi_enable(rq->vq, &rq->napi);
2071 static int virtnet_tx_resize(struct virtnet_info *vi,
2072 struct send_queue *sq, u32 ring_num)
2074 bool running = netif_running(vi->dev);
2075 struct netdev_queue *txq;
2078 qindex = sq - vi->sq;
2081 virtnet_napi_tx_disable(&sq->napi);
2083 txq = netdev_get_tx_queue(vi->dev, qindex);
2085 /* 1. wait all ximt complete
2086 * 2. fix the race of netif_stop_subqueue() vs netif_start_subqueue()
2088 __netif_tx_lock_bh(txq);
2090 /* Prevent rx poll from accessing sq. */
2093 /* Prevent the upper layer from trying to send packets. */
2094 netif_stop_subqueue(vi->dev, qindex);
2096 __netif_tx_unlock_bh(txq);
2098 err = virtqueue_resize(sq->vq, ring_num, virtnet_sq_free_unused_buf);
2100 netdev_err(vi->dev, "resize tx fail: tx queue index: %d err: %d\n", qindex, err);
2102 __netif_tx_lock_bh(txq);
2104 netif_tx_wake_queue(txq);
2105 __netif_tx_unlock_bh(txq);
2108 virtnet_napi_tx_enable(vi, sq->vq, &sq->napi);
2113 * Send command via the control virtqueue and check status. Commands
2114 * supported by the hypervisor, as indicated by feature bits, should
2115 * never fail unless improperly formatted.
2117 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
2118 struct scatterlist *out)
2120 struct scatterlist *sgs[4], hdr, stat;
2121 unsigned out_num = 0, tmp;
2124 /* Caller should know better */
2125 BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
2127 vi->ctrl->status = ~0;
2128 vi->ctrl->hdr.class = class;
2129 vi->ctrl->hdr.cmd = cmd;
2131 sg_init_one(&hdr, &vi->ctrl->hdr, sizeof(vi->ctrl->hdr));
2132 sgs[out_num++] = &hdr;
2135 sgs[out_num++] = out;
2137 /* Add return status. */
2138 sg_init_one(&stat, &vi->ctrl->status, sizeof(vi->ctrl->status));
2139 sgs[out_num] = &stat;
2141 BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
2142 ret = virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
2144 dev_warn(&vi->vdev->dev,
2145 "Failed to add sgs for command vq: %d\n.", ret);
2149 if (unlikely(!virtqueue_kick(vi->cvq)))
2150 return vi->ctrl->status == VIRTIO_NET_OK;
2152 /* Spin for a response, the kick causes an ioport write, trapping
2153 * into the hypervisor, so the request should be handled immediately.
2155 while (!virtqueue_get_buf(vi->cvq, &tmp) &&
2156 !virtqueue_is_broken(vi->cvq))
2159 return vi->ctrl->status == VIRTIO_NET_OK;
2162 static int virtnet_set_mac_address(struct net_device *dev, void *p)
2164 struct virtnet_info *vi = netdev_priv(dev);
2165 struct virtio_device *vdev = vi->vdev;
2167 struct sockaddr *addr;
2168 struct scatterlist sg;
2170 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
2173 addr = kmemdup(p, sizeof(*addr), GFP_KERNEL);
2177 ret = eth_prepare_mac_addr_change(dev, addr);
2181 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
2182 sg_init_one(&sg, addr->sa_data, dev->addr_len);
2183 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
2184 VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
2185 dev_warn(&vdev->dev,
2186 "Failed to set mac address by vq command.\n");
2190 } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
2191 !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
2194 /* Naturally, this has an atomicity problem. */
2195 for (i = 0; i < dev->addr_len; i++)
2196 virtio_cwrite8(vdev,
2197 offsetof(struct virtio_net_config, mac) +
2198 i, addr->sa_data[i]);
2201 eth_commit_mac_addr_change(dev, p);
2209 static void virtnet_stats(struct net_device *dev,
2210 struct rtnl_link_stats64 *tot)
2212 struct virtnet_info *vi = netdev_priv(dev);
2216 for (i = 0; i < vi->max_queue_pairs; i++) {
2217 u64 tpackets, tbytes, terrors, rpackets, rbytes, rdrops;
2218 struct receive_queue *rq = &vi->rq[i];
2219 struct send_queue *sq = &vi->sq[i];
2222 start = u64_stats_fetch_begin(&sq->stats.syncp);
2223 tpackets = sq->stats.packets;
2224 tbytes = sq->stats.bytes;
2225 terrors = sq->stats.tx_timeouts;
2226 } while (u64_stats_fetch_retry(&sq->stats.syncp, start));
2229 start = u64_stats_fetch_begin(&rq->stats.syncp);
2230 rpackets = rq->stats.packets;
2231 rbytes = rq->stats.bytes;
2232 rdrops = rq->stats.drops;
2233 } while (u64_stats_fetch_retry(&rq->stats.syncp, start));
2235 tot->rx_packets += rpackets;
2236 tot->tx_packets += tpackets;
2237 tot->rx_bytes += rbytes;
2238 tot->tx_bytes += tbytes;
2239 tot->rx_dropped += rdrops;
2240 tot->tx_errors += terrors;
2243 tot->tx_dropped = dev->stats.tx_dropped;
2244 tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
2245 tot->rx_length_errors = dev->stats.rx_length_errors;
2246 tot->rx_frame_errors = dev->stats.rx_frame_errors;
2249 static void virtnet_ack_link_announce(struct virtnet_info *vi)
2252 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
2253 VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
2254 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
2258 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
2260 struct scatterlist sg;
2261 struct net_device *dev = vi->dev;
2263 if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
2266 vi->ctrl->mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
2267 sg_init_one(&sg, &vi->ctrl->mq, sizeof(vi->ctrl->mq));
2269 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
2270 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
2271 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
2275 vi->curr_queue_pairs = queue_pairs;
2276 /* virtnet_open() will refill when device is going to up. */
2277 if (dev->flags & IFF_UP)
2278 schedule_delayed_work(&vi->refill, 0);
2284 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
2289 err = _virtnet_set_queues(vi, queue_pairs);
2294 static int virtnet_close(struct net_device *dev)
2296 struct virtnet_info *vi = netdev_priv(dev);
2299 /* Make sure NAPI doesn't schedule refill work */
2300 disable_delayed_refill(vi);
2301 /* Make sure refill_work doesn't re-enable napi! */
2302 cancel_delayed_work_sync(&vi->refill);
2304 for (i = 0; i < vi->max_queue_pairs; i++) {
2305 virtnet_napi_tx_disable(&vi->sq[i].napi);
2306 napi_disable(&vi->rq[i].napi);
2307 xdp_rxq_info_unreg(&vi->rq[i].xdp_rxq);
2313 static void virtnet_set_rx_mode(struct net_device *dev)
2315 struct virtnet_info *vi = netdev_priv(dev);
2316 struct scatterlist sg[2];
2317 struct virtio_net_ctrl_mac *mac_data;
2318 struct netdev_hw_addr *ha;
2324 /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
2325 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
2328 vi->ctrl->promisc = ((dev->flags & IFF_PROMISC) != 0);
2329 vi->ctrl->allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
2331 sg_init_one(sg, &vi->ctrl->promisc, sizeof(vi->ctrl->promisc));
2333 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
2334 VIRTIO_NET_CTRL_RX_PROMISC, sg))
2335 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
2336 vi->ctrl->promisc ? "en" : "dis");
2338 sg_init_one(sg, &vi->ctrl->allmulti, sizeof(vi->ctrl->allmulti));
2340 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
2341 VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
2342 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
2343 vi->ctrl->allmulti ? "en" : "dis");
2345 uc_count = netdev_uc_count(dev);
2346 mc_count = netdev_mc_count(dev);
2347 /* MAC filter - use one buffer for both lists */
2348 buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
2349 (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
2354 sg_init_table(sg, 2);
2356 /* Store the unicast list and count in the front of the buffer */
2357 mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
2359 netdev_for_each_uc_addr(ha, dev)
2360 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
2362 sg_set_buf(&sg[0], mac_data,
2363 sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
2365 /* multicast list and count fill the end */
2366 mac_data = (void *)&mac_data->macs[uc_count][0];
2368 mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
2370 netdev_for_each_mc_addr(ha, dev)
2371 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
2373 sg_set_buf(&sg[1], mac_data,
2374 sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
2376 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
2377 VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
2378 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
2383 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
2384 __be16 proto, u16 vid)
2386 struct virtnet_info *vi = netdev_priv(dev);
2387 struct scatterlist sg;
2389 vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
2390 sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
2392 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
2393 VIRTIO_NET_CTRL_VLAN_ADD, &sg))
2394 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
2398 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
2399 __be16 proto, u16 vid)
2401 struct virtnet_info *vi = netdev_priv(dev);
2402 struct scatterlist sg;
2404 vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
2405 sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
2407 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
2408 VIRTIO_NET_CTRL_VLAN_DEL, &sg))
2409 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
2413 static void virtnet_clean_affinity(struct virtnet_info *vi)
2417 if (vi->affinity_hint_set) {
2418 for (i = 0; i < vi->max_queue_pairs; i++) {
2419 virtqueue_set_affinity(vi->rq[i].vq, NULL);
2420 virtqueue_set_affinity(vi->sq[i].vq, NULL);
2423 vi->affinity_hint_set = false;
2427 static void virtnet_set_affinity(struct virtnet_info *vi)
2436 if (!zalloc_cpumask_var(&mask, GFP_KERNEL)) {
2437 virtnet_clean_affinity(vi);
2441 num_cpu = num_online_cpus();
2442 stride = max_t(int, num_cpu / vi->curr_queue_pairs, 1);
2443 stragglers = num_cpu >= vi->curr_queue_pairs ?
2444 num_cpu % vi->curr_queue_pairs :
2446 cpu = cpumask_first(cpu_online_mask);
2448 for (i = 0; i < vi->curr_queue_pairs; i++) {
2449 group_size = stride + (i < stragglers ? 1 : 0);
2451 for (j = 0; j < group_size; j++) {
2452 cpumask_set_cpu(cpu, mask);
2453 cpu = cpumask_next_wrap(cpu, cpu_online_mask,
2456 virtqueue_set_affinity(vi->rq[i].vq, mask);
2457 virtqueue_set_affinity(vi->sq[i].vq, mask);
2458 __netif_set_xps_queue(vi->dev, cpumask_bits(mask), i, XPS_CPUS);
2459 cpumask_clear(mask);
2462 vi->affinity_hint_set = true;
2463 free_cpumask_var(mask);
2466 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
2468 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2470 virtnet_set_affinity(vi);
2474 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
2476 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2478 virtnet_set_affinity(vi);
2482 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
2484 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2487 virtnet_clean_affinity(vi);
2491 static enum cpuhp_state virtionet_online;
2493 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
2497 ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
2500 ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2504 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2508 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
2510 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2511 cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2515 static void virtnet_get_ringparam(struct net_device *dev,
2516 struct ethtool_ringparam *ring,
2517 struct kernel_ethtool_ringparam *kernel_ring,
2518 struct netlink_ext_ack *extack)
2520 struct virtnet_info *vi = netdev_priv(dev);
2522 ring->rx_max_pending = vi->rq[0].vq->num_max;
2523 ring->tx_max_pending = vi->sq[0].vq->num_max;
2524 ring->rx_pending = virtqueue_get_vring_size(vi->rq[0].vq);
2525 ring->tx_pending = virtqueue_get_vring_size(vi->sq[0].vq);
2528 static int virtnet_set_ringparam(struct net_device *dev,
2529 struct ethtool_ringparam *ring,
2530 struct kernel_ethtool_ringparam *kernel_ring,
2531 struct netlink_ext_ack *extack)
2533 struct virtnet_info *vi = netdev_priv(dev);
2534 u32 rx_pending, tx_pending;
2535 struct receive_queue *rq;
2536 struct send_queue *sq;
2539 if (ring->rx_mini_pending || ring->rx_jumbo_pending)
2542 rx_pending = virtqueue_get_vring_size(vi->rq[0].vq);
2543 tx_pending = virtqueue_get_vring_size(vi->sq[0].vq);
2545 if (ring->rx_pending == rx_pending &&
2546 ring->tx_pending == tx_pending)
2549 if (ring->rx_pending > vi->rq[0].vq->num_max)
2552 if (ring->tx_pending > vi->sq[0].vq->num_max)
2555 for (i = 0; i < vi->max_queue_pairs; i++) {
2559 if (ring->tx_pending != tx_pending) {
2560 err = virtnet_tx_resize(vi, sq, ring->tx_pending);
2565 if (ring->rx_pending != rx_pending) {
2566 err = virtnet_rx_resize(vi, rq, ring->rx_pending);
2575 static bool virtnet_commit_rss_command(struct virtnet_info *vi)
2577 struct net_device *dev = vi->dev;
2578 struct scatterlist sgs[4];
2579 unsigned int sg_buf_size;
2582 sg_init_table(sgs, 4);
2584 sg_buf_size = offsetof(struct virtio_net_ctrl_rss, indirection_table);
2585 sg_set_buf(&sgs[0], &vi->ctrl->rss, sg_buf_size);
2587 sg_buf_size = sizeof(uint16_t) * (vi->ctrl->rss.indirection_table_mask + 1);
2588 sg_set_buf(&sgs[1], vi->ctrl->rss.indirection_table, sg_buf_size);
2590 sg_buf_size = offsetof(struct virtio_net_ctrl_rss, key)
2591 - offsetof(struct virtio_net_ctrl_rss, max_tx_vq);
2592 sg_set_buf(&sgs[2], &vi->ctrl->rss.max_tx_vq, sg_buf_size);
2594 sg_buf_size = vi->rss_key_size;
2595 sg_set_buf(&sgs[3], vi->ctrl->rss.key, sg_buf_size);
2597 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
2598 vi->has_rss ? VIRTIO_NET_CTRL_MQ_RSS_CONFIG
2599 : VIRTIO_NET_CTRL_MQ_HASH_CONFIG, sgs)) {
2600 dev_warn(&dev->dev, "VIRTIONET issue with committing RSS sgs\n");
2606 static void virtnet_init_default_rss(struct virtnet_info *vi)
2611 vi->ctrl->rss.hash_types = vi->rss_hash_types_supported;
2612 vi->rss_hash_types_saved = vi->rss_hash_types_supported;
2613 vi->ctrl->rss.indirection_table_mask = vi->rss_indir_table_size
2614 ? vi->rss_indir_table_size - 1 : 0;
2615 vi->ctrl->rss.unclassified_queue = 0;
2617 for (; i < vi->rss_indir_table_size; ++i) {
2618 indir_val = ethtool_rxfh_indir_default(i, vi->curr_queue_pairs);
2619 vi->ctrl->rss.indirection_table[i] = indir_val;
2622 vi->ctrl->rss.max_tx_vq = vi->curr_queue_pairs;
2623 vi->ctrl->rss.hash_key_length = vi->rss_key_size;
2625 netdev_rss_key_fill(vi->ctrl->rss.key, vi->rss_key_size);
2628 static void virtnet_get_hashflow(const struct virtnet_info *vi, struct ethtool_rxnfc *info)
2631 switch (info->flow_type) {
2633 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_TCPv4) {
2634 info->data = RXH_IP_SRC | RXH_IP_DST |
2635 RXH_L4_B_0_1 | RXH_L4_B_2_3;
2636 } else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv4) {
2637 info->data = RXH_IP_SRC | RXH_IP_DST;
2641 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_TCPv6) {
2642 info->data = RXH_IP_SRC | RXH_IP_DST |
2643 RXH_L4_B_0_1 | RXH_L4_B_2_3;
2644 } else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv6) {
2645 info->data = RXH_IP_SRC | RXH_IP_DST;
2649 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_UDPv4) {
2650 info->data = RXH_IP_SRC | RXH_IP_DST |
2651 RXH_L4_B_0_1 | RXH_L4_B_2_3;
2652 } else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv4) {
2653 info->data = RXH_IP_SRC | RXH_IP_DST;
2657 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_UDPv6) {
2658 info->data = RXH_IP_SRC | RXH_IP_DST |
2659 RXH_L4_B_0_1 | RXH_L4_B_2_3;
2660 } else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv6) {
2661 info->data = RXH_IP_SRC | RXH_IP_DST;
2665 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv4)
2666 info->data = RXH_IP_SRC | RXH_IP_DST;
2670 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv6)
2671 info->data = RXH_IP_SRC | RXH_IP_DST;
2680 static bool virtnet_set_hashflow(struct virtnet_info *vi, struct ethtool_rxnfc *info)
2682 u32 new_hashtypes = vi->rss_hash_types_saved;
2683 bool is_disable = info->data & RXH_DISCARD;
2684 bool is_l4 = info->data == (RXH_IP_SRC | RXH_IP_DST | RXH_L4_B_0_1 | RXH_L4_B_2_3);
2686 /* supports only 'sd', 'sdfn' and 'r' */
2687 if (!((info->data == (RXH_IP_SRC | RXH_IP_DST)) | is_l4 | is_disable))
2690 switch (info->flow_type) {
2692 new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv4 | VIRTIO_NET_RSS_HASH_TYPE_TCPv4);
2694 new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv4
2695 | (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_TCPv4 : 0);
2698 new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv4 | VIRTIO_NET_RSS_HASH_TYPE_UDPv4);
2700 new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv4
2701 | (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_UDPv4 : 0);
2704 new_hashtypes &= ~VIRTIO_NET_RSS_HASH_TYPE_IPv4;
2706 new_hashtypes = VIRTIO_NET_RSS_HASH_TYPE_IPv4;
2709 new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv6 | VIRTIO_NET_RSS_HASH_TYPE_TCPv6);
2711 new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv6
2712 | (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_TCPv6 : 0);
2715 new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv6 | VIRTIO_NET_RSS_HASH_TYPE_UDPv6);
2717 new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv6
2718 | (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_UDPv6 : 0);
2721 new_hashtypes &= ~VIRTIO_NET_RSS_HASH_TYPE_IPv6;
2723 new_hashtypes = VIRTIO_NET_RSS_HASH_TYPE_IPv6;
2726 /* unsupported flow */
2730 /* if unsupported hashtype was set */
2731 if (new_hashtypes != (new_hashtypes & vi->rss_hash_types_supported))
2734 if (new_hashtypes != vi->rss_hash_types_saved) {
2735 vi->rss_hash_types_saved = new_hashtypes;
2736 vi->ctrl->rss.hash_types = vi->rss_hash_types_saved;
2737 if (vi->dev->features & NETIF_F_RXHASH)
2738 return virtnet_commit_rss_command(vi);
2744 static void virtnet_get_drvinfo(struct net_device *dev,
2745 struct ethtool_drvinfo *info)
2747 struct virtnet_info *vi = netdev_priv(dev);
2748 struct virtio_device *vdev = vi->vdev;
2750 strscpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
2751 strscpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
2752 strscpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
2756 /* TODO: Eliminate OOO packets during switching */
2757 static int virtnet_set_channels(struct net_device *dev,
2758 struct ethtool_channels *channels)
2760 struct virtnet_info *vi = netdev_priv(dev);
2761 u16 queue_pairs = channels->combined_count;
2764 /* We don't support separate rx/tx channels.
2765 * We don't allow setting 'other' channels.
2767 if (channels->rx_count || channels->tx_count || channels->other_count)
2770 if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
2773 /* For now we don't support modifying channels while XDP is loaded
2774 * also when XDP is loaded all RX queues have XDP programs so we only
2775 * need to check a single RX queue.
2777 if (vi->rq[0].xdp_prog)
2781 err = _virtnet_set_queues(vi, queue_pairs);
2786 virtnet_set_affinity(vi);
2789 netif_set_real_num_tx_queues(dev, queue_pairs);
2790 netif_set_real_num_rx_queues(dev, queue_pairs);
2795 static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *data)
2797 struct virtnet_info *vi = netdev_priv(dev);
2801 switch (stringset) {
2803 for (i = 0; i < vi->curr_queue_pairs; i++) {
2804 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++)
2805 ethtool_sprintf(&p, "rx_queue_%u_%s", i,
2806 virtnet_rq_stats_desc[j].desc);
2809 for (i = 0; i < vi->curr_queue_pairs; i++) {
2810 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++)
2811 ethtool_sprintf(&p, "tx_queue_%u_%s", i,
2812 virtnet_sq_stats_desc[j].desc);
2818 static int virtnet_get_sset_count(struct net_device *dev, int sset)
2820 struct virtnet_info *vi = netdev_priv(dev);
2824 return vi->curr_queue_pairs * (VIRTNET_RQ_STATS_LEN +
2825 VIRTNET_SQ_STATS_LEN);
2831 static void virtnet_get_ethtool_stats(struct net_device *dev,
2832 struct ethtool_stats *stats, u64 *data)
2834 struct virtnet_info *vi = netdev_priv(dev);
2835 unsigned int idx = 0, start, i, j;
2836 const u8 *stats_base;
2839 for (i = 0; i < vi->curr_queue_pairs; i++) {
2840 struct receive_queue *rq = &vi->rq[i];
2842 stats_base = (u8 *)&rq->stats;
2844 start = u64_stats_fetch_begin(&rq->stats.syncp);
2845 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
2846 offset = virtnet_rq_stats_desc[j].offset;
2847 data[idx + j] = *(u64 *)(stats_base + offset);
2849 } while (u64_stats_fetch_retry(&rq->stats.syncp, start));
2850 idx += VIRTNET_RQ_STATS_LEN;
2853 for (i = 0; i < vi->curr_queue_pairs; i++) {
2854 struct send_queue *sq = &vi->sq[i];
2856 stats_base = (u8 *)&sq->stats;
2858 start = u64_stats_fetch_begin(&sq->stats.syncp);
2859 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
2860 offset = virtnet_sq_stats_desc[j].offset;
2861 data[idx + j] = *(u64 *)(stats_base + offset);
2863 } while (u64_stats_fetch_retry(&sq->stats.syncp, start));
2864 idx += VIRTNET_SQ_STATS_LEN;
2868 static void virtnet_get_channels(struct net_device *dev,
2869 struct ethtool_channels *channels)
2871 struct virtnet_info *vi = netdev_priv(dev);
2873 channels->combined_count = vi->curr_queue_pairs;
2874 channels->max_combined = vi->max_queue_pairs;
2875 channels->max_other = 0;
2876 channels->rx_count = 0;
2877 channels->tx_count = 0;
2878 channels->other_count = 0;
2881 static int virtnet_set_link_ksettings(struct net_device *dev,
2882 const struct ethtool_link_ksettings *cmd)
2884 struct virtnet_info *vi = netdev_priv(dev);
2886 return ethtool_virtdev_set_link_ksettings(dev, cmd,
2887 &vi->speed, &vi->duplex);
2890 static int virtnet_get_link_ksettings(struct net_device *dev,
2891 struct ethtool_link_ksettings *cmd)
2893 struct virtnet_info *vi = netdev_priv(dev);
2895 cmd->base.speed = vi->speed;
2896 cmd->base.duplex = vi->duplex;
2897 cmd->base.port = PORT_OTHER;
2902 static int virtnet_send_notf_coal_cmds(struct virtnet_info *vi,
2903 struct ethtool_coalesce *ec)
2905 struct scatterlist sgs_tx, sgs_rx;
2906 struct virtio_net_ctrl_coal_tx coal_tx;
2907 struct virtio_net_ctrl_coal_rx coal_rx;
2909 coal_tx.tx_usecs = cpu_to_le32(ec->tx_coalesce_usecs);
2910 coal_tx.tx_max_packets = cpu_to_le32(ec->tx_max_coalesced_frames);
2911 sg_init_one(&sgs_tx, &coal_tx, sizeof(coal_tx));
2913 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_NOTF_COAL,
2914 VIRTIO_NET_CTRL_NOTF_COAL_TX_SET,
2918 /* Save parameters */
2919 vi->tx_usecs = ec->tx_coalesce_usecs;
2920 vi->tx_max_packets = ec->tx_max_coalesced_frames;
2922 coal_rx.rx_usecs = cpu_to_le32(ec->rx_coalesce_usecs);
2923 coal_rx.rx_max_packets = cpu_to_le32(ec->rx_max_coalesced_frames);
2924 sg_init_one(&sgs_rx, &coal_rx, sizeof(coal_rx));
2926 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_NOTF_COAL,
2927 VIRTIO_NET_CTRL_NOTF_COAL_RX_SET,
2931 /* Save parameters */
2932 vi->rx_usecs = ec->rx_coalesce_usecs;
2933 vi->rx_max_packets = ec->rx_max_coalesced_frames;
2938 static int virtnet_coal_params_supported(struct ethtool_coalesce *ec)
2940 /* usecs coalescing is supported only if VIRTIO_NET_F_NOTF_COAL
2941 * feature is negotiated.
2943 if (ec->rx_coalesce_usecs || ec->tx_coalesce_usecs)
2946 if (ec->tx_max_coalesced_frames > 1 ||
2947 ec->rx_max_coalesced_frames != 1)
2953 static int virtnet_set_coalesce(struct net_device *dev,
2954 struct ethtool_coalesce *ec,
2955 struct kernel_ethtool_coalesce *kernel_coal,
2956 struct netlink_ext_ack *extack)
2958 struct virtnet_info *vi = netdev_priv(dev);
2959 int ret, i, napi_weight;
2960 bool update_napi = false;
2962 /* Can't change NAPI weight if the link is up */
2963 napi_weight = ec->tx_max_coalesced_frames ? NAPI_POLL_WEIGHT : 0;
2964 if (napi_weight ^ vi->sq[0].napi.weight) {
2965 if (dev->flags & IFF_UP)
2971 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_NOTF_COAL))
2972 ret = virtnet_send_notf_coal_cmds(vi, ec);
2974 ret = virtnet_coal_params_supported(ec);
2980 for (i = 0; i < vi->max_queue_pairs; i++)
2981 vi->sq[i].napi.weight = napi_weight;
2987 static int virtnet_get_coalesce(struct net_device *dev,
2988 struct ethtool_coalesce *ec,
2989 struct kernel_ethtool_coalesce *kernel_coal,
2990 struct netlink_ext_ack *extack)
2992 struct virtnet_info *vi = netdev_priv(dev);
2994 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_NOTF_COAL)) {
2995 ec->rx_coalesce_usecs = vi->rx_usecs;
2996 ec->tx_coalesce_usecs = vi->tx_usecs;
2997 ec->tx_max_coalesced_frames = vi->tx_max_packets;
2998 ec->rx_max_coalesced_frames = vi->rx_max_packets;
3000 ec->rx_max_coalesced_frames = 1;
3002 if (vi->sq[0].napi.weight)
3003 ec->tx_max_coalesced_frames = 1;
3009 static void virtnet_init_settings(struct net_device *dev)
3011 struct virtnet_info *vi = netdev_priv(dev);
3013 vi->speed = SPEED_UNKNOWN;
3014 vi->duplex = DUPLEX_UNKNOWN;
3017 static void virtnet_update_settings(struct virtnet_info *vi)
3022 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_SPEED_DUPLEX))
3025 virtio_cread_le(vi->vdev, struct virtio_net_config, speed, &speed);
3027 if (ethtool_validate_speed(speed))
3030 virtio_cread_le(vi->vdev, struct virtio_net_config, duplex, &duplex);
3032 if (ethtool_validate_duplex(duplex))
3033 vi->duplex = duplex;
3036 static u32 virtnet_get_rxfh_key_size(struct net_device *dev)
3038 return ((struct virtnet_info *)netdev_priv(dev))->rss_key_size;
3041 static u32 virtnet_get_rxfh_indir_size(struct net_device *dev)
3043 return ((struct virtnet_info *)netdev_priv(dev))->rss_indir_table_size;
3046 static int virtnet_get_rxfh(struct net_device *dev, u32 *indir, u8 *key, u8 *hfunc)
3048 struct virtnet_info *vi = netdev_priv(dev);
3052 for (i = 0; i < vi->rss_indir_table_size; ++i)
3053 indir[i] = vi->ctrl->rss.indirection_table[i];
3057 memcpy(key, vi->ctrl->rss.key, vi->rss_key_size);
3060 *hfunc = ETH_RSS_HASH_TOP;
3065 static int virtnet_set_rxfh(struct net_device *dev, const u32 *indir, const u8 *key, const u8 hfunc)
3067 struct virtnet_info *vi = netdev_priv(dev);
3070 if (hfunc != ETH_RSS_HASH_NO_CHANGE && hfunc != ETH_RSS_HASH_TOP)
3074 for (i = 0; i < vi->rss_indir_table_size; ++i)
3075 vi->ctrl->rss.indirection_table[i] = indir[i];
3078 memcpy(vi->ctrl->rss.key, key, vi->rss_key_size);
3080 virtnet_commit_rss_command(vi);
3085 static int virtnet_get_rxnfc(struct net_device *dev, struct ethtool_rxnfc *info, u32 *rule_locs)
3087 struct virtnet_info *vi = netdev_priv(dev);
3090 switch (info->cmd) {
3091 case ETHTOOL_GRXRINGS:
3092 info->data = vi->curr_queue_pairs;
3095 virtnet_get_hashflow(vi, info);
3104 static int virtnet_set_rxnfc(struct net_device *dev, struct ethtool_rxnfc *info)
3106 struct virtnet_info *vi = netdev_priv(dev);
3109 switch (info->cmd) {
3111 if (!virtnet_set_hashflow(vi, info))
3122 static const struct ethtool_ops virtnet_ethtool_ops = {
3123 .supported_coalesce_params = ETHTOOL_COALESCE_MAX_FRAMES |
3124 ETHTOOL_COALESCE_USECS,
3125 .get_drvinfo = virtnet_get_drvinfo,
3126 .get_link = ethtool_op_get_link,
3127 .get_ringparam = virtnet_get_ringparam,
3128 .set_ringparam = virtnet_set_ringparam,
3129 .get_strings = virtnet_get_strings,
3130 .get_sset_count = virtnet_get_sset_count,
3131 .get_ethtool_stats = virtnet_get_ethtool_stats,
3132 .set_channels = virtnet_set_channels,
3133 .get_channels = virtnet_get_channels,
3134 .get_ts_info = ethtool_op_get_ts_info,
3135 .get_link_ksettings = virtnet_get_link_ksettings,
3136 .set_link_ksettings = virtnet_set_link_ksettings,
3137 .set_coalesce = virtnet_set_coalesce,
3138 .get_coalesce = virtnet_get_coalesce,
3139 .get_rxfh_key_size = virtnet_get_rxfh_key_size,
3140 .get_rxfh_indir_size = virtnet_get_rxfh_indir_size,
3141 .get_rxfh = virtnet_get_rxfh,
3142 .set_rxfh = virtnet_set_rxfh,
3143 .get_rxnfc = virtnet_get_rxnfc,
3144 .set_rxnfc = virtnet_set_rxnfc,
3147 static void virtnet_freeze_down(struct virtio_device *vdev)
3149 struct virtnet_info *vi = vdev->priv;
3151 /* Make sure no work handler is accessing the device */
3152 flush_work(&vi->config_work);
3154 netif_tx_lock_bh(vi->dev);
3155 netif_device_detach(vi->dev);
3156 netif_tx_unlock_bh(vi->dev);
3157 if (netif_running(vi->dev))
3158 virtnet_close(vi->dev);
3161 static int init_vqs(struct virtnet_info *vi);
3163 static int virtnet_restore_up(struct virtio_device *vdev)
3165 struct virtnet_info *vi = vdev->priv;
3172 virtio_device_ready(vdev);
3174 enable_delayed_refill(vi);
3176 if (netif_running(vi->dev)) {
3177 err = virtnet_open(vi->dev);
3182 netif_tx_lock_bh(vi->dev);
3183 netif_device_attach(vi->dev);
3184 netif_tx_unlock_bh(vi->dev);
3188 static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads)
3190 struct scatterlist sg;
3191 vi->ctrl->offloads = cpu_to_virtio64(vi->vdev, offloads);
3193 sg_init_one(&sg, &vi->ctrl->offloads, sizeof(vi->ctrl->offloads));
3195 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
3196 VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) {
3197 dev_warn(&vi->dev->dev, "Fail to set guest offload.\n");
3204 static int virtnet_clear_guest_offloads(struct virtnet_info *vi)
3208 if (!vi->guest_offloads)
3211 return virtnet_set_guest_offloads(vi, offloads);
3214 static int virtnet_restore_guest_offloads(struct virtnet_info *vi)
3216 u64 offloads = vi->guest_offloads;
3218 if (!vi->guest_offloads)
3221 return virtnet_set_guest_offloads(vi, offloads);
3224 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
3225 struct netlink_ext_ack *extack)
3227 unsigned int room = SKB_DATA_ALIGN(VIRTIO_XDP_HEADROOM +
3228 sizeof(struct skb_shared_info));
3229 unsigned int max_sz = PAGE_SIZE - room - ETH_HLEN;
3230 struct virtnet_info *vi = netdev_priv(dev);
3231 struct bpf_prog *old_prog;
3232 u16 xdp_qp = 0, curr_qp;
3235 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)
3236 && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3237 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
3238 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
3239 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) ||
3240 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM) ||
3241 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_USO4) ||
3242 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_USO6))) {
3243 NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing GRO_HW/CSUM, disable GRO_HW/CSUM first");
3247 if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
3248 NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required");
3252 if (prog && !prog->aux->xdp_has_frags && dev->mtu > max_sz) {
3253 NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP without frags");
3254 netdev_warn(dev, "single-buffer XDP requires MTU less than %u\n", max_sz);
3258 curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
3260 xdp_qp = nr_cpu_ids;
3262 /* XDP requires extra queues for XDP_TX */
3263 if (curr_qp + xdp_qp > vi->max_queue_pairs) {
3264 netdev_warn_once(dev, "XDP request %i queues but max is %i. XDP_TX and XDP_REDIRECT will operate in a slower locked tx mode.\n",
3265 curr_qp + xdp_qp, vi->max_queue_pairs);
3269 old_prog = rtnl_dereference(vi->rq[0].xdp_prog);
3270 if (!prog && !old_prog)
3274 bpf_prog_add(prog, vi->max_queue_pairs - 1);
3276 /* Make sure NAPI is not using any XDP TX queues for RX. */
3277 if (netif_running(dev)) {
3278 for (i = 0; i < vi->max_queue_pairs; i++) {
3279 napi_disable(&vi->rq[i].napi);
3280 virtnet_napi_tx_disable(&vi->sq[i].napi);
3285 for (i = 0; i < vi->max_queue_pairs; i++) {
3286 rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
3288 virtnet_restore_guest_offloads(vi);
3293 err = _virtnet_set_queues(vi, curr_qp + xdp_qp);
3296 netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
3297 vi->xdp_queue_pairs = xdp_qp;
3300 vi->xdp_enabled = true;
3301 for (i = 0; i < vi->max_queue_pairs; i++) {
3302 rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
3303 if (i == 0 && !old_prog)
3304 virtnet_clear_guest_offloads(vi);
3307 xdp_features_set_redirect_target(dev, true);
3309 xdp_features_clear_redirect_target(dev);
3310 vi->xdp_enabled = false;
3313 for (i = 0; i < vi->max_queue_pairs; i++) {
3315 bpf_prog_put(old_prog);
3316 if (netif_running(dev)) {
3317 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
3318 virtnet_napi_tx_enable(vi, vi->sq[i].vq,
3327 virtnet_clear_guest_offloads(vi);
3328 for (i = 0; i < vi->max_queue_pairs; i++)
3329 rcu_assign_pointer(vi->rq[i].xdp_prog, old_prog);
3332 if (netif_running(dev)) {
3333 for (i = 0; i < vi->max_queue_pairs; i++) {
3334 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
3335 virtnet_napi_tx_enable(vi, vi->sq[i].vq,
3340 bpf_prog_sub(prog, vi->max_queue_pairs - 1);
3344 static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp)
3346 switch (xdp->command) {
3347 case XDP_SETUP_PROG:
3348 return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
3354 static int virtnet_get_phys_port_name(struct net_device *dev, char *buf,
3357 struct virtnet_info *vi = netdev_priv(dev);
3360 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
3363 ret = snprintf(buf, len, "sby");
3370 static int virtnet_set_features(struct net_device *dev,
3371 netdev_features_t features)
3373 struct virtnet_info *vi = netdev_priv(dev);
3377 if ((dev->features ^ features) & NETIF_F_GRO_HW) {
3378 if (vi->xdp_enabled)
3381 if (features & NETIF_F_GRO_HW)
3382 offloads = vi->guest_offloads_capable;
3384 offloads = vi->guest_offloads_capable &
3385 ~GUEST_OFFLOAD_GRO_HW_MASK;
3387 err = virtnet_set_guest_offloads(vi, offloads);
3390 vi->guest_offloads = offloads;
3393 if ((dev->features ^ features) & NETIF_F_RXHASH) {
3394 if (features & NETIF_F_RXHASH)
3395 vi->ctrl->rss.hash_types = vi->rss_hash_types_saved;
3397 vi->ctrl->rss.hash_types = VIRTIO_NET_HASH_REPORT_NONE;
3399 if (!virtnet_commit_rss_command(vi))
3406 static void virtnet_tx_timeout(struct net_device *dev, unsigned int txqueue)
3408 struct virtnet_info *priv = netdev_priv(dev);
3409 struct send_queue *sq = &priv->sq[txqueue];
3410 struct netdev_queue *txq = netdev_get_tx_queue(dev, txqueue);
3412 u64_stats_update_begin(&sq->stats.syncp);
3413 sq->stats.tx_timeouts++;
3414 u64_stats_update_end(&sq->stats.syncp);
3416 netdev_err(dev, "TX timeout on queue: %u, sq: %s, vq: 0x%x, name: %s, %u usecs ago\n",
3417 txqueue, sq->name, sq->vq->index, sq->vq->name,
3418 jiffies_to_usecs(jiffies - READ_ONCE(txq->trans_start)));
3421 static const struct net_device_ops virtnet_netdev = {
3422 .ndo_open = virtnet_open,
3423 .ndo_stop = virtnet_close,
3424 .ndo_start_xmit = start_xmit,
3425 .ndo_validate_addr = eth_validate_addr,
3426 .ndo_set_mac_address = virtnet_set_mac_address,
3427 .ndo_set_rx_mode = virtnet_set_rx_mode,
3428 .ndo_get_stats64 = virtnet_stats,
3429 .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
3430 .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
3431 .ndo_bpf = virtnet_xdp,
3432 .ndo_xdp_xmit = virtnet_xdp_xmit,
3433 .ndo_features_check = passthru_features_check,
3434 .ndo_get_phys_port_name = virtnet_get_phys_port_name,
3435 .ndo_set_features = virtnet_set_features,
3436 .ndo_tx_timeout = virtnet_tx_timeout,
3439 static void virtnet_config_changed_work(struct work_struct *work)
3441 struct virtnet_info *vi =
3442 container_of(work, struct virtnet_info, config_work);
3445 if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
3446 struct virtio_net_config, status, &v) < 0)
3449 if (v & VIRTIO_NET_S_ANNOUNCE) {
3450 netdev_notify_peers(vi->dev);
3451 virtnet_ack_link_announce(vi);
3454 /* Ignore unknown (future) status bits */
3455 v &= VIRTIO_NET_S_LINK_UP;
3457 if (vi->status == v)
3462 if (vi->status & VIRTIO_NET_S_LINK_UP) {
3463 virtnet_update_settings(vi);
3464 netif_carrier_on(vi->dev);
3465 netif_tx_wake_all_queues(vi->dev);
3467 netif_carrier_off(vi->dev);
3468 netif_tx_stop_all_queues(vi->dev);
3472 static void virtnet_config_changed(struct virtio_device *vdev)
3474 struct virtnet_info *vi = vdev->priv;
3476 schedule_work(&vi->config_work);
3479 static void virtnet_free_queues(struct virtnet_info *vi)
3483 for (i = 0; i < vi->max_queue_pairs; i++) {
3484 __netif_napi_del(&vi->rq[i].napi);
3485 __netif_napi_del(&vi->sq[i].napi);
3488 /* We called __netif_napi_del(),
3489 * we need to respect an RCU grace period before freeing vi->rq
3498 static void _free_receive_bufs(struct virtnet_info *vi)
3500 struct bpf_prog *old_prog;
3503 for (i = 0; i < vi->max_queue_pairs; i++) {
3504 while (vi->rq[i].pages)
3505 __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
3507 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
3508 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
3510 bpf_prog_put(old_prog);
3514 static void free_receive_bufs(struct virtnet_info *vi)
3517 _free_receive_bufs(vi);
3521 static void free_receive_page_frags(struct virtnet_info *vi)
3524 for (i = 0; i < vi->max_queue_pairs; i++)
3525 if (vi->rq[i].alloc_frag.page)
3526 put_page(vi->rq[i].alloc_frag.page);
3529 static void virtnet_sq_free_unused_buf(struct virtqueue *vq, void *buf)
3531 if (!is_xdp_frame(buf))
3534 xdp_return_frame(ptr_to_xdp(buf));
3537 static void virtnet_rq_free_unused_buf(struct virtqueue *vq, void *buf)
3539 struct virtnet_info *vi = vq->vdev->priv;
3542 if (vi->mergeable_rx_bufs)
3543 put_page(virt_to_head_page(buf));
3544 else if (vi->big_packets)
3545 give_pages(&vi->rq[i], buf);
3547 put_page(virt_to_head_page(buf));
3550 static void free_unused_bufs(struct virtnet_info *vi)
3555 for (i = 0; i < vi->max_queue_pairs; i++) {
3556 struct virtqueue *vq = vi->sq[i].vq;
3557 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
3558 virtnet_sq_free_unused_buf(vq, buf);
3561 for (i = 0; i < vi->max_queue_pairs; i++) {
3562 struct virtqueue *vq = vi->rq[i].vq;
3563 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
3564 virtnet_rq_free_unused_buf(vq, buf);
3568 static void virtnet_del_vqs(struct virtnet_info *vi)
3570 struct virtio_device *vdev = vi->vdev;
3572 virtnet_clean_affinity(vi);
3574 vdev->config->del_vqs(vdev);
3576 virtnet_free_queues(vi);
3579 /* How large should a single buffer be so a queue full of these can fit at
3580 * least one full packet?
3581 * Logic below assumes the mergeable buffer header is used.
3583 static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq)
3585 const unsigned int hdr_len = vi->hdr_len;
3586 unsigned int rq_size = virtqueue_get_vring_size(vq);
3587 unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu;
3588 unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len;
3589 unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size);
3591 return max(max(min_buf_len, hdr_len) - hdr_len,
3592 (unsigned int)GOOD_PACKET_LEN);
3595 static int virtnet_find_vqs(struct virtnet_info *vi)
3597 vq_callback_t **callbacks;
3598 struct virtqueue **vqs;
3604 /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
3605 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
3606 * possible control vq.
3608 total_vqs = vi->max_queue_pairs * 2 +
3609 virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
3611 /* Allocate space for find_vqs parameters */
3612 vqs = kcalloc(total_vqs, sizeof(*vqs), GFP_KERNEL);
3615 callbacks = kmalloc_array(total_vqs, sizeof(*callbacks), GFP_KERNEL);
3618 names = kmalloc_array(total_vqs, sizeof(*names), GFP_KERNEL);
3621 if (!vi->big_packets || vi->mergeable_rx_bufs) {
3622 ctx = kcalloc(total_vqs, sizeof(*ctx), GFP_KERNEL);
3629 /* Parameters for control virtqueue, if any */
3631 callbacks[total_vqs - 1] = NULL;
3632 names[total_vqs - 1] = "control";
3635 /* Allocate/initialize parameters for send/receive virtqueues */
3636 for (i = 0; i < vi->max_queue_pairs; i++) {
3637 callbacks[rxq2vq(i)] = skb_recv_done;
3638 callbacks[txq2vq(i)] = skb_xmit_done;
3639 sprintf(vi->rq[i].name, "input.%d", i);
3640 sprintf(vi->sq[i].name, "output.%d", i);
3641 names[rxq2vq(i)] = vi->rq[i].name;
3642 names[txq2vq(i)] = vi->sq[i].name;
3644 ctx[rxq2vq(i)] = true;
3647 ret = virtio_find_vqs_ctx(vi->vdev, total_vqs, vqs, callbacks,
3653 vi->cvq = vqs[total_vqs - 1];
3654 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
3655 vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
3658 for (i = 0; i < vi->max_queue_pairs; i++) {
3659 vi->rq[i].vq = vqs[rxq2vq(i)];
3660 vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq);
3661 vi->sq[i].vq = vqs[txq2vq(i)];
3664 /* run here: ret == 0. */
3679 static int virtnet_alloc_queues(struct virtnet_info *vi)
3684 vi->ctrl = kzalloc(sizeof(*vi->ctrl), GFP_KERNEL);
3690 vi->sq = kcalloc(vi->max_queue_pairs, sizeof(*vi->sq), GFP_KERNEL);
3693 vi->rq = kcalloc(vi->max_queue_pairs, sizeof(*vi->rq), GFP_KERNEL);
3697 INIT_DELAYED_WORK(&vi->refill, refill_work);
3698 for (i = 0; i < vi->max_queue_pairs; i++) {
3699 vi->rq[i].pages = NULL;
3700 netif_napi_add_weight(vi->dev, &vi->rq[i].napi, virtnet_poll,
3702 netif_napi_add_tx_weight(vi->dev, &vi->sq[i].napi,
3704 napi_tx ? napi_weight : 0);
3706 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
3707 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
3708 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
3710 u64_stats_init(&vi->rq[i].stats.syncp);
3711 u64_stats_init(&vi->sq[i].stats.syncp);
3724 static int init_vqs(struct virtnet_info *vi)
3728 /* Allocate send & receive queues */
3729 ret = virtnet_alloc_queues(vi);
3733 ret = virtnet_find_vqs(vi);
3738 virtnet_set_affinity(vi);
3744 virtnet_free_queues(vi);
3750 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
3753 struct virtnet_info *vi = netdev_priv(queue->dev);
3754 unsigned int queue_index = get_netdev_rx_queue_index(queue);
3755 unsigned int headroom = virtnet_get_headroom(vi);
3756 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
3757 struct ewma_pkt_len *avg;
3759 BUG_ON(queue_index >= vi->max_queue_pairs);
3760 avg = &vi->rq[queue_index].mrg_avg_pkt_len;
3761 return sprintf(buf, "%u\n",
3762 get_mergeable_buf_len(&vi->rq[queue_index], avg,
3763 SKB_DATA_ALIGN(headroom + tailroom)));
3766 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
3767 __ATTR_RO(mergeable_rx_buffer_size);
3769 static struct attribute *virtio_net_mrg_rx_attrs[] = {
3770 &mergeable_rx_buffer_size_attribute.attr,
3774 static const struct attribute_group virtio_net_mrg_rx_group = {
3775 .name = "virtio_net",
3776 .attrs = virtio_net_mrg_rx_attrs
3780 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
3782 const char *fname, const char *dname)
3784 if (!virtio_has_feature(vdev, fbit))
3787 dev_err(&vdev->dev, "device advertises feature %s but not %s",
3793 #define VIRTNET_FAIL_ON(vdev, fbit, dbit) \
3794 virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
3796 static bool virtnet_validate_features(struct virtio_device *vdev)
3798 if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
3799 (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
3800 "VIRTIO_NET_F_CTRL_VQ") ||
3801 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
3802 "VIRTIO_NET_F_CTRL_VQ") ||
3803 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
3804 "VIRTIO_NET_F_CTRL_VQ") ||
3805 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
3806 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
3807 "VIRTIO_NET_F_CTRL_VQ") ||
3808 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_RSS,
3809 "VIRTIO_NET_F_CTRL_VQ") ||
3810 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_HASH_REPORT,
3811 "VIRTIO_NET_F_CTRL_VQ") ||
3812 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_NOTF_COAL,
3813 "VIRTIO_NET_F_CTRL_VQ"))) {
3820 #define MIN_MTU ETH_MIN_MTU
3821 #define MAX_MTU ETH_MAX_MTU
3823 static int virtnet_validate(struct virtio_device *vdev)
3825 if (!vdev->config->get) {
3826 dev_err(&vdev->dev, "%s failure: config access disabled\n",
3831 if (!virtnet_validate_features(vdev))
3834 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
3835 int mtu = virtio_cread16(vdev,
3836 offsetof(struct virtio_net_config,
3839 __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
3842 if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY) &&
3843 !virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
3844 dev_warn(&vdev->dev, "device advertises feature VIRTIO_NET_F_STANDBY but not VIRTIO_NET_F_MAC, disabling standby");
3845 __virtio_clear_bit(vdev, VIRTIO_NET_F_STANDBY);
3851 static bool virtnet_check_guest_gso(const struct virtnet_info *vi)
3853 return virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3854 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
3855 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
3856 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) ||
3857 (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_USO4) &&
3858 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_USO6));
3861 static void virtnet_set_big_packets(struct virtnet_info *vi, const int mtu)
3863 bool guest_gso = virtnet_check_guest_gso(vi);
3865 /* If device can receive ANY guest GSO packets, regardless of mtu,
3866 * allocate packets of maximum size, otherwise limit it to only
3867 * mtu size worth only.
3869 if (mtu > ETH_DATA_LEN || guest_gso) {
3870 vi->big_packets = true;
3871 vi->big_packets_num_skbfrags = guest_gso ? MAX_SKB_FRAGS : DIV_ROUND_UP(mtu, PAGE_SIZE);
3875 static int virtnet_probe(struct virtio_device *vdev)
3877 int i, err = -ENOMEM;
3878 struct net_device *dev;
3879 struct virtnet_info *vi;
3880 u16 max_queue_pairs;
3883 /* Find if host supports multiqueue/rss virtio_net device */
3884 max_queue_pairs = 1;
3885 if (virtio_has_feature(vdev, VIRTIO_NET_F_MQ) || virtio_has_feature(vdev, VIRTIO_NET_F_RSS))
3887 virtio_cread16(vdev, offsetof(struct virtio_net_config, max_virtqueue_pairs));
3889 /* We need at least 2 queue's */
3890 if (max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
3891 max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
3892 !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
3893 max_queue_pairs = 1;
3895 /* Allocate ourselves a network device with room for our info */
3896 dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
3900 /* Set up network device as normal. */
3901 dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE |
3902 IFF_TX_SKB_NO_LINEAR;
3903 dev->netdev_ops = &virtnet_netdev;
3904 dev->features = NETIF_F_HIGHDMA;
3906 dev->ethtool_ops = &virtnet_ethtool_ops;
3907 SET_NETDEV_DEV(dev, &vdev->dev);
3909 /* Do we support "hardware" checksums? */
3910 if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
3911 /* This opens up the world of extra features. */
3912 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
3914 dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
3916 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
3917 dev->hw_features |= NETIF_F_TSO
3918 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
3920 /* Individual feature bits: what can host handle? */
3921 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
3922 dev->hw_features |= NETIF_F_TSO;
3923 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
3924 dev->hw_features |= NETIF_F_TSO6;
3925 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
3926 dev->hw_features |= NETIF_F_TSO_ECN;
3927 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_USO))
3928 dev->hw_features |= NETIF_F_GSO_UDP_L4;
3930 dev->features |= NETIF_F_GSO_ROBUST;
3933 dev->features |= dev->hw_features & NETIF_F_ALL_TSO;
3934 /* (!csum && gso) case will be fixed by register_netdev() */
3936 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
3937 dev->features |= NETIF_F_RXCSUM;
3938 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3939 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6))
3940 dev->features |= NETIF_F_GRO_HW;
3941 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS))
3942 dev->hw_features |= NETIF_F_GRO_HW;
3944 dev->vlan_features = dev->features;
3945 dev->xdp_features = NETDEV_XDP_ACT_BASIC | NETDEV_XDP_ACT_REDIRECT;
3947 /* MTU range: 68 - 65535 */
3948 dev->min_mtu = MIN_MTU;
3949 dev->max_mtu = MAX_MTU;
3951 /* Configuration may specify what MAC to use. Otherwise random. */
3952 if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
3955 virtio_cread_bytes(vdev,
3956 offsetof(struct virtio_net_config, mac),
3958 eth_hw_addr_set(dev, addr);
3960 eth_hw_addr_random(dev);
3961 dev_info(&vdev->dev, "Assigned random MAC address %pM\n",
3965 /* Set up our device-specific information */
3966 vi = netdev_priv(dev);
3971 INIT_WORK(&vi->config_work, virtnet_config_changed_work);
3972 spin_lock_init(&vi->refill_lock);
3974 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF)) {
3975 vi->mergeable_rx_bufs = true;
3976 dev->xdp_features |= NETDEV_XDP_ACT_RX_SG;
3979 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_NOTF_COAL)) {
3982 vi->tx_max_packets = 0;
3983 vi->rx_max_packets = 0;
3986 if (virtio_has_feature(vdev, VIRTIO_NET_F_HASH_REPORT))
3987 vi->has_rss_hash_report = true;
3989 if (virtio_has_feature(vdev, VIRTIO_NET_F_RSS))
3992 if (vi->has_rss || vi->has_rss_hash_report) {
3993 vi->rss_indir_table_size =
3994 virtio_cread16(vdev, offsetof(struct virtio_net_config,
3995 rss_max_indirection_table_length));
3997 virtio_cread8(vdev, offsetof(struct virtio_net_config, rss_max_key_size));
3999 vi->rss_hash_types_supported =
4000 virtio_cread32(vdev, offsetof(struct virtio_net_config, supported_hash_types));
4001 vi->rss_hash_types_supported &=
4002 ~(VIRTIO_NET_RSS_HASH_TYPE_IP_EX |
4003 VIRTIO_NET_RSS_HASH_TYPE_TCP_EX |
4004 VIRTIO_NET_RSS_HASH_TYPE_UDP_EX);
4006 dev->hw_features |= NETIF_F_RXHASH;
4009 if (vi->has_rss_hash_report)
4010 vi->hdr_len = sizeof(struct virtio_net_hdr_v1_hash);
4011 else if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
4012 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
4013 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
4015 vi->hdr_len = sizeof(struct virtio_net_hdr);
4017 if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
4018 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
4019 vi->any_header_sg = true;
4021 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
4024 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
4025 mtu = virtio_cread16(vdev,
4026 offsetof(struct virtio_net_config,
4028 if (mtu < dev->min_mtu) {
4029 /* Should never trigger: MTU was previously validated
4030 * in virtnet_validate.
4033 "device MTU appears to have changed it is now %d < %d",
4043 virtnet_set_big_packets(vi, mtu);
4045 if (vi->any_header_sg)
4046 dev->needed_headroom = vi->hdr_len;
4048 /* Enable multiqueue by default */
4049 if (num_online_cpus() >= max_queue_pairs)
4050 vi->curr_queue_pairs = max_queue_pairs;
4052 vi->curr_queue_pairs = num_online_cpus();
4053 vi->max_queue_pairs = max_queue_pairs;
4055 /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
4061 if (vi->mergeable_rx_bufs)
4062 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
4064 netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
4065 netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
4067 virtnet_init_settings(dev);
4069 if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY)) {
4070 vi->failover = net_failover_create(vi->dev);
4071 if (IS_ERR(vi->failover)) {
4072 err = PTR_ERR(vi->failover);
4077 if (vi->has_rss || vi->has_rss_hash_report)
4078 virtnet_init_default_rss(vi);
4080 /* serialize netdev register + virtio_device_ready() with ndo_open() */
4083 err = register_netdevice(dev);
4085 pr_debug("virtio_net: registering device failed\n");
4090 virtio_device_ready(vdev);
4092 /* a random MAC address has been assigned, notify the device.
4093 * We don't fail probe if VIRTIO_NET_F_CTRL_MAC_ADDR is not there
4094 * because many devices work fine without getting MAC explicitly
4096 if (!virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
4097 virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
4098 struct scatterlist sg;
4100 sg_init_one(&sg, dev->dev_addr, dev->addr_len);
4101 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
4102 VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
4103 pr_debug("virtio_net: setting MAC address failed\n");
4106 goto free_unregister_netdev;
4112 err = virtnet_cpu_notif_add(vi);
4114 pr_debug("virtio_net: registering cpu notifier failed\n");
4115 goto free_unregister_netdev;
4118 virtnet_set_queues(vi, vi->curr_queue_pairs);
4120 /* Assume link up if device can't report link status,
4121 otherwise get link status from config. */
4122 netif_carrier_off(dev);
4123 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
4124 schedule_work(&vi->config_work);
4126 vi->status = VIRTIO_NET_S_LINK_UP;
4127 virtnet_update_settings(vi);
4128 netif_carrier_on(dev);
4131 for (i = 0; i < ARRAY_SIZE(guest_offloads); i++)
4132 if (virtio_has_feature(vi->vdev, guest_offloads[i]))
4133 set_bit(guest_offloads[i], &vi->guest_offloads);
4134 vi->guest_offloads_capable = vi->guest_offloads;
4136 pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
4137 dev->name, max_queue_pairs);
4141 free_unregister_netdev:
4142 unregister_netdev(dev);
4144 net_failover_destroy(vi->failover);
4146 virtio_reset_device(vdev);
4147 cancel_delayed_work_sync(&vi->refill);
4148 free_receive_page_frags(vi);
4149 virtnet_del_vqs(vi);
4155 static void remove_vq_common(struct virtnet_info *vi)
4157 virtio_reset_device(vi->vdev);
4159 /* Free unused buffers in both send and recv, if any. */
4160 free_unused_bufs(vi);
4162 free_receive_bufs(vi);
4164 free_receive_page_frags(vi);
4166 virtnet_del_vqs(vi);
4169 static void virtnet_remove(struct virtio_device *vdev)
4171 struct virtnet_info *vi = vdev->priv;
4173 virtnet_cpu_notif_remove(vi);
4175 /* Make sure no work handler is accessing the device. */
4176 flush_work(&vi->config_work);
4178 unregister_netdev(vi->dev);
4180 net_failover_destroy(vi->failover);
4182 remove_vq_common(vi);
4184 free_netdev(vi->dev);
4187 static __maybe_unused int virtnet_freeze(struct virtio_device *vdev)
4189 struct virtnet_info *vi = vdev->priv;
4191 virtnet_cpu_notif_remove(vi);
4192 virtnet_freeze_down(vdev);
4193 remove_vq_common(vi);
4198 static __maybe_unused int virtnet_restore(struct virtio_device *vdev)
4200 struct virtnet_info *vi = vdev->priv;
4203 err = virtnet_restore_up(vdev);
4206 virtnet_set_queues(vi, vi->curr_queue_pairs);
4208 err = virtnet_cpu_notif_add(vi);
4210 virtnet_freeze_down(vdev);
4211 remove_vq_common(vi);
4218 static struct virtio_device_id id_table[] = {
4219 { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
4223 #define VIRTNET_FEATURES \
4224 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
4226 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
4227 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
4228 VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
4229 VIRTIO_NET_F_HOST_USO, VIRTIO_NET_F_GUEST_USO4, VIRTIO_NET_F_GUEST_USO6, \
4230 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
4231 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
4232 VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
4233 VIRTIO_NET_F_CTRL_MAC_ADDR, \
4234 VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \
4235 VIRTIO_NET_F_SPEED_DUPLEX, VIRTIO_NET_F_STANDBY, \
4236 VIRTIO_NET_F_RSS, VIRTIO_NET_F_HASH_REPORT, VIRTIO_NET_F_NOTF_COAL, \
4237 VIRTIO_NET_F_GUEST_HDRLEN
4239 static unsigned int features[] = {
4243 static unsigned int features_legacy[] = {
4246 VIRTIO_F_ANY_LAYOUT,
4249 static struct virtio_driver virtio_net_driver = {
4250 .feature_table = features,
4251 .feature_table_size = ARRAY_SIZE(features),
4252 .feature_table_legacy = features_legacy,
4253 .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
4254 .driver.name = KBUILD_MODNAME,
4255 .driver.owner = THIS_MODULE,
4256 .id_table = id_table,
4257 .validate = virtnet_validate,
4258 .probe = virtnet_probe,
4259 .remove = virtnet_remove,
4260 .config_changed = virtnet_config_changed,
4261 #ifdef CONFIG_PM_SLEEP
4262 .freeze = virtnet_freeze,
4263 .restore = virtnet_restore,
4267 static __init int virtio_net_driver_init(void)
4271 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
4273 virtnet_cpu_down_prep);
4276 virtionet_online = ret;
4277 ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
4278 NULL, virtnet_cpu_dead);
4281 ret = register_virtio_driver(&virtio_net_driver);
4286 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
4288 cpuhp_remove_multi_state(virtionet_online);
4292 module_init(virtio_net_driver_init);
4294 static __exit void virtio_net_driver_exit(void)
4296 unregister_virtio_driver(&virtio_net_driver);
4297 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
4298 cpuhp_remove_multi_state(virtionet_online);
4300 module_exit(virtio_net_driver_exit);
4302 MODULE_DEVICE_TABLE(virtio, id_table);
4303 MODULE_DESCRIPTION("Virtio network driver");
4304 MODULE_LICENSE("GPL");