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 <linux/dim.h>
23 #include <net/route.h>
25 #include <net/net_failover.h>
26 #include <net/netdev_rx_queue.h>
28 static int napi_weight = NAPI_POLL_WEIGHT;
29 module_param(napi_weight, int, 0444);
31 static bool csum = true, gso = true, napi_tx = true;
32 module_param(csum, bool, 0444);
33 module_param(gso, bool, 0444);
34 module_param(napi_tx, bool, 0644);
36 /* FIXME: MTU in config. */
37 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
38 #define GOOD_COPY_LEN 128
40 #define VIRTNET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
42 /* Amount of XDP headroom to prepend to packets for use by xdp_adjust_head */
43 #define VIRTIO_XDP_HEADROOM 256
45 /* Separating two types of XDP xmit */
46 #define VIRTIO_XDP_TX BIT(0)
47 #define VIRTIO_XDP_REDIR BIT(1)
49 #define VIRTIO_XDP_FLAG BIT(0)
51 /* RX packet size EWMA. The average packet size is used to determine the packet
52 * buffer size when refilling RX rings. As the entire RX ring may be refilled
53 * at once, the weight is chosen so that the EWMA will be insensitive to short-
54 * term, transient changes in packet size.
56 DECLARE_EWMA(pkt_len, 0, 64)
58 #define VIRTNET_DRIVER_VERSION "1.0.0"
60 static const unsigned long guest_offloads[] = {
61 VIRTIO_NET_F_GUEST_TSO4,
62 VIRTIO_NET_F_GUEST_TSO6,
63 VIRTIO_NET_F_GUEST_ECN,
64 VIRTIO_NET_F_GUEST_UFO,
65 VIRTIO_NET_F_GUEST_CSUM,
66 VIRTIO_NET_F_GUEST_USO4,
67 VIRTIO_NET_F_GUEST_USO6,
68 VIRTIO_NET_F_GUEST_HDRLEN
71 #define GUEST_OFFLOAD_GRO_HW_MASK ((1ULL << VIRTIO_NET_F_GUEST_TSO4) | \
72 (1ULL << VIRTIO_NET_F_GUEST_TSO6) | \
73 (1ULL << VIRTIO_NET_F_GUEST_ECN) | \
74 (1ULL << VIRTIO_NET_F_GUEST_UFO) | \
75 (1ULL << VIRTIO_NET_F_GUEST_USO4) | \
76 (1ULL << VIRTIO_NET_F_GUEST_USO6))
78 struct virtnet_stat_desc {
79 char desc[ETH_GSTRING_LEN];
83 struct virtnet_sq_stats {
84 struct u64_stats_sync syncp;
88 u64_stats_t xdp_tx_drops;
90 u64_stats_t tx_timeouts;
93 struct virtnet_rq_stats {
94 struct u64_stats_sync syncp;
98 u64_stats_t xdp_packets;
100 u64_stats_t xdp_redirects;
101 u64_stats_t xdp_drops;
105 #define VIRTNET_SQ_STAT(m) offsetof(struct virtnet_sq_stats, m)
106 #define VIRTNET_RQ_STAT(m) offsetof(struct virtnet_rq_stats, m)
108 static const struct virtnet_stat_desc virtnet_sq_stats_desc[] = {
109 { "packets", VIRTNET_SQ_STAT(packets) },
110 { "bytes", VIRTNET_SQ_STAT(bytes) },
111 { "xdp_tx", VIRTNET_SQ_STAT(xdp_tx) },
112 { "xdp_tx_drops", VIRTNET_SQ_STAT(xdp_tx_drops) },
113 { "kicks", VIRTNET_SQ_STAT(kicks) },
114 { "tx_timeouts", VIRTNET_SQ_STAT(tx_timeouts) },
117 static const struct virtnet_stat_desc virtnet_rq_stats_desc[] = {
118 { "packets", VIRTNET_RQ_STAT(packets) },
119 { "bytes", VIRTNET_RQ_STAT(bytes) },
120 { "drops", VIRTNET_RQ_STAT(drops) },
121 { "xdp_packets", VIRTNET_RQ_STAT(xdp_packets) },
122 { "xdp_tx", VIRTNET_RQ_STAT(xdp_tx) },
123 { "xdp_redirects", VIRTNET_RQ_STAT(xdp_redirects) },
124 { "xdp_drops", VIRTNET_RQ_STAT(xdp_drops) },
125 { "kicks", VIRTNET_RQ_STAT(kicks) },
128 #define VIRTNET_SQ_STATS_LEN ARRAY_SIZE(virtnet_sq_stats_desc)
129 #define VIRTNET_RQ_STATS_LEN ARRAY_SIZE(virtnet_rq_stats_desc)
131 struct virtnet_interrupt_coalesce {
136 /* The dma information of pages allocated at a time. */
137 struct virtnet_rq_dma {
144 /* Internal representation of a send virtqueue */
146 /* Virtqueue associated with this send _queue */
147 struct virtqueue *vq;
149 /* TX: fragments + linear part + virtio header */
150 struct scatterlist sg[MAX_SKB_FRAGS + 2];
152 /* Name of the send queue: output.$index */
155 struct virtnet_sq_stats stats;
157 struct virtnet_interrupt_coalesce intr_coal;
159 struct napi_struct napi;
161 /* Record whether sq is in reset state. */
165 /* Internal representation of a receive virtqueue */
166 struct receive_queue {
167 /* Virtqueue associated with this receive_queue */
168 struct virtqueue *vq;
170 struct napi_struct napi;
172 struct bpf_prog __rcu *xdp_prog;
174 struct virtnet_rq_stats stats;
176 /* The number of rx notifications */
179 /* Is dynamic interrupt moderation enabled? */
182 /* Dynamic Interrupt Moderation */
187 struct virtnet_interrupt_coalesce intr_coal;
189 /* Chain pages by the private ptr. */
192 /* Average packet length for mergeable receive buffers. */
193 struct ewma_pkt_len mrg_avg_pkt_len;
195 /* Page frag for packet buffer allocation. */
196 struct page_frag alloc_frag;
198 /* RX: fragments + linear part + virtio header */
199 struct scatterlist sg[MAX_SKB_FRAGS + 2];
201 /* Min single buffer size for mergeable buffers case. */
202 unsigned int min_buf_len;
204 /* Name of this receive queue: input.$index */
207 struct xdp_rxq_info xdp_rxq;
209 /* Record the last dma info to free after new pages is allocated. */
210 struct virtnet_rq_dma *last_dma;
216 /* This structure can contain rss message with maximum settings for indirection table and keysize
217 * Note, that default structure that describes RSS configuration virtio_net_rss_config
218 * contains same info but can't handle table values.
219 * In any case, structure would be passed to virtio hw through sg_buf split by parts
220 * because table sizes may be differ according to the device configuration.
222 #define VIRTIO_NET_RSS_MAX_KEY_SIZE 40
223 #define VIRTIO_NET_RSS_MAX_TABLE_LEN 128
224 struct virtio_net_ctrl_rss {
226 u16 indirection_table_mask;
227 u16 unclassified_queue;
228 u16 indirection_table[VIRTIO_NET_RSS_MAX_TABLE_LEN];
231 u8 key[VIRTIO_NET_RSS_MAX_KEY_SIZE];
234 /* Control VQ buffers: protected by the rtnl lock */
236 struct virtio_net_ctrl_hdr hdr;
237 virtio_net_ctrl_ack status;
238 struct virtio_net_ctrl_mq mq;
243 struct virtio_net_ctrl_rss rss;
244 struct virtio_net_ctrl_coal_tx coal_tx;
245 struct virtio_net_ctrl_coal_rx coal_rx;
246 struct virtio_net_ctrl_coal_vq coal_vq;
249 struct virtnet_info {
250 struct virtio_device *vdev;
251 struct virtqueue *cvq;
252 struct net_device *dev;
253 struct send_queue *sq;
254 struct receive_queue *rq;
257 /* Max # of queue pairs supported by the device */
260 /* # of queue pairs currently used by the driver */
261 u16 curr_queue_pairs;
263 /* # of XDP queue pairs currently used by the driver */
266 /* xdp_queue_pairs may be 0, when xdp is already loaded. So add this. */
269 /* I like... big packets and I cannot lie! */
272 /* number of sg entries allocated for big packets */
273 unsigned int big_packets_num_skbfrags;
275 /* Host will merge rx buffers for big packets (shake it! shake it!) */
276 bool mergeable_rx_bufs;
278 /* Host supports rss and/or hash report */
280 bool has_rss_hash_report;
282 u16 rss_indir_table_size;
283 u32 rss_hash_types_supported;
284 u32 rss_hash_types_saved;
286 /* Has control virtqueue */
289 /* Host can handle any s/g split between our header and packet data */
292 /* Packet virtio header size */
295 /* Work struct for delayed refilling if we run low on memory. */
296 struct delayed_work refill;
298 /* Is delayed refill enabled? */
301 /* The lock to synchronize the access to refill_enabled */
302 spinlock_t refill_lock;
304 /* Work struct for config space updates */
305 struct work_struct config_work;
307 /* Does the affinity hint is set for virtqueues? */
308 bool affinity_hint_set;
310 /* CPU hotplug instances for online & dead */
311 struct hlist_node node;
312 struct hlist_node node_dead;
314 struct control_buf *ctrl;
316 /* Ethtool settings */
320 /* Is rx dynamic interrupt moderation enabled? */
323 /* Interrupt coalescing settings */
324 struct virtnet_interrupt_coalesce intr_coal_tx;
325 struct virtnet_interrupt_coalesce intr_coal_rx;
327 unsigned long guest_offloads;
328 unsigned long guest_offloads_capable;
330 /* failover when STANDBY feature enabled */
331 struct failover *failover;
334 struct padded_vnet_hdr {
335 struct virtio_net_hdr_v1_hash hdr;
337 * hdr is in a separate sg buffer, and data sg buffer shares same page
338 * with this header sg. This padding makes next sg 16 byte aligned
344 struct virtio_net_common_hdr {
346 struct virtio_net_hdr hdr;
347 struct virtio_net_hdr_mrg_rxbuf mrg_hdr;
348 struct virtio_net_hdr_v1_hash hash_v1_hdr;
352 static void virtnet_sq_free_unused_buf(struct virtqueue *vq, void *buf);
354 static bool is_xdp_frame(void *ptr)
356 return (unsigned long)ptr & VIRTIO_XDP_FLAG;
359 static void *xdp_to_ptr(struct xdp_frame *ptr)
361 return (void *)((unsigned long)ptr | VIRTIO_XDP_FLAG);
364 static struct xdp_frame *ptr_to_xdp(void *ptr)
366 return (struct xdp_frame *)((unsigned long)ptr & ~VIRTIO_XDP_FLAG);
369 /* Converting between virtqueue no. and kernel tx/rx queue no.
370 * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
372 static int vq2txq(struct virtqueue *vq)
374 return (vq->index - 1) / 2;
377 static int txq2vq(int txq)
382 static int vq2rxq(struct virtqueue *vq)
384 return vq->index / 2;
387 static int rxq2vq(int rxq)
392 static inline struct virtio_net_common_hdr *
393 skb_vnet_common_hdr(struct sk_buff *skb)
395 return (struct virtio_net_common_hdr *)skb->cb;
399 * private is used to chain pages for big packets, put the whole
400 * most recent used list in the beginning for reuse
402 static void give_pages(struct receive_queue *rq, struct page *page)
406 /* Find end of list, sew whole thing into vi->rq.pages. */
407 for (end = page; end->private; end = (struct page *)end->private);
408 end->private = (unsigned long)rq->pages;
412 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
414 struct page *p = rq->pages;
417 rq->pages = (struct page *)p->private;
418 /* clear private here, it is used to chain pages */
421 p = alloc_page(gfp_mask);
425 static void virtnet_rq_free_buf(struct virtnet_info *vi,
426 struct receive_queue *rq, void *buf)
428 if (vi->mergeable_rx_bufs)
429 put_page(virt_to_head_page(buf));
430 else if (vi->big_packets)
433 put_page(virt_to_head_page(buf));
436 static void enable_delayed_refill(struct virtnet_info *vi)
438 spin_lock_bh(&vi->refill_lock);
439 vi->refill_enabled = true;
440 spin_unlock_bh(&vi->refill_lock);
443 static void disable_delayed_refill(struct virtnet_info *vi)
445 spin_lock_bh(&vi->refill_lock);
446 vi->refill_enabled = false;
447 spin_unlock_bh(&vi->refill_lock);
450 static void virtqueue_napi_schedule(struct napi_struct *napi,
451 struct virtqueue *vq)
453 if (napi_schedule_prep(napi)) {
454 virtqueue_disable_cb(vq);
455 __napi_schedule(napi);
459 static bool virtqueue_napi_complete(struct napi_struct *napi,
460 struct virtqueue *vq, int processed)
464 opaque = virtqueue_enable_cb_prepare(vq);
465 if (napi_complete_done(napi, processed)) {
466 if (unlikely(virtqueue_poll(vq, opaque)))
467 virtqueue_napi_schedule(napi, vq);
471 virtqueue_disable_cb(vq);
477 static void skb_xmit_done(struct virtqueue *vq)
479 struct virtnet_info *vi = vq->vdev->priv;
480 struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi;
482 /* Suppress further interrupts. */
483 virtqueue_disable_cb(vq);
486 virtqueue_napi_schedule(napi, vq);
488 /* We were probably waiting for more output buffers. */
489 netif_wake_subqueue(vi->dev, vq2txq(vq));
492 #define MRG_CTX_HEADER_SHIFT 22
493 static void *mergeable_len_to_ctx(unsigned int truesize,
494 unsigned int headroom)
496 return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) | truesize);
499 static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx)
501 return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT;
504 static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx)
506 return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) - 1);
509 static struct sk_buff *virtnet_build_skb(void *buf, unsigned int buflen,
510 unsigned int headroom,
515 skb = build_skb(buf, buflen);
519 skb_reserve(skb, headroom);
525 /* Called from bottom half context */
526 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
527 struct receive_queue *rq,
528 struct page *page, unsigned int offset,
529 unsigned int len, unsigned int truesize,
530 unsigned int headroom)
533 struct virtio_net_common_hdr *hdr;
534 unsigned int copy, hdr_len, hdr_padded_len;
535 struct page *page_to_free = NULL;
536 int tailroom, shinfo_size;
537 char *p, *hdr_p, *buf;
539 p = page_address(page) + offset;
542 hdr_len = vi->hdr_len;
543 if (vi->mergeable_rx_bufs)
544 hdr_padded_len = hdr_len;
546 hdr_padded_len = sizeof(struct padded_vnet_hdr);
550 offset += hdr_padded_len;
552 tailroom = truesize - headroom - hdr_padded_len - len;
554 shinfo_size = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
556 /* copy small packet so we can reuse these pages */
557 if (!NET_IP_ALIGN && len > GOOD_COPY_LEN && tailroom >= shinfo_size) {
558 skb = virtnet_build_skb(buf, truesize, p - buf, len);
562 page = (struct page *)page->private;
564 give_pages(rq, page);
568 /* copy small packet so we can reuse these pages for small data */
569 skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
573 /* Copy all frame if it fits skb->head, otherwise
574 * we let virtio_net_hdr_to_skb() and GRO pull headers as needed.
576 if (len <= skb_tailroom(skb))
580 skb_put_data(skb, p, copy);
585 if (vi->mergeable_rx_bufs) {
587 skb_add_rx_frag(skb, 0, page, offset, len, truesize);
594 * Verify that we can indeed put this data into a skb.
595 * This is here to handle cases when the device erroneously
596 * tries to receive more than is possible. This is usually
597 * the case of a broken device.
599 if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
600 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
604 BUG_ON(offset >= PAGE_SIZE);
606 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
607 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
608 frag_size, truesize);
610 page = (struct page *)page->private;
615 give_pages(rq, page);
618 hdr = skb_vnet_common_hdr(skb);
619 memcpy(hdr, hdr_p, hdr_len);
621 put_page(page_to_free);
626 static void virtnet_rq_unmap(struct receive_queue *rq, void *buf, u32 len)
628 struct page *page = virt_to_head_page(buf);
629 struct virtnet_rq_dma *dma;
633 head = page_address(page);
639 if (dma->need_sync && len) {
640 offset = buf - (head + sizeof(*dma));
642 virtqueue_dma_sync_single_range_for_cpu(rq->vq, dma->addr,
650 virtqueue_dma_unmap_single_attrs(rq->vq, dma->addr, dma->len,
651 DMA_FROM_DEVICE, DMA_ATTR_SKIP_CPU_SYNC);
655 static void *virtnet_rq_get_buf(struct receive_queue *rq, u32 *len, void **ctx)
659 buf = virtqueue_get_buf_ctx(rq->vq, len, ctx);
660 if (buf && rq->do_dma)
661 virtnet_rq_unmap(rq, buf, *len);
666 static void virtnet_rq_init_one_sg(struct receive_queue *rq, void *buf, u32 len)
668 struct virtnet_rq_dma *dma;
674 sg_init_one(rq->sg, buf, len);
678 head = page_address(rq->alloc_frag.page);
684 addr = dma->addr - sizeof(*dma) + offset;
686 sg_init_table(rq->sg, 1);
687 rq->sg[0].dma_address = addr;
688 rq->sg[0].length = len;
691 static void *virtnet_rq_alloc(struct receive_queue *rq, u32 size, gfp_t gfp)
693 struct page_frag *alloc_frag = &rq->alloc_frag;
694 struct virtnet_rq_dma *dma;
698 if (unlikely(!skb_page_frag_refill(size, alloc_frag, gfp)))
701 head = page_address(alloc_frag->page);
707 if (!alloc_frag->offset) {
709 /* Now, the new page is allocated, the last dma
710 * will not be used. So the dma can be unmapped
713 virtnet_rq_unmap(rq, rq->last_dma, 0);
717 dma->len = alloc_frag->size - sizeof(*dma);
719 addr = virtqueue_dma_map_single_attrs(rq->vq, dma + 1,
720 dma->len, DMA_FROM_DEVICE, 0);
721 if (virtqueue_dma_mapping_error(rq->vq, addr))
725 dma->need_sync = virtqueue_dma_need_sync(rq->vq, addr);
727 /* Add a reference to dma to prevent the entire dma from
728 * being released during error handling. This reference
729 * will be freed after the pages are no longer used.
731 get_page(alloc_frag->page);
733 alloc_frag->offset = sizeof(*dma);
741 buf = head + alloc_frag->offset;
743 get_page(alloc_frag->page);
744 alloc_frag->offset += size;
749 static void virtnet_rq_set_premapped(struct virtnet_info *vi)
753 /* disable for big mode */
754 if (!vi->mergeable_rx_bufs && vi->big_packets)
757 for (i = 0; i < vi->max_queue_pairs; i++) {
758 if (virtqueue_set_dma_premapped(vi->rq[i].vq))
761 vi->rq[i].do_dma = true;
765 static void virtnet_rq_unmap_free_buf(struct virtqueue *vq, void *buf)
767 struct virtnet_info *vi = vq->vdev->priv;
768 struct receive_queue *rq;
774 virtnet_rq_unmap(rq, buf, 0);
776 virtnet_rq_free_buf(vi, rq, buf);
779 static void free_old_xmit_skbs(struct send_queue *sq, bool in_napi)
782 unsigned int packets = 0;
783 unsigned int bytes = 0;
786 while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
787 if (likely(!is_xdp_frame(ptr))) {
788 struct sk_buff *skb = ptr;
790 pr_debug("Sent skb %p\n", skb);
793 napi_consume_skb(skb, in_napi);
795 struct xdp_frame *frame = ptr_to_xdp(ptr);
797 bytes += xdp_get_frame_len(frame);
798 xdp_return_frame(frame);
803 /* Avoid overhead when no packets have been processed
804 * happens when called speculatively from start_xmit.
809 u64_stats_update_begin(&sq->stats.syncp);
810 u64_stats_add(&sq->stats.bytes, bytes);
811 u64_stats_add(&sq->stats.packets, packets);
812 u64_stats_update_end(&sq->stats.syncp);
815 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
817 if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
819 else if (q < vi->curr_queue_pairs)
825 static void check_sq_full_and_disable(struct virtnet_info *vi,
826 struct net_device *dev,
827 struct send_queue *sq)
829 bool use_napi = sq->napi.weight;
834 /* If running out of space, stop queue to avoid getting packets that we
835 * are then unable to transmit.
836 * An alternative would be to force queuing layer to requeue the skb by
837 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
838 * returned in a normal path of operation: it means that driver is not
839 * maintaining the TX queue stop/start state properly, and causes
840 * the stack to do a non-trivial amount of useless work.
841 * Since most packets only take 1 or 2 ring slots, stopping the queue
842 * early means 16 slots are typically wasted.
844 if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
845 netif_stop_subqueue(dev, qnum);
847 if (unlikely(!virtqueue_enable_cb_delayed(sq->vq)))
848 virtqueue_napi_schedule(&sq->napi, sq->vq);
849 } else if (unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
850 /* More just got used, free them then recheck. */
851 free_old_xmit_skbs(sq, false);
852 if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
853 netif_start_subqueue(dev, qnum);
854 virtqueue_disable_cb(sq->vq);
860 static int __virtnet_xdp_xmit_one(struct virtnet_info *vi,
861 struct send_queue *sq,
862 struct xdp_frame *xdpf)
864 struct virtio_net_hdr_mrg_rxbuf *hdr;
865 struct skb_shared_info *shinfo;
869 if (unlikely(xdpf->headroom < vi->hdr_len))
872 if (unlikely(xdp_frame_has_frags(xdpf))) {
873 shinfo = xdp_get_shared_info_from_frame(xdpf);
874 nr_frags = shinfo->nr_frags;
877 /* In wrapping function virtnet_xdp_xmit(), we need to free
878 * up the pending old buffers, where we need to calculate the
879 * position of skb_shared_info in xdp_get_frame_len() and
880 * xdp_return_frame(), which will involve to xdpf->data and
881 * xdpf->headroom. Therefore, we need to update the value of
882 * headroom synchronously here.
884 xdpf->headroom -= vi->hdr_len;
885 xdpf->data -= vi->hdr_len;
886 /* Zero header and leave csum up to XDP layers */
888 memset(hdr, 0, vi->hdr_len);
889 xdpf->len += vi->hdr_len;
891 sg_init_table(sq->sg, nr_frags + 1);
892 sg_set_buf(sq->sg, xdpf->data, xdpf->len);
893 for (i = 0; i < nr_frags; i++) {
894 skb_frag_t *frag = &shinfo->frags[i];
896 sg_set_page(&sq->sg[i + 1], skb_frag_page(frag),
897 skb_frag_size(frag), skb_frag_off(frag));
900 err = virtqueue_add_outbuf(sq->vq, sq->sg, nr_frags + 1,
901 xdp_to_ptr(xdpf), GFP_ATOMIC);
903 return -ENOSPC; /* Caller handle free/refcnt */
908 /* when vi->curr_queue_pairs > nr_cpu_ids, the txq/sq is only used for xdp tx on
909 * the current cpu, so it does not need to be locked.
911 * Here we use marco instead of inline functions because we have to deal with
912 * three issues at the same time: 1. the choice of sq. 2. judge and execute the
913 * lock/unlock of txq 3. make sparse happy. It is difficult for two inline
914 * functions to perfectly solve these three problems at the same time.
916 #define virtnet_xdp_get_sq(vi) ({ \
917 int cpu = smp_processor_id(); \
918 struct netdev_queue *txq; \
919 typeof(vi) v = (vi); \
922 if (v->curr_queue_pairs > nr_cpu_ids) { \
923 qp = v->curr_queue_pairs - v->xdp_queue_pairs; \
925 txq = netdev_get_tx_queue(v->dev, qp); \
926 __netif_tx_acquire(txq); \
928 qp = cpu % v->curr_queue_pairs; \
929 txq = netdev_get_tx_queue(v->dev, qp); \
930 __netif_tx_lock(txq, cpu); \
935 #define virtnet_xdp_put_sq(vi, q) { \
936 struct netdev_queue *txq; \
937 typeof(vi) v = (vi); \
939 txq = netdev_get_tx_queue(v->dev, (q) - v->sq); \
940 if (v->curr_queue_pairs > nr_cpu_ids) \
941 __netif_tx_release(txq); \
943 __netif_tx_unlock(txq); \
946 static int virtnet_xdp_xmit(struct net_device *dev,
947 int n, struct xdp_frame **frames, u32 flags)
949 struct virtnet_info *vi = netdev_priv(dev);
950 struct receive_queue *rq = vi->rq;
951 struct bpf_prog *xdp_prog;
952 struct send_queue *sq;
962 /* Only allow ndo_xdp_xmit if XDP is loaded on dev, as this
963 * indicate XDP resources have been successfully allocated.
965 xdp_prog = rcu_access_pointer(rq->xdp_prog);
969 sq = virtnet_xdp_get_sq(vi);
971 if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK)) {
976 /* Free up any pending old buffers before queueing new ones. */
977 while ((ptr = virtqueue_get_buf(sq->vq, &len)) != NULL) {
978 if (likely(is_xdp_frame(ptr))) {
979 struct xdp_frame *frame = ptr_to_xdp(ptr);
981 bytes += xdp_get_frame_len(frame);
982 xdp_return_frame(frame);
984 struct sk_buff *skb = ptr;
987 napi_consume_skb(skb, false);
992 for (i = 0; i < n; i++) {
993 struct xdp_frame *xdpf = frames[i];
995 if (__virtnet_xdp_xmit_one(vi, sq, xdpf))
1001 if (!is_xdp_raw_buffer_queue(vi, sq - vi->sq))
1002 check_sq_full_and_disable(vi, dev, sq);
1004 if (flags & XDP_XMIT_FLUSH) {
1005 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq))
1009 u64_stats_update_begin(&sq->stats.syncp);
1010 u64_stats_add(&sq->stats.bytes, bytes);
1011 u64_stats_add(&sq->stats.packets, packets);
1012 u64_stats_add(&sq->stats.xdp_tx, n);
1013 u64_stats_add(&sq->stats.xdp_tx_drops, n - nxmit);
1014 u64_stats_add(&sq->stats.kicks, kicks);
1015 u64_stats_update_end(&sq->stats.syncp);
1017 virtnet_xdp_put_sq(vi, sq);
1021 static void put_xdp_frags(struct xdp_buff *xdp)
1023 struct skb_shared_info *shinfo;
1024 struct page *xdp_page;
1027 if (xdp_buff_has_frags(xdp)) {
1028 shinfo = xdp_get_shared_info_from_buff(xdp);
1029 for (i = 0; i < shinfo->nr_frags; i++) {
1030 xdp_page = skb_frag_page(&shinfo->frags[i]);
1036 static int virtnet_xdp_handler(struct bpf_prog *xdp_prog, struct xdp_buff *xdp,
1037 struct net_device *dev,
1038 unsigned int *xdp_xmit,
1039 struct virtnet_rq_stats *stats)
1041 struct xdp_frame *xdpf;
1045 act = bpf_prog_run_xdp(xdp_prog, xdp);
1046 u64_stats_inc(&stats->xdp_packets);
1053 u64_stats_inc(&stats->xdp_tx);
1054 xdpf = xdp_convert_buff_to_frame(xdp);
1055 if (unlikely(!xdpf)) {
1056 netdev_dbg(dev, "convert buff to frame failed for xdp\n");
1060 err = virtnet_xdp_xmit(dev, 1, &xdpf, 0);
1061 if (unlikely(!err)) {
1062 xdp_return_frame_rx_napi(xdpf);
1063 } else if (unlikely(err < 0)) {
1064 trace_xdp_exception(dev, xdp_prog, act);
1067 *xdp_xmit |= VIRTIO_XDP_TX;
1071 u64_stats_inc(&stats->xdp_redirects);
1072 err = xdp_do_redirect(dev, xdp, xdp_prog);
1076 *xdp_xmit |= VIRTIO_XDP_REDIR;
1080 bpf_warn_invalid_xdp_action(dev, xdp_prog, act);
1083 trace_xdp_exception(dev, xdp_prog, act);
1090 static unsigned int virtnet_get_headroom(struct virtnet_info *vi)
1092 return vi->xdp_enabled ? VIRTIO_XDP_HEADROOM : 0;
1095 /* We copy the packet for XDP in the following cases:
1097 * 1) Packet is scattered across multiple rx buffers.
1098 * 2) Headroom space is insufficient.
1100 * This is inefficient but it's a temporary condition that
1101 * we hit right after XDP is enabled and until queue is refilled
1102 * with large buffers with sufficient headroom - so it should affect
1103 * at most queue size packets.
1104 * Afterwards, the conditions to enable
1105 * XDP should preclude the underlying device from sending packets
1106 * across multiple buffers (num_buf > 1), and we make sure buffers
1107 * have enough headroom.
1109 static struct page *xdp_linearize_page(struct receive_queue *rq,
1116 int tailroom = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1119 if (page_off + *len + tailroom > PAGE_SIZE)
1122 page = alloc_page(GFP_ATOMIC);
1126 memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
1129 while (--*num_buf) {
1130 unsigned int buflen;
1134 buf = virtnet_rq_get_buf(rq, &buflen, NULL);
1138 p = virt_to_head_page(buf);
1139 off = buf - page_address(p);
1141 /* guard against a misconfigured or uncooperative backend that
1142 * is sending packet larger than the MTU.
1144 if ((page_off + buflen + tailroom) > PAGE_SIZE) {
1149 memcpy(page_address(page) + page_off,
1150 page_address(p) + off, buflen);
1155 /* Headroom does not contribute to packet length */
1156 *len = page_off - VIRTIO_XDP_HEADROOM;
1159 __free_pages(page, 0);
1163 static struct sk_buff *receive_small_build_skb(struct virtnet_info *vi,
1164 unsigned int xdp_headroom,
1168 unsigned int header_offset;
1169 unsigned int headroom;
1170 unsigned int buflen;
1171 struct sk_buff *skb;
1173 header_offset = VIRTNET_RX_PAD + xdp_headroom;
1174 headroom = vi->hdr_len + header_offset;
1175 buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
1176 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1178 skb = virtnet_build_skb(buf, buflen, headroom, len);
1182 buf += header_offset;
1183 memcpy(skb_vnet_common_hdr(skb), buf, vi->hdr_len);
1188 static struct sk_buff *receive_small_xdp(struct net_device *dev,
1189 struct virtnet_info *vi,
1190 struct receive_queue *rq,
1191 struct bpf_prog *xdp_prog,
1193 unsigned int xdp_headroom,
1195 unsigned int *xdp_xmit,
1196 struct virtnet_rq_stats *stats)
1198 unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom;
1199 unsigned int headroom = vi->hdr_len + header_offset;
1200 struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset;
1201 struct page *page = virt_to_head_page(buf);
1202 struct page *xdp_page;
1203 unsigned int buflen;
1204 struct xdp_buff xdp;
1205 struct sk_buff *skb;
1206 unsigned int metasize = 0;
1209 if (unlikely(hdr->hdr.gso_type))
1212 buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
1213 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1215 if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) {
1216 int offset = buf - page_address(page) + header_offset;
1217 unsigned int tlen = len + vi->hdr_len;
1220 xdp_headroom = virtnet_get_headroom(vi);
1221 header_offset = VIRTNET_RX_PAD + xdp_headroom;
1222 headroom = vi->hdr_len + header_offset;
1223 buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
1224 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1225 xdp_page = xdp_linearize_page(rq, &num_buf, page,
1226 offset, header_offset,
1231 buf = page_address(xdp_page);
1236 xdp_init_buff(&xdp, buflen, &rq->xdp_rxq);
1237 xdp_prepare_buff(&xdp, buf + VIRTNET_RX_PAD + vi->hdr_len,
1238 xdp_headroom, len, true);
1240 act = virtnet_xdp_handler(xdp_prog, &xdp, dev, xdp_xmit, stats);
1244 /* Recalculate length in case bpf program changed it */
1245 len = xdp.data_end - xdp.data;
1246 metasize = xdp.data - xdp.data_meta;
1257 skb = virtnet_build_skb(buf, buflen, xdp.data - buf, len);
1262 skb_metadata_set(skb, metasize);
1267 u64_stats_inc(&stats->xdp_drops);
1269 u64_stats_inc(&stats->drops);
1275 static struct sk_buff *receive_small(struct net_device *dev,
1276 struct virtnet_info *vi,
1277 struct receive_queue *rq,
1278 void *buf, void *ctx,
1280 unsigned int *xdp_xmit,
1281 struct virtnet_rq_stats *stats)
1283 unsigned int xdp_headroom = (unsigned long)ctx;
1284 struct page *page = virt_to_head_page(buf);
1285 struct sk_buff *skb;
1288 u64_stats_add(&stats->bytes, len);
1290 if (unlikely(len > GOOD_PACKET_LEN)) {
1291 pr_debug("%s: rx error: len %u exceeds max size %d\n",
1292 dev->name, len, GOOD_PACKET_LEN);
1293 DEV_STATS_INC(dev, rx_length_errors);
1297 if (unlikely(vi->xdp_enabled)) {
1298 struct bpf_prog *xdp_prog;
1301 xdp_prog = rcu_dereference(rq->xdp_prog);
1303 skb = receive_small_xdp(dev, vi, rq, xdp_prog, buf,
1304 xdp_headroom, len, xdp_xmit,
1312 skb = receive_small_build_skb(vi, xdp_headroom, buf, len);
1317 u64_stats_inc(&stats->drops);
1322 static struct sk_buff *receive_big(struct net_device *dev,
1323 struct virtnet_info *vi,
1324 struct receive_queue *rq,
1327 struct virtnet_rq_stats *stats)
1329 struct page *page = buf;
1330 struct sk_buff *skb =
1331 page_to_skb(vi, rq, page, 0, len, PAGE_SIZE, 0);
1333 u64_stats_add(&stats->bytes, len - vi->hdr_len);
1340 u64_stats_inc(&stats->drops);
1341 give_pages(rq, page);
1345 static void mergeable_buf_free(struct receive_queue *rq, int num_buf,
1346 struct net_device *dev,
1347 struct virtnet_rq_stats *stats)
1353 while (num_buf-- > 1) {
1354 buf = virtnet_rq_get_buf(rq, &len, NULL);
1355 if (unlikely(!buf)) {
1356 pr_debug("%s: rx error: %d buffers missing\n",
1357 dev->name, num_buf);
1358 DEV_STATS_INC(dev, rx_length_errors);
1361 u64_stats_add(&stats->bytes, len);
1362 page = virt_to_head_page(buf);
1367 /* Why not use xdp_build_skb_from_frame() ?
1368 * XDP core assumes that xdp frags are PAGE_SIZE in length, while in
1369 * virtio-net there are 2 points that do not match its requirements:
1370 * 1. The size of the prefilled buffer is not fixed before xdp is set.
1371 * 2. xdp_build_skb_from_frame() does more checks that we don't need,
1372 * like eth_type_trans() (which virtio-net does in receive_buf()).
1374 static struct sk_buff *build_skb_from_xdp_buff(struct net_device *dev,
1375 struct virtnet_info *vi,
1376 struct xdp_buff *xdp,
1377 unsigned int xdp_frags_truesz)
1379 struct skb_shared_info *sinfo = xdp_get_shared_info_from_buff(xdp);
1380 unsigned int headroom, data_len;
1381 struct sk_buff *skb;
1385 if (unlikely(xdp->data_end > xdp_data_hard_end(xdp))) {
1386 pr_debug("Error building skb as missing reserved tailroom for xdp");
1390 if (unlikely(xdp_buff_has_frags(xdp)))
1391 nr_frags = sinfo->nr_frags;
1393 skb = build_skb(xdp->data_hard_start, xdp->frame_sz);
1397 headroom = xdp->data - xdp->data_hard_start;
1398 data_len = xdp->data_end - xdp->data;
1399 skb_reserve(skb, headroom);
1400 __skb_put(skb, data_len);
1402 metasize = xdp->data - xdp->data_meta;
1403 metasize = metasize > 0 ? metasize : 0;
1405 skb_metadata_set(skb, metasize);
1407 if (unlikely(xdp_buff_has_frags(xdp)))
1408 xdp_update_skb_shared_info(skb, nr_frags,
1409 sinfo->xdp_frags_size,
1411 xdp_buff_is_frag_pfmemalloc(xdp));
1416 /* TODO: build xdp in big mode */
1417 static int virtnet_build_xdp_buff_mrg(struct net_device *dev,
1418 struct virtnet_info *vi,
1419 struct receive_queue *rq,
1420 struct xdp_buff *xdp,
1423 unsigned int frame_sz,
1425 unsigned int *xdp_frags_truesize,
1426 struct virtnet_rq_stats *stats)
1428 struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
1429 unsigned int headroom, tailroom, room;
1430 unsigned int truesize, cur_frag_size;
1431 struct skb_shared_info *shinfo;
1432 unsigned int xdp_frags_truesz = 0;
1438 xdp_init_buff(xdp, frame_sz, &rq->xdp_rxq);
1439 xdp_prepare_buff(xdp, buf - VIRTIO_XDP_HEADROOM,
1440 VIRTIO_XDP_HEADROOM + vi->hdr_len, len - vi->hdr_len, true);
1446 /* If we want to build multi-buffer xdp, we need
1447 * to specify that the flags of xdp_buff have the
1448 * XDP_FLAGS_HAS_FRAG bit.
1450 if (!xdp_buff_has_frags(xdp))
1451 xdp_buff_set_frags_flag(xdp);
1453 shinfo = xdp_get_shared_info_from_buff(xdp);
1454 shinfo->nr_frags = 0;
1455 shinfo->xdp_frags_size = 0;
1458 if (*num_buf > MAX_SKB_FRAGS + 1)
1461 while (--*num_buf > 0) {
1462 buf = virtnet_rq_get_buf(rq, &len, &ctx);
1463 if (unlikely(!buf)) {
1464 pr_debug("%s: rx error: %d buffers out of %d missing\n",
1465 dev->name, *num_buf,
1466 virtio16_to_cpu(vi->vdev, hdr->num_buffers));
1467 DEV_STATS_INC(dev, rx_length_errors);
1471 u64_stats_add(&stats->bytes, len);
1472 page = virt_to_head_page(buf);
1473 offset = buf - page_address(page);
1475 truesize = mergeable_ctx_to_truesize(ctx);
1476 headroom = mergeable_ctx_to_headroom(ctx);
1477 tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1478 room = SKB_DATA_ALIGN(headroom + tailroom);
1480 cur_frag_size = truesize;
1481 xdp_frags_truesz += cur_frag_size;
1482 if (unlikely(len > truesize - room || cur_frag_size > PAGE_SIZE)) {
1484 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
1485 dev->name, len, (unsigned long)(truesize - room));
1486 DEV_STATS_INC(dev, rx_length_errors);
1490 frag = &shinfo->frags[shinfo->nr_frags++];
1491 skb_frag_fill_page_desc(frag, page, offset, len);
1492 if (page_is_pfmemalloc(page))
1493 xdp_buff_set_frag_pfmemalloc(xdp);
1495 shinfo->xdp_frags_size += len;
1498 *xdp_frags_truesize = xdp_frags_truesz;
1506 static void *mergeable_xdp_get_buf(struct virtnet_info *vi,
1507 struct receive_queue *rq,
1508 struct bpf_prog *xdp_prog,
1510 unsigned int *frame_sz,
1515 struct virtio_net_hdr_mrg_rxbuf *hdr)
1517 unsigned int truesize = mergeable_ctx_to_truesize(ctx);
1518 unsigned int headroom = mergeable_ctx_to_headroom(ctx);
1519 struct page *xdp_page;
1520 unsigned int xdp_room;
1522 /* Transient failure which in theory could occur if
1523 * in-flight packets from before XDP was enabled reach
1524 * the receive path after XDP is loaded.
1526 if (unlikely(hdr->hdr.gso_type))
1529 /* Now XDP core assumes frag size is PAGE_SIZE, but buffers
1530 * with headroom may add hole in truesize, which
1531 * make their length exceed PAGE_SIZE. So we disabled the
1532 * hole mechanism for xdp. See add_recvbuf_mergeable().
1534 *frame_sz = truesize;
1536 if (likely(headroom >= virtnet_get_headroom(vi) &&
1537 (*num_buf == 1 || xdp_prog->aux->xdp_has_frags))) {
1538 return page_address(*page) + offset;
1541 /* This happens when headroom is not enough because
1542 * of the buffer was prefilled before XDP is set.
1543 * This should only happen for the first several packets.
1544 * In fact, vq reset can be used here to help us clean up
1545 * the prefilled buffers, but many existing devices do not
1546 * support it, and we don't want to bother users who are
1547 * using xdp normally.
1549 if (!xdp_prog->aux->xdp_has_frags) {
1550 /* linearize data for XDP */
1551 xdp_page = xdp_linearize_page(rq, num_buf,
1553 VIRTIO_XDP_HEADROOM,
1558 xdp_room = SKB_DATA_ALIGN(VIRTIO_XDP_HEADROOM +
1559 sizeof(struct skb_shared_info));
1560 if (*len + xdp_room > PAGE_SIZE)
1563 xdp_page = alloc_page(GFP_ATOMIC);
1567 memcpy(page_address(xdp_page) + VIRTIO_XDP_HEADROOM,
1568 page_address(*page) + offset, *len);
1571 *frame_sz = PAGE_SIZE;
1577 return page_address(*page) + VIRTIO_XDP_HEADROOM;
1580 static struct sk_buff *receive_mergeable_xdp(struct net_device *dev,
1581 struct virtnet_info *vi,
1582 struct receive_queue *rq,
1583 struct bpf_prog *xdp_prog,
1587 unsigned int *xdp_xmit,
1588 struct virtnet_rq_stats *stats)
1590 struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
1591 int num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
1592 struct page *page = virt_to_head_page(buf);
1593 int offset = buf - page_address(page);
1594 unsigned int xdp_frags_truesz = 0;
1595 struct sk_buff *head_skb;
1596 unsigned int frame_sz;
1597 struct xdp_buff xdp;
1602 data = mergeable_xdp_get_buf(vi, rq, xdp_prog, ctx, &frame_sz, &num_buf, &page,
1604 if (unlikely(!data))
1607 err = virtnet_build_xdp_buff_mrg(dev, vi, rq, &xdp, data, len, frame_sz,
1608 &num_buf, &xdp_frags_truesz, stats);
1612 act = virtnet_xdp_handler(xdp_prog, &xdp, dev, xdp_xmit, stats);
1616 head_skb = build_skb_from_xdp_buff(dev, vi, &xdp, xdp_frags_truesz);
1617 if (unlikely(!head_skb))
1629 put_xdp_frags(&xdp);
1633 mergeable_buf_free(rq, num_buf, dev, stats);
1635 u64_stats_inc(&stats->xdp_drops);
1636 u64_stats_inc(&stats->drops);
1640 static struct sk_buff *receive_mergeable(struct net_device *dev,
1641 struct virtnet_info *vi,
1642 struct receive_queue *rq,
1646 unsigned int *xdp_xmit,
1647 struct virtnet_rq_stats *stats)
1649 struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
1650 int num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
1651 struct page *page = virt_to_head_page(buf);
1652 int offset = buf - page_address(page);
1653 struct sk_buff *head_skb, *curr_skb;
1654 unsigned int truesize = mergeable_ctx_to_truesize(ctx);
1655 unsigned int headroom = mergeable_ctx_to_headroom(ctx);
1656 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1657 unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
1660 u64_stats_add(&stats->bytes, len - vi->hdr_len);
1662 if (unlikely(len > truesize - room)) {
1663 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
1664 dev->name, len, (unsigned long)(truesize - room));
1665 DEV_STATS_INC(dev, rx_length_errors);
1669 if (unlikely(vi->xdp_enabled)) {
1670 struct bpf_prog *xdp_prog;
1673 xdp_prog = rcu_dereference(rq->xdp_prog);
1675 head_skb = receive_mergeable_xdp(dev, vi, rq, xdp_prog, buf, ctx,
1676 len, xdp_xmit, stats);
1683 head_skb = page_to_skb(vi, rq, page, offset, len, truesize, headroom);
1684 curr_skb = head_skb;
1686 if (unlikely(!curr_skb))
1691 buf = virtnet_rq_get_buf(rq, &len, &ctx);
1692 if (unlikely(!buf)) {
1693 pr_debug("%s: rx error: %d buffers out of %d missing\n",
1695 virtio16_to_cpu(vi->vdev,
1697 DEV_STATS_INC(dev, rx_length_errors);
1701 u64_stats_add(&stats->bytes, len);
1702 page = virt_to_head_page(buf);
1704 truesize = mergeable_ctx_to_truesize(ctx);
1705 headroom = mergeable_ctx_to_headroom(ctx);
1706 tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1707 room = SKB_DATA_ALIGN(headroom + tailroom);
1708 if (unlikely(len > truesize - room)) {
1709 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
1710 dev->name, len, (unsigned long)(truesize - room));
1711 DEV_STATS_INC(dev, rx_length_errors);
1715 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
1716 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
1717 struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
1719 if (unlikely(!nskb))
1721 if (curr_skb == head_skb)
1722 skb_shinfo(curr_skb)->frag_list = nskb;
1724 curr_skb->next = nskb;
1726 head_skb->truesize += nskb->truesize;
1729 if (curr_skb != head_skb) {
1730 head_skb->data_len += len;
1731 head_skb->len += len;
1732 head_skb->truesize += truesize;
1734 offset = buf - page_address(page);
1735 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
1737 skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
1740 skb_add_rx_frag(curr_skb, num_skb_frags, page,
1741 offset, len, truesize);
1745 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
1750 mergeable_buf_free(rq, num_buf, dev, stats);
1753 u64_stats_inc(&stats->drops);
1754 dev_kfree_skb(head_skb);
1758 static void virtio_skb_set_hash(const struct virtio_net_hdr_v1_hash *hdr_hash,
1759 struct sk_buff *skb)
1761 enum pkt_hash_types rss_hash_type;
1763 if (!hdr_hash || !skb)
1766 switch (__le16_to_cpu(hdr_hash->hash_report)) {
1767 case VIRTIO_NET_HASH_REPORT_TCPv4:
1768 case VIRTIO_NET_HASH_REPORT_UDPv4:
1769 case VIRTIO_NET_HASH_REPORT_TCPv6:
1770 case VIRTIO_NET_HASH_REPORT_UDPv6:
1771 case VIRTIO_NET_HASH_REPORT_TCPv6_EX:
1772 case VIRTIO_NET_HASH_REPORT_UDPv6_EX:
1773 rss_hash_type = PKT_HASH_TYPE_L4;
1775 case VIRTIO_NET_HASH_REPORT_IPv4:
1776 case VIRTIO_NET_HASH_REPORT_IPv6:
1777 case VIRTIO_NET_HASH_REPORT_IPv6_EX:
1778 rss_hash_type = PKT_HASH_TYPE_L3;
1780 case VIRTIO_NET_HASH_REPORT_NONE:
1782 rss_hash_type = PKT_HASH_TYPE_NONE;
1784 skb_set_hash(skb, __le32_to_cpu(hdr_hash->hash_value), rss_hash_type);
1787 static void receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
1788 void *buf, unsigned int len, void **ctx,
1789 unsigned int *xdp_xmit,
1790 struct virtnet_rq_stats *stats)
1792 struct net_device *dev = vi->dev;
1793 struct sk_buff *skb;
1794 struct virtio_net_common_hdr *hdr;
1796 if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
1797 pr_debug("%s: short packet %i\n", dev->name, len);
1798 DEV_STATS_INC(dev, rx_length_errors);
1799 virtnet_rq_free_buf(vi, rq, buf);
1803 if (vi->mergeable_rx_bufs)
1804 skb = receive_mergeable(dev, vi, rq, buf, ctx, len, xdp_xmit,
1806 else if (vi->big_packets)
1807 skb = receive_big(dev, vi, rq, buf, len, stats);
1809 skb = receive_small(dev, vi, rq, buf, ctx, len, xdp_xmit, stats);
1814 hdr = skb_vnet_common_hdr(skb);
1815 if (dev->features & NETIF_F_RXHASH && vi->has_rss_hash_report)
1816 virtio_skb_set_hash(&hdr->hash_v1_hdr, skb);
1818 if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
1819 skb->ip_summed = CHECKSUM_UNNECESSARY;
1821 if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
1822 virtio_is_little_endian(vi->vdev))) {
1823 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
1824 dev->name, hdr->hdr.gso_type,
1829 skb_record_rx_queue(skb, vq2rxq(rq->vq));
1830 skb->protocol = eth_type_trans(skb, dev);
1831 pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
1832 ntohs(skb->protocol), skb->len, skb->pkt_type);
1834 napi_gro_receive(&rq->napi, skb);
1838 DEV_STATS_INC(dev, rx_frame_errors);
1842 /* Unlike mergeable buffers, all buffers are allocated to the
1843 * same size, except for the headroom. For this reason we do
1844 * not need to use mergeable_len_to_ctx here - it is enough
1845 * to store the headroom as the context ignoring the truesize.
1847 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
1851 unsigned int xdp_headroom = virtnet_get_headroom(vi);
1852 void *ctx = (void *)(unsigned long)xdp_headroom;
1853 int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
1856 len = SKB_DATA_ALIGN(len) +
1857 SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
1859 buf = virtnet_rq_alloc(rq, len, gfp);
1863 virtnet_rq_init_one_sg(rq, buf + VIRTNET_RX_PAD + xdp_headroom,
1864 vi->hdr_len + GOOD_PACKET_LEN);
1866 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1869 virtnet_rq_unmap(rq, buf, 0);
1870 put_page(virt_to_head_page(buf));
1876 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
1879 struct page *first, *list = NULL;
1883 sg_init_table(rq->sg, vi->big_packets_num_skbfrags + 2);
1885 /* page in rq->sg[vi->big_packets_num_skbfrags + 1] is list tail */
1886 for (i = vi->big_packets_num_skbfrags + 1; i > 1; --i) {
1887 first = get_a_page(rq, gfp);
1890 give_pages(rq, list);
1893 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
1895 /* chain new page in list head to match sg */
1896 first->private = (unsigned long)list;
1900 first = get_a_page(rq, gfp);
1902 give_pages(rq, list);
1905 p = page_address(first);
1907 /* rq->sg[0], rq->sg[1] share the same page */
1908 /* a separated rq->sg[0] for header - required in case !any_header_sg */
1909 sg_set_buf(&rq->sg[0], p, vi->hdr_len);
1911 /* rq->sg[1] for data packet, from offset */
1912 offset = sizeof(struct padded_vnet_hdr);
1913 sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
1915 /* chain first in list head */
1916 first->private = (unsigned long)list;
1917 err = virtqueue_add_inbuf(rq->vq, rq->sg, vi->big_packets_num_skbfrags + 2,
1920 give_pages(rq, first);
1925 static unsigned int get_mergeable_buf_len(struct receive_queue *rq,
1926 struct ewma_pkt_len *avg_pkt_len,
1929 struct virtnet_info *vi = rq->vq->vdev->priv;
1930 const size_t hdr_len = vi->hdr_len;
1934 return PAGE_SIZE - room;
1936 len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
1937 rq->min_buf_len, PAGE_SIZE - hdr_len);
1939 return ALIGN(len, L1_CACHE_BYTES);
1942 static int add_recvbuf_mergeable(struct virtnet_info *vi,
1943 struct receive_queue *rq, gfp_t gfp)
1945 struct page_frag *alloc_frag = &rq->alloc_frag;
1946 unsigned int headroom = virtnet_get_headroom(vi);
1947 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
1948 unsigned int room = SKB_DATA_ALIGN(headroom + tailroom);
1949 unsigned int len, hole;
1954 /* Extra tailroom is needed to satisfy XDP's assumption. This
1955 * means rx frags coalescing won't work, but consider we've
1956 * disabled GSO for XDP, it won't be a big issue.
1958 len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len, room);
1960 buf = virtnet_rq_alloc(rq, len + room, gfp);
1964 buf += headroom; /* advance address leaving hole at front of pkt */
1965 hole = alloc_frag->size - alloc_frag->offset;
1966 if (hole < len + room) {
1967 /* To avoid internal fragmentation, if there is very likely not
1968 * enough space for another buffer, add the remaining space to
1969 * the current buffer.
1970 * XDP core assumes that frame_size of xdp_buff and the length
1971 * of the frag are PAGE_SIZE, so we disable the hole mechanism.
1975 alloc_frag->offset += hole;
1978 virtnet_rq_init_one_sg(rq, buf, len);
1980 ctx = mergeable_len_to_ctx(len + room, headroom);
1981 err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
1984 virtnet_rq_unmap(rq, buf, 0);
1985 put_page(virt_to_head_page(buf));
1992 * Returns false if we couldn't fill entirely (OOM).
1994 * Normally run in the receive path, but can also be run from ndo_open
1995 * before we're receiving packets, or from refill_work which is
1996 * careful to disable receiving (using napi_disable).
1998 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
2005 if (vi->mergeable_rx_bufs)
2006 err = add_recvbuf_mergeable(vi, rq, gfp);
2007 else if (vi->big_packets)
2008 err = add_recvbuf_big(vi, rq, gfp);
2010 err = add_recvbuf_small(vi, rq, gfp);
2012 oom = err == -ENOMEM;
2015 } while (rq->vq->num_free);
2016 if (virtqueue_kick_prepare(rq->vq) && virtqueue_notify(rq->vq)) {
2017 unsigned long flags;
2019 flags = u64_stats_update_begin_irqsave(&rq->stats.syncp);
2020 u64_stats_inc(&rq->stats.kicks);
2021 u64_stats_update_end_irqrestore(&rq->stats.syncp, flags);
2027 static void skb_recv_done(struct virtqueue *rvq)
2029 struct virtnet_info *vi = rvq->vdev->priv;
2030 struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
2033 virtqueue_napi_schedule(&rq->napi, rvq);
2036 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
2040 /* If all buffers were filled by other side before we napi_enabled, we
2041 * won't get another interrupt, so process any outstanding packets now.
2042 * Call local_bh_enable after to trigger softIRQ processing.
2045 virtqueue_napi_schedule(napi, vq);
2049 static void virtnet_napi_tx_enable(struct virtnet_info *vi,
2050 struct virtqueue *vq,
2051 struct napi_struct *napi)
2056 /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
2057 * enable the feature if this is likely affine with the transmit path.
2059 if (!vi->affinity_hint_set) {
2064 return virtnet_napi_enable(vq, napi);
2067 static void virtnet_napi_tx_disable(struct napi_struct *napi)
2073 static void refill_work(struct work_struct *work)
2075 struct virtnet_info *vi =
2076 container_of(work, struct virtnet_info, refill.work);
2080 for (i = 0; i < vi->curr_queue_pairs; i++) {
2081 struct receive_queue *rq = &vi->rq[i];
2083 napi_disable(&rq->napi);
2084 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
2085 virtnet_napi_enable(rq->vq, &rq->napi);
2087 /* In theory, this can happen: if we don't get any buffers in
2088 * we will *never* try to fill again.
2091 schedule_delayed_work(&vi->refill, HZ/2);
2095 static int virtnet_receive(struct receive_queue *rq, int budget,
2096 unsigned int *xdp_xmit)
2098 struct virtnet_info *vi = rq->vq->vdev->priv;
2099 struct virtnet_rq_stats stats = {};
2105 if (!vi->big_packets || vi->mergeable_rx_bufs) {
2108 while (packets < budget &&
2109 (buf = virtnet_rq_get_buf(rq, &len, &ctx))) {
2110 receive_buf(vi, rq, buf, len, ctx, xdp_xmit, &stats);
2114 while (packets < budget &&
2115 (buf = virtnet_rq_get_buf(rq, &len, NULL)) != NULL) {
2116 receive_buf(vi, rq, buf, len, NULL, xdp_xmit, &stats);
2121 if (rq->vq->num_free > min((unsigned int)budget, virtqueue_get_vring_size(rq->vq)) / 2) {
2122 if (!try_fill_recv(vi, rq, GFP_ATOMIC)) {
2123 spin_lock(&vi->refill_lock);
2124 if (vi->refill_enabled)
2125 schedule_delayed_work(&vi->refill, 0);
2126 spin_unlock(&vi->refill_lock);
2130 u64_stats_set(&stats.packets, packets);
2131 u64_stats_update_begin(&rq->stats.syncp);
2132 for (i = 0; i < VIRTNET_RQ_STATS_LEN; i++) {
2133 size_t offset = virtnet_rq_stats_desc[i].offset;
2134 u64_stats_t *item, *src;
2136 item = (u64_stats_t *)((u8 *)&rq->stats + offset);
2137 src = (u64_stats_t *)((u8 *)&stats + offset);
2138 u64_stats_add(item, u64_stats_read(src));
2140 u64_stats_update_end(&rq->stats.syncp);
2145 static void virtnet_poll_cleantx(struct receive_queue *rq)
2147 struct virtnet_info *vi = rq->vq->vdev->priv;
2148 unsigned int index = vq2rxq(rq->vq);
2149 struct send_queue *sq = &vi->sq[index];
2150 struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index);
2152 if (!sq->napi.weight || is_xdp_raw_buffer_queue(vi, index))
2155 if (__netif_tx_trylock(txq)) {
2157 __netif_tx_unlock(txq);
2162 virtqueue_disable_cb(sq->vq);
2163 free_old_xmit_skbs(sq, true);
2164 } while (unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
2166 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
2167 netif_tx_wake_queue(txq);
2169 __netif_tx_unlock(txq);
2173 static void virtnet_rx_dim_update(struct virtnet_info *vi, struct receive_queue *rq)
2175 struct dim_sample cur_sample = {};
2177 if (!rq->packets_in_napi)
2180 u64_stats_update_begin(&rq->stats.syncp);
2181 dim_update_sample(rq->calls,
2182 u64_stats_read(&rq->stats.packets),
2183 u64_stats_read(&rq->stats.bytes),
2185 u64_stats_update_end(&rq->stats.syncp);
2187 net_dim(&rq->dim, cur_sample);
2188 rq->packets_in_napi = 0;
2191 static int virtnet_poll(struct napi_struct *napi, int budget)
2193 struct receive_queue *rq =
2194 container_of(napi, struct receive_queue, napi);
2195 struct virtnet_info *vi = rq->vq->vdev->priv;
2196 struct send_queue *sq;
2197 unsigned int received;
2198 unsigned int xdp_xmit = 0;
2201 virtnet_poll_cleantx(rq);
2203 received = virtnet_receive(rq, budget, &xdp_xmit);
2204 rq->packets_in_napi += received;
2206 if (xdp_xmit & VIRTIO_XDP_REDIR)
2209 /* Out of packets? */
2210 if (received < budget) {
2211 napi_complete = virtqueue_napi_complete(napi, rq->vq, received);
2212 if (napi_complete && rq->dim_enabled)
2213 virtnet_rx_dim_update(vi, rq);
2216 if (xdp_xmit & VIRTIO_XDP_TX) {
2217 sq = virtnet_xdp_get_sq(vi);
2218 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
2219 u64_stats_update_begin(&sq->stats.syncp);
2220 u64_stats_inc(&sq->stats.kicks);
2221 u64_stats_update_end(&sq->stats.syncp);
2223 virtnet_xdp_put_sq(vi, sq);
2229 static void virtnet_disable_queue_pair(struct virtnet_info *vi, int qp_index)
2231 virtnet_napi_tx_disable(&vi->sq[qp_index].napi);
2232 napi_disable(&vi->rq[qp_index].napi);
2233 xdp_rxq_info_unreg(&vi->rq[qp_index].xdp_rxq);
2236 static int virtnet_enable_queue_pair(struct virtnet_info *vi, int qp_index)
2238 struct net_device *dev = vi->dev;
2241 err = xdp_rxq_info_reg(&vi->rq[qp_index].xdp_rxq, dev, qp_index,
2242 vi->rq[qp_index].napi.napi_id);
2246 err = xdp_rxq_info_reg_mem_model(&vi->rq[qp_index].xdp_rxq,
2247 MEM_TYPE_PAGE_SHARED, NULL);
2249 goto err_xdp_reg_mem_model;
2251 virtnet_napi_enable(vi->rq[qp_index].vq, &vi->rq[qp_index].napi);
2252 virtnet_napi_tx_enable(vi, vi->sq[qp_index].vq, &vi->sq[qp_index].napi);
2256 err_xdp_reg_mem_model:
2257 xdp_rxq_info_unreg(&vi->rq[qp_index].xdp_rxq);
2261 static int virtnet_open(struct net_device *dev)
2263 struct virtnet_info *vi = netdev_priv(dev);
2266 enable_delayed_refill(vi);
2268 for (i = 0; i < vi->max_queue_pairs; i++) {
2269 if (i < vi->curr_queue_pairs)
2270 /* Make sure we have some buffers: if oom use wq. */
2271 if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
2272 schedule_delayed_work(&vi->refill, 0);
2274 err = virtnet_enable_queue_pair(vi, i);
2282 disable_delayed_refill(vi);
2283 cancel_delayed_work_sync(&vi->refill);
2285 for (i--; i >= 0; i--) {
2286 virtnet_disable_queue_pair(vi, i);
2287 cancel_work_sync(&vi->rq[i].dim.work);
2293 static int virtnet_poll_tx(struct napi_struct *napi, int budget)
2295 struct send_queue *sq = container_of(napi, struct send_queue, napi);
2296 struct virtnet_info *vi = sq->vq->vdev->priv;
2297 unsigned int index = vq2txq(sq->vq);
2298 struct netdev_queue *txq;
2302 if (unlikely(is_xdp_raw_buffer_queue(vi, index))) {
2303 /* We don't need to enable cb for XDP */
2304 napi_complete_done(napi, 0);
2308 txq = netdev_get_tx_queue(vi->dev, index);
2309 __netif_tx_lock(txq, raw_smp_processor_id());
2310 virtqueue_disable_cb(sq->vq);
2311 free_old_xmit_skbs(sq, true);
2313 if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
2314 netif_tx_wake_queue(txq);
2316 opaque = virtqueue_enable_cb_prepare(sq->vq);
2318 done = napi_complete_done(napi, 0);
2321 virtqueue_disable_cb(sq->vq);
2323 __netif_tx_unlock(txq);
2326 if (unlikely(virtqueue_poll(sq->vq, opaque))) {
2327 if (napi_schedule_prep(napi)) {
2328 __netif_tx_lock(txq, raw_smp_processor_id());
2329 virtqueue_disable_cb(sq->vq);
2330 __netif_tx_unlock(txq);
2331 __napi_schedule(napi);
2339 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
2341 struct virtio_net_hdr_mrg_rxbuf *hdr;
2342 const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
2343 struct virtnet_info *vi = sq->vq->vdev->priv;
2345 unsigned hdr_len = vi->hdr_len;
2348 pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
2350 can_push = vi->any_header_sg &&
2351 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
2352 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
2353 /* Even if we can, don't push here yet as this would skew
2354 * csum_start offset below. */
2356 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
2358 hdr = &skb_vnet_common_hdr(skb)->mrg_hdr;
2360 if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
2361 virtio_is_little_endian(vi->vdev), false,
2365 if (vi->mergeable_rx_bufs)
2366 hdr->num_buffers = 0;
2368 sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
2370 __skb_push(skb, hdr_len);
2371 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
2372 if (unlikely(num_sg < 0))
2374 /* Pull header back to avoid skew in tx bytes calculations. */
2375 __skb_pull(skb, hdr_len);
2377 sg_set_buf(sq->sg, hdr, hdr_len);
2378 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
2379 if (unlikely(num_sg < 0))
2383 return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
2386 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
2388 struct virtnet_info *vi = netdev_priv(dev);
2389 int qnum = skb_get_queue_mapping(skb);
2390 struct send_queue *sq = &vi->sq[qnum];
2392 struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
2393 bool kick = !netdev_xmit_more();
2394 bool use_napi = sq->napi.weight;
2396 /* Free up any pending old buffers before queueing new ones. */
2399 virtqueue_disable_cb(sq->vq);
2401 free_old_xmit_skbs(sq, false);
2403 } while (use_napi && kick &&
2404 unlikely(!virtqueue_enable_cb_delayed(sq->vq)));
2406 /* timestamp packet in software */
2407 skb_tx_timestamp(skb);
2409 /* Try to transmit */
2410 err = xmit_skb(sq, skb);
2412 /* This should not happen! */
2413 if (unlikely(err)) {
2414 DEV_STATS_INC(dev, tx_fifo_errors);
2415 if (net_ratelimit())
2417 "Unexpected TXQ (%d) queue failure: %d\n",
2419 DEV_STATS_INC(dev, tx_dropped);
2420 dev_kfree_skb_any(skb);
2421 return NETDEV_TX_OK;
2424 /* Don't wait up for transmitted skbs to be freed. */
2430 check_sq_full_and_disable(vi, dev, sq);
2432 if (kick || netif_xmit_stopped(txq)) {
2433 if (virtqueue_kick_prepare(sq->vq) && virtqueue_notify(sq->vq)) {
2434 u64_stats_update_begin(&sq->stats.syncp);
2435 u64_stats_inc(&sq->stats.kicks);
2436 u64_stats_update_end(&sq->stats.syncp);
2440 return NETDEV_TX_OK;
2443 static int virtnet_rx_resize(struct virtnet_info *vi,
2444 struct receive_queue *rq, u32 ring_num)
2446 bool running = netif_running(vi->dev);
2449 qindex = rq - vi->rq;
2452 napi_disable(&rq->napi);
2453 cancel_work_sync(&rq->dim.work);
2456 err = virtqueue_resize(rq->vq, ring_num, virtnet_rq_unmap_free_buf);
2458 netdev_err(vi->dev, "resize rx fail: rx queue index: %d err: %d\n", qindex, err);
2460 if (!try_fill_recv(vi, rq, GFP_KERNEL))
2461 schedule_delayed_work(&vi->refill, 0);
2464 virtnet_napi_enable(rq->vq, &rq->napi);
2468 static int virtnet_tx_resize(struct virtnet_info *vi,
2469 struct send_queue *sq, u32 ring_num)
2471 bool running = netif_running(vi->dev);
2472 struct netdev_queue *txq;
2475 qindex = sq - vi->sq;
2478 virtnet_napi_tx_disable(&sq->napi);
2480 txq = netdev_get_tx_queue(vi->dev, qindex);
2482 /* 1. wait all ximt complete
2483 * 2. fix the race of netif_stop_subqueue() vs netif_start_subqueue()
2485 __netif_tx_lock_bh(txq);
2487 /* Prevent rx poll from accessing sq. */
2490 /* Prevent the upper layer from trying to send packets. */
2491 netif_stop_subqueue(vi->dev, qindex);
2493 __netif_tx_unlock_bh(txq);
2495 err = virtqueue_resize(sq->vq, ring_num, virtnet_sq_free_unused_buf);
2497 netdev_err(vi->dev, "resize tx fail: tx queue index: %d err: %d\n", qindex, err);
2499 __netif_tx_lock_bh(txq);
2501 netif_tx_wake_queue(txq);
2502 __netif_tx_unlock_bh(txq);
2505 virtnet_napi_tx_enable(vi, sq->vq, &sq->napi);
2510 * Send command via the control virtqueue and check status. Commands
2511 * supported by the hypervisor, as indicated by feature bits, should
2512 * never fail unless improperly formatted.
2514 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
2515 struct scatterlist *out)
2517 struct scatterlist *sgs[4], hdr, stat;
2518 unsigned out_num = 0, tmp;
2521 /* Caller should know better */
2522 BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
2524 vi->ctrl->status = ~0;
2525 vi->ctrl->hdr.class = class;
2526 vi->ctrl->hdr.cmd = cmd;
2528 sg_init_one(&hdr, &vi->ctrl->hdr, sizeof(vi->ctrl->hdr));
2529 sgs[out_num++] = &hdr;
2532 sgs[out_num++] = out;
2534 /* Add return status. */
2535 sg_init_one(&stat, &vi->ctrl->status, sizeof(vi->ctrl->status));
2536 sgs[out_num] = &stat;
2538 BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
2539 ret = virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
2541 dev_warn(&vi->vdev->dev,
2542 "Failed to add sgs for command vq: %d\n.", ret);
2546 if (unlikely(!virtqueue_kick(vi->cvq)))
2547 return vi->ctrl->status == VIRTIO_NET_OK;
2549 /* Spin for a response, the kick causes an ioport write, trapping
2550 * into the hypervisor, so the request should be handled immediately.
2552 while (!virtqueue_get_buf(vi->cvq, &tmp) &&
2553 !virtqueue_is_broken(vi->cvq))
2556 return vi->ctrl->status == VIRTIO_NET_OK;
2559 static int virtnet_set_mac_address(struct net_device *dev, void *p)
2561 struct virtnet_info *vi = netdev_priv(dev);
2562 struct virtio_device *vdev = vi->vdev;
2564 struct sockaddr *addr;
2565 struct scatterlist sg;
2567 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
2570 addr = kmemdup(p, sizeof(*addr), GFP_KERNEL);
2574 ret = eth_prepare_mac_addr_change(dev, addr);
2578 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
2579 sg_init_one(&sg, addr->sa_data, dev->addr_len);
2580 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
2581 VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
2582 dev_warn(&vdev->dev,
2583 "Failed to set mac address by vq command.\n");
2587 } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
2588 !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
2591 /* Naturally, this has an atomicity problem. */
2592 for (i = 0; i < dev->addr_len; i++)
2593 virtio_cwrite8(vdev,
2594 offsetof(struct virtio_net_config, mac) +
2595 i, addr->sa_data[i]);
2598 eth_commit_mac_addr_change(dev, p);
2606 static void virtnet_stats(struct net_device *dev,
2607 struct rtnl_link_stats64 *tot)
2609 struct virtnet_info *vi = netdev_priv(dev);
2613 for (i = 0; i < vi->max_queue_pairs; i++) {
2614 u64 tpackets, tbytes, terrors, rpackets, rbytes, rdrops;
2615 struct receive_queue *rq = &vi->rq[i];
2616 struct send_queue *sq = &vi->sq[i];
2619 start = u64_stats_fetch_begin(&sq->stats.syncp);
2620 tpackets = u64_stats_read(&sq->stats.packets);
2621 tbytes = u64_stats_read(&sq->stats.bytes);
2622 terrors = u64_stats_read(&sq->stats.tx_timeouts);
2623 } while (u64_stats_fetch_retry(&sq->stats.syncp, start));
2626 start = u64_stats_fetch_begin(&rq->stats.syncp);
2627 rpackets = u64_stats_read(&rq->stats.packets);
2628 rbytes = u64_stats_read(&rq->stats.bytes);
2629 rdrops = u64_stats_read(&rq->stats.drops);
2630 } while (u64_stats_fetch_retry(&rq->stats.syncp, start));
2632 tot->rx_packets += rpackets;
2633 tot->tx_packets += tpackets;
2634 tot->rx_bytes += rbytes;
2635 tot->tx_bytes += tbytes;
2636 tot->rx_dropped += rdrops;
2637 tot->tx_errors += terrors;
2640 tot->tx_dropped = DEV_STATS_READ(dev, tx_dropped);
2641 tot->tx_fifo_errors = DEV_STATS_READ(dev, tx_fifo_errors);
2642 tot->rx_length_errors = DEV_STATS_READ(dev, rx_length_errors);
2643 tot->rx_frame_errors = DEV_STATS_READ(dev, rx_frame_errors);
2646 static void virtnet_ack_link_announce(struct virtnet_info *vi)
2649 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
2650 VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
2651 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
2655 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
2657 struct scatterlist sg;
2658 struct net_device *dev = vi->dev;
2660 if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
2663 vi->ctrl->mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
2664 sg_init_one(&sg, &vi->ctrl->mq, sizeof(vi->ctrl->mq));
2666 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
2667 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
2668 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
2672 vi->curr_queue_pairs = queue_pairs;
2673 /* virtnet_open() will refill when device is going to up. */
2674 if (dev->flags & IFF_UP)
2675 schedule_delayed_work(&vi->refill, 0);
2681 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
2686 err = _virtnet_set_queues(vi, queue_pairs);
2691 static int virtnet_close(struct net_device *dev)
2693 struct virtnet_info *vi = netdev_priv(dev);
2696 /* Make sure NAPI doesn't schedule refill work */
2697 disable_delayed_refill(vi);
2698 /* Make sure refill_work doesn't re-enable napi! */
2699 cancel_delayed_work_sync(&vi->refill);
2701 for (i = 0; i < vi->max_queue_pairs; i++) {
2702 virtnet_disable_queue_pair(vi, i);
2703 cancel_work_sync(&vi->rq[i].dim.work);
2709 static void virtnet_set_rx_mode(struct net_device *dev)
2711 struct virtnet_info *vi = netdev_priv(dev);
2712 struct scatterlist sg[2];
2713 struct virtio_net_ctrl_mac *mac_data;
2714 struct netdev_hw_addr *ha;
2720 /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
2721 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
2724 vi->ctrl->promisc = ((dev->flags & IFF_PROMISC) != 0);
2725 vi->ctrl->allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
2727 sg_init_one(sg, &vi->ctrl->promisc, sizeof(vi->ctrl->promisc));
2729 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
2730 VIRTIO_NET_CTRL_RX_PROMISC, sg))
2731 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
2732 vi->ctrl->promisc ? "en" : "dis");
2734 sg_init_one(sg, &vi->ctrl->allmulti, sizeof(vi->ctrl->allmulti));
2736 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
2737 VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
2738 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
2739 vi->ctrl->allmulti ? "en" : "dis");
2741 uc_count = netdev_uc_count(dev);
2742 mc_count = netdev_mc_count(dev);
2743 /* MAC filter - use one buffer for both lists */
2744 buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
2745 (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
2750 sg_init_table(sg, 2);
2752 /* Store the unicast list and count in the front of the buffer */
2753 mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
2755 netdev_for_each_uc_addr(ha, dev)
2756 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
2758 sg_set_buf(&sg[0], mac_data,
2759 sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
2761 /* multicast list and count fill the end */
2762 mac_data = (void *)&mac_data->macs[uc_count][0];
2764 mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
2766 netdev_for_each_mc_addr(ha, dev)
2767 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
2769 sg_set_buf(&sg[1], mac_data,
2770 sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
2772 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
2773 VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
2774 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
2779 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
2780 __be16 proto, u16 vid)
2782 struct virtnet_info *vi = netdev_priv(dev);
2783 struct scatterlist sg;
2785 vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
2786 sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
2788 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
2789 VIRTIO_NET_CTRL_VLAN_ADD, &sg))
2790 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
2794 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
2795 __be16 proto, u16 vid)
2797 struct virtnet_info *vi = netdev_priv(dev);
2798 struct scatterlist sg;
2800 vi->ctrl->vid = cpu_to_virtio16(vi->vdev, vid);
2801 sg_init_one(&sg, &vi->ctrl->vid, sizeof(vi->ctrl->vid));
2803 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
2804 VIRTIO_NET_CTRL_VLAN_DEL, &sg))
2805 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
2809 static void virtnet_clean_affinity(struct virtnet_info *vi)
2813 if (vi->affinity_hint_set) {
2814 for (i = 0; i < vi->max_queue_pairs; i++) {
2815 virtqueue_set_affinity(vi->rq[i].vq, NULL);
2816 virtqueue_set_affinity(vi->sq[i].vq, NULL);
2819 vi->affinity_hint_set = false;
2823 static void virtnet_set_affinity(struct virtnet_info *vi)
2832 if (!zalloc_cpumask_var(&mask, GFP_KERNEL)) {
2833 virtnet_clean_affinity(vi);
2837 num_cpu = num_online_cpus();
2838 stride = max_t(int, num_cpu / vi->curr_queue_pairs, 1);
2839 stragglers = num_cpu >= vi->curr_queue_pairs ?
2840 num_cpu % vi->curr_queue_pairs :
2842 cpu = cpumask_first(cpu_online_mask);
2844 for (i = 0; i < vi->curr_queue_pairs; i++) {
2845 group_size = stride + (i < stragglers ? 1 : 0);
2847 for (j = 0; j < group_size; j++) {
2848 cpumask_set_cpu(cpu, mask);
2849 cpu = cpumask_next_wrap(cpu, cpu_online_mask,
2852 virtqueue_set_affinity(vi->rq[i].vq, mask);
2853 virtqueue_set_affinity(vi->sq[i].vq, mask);
2854 __netif_set_xps_queue(vi->dev, cpumask_bits(mask), i, XPS_CPUS);
2855 cpumask_clear(mask);
2858 vi->affinity_hint_set = true;
2859 free_cpumask_var(mask);
2862 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
2864 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2866 virtnet_set_affinity(vi);
2870 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
2872 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2874 virtnet_set_affinity(vi);
2878 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
2880 struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
2883 virtnet_clean_affinity(vi);
2887 static enum cpuhp_state virtionet_online;
2889 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
2893 ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
2896 ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2900 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2904 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
2906 cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
2907 cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
2911 static int virtnet_send_ctrl_coal_vq_cmd(struct virtnet_info *vi,
2912 u16 vqn, u32 max_usecs, u32 max_packets)
2914 struct scatterlist sgs;
2916 vi->ctrl->coal_vq.vqn = cpu_to_le16(vqn);
2917 vi->ctrl->coal_vq.coal.max_usecs = cpu_to_le32(max_usecs);
2918 vi->ctrl->coal_vq.coal.max_packets = cpu_to_le32(max_packets);
2919 sg_init_one(&sgs, &vi->ctrl->coal_vq, sizeof(vi->ctrl->coal_vq));
2921 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_NOTF_COAL,
2922 VIRTIO_NET_CTRL_NOTF_COAL_VQ_SET,
2929 static int virtnet_send_rx_ctrl_coal_vq_cmd(struct virtnet_info *vi,
2930 u16 queue, u32 max_usecs,
2935 err = virtnet_send_ctrl_coal_vq_cmd(vi, rxq2vq(queue),
2936 max_usecs, max_packets);
2940 vi->rq[queue].intr_coal.max_usecs = max_usecs;
2941 vi->rq[queue].intr_coal.max_packets = max_packets;
2946 static int virtnet_send_tx_ctrl_coal_vq_cmd(struct virtnet_info *vi,
2947 u16 queue, u32 max_usecs,
2952 err = virtnet_send_ctrl_coal_vq_cmd(vi, txq2vq(queue),
2953 max_usecs, max_packets);
2957 vi->sq[queue].intr_coal.max_usecs = max_usecs;
2958 vi->sq[queue].intr_coal.max_packets = max_packets;
2963 static void virtnet_get_ringparam(struct net_device *dev,
2964 struct ethtool_ringparam *ring,
2965 struct kernel_ethtool_ringparam *kernel_ring,
2966 struct netlink_ext_ack *extack)
2968 struct virtnet_info *vi = netdev_priv(dev);
2970 ring->rx_max_pending = vi->rq[0].vq->num_max;
2971 ring->tx_max_pending = vi->sq[0].vq->num_max;
2972 ring->rx_pending = virtqueue_get_vring_size(vi->rq[0].vq);
2973 ring->tx_pending = virtqueue_get_vring_size(vi->sq[0].vq);
2976 static int virtnet_set_ringparam(struct net_device *dev,
2977 struct ethtool_ringparam *ring,
2978 struct kernel_ethtool_ringparam *kernel_ring,
2979 struct netlink_ext_ack *extack)
2981 struct virtnet_info *vi = netdev_priv(dev);
2982 u32 rx_pending, tx_pending;
2983 struct receive_queue *rq;
2984 struct send_queue *sq;
2987 if (ring->rx_mini_pending || ring->rx_jumbo_pending)
2990 rx_pending = virtqueue_get_vring_size(vi->rq[0].vq);
2991 tx_pending = virtqueue_get_vring_size(vi->sq[0].vq);
2993 if (ring->rx_pending == rx_pending &&
2994 ring->tx_pending == tx_pending)
2997 if (ring->rx_pending > vi->rq[0].vq->num_max)
3000 if (ring->tx_pending > vi->sq[0].vq->num_max)
3003 for (i = 0; i < vi->max_queue_pairs; i++) {
3007 if (ring->tx_pending != tx_pending) {
3008 err = virtnet_tx_resize(vi, sq, ring->tx_pending);
3012 /* Upon disabling and re-enabling a transmit virtqueue, the device must
3013 * set the coalescing parameters of the virtqueue to those configured
3014 * through the VIRTIO_NET_CTRL_NOTF_COAL_TX_SET command, or, if the driver
3015 * did not set any TX coalescing parameters, to 0.
3017 err = virtnet_send_tx_ctrl_coal_vq_cmd(vi, i,
3018 vi->intr_coal_tx.max_usecs,
3019 vi->intr_coal_tx.max_packets);
3024 if (ring->rx_pending != rx_pending) {
3025 err = virtnet_rx_resize(vi, rq, ring->rx_pending);
3029 /* The reason is same as the transmit virtqueue reset */
3030 err = virtnet_send_rx_ctrl_coal_vq_cmd(vi, i,
3031 vi->intr_coal_rx.max_usecs,
3032 vi->intr_coal_rx.max_packets);
3041 static bool virtnet_commit_rss_command(struct virtnet_info *vi)
3043 struct net_device *dev = vi->dev;
3044 struct scatterlist sgs[4];
3045 unsigned int sg_buf_size;
3048 sg_init_table(sgs, 4);
3050 sg_buf_size = offsetof(struct virtio_net_ctrl_rss, indirection_table);
3051 sg_set_buf(&sgs[0], &vi->ctrl->rss, sg_buf_size);
3053 sg_buf_size = sizeof(uint16_t) * (vi->ctrl->rss.indirection_table_mask + 1);
3054 sg_set_buf(&sgs[1], vi->ctrl->rss.indirection_table, sg_buf_size);
3056 sg_buf_size = offsetof(struct virtio_net_ctrl_rss, key)
3057 - offsetof(struct virtio_net_ctrl_rss, max_tx_vq);
3058 sg_set_buf(&sgs[2], &vi->ctrl->rss.max_tx_vq, sg_buf_size);
3060 sg_buf_size = vi->rss_key_size;
3061 sg_set_buf(&sgs[3], vi->ctrl->rss.key, sg_buf_size);
3063 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
3064 vi->has_rss ? VIRTIO_NET_CTRL_MQ_RSS_CONFIG
3065 : VIRTIO_NET_CTRL_MQ_HASH_CONFIG, sgs)) {
3066 dev_warn(&dev->dev, "VIRTIONET issue with committing RSS sgs\n");
3072 static void virtnet_init_default_rss(struct virtnet_info *vi)
3077 vi->ctrl->rss.hash_types = vi->rss_hash_types_supported;
3078 vi->rss_hash_types_saved = vi->rss_hash_types_supported;
3079 vi->ctrl->rss.indirection_table_mask = vi->rss_indir_table_size
3080 ? vi->rss_indir_table_size - 1 : 0;
3081 vi->ctrl->rss.unclassified_queue = 0;
3083 for (; i < vi->rss_indir_table_size; ++i) {
3084 indir_val = ethtool_rxfh_indir_default(i, vi->curr_queue_pairs);
3085 vi->ctrl->rss.indirection_table[i] = indir_val;
3088 vi->ctrl->rss.max_tx_vq = vi->has_rss ? vi->curr_queue_pairs : 0;
3089 vi->ctrl->rss.hash_key_length = vi->rss_key_size;
3091 netdev_rss_key_fill(vi->ctrl->rss.key, vi->rss_key_size);
3094 static void virtnet_get_hashflow(const struct virtnet_info *vi, struct ethtool_rxnfc *info)
3097 switch (info->flow_type) {
3099 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_TCPv4) {
3100 info->data = RXH_IP_SRC | RXH_IP_DST |
3101 RXH_L4_B_0_1 | RXH_L4_B_2_3;
3102 } else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv4) {
3103 info->data = RXH_IP_SRC | RXH_IP_DST;
3107 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_TCPv6) {
3108 info->data = RXH_IP_SRC | RXH_IP_DST |
3109 RXH_L4_B_0_1 | RXH_L4_B_2_3;
3110 } else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv6) {
3111 info->data = RXH_IP_SRC | RXH_IP_DST;
3115 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_UDPv4) {
3116 info->data = RXH_IP_SRC | RXH_IP_DST |
3117 RXH_L4_B_0_1 | RXH_L4_B_2_3;
3118 } else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv4) {
3119 info->data = RXH_IP_SRC | RXH_IP_DST;
3123 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_UDPv6) {
3124 info->data = RXH_IP_SRC | RXH_IP_DST |
3125 RXH_L4_B_0_1 | RXH_L4_B_2_3;
3126 } else if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv6) {
3127 info->data = RXH_IP_SRC | RXH_IP_DST;
3131 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv4)
3132 info->data = RXH_IP_SRC | RXH_IP_DST;
3136 if (vi->rss_hash_types_saved & VIRTIO_NET_RSS_HASH_TYPE_IPv6)
3137 info->data = RXH_IP_SRC | RXH_IP_DST;
3146 static bool virtnet_set_hashflow(struct virtnet_info *vi, struct ethtool_rxnfc *info)
3148 u32 new_hashtypes = vi->rss_hash_types_saved;
3149 bool is_disable = info->data & RXH_DISCARD;
3150 bool is_l4 = info->data == (RXH_IP_SRC | RXH_IP_DST | RXH_L4_B_0_1 | RXH_L4_B_2_3);
3152 /* supports only 'sd', 'sdfn' and 'r' */
3153 if (!((info->data == (RXH_IP_SRC | RXH_IP_DST)) | is_l4 | is_disable))
3156 switch (info->flow_type) {
3158 new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv4 | VIRTIO_NET_RSS_HASH_TYPE_TCPv4);
3160 new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv4
3161 | (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_TCPv4 : 0);
3164 new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv4 | VIRTIO_NET_RSS_HASH_TYPE_UDPv4);
3166 new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv4
3167 | (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_UDPv4 : 0);
3170 new_hashtypes &= ~VIRTIO_NET_RSS_HASH_TYPE_IPv4;
3172 new_hashtypes = VIRTIO_NET_RSS_HASH_TYPE_IPv4;
3175 new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv6 | VIRTIO_NET_RSS_HASH_TYPE_TCPv6);
3177 new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv6
3178 | (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_TCPv6 : 0);
3181 new_hashtypes &= ~(VIRTIO_NET_RSS_HASH_TYPE_IPv6 | VIRTIO_NET_RSS_HASH_TYPE_UDPv6);
3183 new_hashtypes |= VIRTIO_NET_RSS_HASH_TYPE_IPv6
3184 | (is_l4 ? VIRTIO_NET_RSS_HASH_TYPE_UDPv6 : 0);
3187 new_hashtypes &= ~VIRTIO_NET_RSS_HASH_TYPE_IPv6;
3189 new_hashtypes = VIRTIO_NET_RSS_HASH_TYPE_IPv6;
3192 /* unsupported flow */
3196 /* if unsupported hashtype was set */
3197 if (new_hashtypes != (new_hashtypes & vi->rss_hash_types_supported))
3200 if (new_hashtypes != vi->rss_hash_types_saved) {
3201 vi->rss_hash_types_saved = new_hashtypes;
3202 vi->ctrl->rss.hash_types = vi->rss_hash_types_saved;
3203 if (vi->dev->features & NETIF_F_RXHASH)
3204 return virtnet_commit_rss_command(vi);
3210 static void virtnet_get_drvinfo(struct net_device *dev,
3211 struct ethtool_drvinfo *info)
3213 struct virtnet_info *vi = netdev_priv(dev);
3214 struct virtio_device *vdev = vi->vdev;
3216 strscpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
3217 strscpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
3218 strscpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
3222 /* TODO: Eliminate OOO packets during switching */
3223 static int virtnet_set_channels(struct net_device *dev,
3224 struct ethtool_channels *channels)
3226 struct virtnet_info *vi = netdev_priv(dev);
3227 u16 queue_pairs = channels->combined_count;
3230 /* We don't support separate rx/tx channels.
3231 * We don't allow setting 'other' channels.
3233 if (channels->rx_count || channels->tx_count || channels->other_count)
3236 if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
3239 /* For now we don't support modifying channels while XDP is loaded
3240 * also when XDP is loaded all RX queues have XDP programs so we only
3241 * need to check a single RX queue.
3243 if (vi->rq[0].xdp_prog)
3247 err = _virtnet_set_queues(vi, queue_pairs);
3252 virtnet_set_affinity(vi);
3255 netif_set_real_num_tx_queues(dev, queue_pairs);
3256 netif_set_real_num_rx_queues(dev, queue_pairs);
3261 static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *data)
3263 struct virtnet_info *vi = netdev_priv(dev);
3267 switch (stringset) {
3269 for (i = 0; i < vi->curr_queue_pairs; i++) {
3270 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++)
3271 ethtool_sprintf(&p, "rx_queue_%u_%s", i,
3272 virtnet_rq_stats_desc[j].desc);
3275 for (i = 0; i < vi->curr_queue_pairs; i++) {
3276 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++)
3277 ethtool_sprintf(&p, "tx_queue_%u_%s", i,
3278 virtnet_sq_stats_desc[j].desc);
3284 static int virtnet_get_sset_count(struct net_device *dev, int sset)
3286 struct virtnet_info *vi = netdev_priv(dev);
3290 return vi->curr_queue_pairs * (VIRTNET_RQ_STATS_LEN +
3291 VIRTNET_SQ_STATS_LEN);
3297 static void virtnet_get_ethtool_stats(struct net_device *dev,
3298 struct ethtool_stats *stats, u64 *data)
3300 struct virtnet_info *vi = netdev_priv(dev);
3301 unsigned int idx = 0, start, i, j;
3302 const u8 *stats_base;
3303 const u64_stats_t *p;
3306 for (i = 0; i < vi->curr_queue_pairs; i++) {
3307 struct receive_queue *rq = &vi->rq[i];
3309 stats_base = (const u8 *)&rq->stats;
3311 start = u64_stats_fetch_begin(&rq->stats.syncp);
3312 for (j = 0; j < VIRTNET_RQ_STATS_LEN; j++) {
3313 offset = virtnet_rq_stats_desc[j].offset;
3314 p = (const u64_stats_t *)(stats_base + offset);
3315 data[idx + j] = u64_stats_read(p);
3317 } while (u64_stats_fetch_retry(&rq->stats.syncp, start));
3318 idx += VIRTNET_RQ_STATS_LEN;
3321 for (i = 0; i < vi->curr_queue_pairs; i++) {
3322 struct send_queue *sq = &vi->sq[i];
3324 stats_base = (const u8 *)&sq->stats;
3326 start = u64_stats_fetch_begin(&sq->stats.syncp);
3327 for (j = 0; j < VIRTNET_SQ_STATS_LEN; j++) {
3328 offset = virtnet_sq_stats_desc[j].offset;
3329 p = (const u64_stats_t *)(stats_base + offset);
3330 data[idx + j] = u64_stats_read(p);
3332 } while (u64_stats_fetch_retry(&sq->stats.syncp, start));
3333 idx += VIRTNET_SQ_STATS_LEN;
3337 static void virtnet_get_channels(struct net_device *dev,
3338 struct ethtool_channels *channels)
3340 struct virtnet_info *vi = netdev_priv(dev);
3342 channels->combined_count = vi->curr_queue_pairs;
3343 channels->max_combined = vi->max_queue_pairs;
3344 channels->max_other = 0;
3345 channels->rx_count = 0;
3346 channels->tx_count = 0;
3347 channels->other_count = 0;
3350 static int virtnet_set_link_ksettings(struct net_device *dev,
3351 const struct ethtool_link_ksettings *cmd)
3353 struct virtnet_info *vi = netdev_priv(dev);
3355 return ethtool_virtdev_set_link_ksettings(dev, cmd,
3356 &vi->speed, &vi->duplex);
3359 static int virtnet_get_link_ksettings(struct net_device *dev,
3360 struct ethtool_link_ksettings *cmd)
3362 struct virtnet_info *vi = netdev_priv(dev);
3364 cmd->base.speed = vi->speed;
3365 cmd->base.duplex = vi->duplex;
3366 cmd->base.port = PORT_OTHER;
3371 static int virtnet_send_tx_notf_coal_cmds(struct virtnet_info *vi,
3372 struct ethtool_coalesce *ec)
3374 struct scatterlist sgs_tx;
3377 vi->ctrl->coal_tx.tx_usecs = cpu_to_le32(ec->tx_coalesce_usecs);
3378 vi->ctrl->coal_tx.tx_max_packets = cpu_to_le32(ec->tx_max_coalesced_frames);
3379 sg_init_one(&sgs_tx, &vi->ctrl->coal_tx, sizeof(vi->ctrl->coal_tx));
3381 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_NOTF_COAL,
3382 VIRTIO_NET_CTRL_NOTF_COAL_TX_SET,
3386 vi->intr_coal_tx.max_usecs = ec->tx_coalesce_usecs;
3387 vi->intr_coal_tx.max_packets = ec->tx_max_coalesced_frames;
3388 for (i = 0; i < vi->max_queue_pairs; i++) {
3389 vi->sq[i].intr_coal.max_usecs = ec->tx_coalesce_usecs;
3390 vi->sq[i].intr_coal.max_packets = ec->tx_max_coalesced_frames;
3396 static int virtnet_send_rx_notf_coal_cmds(struct virtnet_info *vi,
3397 struct ethtool_coalesce *ec)
3399 bool rx_ctrl_dim_on = !!ec->use_adaptive_rx_coalesce;
3400 struct scatterlist sgs_rx;
3403 if (rx_ctrl_dim_on && !virtio_has_feature(vi->vdev, VIRTIO_NET_F_VQ_NOTF_COAL))
3406 if (rx_ctrl_dim_on && (ec->rx_coalesce_usecs != vi->intr_coal_rx.max_usecs ||
3407 ec->rx_max_coalesced_frames != vi->intr_coal_rx.max_packets))
3410 if (rx_ctrl_dim_on && !vi->rx_dim_enabled) {
3411 vi->rx_dim_enabled = true;
3412 for (i = 0; i < vi->max_queue_pairs; i++)
3413 vi->rq[i].dim_enabled = true;
3417 if (!rx_ctrl_dim_on && vi->rx_dim_enabled) {
3418 vi->rx_dim_enabled = false;
3419 for (i = 0; i < vi->max_queue_pairs; i++)
3420 vi->rq[i].dim_enabled = false;
3423 /* Since the per-queue coalescing params can be set,
3424 * we need apply the global new params even if they
3427 vi->ctrl->coal_rx.rx_usecs = cpu_to_le32(ec->rx_coalesce_usecs);
3428 vi->ctrl->coal_rx.rx_max_packets = cpu_to_le32(ec->rx_max_coalesced_frames);
3429 sg_init_one(&sgs_rx, &vi->ctrl->coal_rx, sizeof(vi->ctrl->coal_rx));
3431 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_NOTF_COAL,
3432 VIRTIO_NET_CTRL_NOTF_COAL_RX_SET,
3436 vi->intr_coal_rx.max_usecs = ec->rx_coalesce_usecs;
3437 vi->intr_coal_rx.max_packets = ec->rx_max_coalesced_frames;
3438 for (i = 0; i < vi->max_queue_pairs; i++) {
3439 vi->rq[i].intr_coal.max_usecs = ec->rx_coalesce_usecs;
3440 vi->rq[i].intr_coal.max_packets = ec->rx_max_coalesced_frames;
3446 static int virtnet_send_notf_coal_cmds(struct virtnet_info *vi,
3447 struct ethtool_coalesce *ec)
3451 err = virtnet_send_tx_notf_coal_cmds(vi, ec);
3455 err = virtnet_send_rx_notf_coal_cmds(vi, ec);
3462 static int virtnet_send_rx_notf_coal_vq_cmds(struct virtnet_info *vi,
3463 struct ethtool_coalesce *ec,
3466 bool rx_ctrl_dim_on = !!ec->use_adaptive_rx_coalesce;
3467 bool cur_rx_dim = vi->rq[queue].dim_enabled;
3468 u32 max_usecs, max_packets;
3471 max_usecs = vi->rq[queue].intr_coal.max_usecs;
3472 max_packets = vi->rq[queue].intr_coal.max_packets;
3474 if (rx_ctrl_dim_on && (ec->rx_coalesce_usecs != max_usecs ||
3475 ec->rx_max_coalesced_frames != max_packets))
3478 if (rx_ctrl_dim_on && !cur_rx_dim) {
3479 vi->rq[queue].dim_enabled = true;
3483 if (!rx_ctrl_dim_on && cur_rx_dim)
3484 vi->rq[queue].dim_enabled = false;
3486 /* If no params are updated, userspace ethtool will
3487 * reject the modification.
3489 err = virtnet_send_rx_ctrl_coal_vq_cmd(vi, queue,
3490 ec->rx_coalesce_usecs,
3491 ec->rx_max_coalesced_frames);
3498 static int virtnet_send_notf_coal_vq_cmds(struct virtnet_info *vi,
3499 struct ethtool_coalesce *ec,
3504 err = virtnet_send_rx_notf_coal_vq_cmds(vi, ec, queue);
3508 err = virtnet_send_tx_ctrl_coal_vq_cmd(vi, queue,
3509 ec->tx_coalesce_usecs,
3510 ec->tx_max_coalesced_frames);
3517 static void virtnet_rx_dim_work(struct work_struct *work)
3519 struct dim *dim = container_of(work, struct dim, work);
3520 struct receive_queue *rq = container_of(dim,
3521 struct receive_queue, dim);
3522 struct virtnet_info *vi = rq->vq->vdev->priv;
3523 struct net_device *dev = vi->dev;
3524 struct dim_cq_moder update_moder;
3527 if (!rtnl_trylock())
3530 /* Each rxq's work is queued by "net_dim()->schedule_work()"
3531 * in response to NAPI traffic changes. Note that dim->profile_ix
3532 * for each rxq is updated prior to the queuing action.
3533 * So we only need to traverse and update profiles for all rxqs
3534 * in the work which is holding rtnl_lock.
3536 for (i = 0; i < vi->curr_queue_pairs; i++) {
3541 if (!rq->dim_enabled)
3544 update_moder = net_dim_get_rx_moderation(dim->mode, dim->profile_ix);
3545 if (update_moder.usec != rq->intr_coal.max_usecs ||
3546 update_moder.pkts != rq->intr_coal.max_packets) {
3547 err = virtnet_send_rx_ctrl_coal_vq_cmd(vi, qnum,
3551 pr_debug("%s: Failed to send dim parameters on rxq%d\n",
3553 dim->state = DIM_START_MEASURE;
3560 static int virtnet_coal_params_supported(struct ethtool_coalesce *ec)
3562 /* usecs coalescing is supported only if VIRTIO_NET_F_NOTF_COAL
3563 * or VIRTIO_NET_F_VQ_NOTF_COAL feature is negotiated.
3565 if (ec->rx_coalesce_usecs || ec->tx_coalesce_usecs)
3568 if (ec->tx_max_coalesced_frames > 1 ||
3569 ec->rx_max_coalesced_frames != 1)
3575 static int virtnet_should_update_vq_weight(int dev_flags, int weight,
3576 int vq_weight, bool *should_update)
3578 if (weight ^ vq_weight) {
3579 if (dev_flags & IFF_UP)
3581 *should_update = true;
3587 static int virtnet_set_coalesce(struct net_device *dev,
3588 struct ethtool_coalesce *ec,
3589 struct kernel_ethtool_coalesce *kernel_coal,
3590 struct netlink_ext_ack *extack)
3592 struct virtnet_info *vi = netdev_priv(dev);
3593 int ret, queue_number, napi_weight;
3594 bool update_napi = false;
3596 /* Can't change NAPI weight if the link is up */
3597 napi_weight = ec->tx_max_coalesced_frames ? NAPI_POLL_WEIGHT : 0;
3598 for (queue_number = 0; queue_number < vi->max_queue_pairs; queue_number++) {
3599 ret = virtnet_should_update_vq_weight(dev->flags, napi_weight,
3600 vi->sq[queue_number].napi.weight,
3606 /* All queues that belong to [queue_number, vi->max_queue_pairs] will be
3607 * updated for the sake of simplicity, which might not be necessary
3613 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_NOTF_COAL))
3614 ret = virtnet_send_notf_coal_cmds(vi, ec);
3616 ret = virtnet_coal_params_supported(ec);
3622 for (; queue_number < vi->max_queue_pairs; queue_number++)
3623 vi->sq[queue_number].napi.weight = napi_weight;
3629 static int virtnet_get_coalesce(struct net_device *dev,
3630 struct ethtool_coalesce *ec,
3631 struct kernel_ethtool_coalesce *kernel_coal,
3632 struct netlink_ext_ack *extack)
3634 struct virtnet_info *vi = netdev_priv(dev);
3636 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_NOTF_COAL)) {
3637 ec->rx_coalesce_usecs = vi->intr_coal_rx.max_usecs;
3638 ec->tx_coalesce_usecs = vi->intr_coal_tx.max_usecs;
3639 ec->tx_max_coalesced_frames = vi->intr_coal_tx.max_packets;
3640 ec->rx_max_coalesced_frames = vi->intr_coal_rx.max_packets;
3641 ec->use_adaptive_rx_coalesce = vi->rx_dim_enabled;
3643 ec->rx_max_coalesced_frames = 1;
3645 if (vi->sq[0].napi.weight)
3646 ec->tx_max_coalesced_frames = 1;
3652 static int virtnet_set_per_queue_coalesce(struct net_device *dev,
3654 struct ethtool_coalesce *ec)
3656 struct virtnet_info *vi = netdev_priv(dev);
3657 int ret, napi_weight;
3658 bool update_napi = false;
3660 if (queue >= vi->max_queue_pairs)
3663 /* Can't change NAPI weight if the link is up */
3664 napi_weight = ec->tx_max_coalesced_frames ? NAPI_POLL_WEIGHT : 0;
3665 ret = virtnet_should_update_vq_weight(dev->flags, napi_weight,
3666 vi->sq[queue].napi.weight,
3671 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_VQ_NOTF_COAL))
3672 ret = virtnet_send_notf_coal_vq_cmds(vi, ec, queue);
3674 ret = virtnet_coal_params_supported(ec);
3680 vi->sq[queue].napi.weight = napi_weight;
3685 static int virtnet_get_per_queue_coalesce(struct net_device *dev,
3687 struct ethtool_coalesce *ec)
3689 struct virtnet_info *vi = netdev_priv(dev);
3691 if (queue >= vi->max_queue_pairs)
3694 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_VQ_NOTF_COAL)) {
3695 ec->rx_coalesce_usecs = vi->rq[queue].intr_coal.max_usecs;
3696 ec->tx_coalesce_usecs = vi->sq[queue].intr_coal.max_usecs;
3697 ec->tx_max_coalesced_frames = vi->sq[queue].intr_coal.max_packets;
3698 ec->rx_max_coalesced_frames = vi->rq[queue].intr_coal.max_packets;
3699 ec->use_adaptive_rx_coalesce = vi->rq[queue].dim_enabled;
3701 ec->rx_max_coalesced_frames = 1;
3703 if (vi->sq[queue].napi.weight)
3704 ec->tx_max_coalesced_frames = 1;
3710 static void virtnet_init_settings(struct net_device *dev)
3712 struct virtnet_info *vi = netdev_priv(dev);
3714 vi->speed = SPEED_UNKNOWN;
3715 vi->duplex = DUPLEX_UNKNOWN;
3718 static void virtnet_update_settings(struct virtnet_info *vi)
3723 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_SPEED_DUPLEX))
3726 virtio_cread_le(vi->vdev, struct virtio_net_config, speed, &speed);
3728 if (ethtool_validate_speed(speed))
3731 virtio_cread_le(vi->vdev, struct virtio_net_config, duplex, &duplex);
3733 if (ethtool_validate_duplex(duplex))
3734 vi->duplex = duplex;
3737 static u32 virtnet_get_rxfh_key_size(struct net_device *dev)
3739 return ((struct virtnet_info *)netdev_priv(dev))->rss_key_size;
3742 static u32 virtnet_get_rxfh_indir_size(struct net_device *dev)
3744 return ((struct virtnet_info *)netdev_priv(dev))->rss_indir_table_size;
3747 static int virtnet_get_rxfh(struct net_device *dev,
3748 struct ethtool_rxfh_param *rxfh)
3750 struct virtnet_info *vi = netdev_priv(dev);
3754 for (i = 0; i < vi->rss_indir_table_size; ++i)
3755 rxfh->indir[i] = vi->ctrl->rss.indirection_table[i];
3759 memcpy(rxfh->key, vi->ctrl->rss.key, vi->rss_key_size);
3761 rxfh->hfunc = ETH_RSS_HASH_TOP;
3766 static int virtnet_set_rxfh(struct net_device *dev,
3767 struct ethtool_rxfh_param *rxfh,
3768 struct netlink_ext_ack *extack)
3770 struct virtnet_info *vi = netdev_priv(dev);
3773 if (rxfh->hfunc != ETH_RSS_HASH_NO_CHANGE &&
3774 rxfh->hfunc != ETH_RSS_HASH_TOP)
3778 for (i = 0; i < vi->rss_indir_table_size; ++i)
3779 vi->ctrl->rss.indirection_table[i] = rxfh->indir[i];
3782 memcpy(vi->ctrl->rss.key, rxfh->key, vi->rss_key_size);
3784 virtnet_commit_rss_command(vi);
3789 static int virtnet_get_rxnfc(struct net_device *dev, struct ethtool_rxnfc *info, u32 *rule_locs)
3791 struct virtnet_info *vi = netdev_priv(dev);
3794 switch (info->cmd) {
3795 case ETHTOOL_GRXRINGS:
3796 info->data = vi->curr_queue_pairs;
3799 virtnet_get_hashflow(vi, info);
3808 static int virtnet_set_rxnfc(struct net_device *dev, struct ethtool_rxnfc *info)
3810 struct virtnet_info *vi = netdev_priv(dev);
3813 switch (info->cmd) {
3815 if (!virtnet_set_hashflow(vi, info))
3826 static const struct ethtool_ops virtnet_ethtool_ops = {
3827 .supported_coalesce_params = ETHTOOL_COALESCE_MAX_FRAMES |
3828 ETHTOOL_COALESCE_USECS | ETHTOOL_COALESCE_USE_ADAPTIVE_RX,
3829 .get_drvinfo = virtnet_get_drvinfo,
3830 .get_link = ethtool_op_get_link,
3831 .get_ringparam = virtnet_get_ringparam,
3832 .set_ringparam = virtnet_set_ringparam,
3833 .get_strings = virtnet_get_strings,
3834 .get_sset_count = virtnet_get_sset_count,
3835 .get_ethtool_stats = virtnet_get_ethtool_stats,
3836 .set_channels = virtnet_set_channels,
3837 .get_channels = virtnet_get_channels,
3838 .get_ts_info = ethtool_op_get_ts_info,
3839 .get_link_ksettings = virtnet_get_link_ksettings,
3840 .set_link_ksettings = virtnet_set_link_ksettings,
3841 .set_coalesce = virtnet_set_coalesce,
3842 .get_coalesce = virtnet_get_coalesce,
3843 .set_per_queue_coalesce = virtnet_set_per_queue_coalesce,
3844 .get_per_queue_coalesce = virtnet_get_per_queue_coalesce,
3845 .get_rxfh_key_size = virtnet_get_rxfh_key_size,
3846 .get_rxfh_indir_size = virtnet_get_rxfh_indir_size,
3847 .get_rxfh = virtnet_get_rxfh,
3848 .set_rxfh = virtnet_set_rxfh,
3849 .get_rxnfc = virtnet_get_rxnfc,
3850 .set_rxnfc = virtnet_set_rxnfc,
3853 static void virtnet_freeze_down(struct virtio_device *vdev)
3855 struct virtnet_info *vi = vdev->priv;
3857 /* Make sure no work handler is accessing the device */
3858 flush_work(&vi->config_work);
3860 netif_tx_lock_bh(vi->dev);
3861 netif_device_detach(vi->dev);
3862 netif_tx_unlock_bh(vi->dev);
3863 if (netif_running(vi->dev))
3864 virtnet_close(vi->dev);
3867 static int init_vqs(struct virtnet_info *vi);
3869 static int virtnet_restore_up(struct virtio_device *vdev)
3871 struct virtnet_info *vi = vdev->priv;
3878 virtio_device_ready(vdev);
3880 enable_delayed_refill(vi);
3882 if (netif_running(vi->dev)) {
3883 err = virtnet_open(vi->dev);
3888 netif_tx_lock_bh(vi->dev);
3889 netif_device_attach(vi->dev);
3890 netif_tx_unlock_bh(vi->dev);
3894 static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads)
3896 struct scatterlist sg;
3897 vi->ctrl->offloads = cpu_to_virtio64(vi->vdev, offloads);
3899 sg_init_one(&sg, &vi->ctrl->offloads, sizeof(vi->ctrl->offloads));
3901 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
3902 VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) {
3903 dev_warn(&vi->dev->dev, "Fail to set guest offload.\n");
3910 static int virtnet_clear_guest_offloads(struct virtnet_info *vi)
3914 if (!vi->guest_offloads)
3917 return virtnet_set_guest_offloads(vi, offloads);
3920 static int virtnet_restore_guest_offloads(struct virtnet_info *vi)
3922 u64 offloads = vi->guest_offloads;
3924 if (!vi->guest_offloads)
3927 return virtnet_set_guest_offloads(vi, offloads);
3930 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
3931 struct netlink_ext_ack *extack)
3933 unsigned int room = SKB_DATA_ALIGN(VIRTIO_XDP_HEADROOM +
3934 sizeof(struct skb_shared_info));
3935 unsigned int max_sz = PAGE_SIZE - room - ETH_HLEN;
3936 struct virtnet_info *vi = netdev_priv(dev);
3937 struct bpf_prog *old_prog;
3938 u16 xdp_qp = 0, curr_qp;
3941 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)
3942 && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
3943 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
3944 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
3945 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) ||
3946 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM) ||
3947 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_USO4) ||
3948 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_USO6))) {
3949 NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing GRO_HW/CSUM, disable GRO_HW/CSUM first");
3953 if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
3954 NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required");
3958 if (prog && !prog->aux->xdp_has_frags && dev->mtu > max_sz) {
3959 NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP without frags");
3960 netdev_warn(dev, "single-buffer XDP requires MTU less than %u\n", max_sz);
3964 curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
3966 xdp_qp = nr_cpu_ids;
3968 /* XDP requires extra queues for XDP_TX */
3969 if (curr_qp + xdp_qp > vi->max_queue_pairs) {
3970 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",
3971 curr_qp + xdp_qp, vi->max_queue_pairs);
3975 old_prog = rtnl_dereference(vi->rq[0].xdp_prog);
3976 if (!prog && !old_prog)
3980 bpf_prog_add(prog, vi->max_queue_pairs - 1);
3982 /* Make sure NAPI is not using any XDP TX queues for RX. */
3983 if (netif_running(dev)) {
3984 for (i = 0; i < vi->max_queue_pairs; i++) {
3985 napi_disable(&vi->rq[i].napi);
3986 virtnet_napi_tx_disable(&vi->sq[i].napi);
3991 for (i = 0; i < vi->max_queue_pairs; i++) {
3992 rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
3994 virtnet_restore_guest_offloads(vi);
3999 err = _virtnet_set_queues(vi, curr_qp + xdp_qp);
4002 netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
4003 vi->xdp_queue_pairs = xdp_qp;
4006 vi->xdp_enabled = true;
4007 for (i = 0; i < vi->max_queue_pairs; i++) {
4008 rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
4009 if (i == 0 && !old_prog)
4010 virtnet_clear_guest_offloads(vi);
4013 xdp_features_set_redirect_target(dev, true);
4015 xdp_features_clear_redirect_target(dev);
4016 vi->xdp_enabled = false;
4019 for (i = 0; i < vi->max_queue_pairs; i++) {
4021 bpf_prog_put(old_prog);
4022 if (netif_running(dev)) {
4023 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
4024 virtnet_napi_tx_enable(vi, vi->sq[i].vq,
4033 virtnet_clear_guest_offloads(vi);
4034 for (i = 0; i < vi->max_queue_pairs; i++)
4035 rcu_assign_pointer(vi->rq[i].xdp_prog, old_prog);
4038 if (netif_running(dev)) {
4039 for (i = 0; i < vi->max_queue_pairs; i++) {
4040 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
4041 virtnet_napi_tx_enable(vi, vi->sq[i].vq,
4046 bpf_prog_sub(prog, vi->max_queue_pairs - 1);
4050 static int virtnet_xdp(struct net_device *dev, struct netdev_bpf *xdp)
4052 switch (xdp->command) {
4053 case XDP_SETUP_PROG:
4054 return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
4060 static int virtnet_get_phys_port_name(struct net_device *dev, char *buf,
4063 struct virtnet_info *vi = netdev_priv(dev);
4066 if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_STANDBY))
4069 ret = snprintf(buf, len, "sby");
4076 static int virtnet_set_features(struct net_device *dev,
4077 netdev_features_t features)
4079 struct virtnet_info *vi = netdev_priv(dev);
4083 if ((dev->features ^ features) & NETIF_F_GRO_HW) {
4084 if (vi->xdp_enabled)
4087 if (features & NETIF_F_GRO_HW)
4088 offloads = vi->guest_offloads_capable;
4090 offloads = vi->guest_offloads_capable &
4091 ~GUEST_OFFLOAD_GRO_HW_MASK;
4093 err = virtnet_set_guest_offloads(vi, offloads);
4096 vi->guest_offloads = offloads;
4099 if ((dev->features ^ features) & NETIF_F_RXHASH) {
4100 if (features & NETIF_F_RXHASH)
4101 vi->ctrl->rss.hash_types = vi->rss_hash_types_saved;
4103 vi->ctrl->rss.hash_types = VIRTIO_NET_HASH_REPORT_NONE;
4105 if (!virtnet_commit_rss_command(vi))
4112 static void virtnet_tx_timeout(struct net_device *dev, unsigned int txqueue)
4114 struct virtnet_info *priv = netdev_priv(dev);
4115 struct send_queue *sq = &priv->sq[txqueue];
4116 struct netdev_queue *txq = netdev_get_tx_queue(dev, txqueue);
4118 u64_stats_update_begin(&sq->stats.syncp);
4119 u64_stats_inc(&sq->stats.tx_timeouts);
4120 u64_stats_update_end(&sq->stats.syncp);
4122 netdev_err(dev, "TX timeout on queue: %u, sq: %s, vq: 0x%x, name: %s, %u usecs ago\n",
4123 txqueue, sq->name, sq->vq->index, sq->vq->name,
4124 jiffies_to_usecs(jiffies - READ_ONCE(txq->trans_start)));
4127 static const struct net_device_ops virtnet_netdev = {
4128 .ndo_open = virtnet_open,
4129 .ndo_stop = virtnet_close,
4130 .ndo_start_xmit = start_xmit,
4131 .ndo_validate_addr = eth_validate_addr,
4132 .ndo_set_mac_address = virtnet_set_mac_address,
4133 .ndo_set_rx_mode = virtnet_set_rx_mode,
4134 .ndo_get_stats64 = virtnet_stats,
4135 .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
4136 .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
4137 .ndo_bpf = virtnet_xdp,
4138 .ndo_xdp_xmit = virtnet_xdp_xmit,
4139 .ndo_features_check = passthru_features_check,
4140 .ndo_get_phys_port_name = virtnet_get_phys_port_name,
4141 .ndo_set_features = virtnet_set_features,
4142 .ndo_tx_timeout = virtnet_tx_timeout,
4145 static void virtnet_config_changed_work(struct work_struct *work)
4147 struct virtnet_info *vi =
4148 container_of(work, struct virtnet_info, config_work);
4151 if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
4152 struct virtio_net_config, status, &v) < 0)
4155 if (v & VIRTIO_NET_S_ANNOUNCE) {
4156 netdev_notify_peers(vi->dev);
4157 virtnet_ack_link_announce(vi);
4160 /* Ignore unknown (future) status bits */
4161 v &= VIRTIO_NET_S_LINK_UP;
4163 if (vi->status == v)
4168 if (vi->status & VIRTIO_NET_S_LINK_UP) {
4169 virtnet_update_settings(vi);
4170 netif_carrier_on(vi->dev);
4171 netif_tx_wake_all_queues(vi->dev);
4173 netif_carrier_off(vi->dev);
4174 netif_tx_stop_all_queues(vi->dev);
4178 static void virtnet_config_changed(struct virtio_device *vdev)
4180 struct virtnet_info *vi = vdev->priv;
4182 schedule_work(&vi->config_work);
4185 static void virtnet_free_queues(struct virtnet_info *vi)
4189 for (i = 0; i < vi->max_queue_pairs; i++) {
4190 __netif_napi_del(&vi->rq[i].napi);
4191 __netif_napi_del(&vi->sq[i].napi);
4194 /* We called __netif_napi_del(),
4195 * we need to respect an RCU grace period before freeing vi->rq
4204 static void _free_receive_bufs(struct virtnet_info *vi)
4206 struct bpf_prog *old_prog;
4209 for (i = 0; i < vi->max_queue_pairs; i++) {
4210 while (vi->rq[i].pages)
4211 __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
4213 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
4214 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
4216 bpf_prog_put(old_prog);
4220 static void free_receive_bufs(struct virtnet_info *vi)
4223 _free_receive_bufs(vi);
4227 static void free_receive_page_frags(struct virtnet_info *vi)
4230 for (i = 0; i < vi->max_queue_pairs; i++)
4231 if (vi->rq[i].alloc_frag.page) {
4232 if (vi->rq[i].do_dma && vi->rq[i].last_dma)
4233 virtnet_rq_unmap(&vi->rq[i], vi->rq[i].last_dma, 0);
4234 put_page(vi->rq[i].alloc_frag.page);
4238 static void virtnet_sq_free_unused_buf(struct virtqueue *vq, void *buf)
4240 if (!is_xdp_frame(buf))
4243 xdp_return_frame(ptr_to_xdp(buf));
4246 static void free_unused_bufs(struct virtnet_info *vi)
4251 for (i = 0; i < vi->max_queue_pairs; i++) {
4252 struct virtqueue *vq = vi->sq[i].vq;
4253 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
4254 virtnet_sq_free_unused_buf(vq, buf);
4258 for (i = 0; i < vi->max_queue_pairs; i++) {
4259 struct virtqueue *vq = vi->rq[i].vq;
4261 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL)
4262 virtnet_rq_unmap_free_buf(vq, buf);
4267 static void virtnet_del_vqs(struct virtnet_info *vi)
4269 struct virtio_device *vdev = vi->vdev;
4271 virtnet_clean_affinity(vi);
4273 vdev->config->del_vqs(vdev);
4275 virtnet_free_queues(vi);
4278 /* How large should a single buffer be so a queue full of these can fit at
4279 * least one full packet?
4280 * Logic below assumes the mergeable buffer header is used.
4282 static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq)
4284 const unsigned int hdr_len = vi->hdr_len;
4285 unsigned int rq_size = virtqueue_get_vring_size(vq);
4286 unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu;
4287 unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len;
4288 unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size);
4290 return max(max(min_buf_len, hdr_len) - hdr_len,
4291 (unsigned int)GOOD_PACKET_LEN);
4294 static int virtnet_find_vqs(struct virtnet_info *vi)
4296 vq_callback_t **callbacks;
4297 struct virtqueue **vqs;
4303 /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
4304 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
4305 * possible control vq.
4307 total_vqs = vi->max_queue_pairs * 2 +
4308 virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
4310 /* Allocate space for find_vqs parameters */
4311 vqs = kcalloc(total_vqs, sizeof(*vqs), GFP_KERNEL);
4314 callbacks = kmalloc_array(total_vqs, sizeof(*callbacks), GFP_KERNEL);
4317 names = kmalloc_array(total_vqs, sizeof(*names), GFP_KERNEL);
4320 if (!vi->big_packets || vi->mergeable_rx_bufs) {
4321 ctx = kcalloc(total_vqs, sizeof(*ctx), GFP_KERNEL);
4328 /* Parameters for control virtqueue, if any */
4330 callbacks[total_vqs - 1] = NULL;
4331 names[total_vqs - 1] = "control";
4334 /* Allocate/initialize parameters for send/receive virtqueues */
4335 for (i = 0; i < vi->max_queue_pairs; i++) {
4336 callbacks[rxq2vq(i)] = skb_recv_done;
4337 callbacks[txq2vq(i)] = skb_xmit_done;
4338 sprintf(vi->rq[i].name, "input.%d", i);
4339 sprintf(vi->sq[i].name, "output.%d", i);
4340 names[rxq2vq(i)] = vi->rq[i].name;
4341 names[txq2vq(i)] = vi->sq[i].name;
4343 ctx[rxq2vq(i)] = true;
4346 ret = virtio_find_vqs_ctx(vi->vdev, total_vqs, vqs, callbacks,
4352 vi->cvq = vqs[total_vqs - 1];
4353 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
4354 vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
4357 for (i = 0; i < vi->max_queue_pairs; i++) {
4358 vi->rq[i].vq = vqs[rxq2vq(i)];
4359 vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq);
4360 vi->sq[i].vq = vqs[txq2vq(i)];
4363 /* run here: ret == 0. */
4378 static int virtnet_alloc_queues(struct virtnet_info *vi)
4383 vi->ctrl = kzalloc(sizeof(*vi->ctrl), GFP_KERNEL);
4389 vi->sq = kcalloc(vi->max_queue_pairs, sizeof(*vi->sq), GFP_KERNEL);
4392 vi->rq = kcalloc(vi->max_queue_pairs, sizeof(*vi->rq), GFP_KERNEL);
4396 INIT_DELAYED_WORK(&vi->refill, refill_work);
4397 for (i = 0; i < vi->max_queue_pairs; i++) {
4398 vi->rq[i].pages = NULL;
4399 netif_napi_add_weight(vi->dev, &vi->rq[i].napi, virtnet_poll,
4401 netif_napi_add_tx_weight(vi->dev, &vi->sq[i].napi,
4403 napi_tx ? napi_weight : 0);
4405 INIT_WORK(&vi->rq[i].dim.work, virtnet_rx_dim_work);
4406 vi->rq[i].dim.mode = DIM_CQ_PERIOD_MODE_START_FROM_EQE;
4408 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
4409 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
4410 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
4412 u64_stats_init(&vi->rq[i].stats.syncp);
4413 u64_stats_init(&vi->sq[i].stats.syncp);
4426 static int init_vqs(struct virtnet_info *vi)
4430 /* Allocate send & receive queues */
4431 ret = virtnet_alloc_queues(vi);
4435 ret = virtnet_find_vqs(vi);
4439 virtnet_rq_set_premapped(vi);
4442 virtnet_set_affinity(vi);
4448 virtnet_free_queues(vi);
4454 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
4457 struct virtnet_info *vi = netdev_priv(queue->dev);
4458 unsigned int queue_index = get_netdev_rx_queue_index(queue);
4459 unsigned int headroom = virtnet_get_headroom(vi);
4460 unsigned int tailroom = headroom ? sizeof(struct skb_shared_info) : 0;
4461 struct ewma_pkt_len *avg;
4463 BUG_ON(queue_index >= vi->max_queue_pairs);
4464 avg = &vi->rq[queue_index].mrg_avg_pkt_len;
4465 return sprintf(buf, "%u\n",
4466 get_mergeable_buf_len(&vi->rq[queue_index], avg,
4467 SKB_DATA_ALIGN(headroom + tailroom)));
4470 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
4471 __ATTR_RO(mergeable_rx_buffer_size);
4473 static struct attribute *virtio_net_mrg_rx_attrs[] = {
4474 &mergeable_rx_buffer_size_attribute.attr,
4478 static const struct attribute_group virtio_net_mrg_rx_group = {
4479 .name = "virtio_net",
4480 .attrs = virtio_net_mrg_rx_attrs
4484 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
4486 const char *fname, const char *dname)
4488 if (!virtio_has_feature(vdev, fbit))
4491 dev_err(&vdev->dev, "device advertises feature %s but not %s",
4497 #define VIRTNET_FAIL_ON(vdev, fbit, dbit) \
4498 virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
4500 static bool virtnet_validate_features(struct virtio_device *vdev)
4502 if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
4503 (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
4504 "VIRTIO_NET_F_CTRL_VQ") ||
4505 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
4506 "VIRTIO_NET_F_CTRL_VQ") ||
4507 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
4508 "VIRTIO_NET_F_CTRL_VQ") ||
4509 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
4510 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
4511 "VIRTIO_NET_F_CTRL_VQ") ||
4512 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_RSS,
4513 "VIRTIO_NET_F_CTRL_VQ") ||
4514 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_HASH_REPORT,
4515 "VIRTIO_NET_F_CTRL_VQ") ||
4516 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_NOTF_COAL,
4517 "VIRTIO_NET_F_CTRL_VQ") ||
4518 VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_VQ_NOTF_COAL,
4519 "VIRTIO_NET_F_CTRL_VQ"))) {
4526 #define MIN_MTU ETH_MIN_MTU
4527 #define MAX_MTU ETH_MAX_MTU
4529 static int virtnet_validate(struct virtio_device *vdev)
4531 if (!vdev->config->get) {
4532 dev_err(&vdev->dev, "%s failure: config access disabled\n",
4537 if (!virtnet_validate_features(vdev))
4540 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
4541 int mtu = virtio_cread16(vdev,
4542 offsetof(struct virtio_net_config,
4545 __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
4548 if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY) &&
4549 !virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
4550 dev_warn(&vdev->dev, "device advertises feature VIRTIO_NET_F_STANDBY but not VIRTIO_NET_F_MAC, disabling standby");
4551 __virtio_clear_bit(vdev, VIRTIO_NET_F_STANDBY);
4557 static bool virtnet_check_guest_gso(const struct virtnet_info *vi)
4559 return virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
4560 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
4561 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
4562 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO) ||
4563 (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_USO4) &&
4564 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_USO6));
4567 static void virtnet_set_big_packets(struct virtnet_info *vi, const int mtu)
4569 bool guest_gso = virtnet_check_guest_gso(vi);
4571 /* If device can receive ANY guest GSO packets, regardless of mtu,
4572 * allocate packets of maximum size, otherwise limit it to only
4573 * mtu size worth only.
4575 if (mtu > ETH_DATA_LEN || guest_gso) {
4576 vi->big_packets = true;
4577 vi->big_packets_num_skbfrags = guest_gso ? MAX_SKB_FRAGS : DIV_ROUND_UP(mtu, PAGE_SIZE);
4581 static int virtnet_probe(struct virtio_device *vdev)
4583 int i, err = -ENOMEM;
4584 struct net_device *dev;
4585 struct virtnet_info *vi;
4586 u16 max_queue_pairs;
4589 /* Find if host supports multiqueue/rss virtio_net device */
4590 max_queue_pairs = 1;
4591 if (virtio_has_feature(vdev, VIRTIO_NET_F_MQ) || virtio_has_feature(vdev, VIRTIO_NET_F_RSS))
4593 virtio_cread16(vdev, offsetof(struct virtio_net_config, max_virtqueue_pairs));
4595 /* We need at least 2 queue's */
4596 if (max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
4597 max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
4598 !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
4599 max_queue_pairs = 1;
4601 /* Allocate ourselves a network device with room for our info */
4602 dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
4606 /* Set up network device as normal. */
4607 dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE |
4608 IFF_TX_SKB_NO_LINEAR;
4609 dev->netdev_ops = &virtnet_netdev;
4610 dev->features = NETIF_F_HIGHDMA;
4612 dev->ethtool_ops = &virtnet_ethtool_ops;
4613 SET_NETDEV_DEV(dev, &vdev->dev);
4615 /* Do we support "hardware" checksums? */
4616 if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
4617 /* This opens up the world of extra features. */
4618 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
4620 dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
4622 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
4623 dev->hw_features |= NETIF_F_TSO
4624 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
4626 /* Individual feature bits: what can host handle? */
4627 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
4628 dev->hw_features |= NETIF_F_TSO;
4629 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
4630 dev->hw_features |= NETIF_F_TSO6;
4631 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
4632 dev->hw_features |= NETIF_F_TSO_ECN;
4633 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_USO))
4634 dev->hw_features |= NETIF_F_GSO_UDP_L4;
4636 dev->features |= NETIF_F_GSO_ROBUST;
4639 dev->features |= dev->hw_features & NETIF_F_ALL_TSO;
4640 /* (!csum && gso) case will be fixed by register_netdev() */
4642 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
4643 dev->features |= NETIF_F_RXCSUM;
4644 if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
4645 virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6))
4646 dev->features |= NETIF_F_GRO_HW;
4647 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS))
4648 dev->hw_features |= NETIF_F_GRO_HW;
4650 dev->vlan_features = dev->features;
4651 dev->xdp_features = NETDEV_XDP_ACT_BASIC | NETDEV_XDP_ACT_REDIRECT;
4653 /* MTU range: 68 - 65535 */
4654 dev->min_mtu = MIN_MTU;
4655 dev->max_mtu = MAX_MTU;
4657 /* Configuration may specify what MAC to use. Otherwise random. */
4658 if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC)) {
4661 virtio_cread_bytes(vdev,
4662 offsetof(struct virtio_net_config, mac),
4664 eth_hw_addr_set(dev, addr);
4666 eth_hw_addr_random(dev);
4667 dev_info(&vdev->dev, "Assigned random MAC address %pM\n",
4671 /* Set up our device-specific information */
4672 vi = netdev_priv(dev);
4677 INIT_WORK(&vi->config_work, virtnet_config_changed_work);
4678 spin_lock_init(&vi->refill_lock);
4680 if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF)) {
4681 vi->mergeable_rx_bufs = true;
4682 dev->xdp_features |= NETDEV_XDP_ACT_RX_SG;
4685 if (virtio_has_feature(vdev, VIRTIO_NET_F_HASH_REPORT))
4686 vi->has_rss_hash_report = true;
4688 if (virtio_has_feature(vdev, VIRTIO_NET_F_RSS))
4691 if (vi->has_rss || vi->has_rss_hash_report) {
4692 vi->rss_indir_table_size =
4693 virtio_cread16(vdev, offsetof(struct virtio_net_config,
4694 rss_max_indirection_table_length));
4696 virtio_cread8(vdev, offsetof(struct virtio_net_config, rss_max_key_size));
4698 vi->rss_hash_types_supported =
4699 virtio_cread32(vdev, offsetof(struct virtio_net_config, supported_hash_types));
4700 vi->rss_hash_types_supported &=
4701 ~(VIRTIO_NET_RSS_HASH_TYPE_IP_EX |
4702 VIRTIO_NET_RSS_HASH_TYPE_TCP_EX |
4703 VIRTIO_NET_RSS_HASH_TYPE_UDP_EX);
4705 dev->hw_features |= NETIF_F_RXHASH;
4708 if (vi->has_rss_hash_report)
4709 vi->hdr_len = sizeof(struct virtio_net_hdr_v1_hash);
4710 else if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
4711 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
4712 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
4714 vi->hdr_len = sizeof(struct virtio_net_hdr);
4716 if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
4717 virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
4718 vi->any_header_sg = true;
4720 if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
4723 if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
4724 mtu = virtio_cread16(vdev,
4725 offsetof(struct virtio_net_config,
4727 if (mtu < dev->min_mtu) {
4728 /* Should never trigger: MTU was previously validated
4729 * in virtnet_validate.
4732 "device MTU appears to have changed it is now %d < %d",
4742 virtnet_set_big_packets(vi, mtu);
4744 if (vi->any_header_sg)
4745 dev->needed_headroom = vi->hdr_len;
4747 /* Enable multiqueue by default */
4748 if (num_online_cpus() >= max_queue_pairs)
4749 vi->curr_queue_pairs = max_queue_pairs;
4751 vi->curr_queue_pairs = num_online_cpus();
4752 vi->max_queue_pairs = max_queue_pairs;
4754 /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
4759 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_NOTF_COAL)) {
4760 vi->intr_coal_rx.max_usecs = 0;
4761 vi->intr_coal_tx.max_usecs = 0;
4762 vi->intr_coal_rx.max_packets = 0;
4764 /* Keep the default values of the coalescing parameters
4765 * aligned with the default napi_tx state.
4767 if (vi->sq[0].napi.weight)
4768 vi->intr_coal_tx.max_packets = 1;
4770 vi->intr_coal_tx.max_packets = 0;
4773 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_VQ_NOTF_COAL)) {
4774 /* The reason is the same as VIRTIO_NET_F_NOTF_COAL. */
4775 for (i = 0; i < vi->max_queue_pairs; i++)
4776 if (vi->sq[i].napi.weight)
4777 vi->sq[i].intr_coal.max_packets = 1;
4781 if (vi->mergeable_rx_bufs)
4782 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
4784 netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
4785 netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
4787 virtnet_init_settings(dev);
4789 if (virtio_has_feature(vdev, VIRTIO_NET_F_STANDBY)) {
4790 vi->failover = net_failover_create(vi->dev);
4791 if (IS_ERR(vi->failover)) {
4792 err = PTR_ERR(vi->failover);
4797 if (vi->has_rss || vi->has_rss_hash_report)
4798 virtnet_init_default_rss(vi);
4800 /* serialize netdev register + virtio_device_ready() with ndo_open() */
4803 err = register_netdevice(dev);
4805 pr_debug("virtio_net: registering device failed\n");
4810 virtio_device_ready(vdev);
4812 _virtnet_set_queues(vi, vi->curr_queue_pairs);
4814 /* a random MAC address has been assigned, notify the device.
4815 * We don't fail probe if VIRTIO_NET_F_CTRL_MAC_ADDR is not there
4816 * because many devices work fine without getting MAC explicitly
4818 if (!virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
4819 virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
4820 struct scatterlist sg;
4822 sg_init_one(&sg, dev->dev_addr, dev->addr_len);
4823 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
4824 VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
4825 pr_debug("virtio_net: setting MAC address failed\n");
4828 goto free_unregister_netdev;
4834 err = virtnet_cpu_notif_add(vi);
4836 pr_debug("virtio_net: registering cpu notifier failed\n");
4837 goto free_unregister_netdev;
4840 /* Assume link up if device can't report link status,
4841 otherwise get link status from config. */
4842 netif_carrier_off(dev);
4843 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
4844 schedule_work(&vi->config_work);
4846 vi->status = VIRTIO_NET_S_LINK_UP;
4847 virtnet_update_settings(vi);
4848 netif_carrier_on(dev);
4851 for (i = 0; i < ARRAY_SIZE(guest_offloads); i++)
4852 if (virtio_has_feature(vi->vdev, guest_offloads[i]))
4853 set_bit(guest_offloads[i], &vi->guest_offloads);
4854 vi->guest_offloads_capable = vi->guest_offloads;
4856 pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
4857 dev->name, max_queue_pairs);
4861 free_unregister_netdev:
4862 unregister_netdev(dev);
4864 net_failover_destroy(vi->failover);
4866 virtio_reset_device(vdev);
4867 cancel_delayed_work_sync(&vi->refill);
4868 free_receive_page_frags(vi);
4869 virtnet_del_vqs(vi);
4875 static void remove_vq_common(struct virtnet_info *vi)
4877 virtio_reset_device(vi->vdev);
4879 /* Free unused buffers in both send and recv, if any. */
4880 free_unused_bufs(vi);
4882 free_receive_bufs(vi);
4884 free_receive_page_frags(vi);
4886 virtnet_del_vqs(vi);
4889 static void virtnet_remove(struct virtio_device *vdev)
4891 struct virtnet_info *vi = vdev->priv;
4893 virtnet_cpu_notif_remove(vi);
4895 /* Make sure no work handler is accessing the device. */
4896 flush_work(&vi->config_work);
4898 unregister_netdev(vi->dev);
4900 net_failover_destroy(vi->failover);
4902 remove_vq_common(vi);
4904 free_netdev(vi->dev);
4907 static __maybe_unused int virtnet_freeze(struct virtio_device *vdev)
4909 struct virtnet_info *vi = vdev->priv;
4911 virtnet_cpu_notif_remove(vi);
4912 virtnet_freeze_down(vdev);
4913 remove_vq_common(vi);
4918 static __maybe_unused int virtnet_restore(struct virtio_device *vdev)
4920 struct virtnet_info *vi = vdev->priv;
4923 err = virtnet_restore_up(vdev);
4926 virtnet_set_queues(vi, vi->curr_queue_pairs);
4928 err = virtnet_cpu_notif_add(vi);
4930 virtnet_freeze_down(vdev);
4931 remove_vq_common(vi);
4938 static struct virtio_device_id id_table[] = {
4939 { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
4943 #define VIRTNET_FEATURES \
4944 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
4946 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
4947 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
4948 VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
4949 VIRTIO_NET_F_HOST_USO, VIRTIO_NET_F_GUEST_USO4, VIRTIO_NET_F_GUEST_USO6, \
4950 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
4951 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
4952 VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
4953 VIRTIO_NET_F_CTRL_MAC_ADDR, \
4954 VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \
4955 VIRTIO_NET_F_SPEED_DUPLEX, VIRTIO_NET_F_STANDBY, \
4956 VIRTIO_NET_F_RSS, VIRTIO_NET_F_HASH_REPORT, VIRTIO_NET_F_NOTF_COAL, \
4957 VIRTIO_NET_F_VQ_NOTF_COAL, \
4958 VIRTIO_NET_F_GUEST_HDRLEN
4960 static unsigned int features[] = {
4964 static unsigned int features_legacy[] = {
4967 VIRTIO_F_ANY_LAYOUT,
4970 static struct virtio_driver virtio_net_driver = {
4971 .feature_table = features,
4972 .feature_table_size = ARRAY_SIZE(features),
4973 .feature_table_legacy = features_legacy,
4974 .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
4975 .driver.name = KBUILD_MODNAME,
4976 .driver.owner = THIS_MODULE,
4977 .id_table = id_table,
4978 .validate = virtnet_validate,
4979 .probe = virtnet_probe,
4980 .remove = virtnet_remove,
4981 .config_changed = virtnet_config_changed,
4982 #ifdef CONFIG_PM_SLEEP
4983 .freeze = virtnet_freeze,
4984 .restore = virtnet_restore,
4988 static __init int virtio_net_driver_init(void)
4992 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
4994 virtnet_cpu_down_prep);
4997 virtionet_online = ret;
4998 ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
4999 NULL, virtnet_cpu_dead);
5002 ret = register_virtio_driver(&virtio_net_driver);
5007 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
5009 cpuhp_remove_multi_state(virtionet_online);
5013 module_init(virtio_net_driver_init);
5015 static __exit void virtio_net_driver_exit(void)
5017 unregister_virtio_driver(&virtio_net_driver);
5018 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
5019 cpuhp_remove_multi_state(virtionet_online);
5021 module_exit(virtio_net_driver_exit);
5023 MODULE_DEVICE_TABLE(virtio, id_table);
5024 MODULE_DESCRIPTION("Virtio network driver");
5025 MODULE_LICENSE("GPL");