1 /* A network driver using virtio.
3 * Copyright 2007 Rusty Russell <rusty@rustcorp.com.au> IBM Corporation
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation; either version 2 of the License, or
8 * (at your option) any later version.
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, see <http://www.gnu.org/licenses/>.
19 #include <linux/netdevice.h>
20 #include <linux/etherdevice.h>
21 #include <linux/ethtool.h>
22 #include <linux/module.h>
23 #include <linux/virtio.h>
24 #include <linux/virtio_net.h>
25 #include <linux/bpf.h>
26 #include <linux/bpf_trace.h>
27 #include <linux/scatterlist.h>
28 #include <linux/if_vlan.h>
29 #include <linux/slab.h>
30 #include <linux/cpu.h>
31 #include <linux/average.h>
32 #include <linux/filter.h>
33 #include <net/route.h>
35 static int napi_weight
= NAPI_POLL_WEIGHT
;
36 module_param(napi_weight
, int, 0444);
38 static bool csum
= true, gso
= true, napi_tx
;
39 module_param(csum
, bool, 0444);
40 module_param(gso
, bool, 0444);
41 module_param(napi_tx
, bool, 0644);
43 /* FIXME: MTU in config. */
44 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
45 #define GOOD_COPY_LEN 128
47 #define VIRTNET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
49 /* Amount of XDP headroom to prepend to packets for use by xdp_adjust_head */
50 #define VIRTIO_XDP_HEADROOM 256
52 /* RX packet size EWMA. The average packet size is used to determine the packet
53 * buffer size when refilling RX rings. As the entire RX ring may be refilled
54 * at once, the weight is chosen so that the EWMA will be insensitive to short-
55 * term, transient changes in packet size.
57 DECLARE_EWMA(pkt_len
, 0, 64)
59 #define VIRTNET_DRIVER_VERSION "1.0.0"
61 static const unsigned long guest_offloads
[] = {
62 VIRTIO_NET_F_GUEST_TSO4
,
63 VIRTIO_NET_F_GUEST_TSO6
,
64 VIRTIO_NET_F_GUEST_ECN
,
65 VIRTIO_NET_F_GUEST_UFO
68 struct virtnet_stats
{
69 struct u64_stats_sync tx_syncp
;
70 struct u64_stats_sync rx_syncp
;
78 /* Internal representation of a send virtqueue */
80 /* Virtqueue associated with this send _queue */
83 /* TX: fragments + linear part + virtio header */
84 struct scatterlist sg
[MAX_SKB_FRAGS
+ 2];
86 /* Name of the send queue: output.$index */
89 struct napi_struct napi
;
92 /* Internal representation of a receive virtqueue */
93 struct receive_queue
{
94 /* Virtqueue associated with this receive_queue */
97 struct napi_struct napi
;
99 struct bpf_prog __rcu
*xdp_prog
;
101 /* Chain pages by the private ptr. */
104 /* Average packet length for mergeable receive buffers. */
105 struct ewma_pkt_len mrg_avg_pkt_len
;
107 /* Page frag for packet buffer allocation. */
108 struct page_frag alloc_frag
;
110 /* RX: fragments + linear part + virtio header */
111 struct scatterlist sg
[MAX_SKB_FRAGS
+ 2];
113 /* Min single buffer size for mergeable buffers case. */
114 unsigned int min_buf_len
;
116 /* Name of this receive queue: input.$index */
120 struct virtnet_info
{
121 struct virtio_device
*vdev
;
122 struct virtqueue
*cvq
;
123 struct net_device
*dev
;
124 struct send_queue
*sq
;
125 struct receive_queue
*rq
;
128 /* Max # of queue pairs supported by the device */
131 /* # of queue pairs currently used by the driver */
132 u16 curr_queue_pairs
;
134 /* # of XDP queue pairs currently used by the driver */
137 /* I like... big packets and I cannot lie! */
140 /* Host will merge rx buffers for big packets (shake it! shake it!) */
141 bool mergeable_rx_bufs
;
143 /* Has control virtqueue */
146 /* Host can handle any s/g split between our header and packet data */
149 /* Packet virtio header size */
152 /* Active statistics */
153 struct virtnet_stats __percpu
*stats
;
155 /* Work struct for refilling if we run low on memory. */
156 struct delayed_work refill
;
158 /* Work struct for config space updates */
159 struct work_struct config_work
;
161 /* Does the affinity hint is set for virtqueues? */
162 bool affinity_hint_set
;
164 /* CPU hotplug instances for online & dead */
165 struct hlist_node node
;
166 struct hlist_node node_dead
;
168 /* Control VQ buffers: protected by the rtnl lock */
169 struct virtio_net_ctrl_hdr ctrl_hdr
;
170 virtio_net_ctrl_ack ctrl_status
;
171 struct virtio_net_ctrl_mq ctrl_mq
;
177 /* Ethtool settings */
181 unsigned long guest_offloads
;
184 struct padded_vnet_hdr
{
185 struct virtio_net_hdr_mrg_rxbuf hdr
;
187 * hdr is in a separate sg buffer, and data sg buffer shares same page
188 * with this header sg. This padding makes next sg 16 byte aligned
194 /* Converting between virtqueue no. and kernel tx/rx queue no.
195 * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
197 static int vq2txq(struct virtqueue
*vq
)
199 return (vq
->index
- 1) / 2;
202 static int txq2vq(int txq
)
207 static int vq2rxq(struct virtqueue
*vq
)
209 return vq
->index
/ 2;
212 static int rxq2vq(int rxq
)
217 static inline struct virtio_net_hdr_mrg_rxbuf
*skb_vnet_hdr(struct sk_buff
*skb
)
219 return (struct virtio_net_hdr_mrg_rxbuf
*)skb
->cb
;
223 * private is used to chain pages for big packets, put the whole
224 * most recent used list in the beginning for reuse
226 static void give_pages(struct receive_queue
*rq
, struct page
*page
)
230 /* Find end of list, sew whole thing into vi->rq.pages. */
231 for (end
= page
; end
->private; end
= (struct page
*)end
->private);
232 end
->private = (unsigned long)rq
->pages
;
236 static struct page
*get_a_page(struct receive_queue
*rq
, gfp_t gfp_mask
)
238 struct page
*p
= rq
->pages
;
241 rq
->pages
= (struct page
*)p
->private;
242 /* clear private here, it is used to chain pages */
245 p
= alloc_page(gfp_mask
);
249 static void virtqueue_napi_schedule(struct napi_struct
*napi
,
250 struct virtqueue
*vq
)
252 if (napi_schedule_prep(napi
)) {
253 virtqueue_disable_cb(vq
);
254 __napi_schedule(napi
);
258 static void virtqueue_napi_complete(struct napi_struct
*napi
,
259 struct virtqueue
*vq
, int processed
)
263 opaque
= virtqueue_enable_cb_prepare(vq
);
264 if (napi_complete_done(napi
, processed
) &&
265 unlikely(virtqueue_poll(vq
, opaque
)))
266 virtqueue_napi_schedule(napi
, vq
);
269 static void skb_xmit_done(struct virtqueue
*vq
)
271 struct virtnet_info
*vi
= vq
->vdev
->priv
;
272 struct napi_struct
*napi
= &vi
->sq
[vq2txq(vq
)].napi
;
274 /* Suppress further interrupts. */
275 virtqueue_disable_cb(vq
);
278 virtqueue_napi_schedule(napi
, vq
);
280 /* We were probably waiting for more output buffers. */
281 netif_wake_subqueue(vi
->dev
, vq2txq(vq
));
284 #define MRG_CTX_HEADER_SHIFT 22
285 static void *mergeable_len_to_ctx(unsigned int truesize
,
286 unsigned int headroom
)
288 return (void *)(unsigned long)((headroom
<< MRG_CTX_HEADER_SHIFT
) | truesize
);
291 static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx
)
293 return (unsigned long)mrg_ctx
>> MRG_CTX_HEADER_SHIFT
;
296 static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx
)
298 return (unsigned long)mrg_ctx
& ((1 << MRG_CTX_HEADER_SHIFT
) - 1);
301 /* Called from bottom half context */
302 static struct sk_buff
*page_to_skb(struct virtnet_info
*vi
,
303 struct receive_queue
*rq
,
304 struct page
*page
, unsigned int offset
,
305 unsigned int len
, unsigned int truesize
)
308 struct virtio_net_hdr_mrg_rxbuf
*hdr
;
309 unsigned int copy
, hdr_len
, hdr_padded_len
;
312 p
= page_address(page
) + offset
;
314 /* copy small packet so we can reuse these pages for small data */
315 skb
= napi_alloc_skb(&rq
->napi
, GOOD_COPY_LEN
);
319 hdr
= skb_vnet_hdr(skb
);
321 hdr_len
= vi
->hdr_len
;
322 if (vi
->mergeable_rx_bufs
)
323 hdr_padded_len
= sizeof(*hdr
);
325 hdr_padded_len
= sizeof(struct padded_vnet_hdr
);
327 memcpy(hdr
, p
, hdr_len
);
330 offset
+= hdr_padded_len
;
334 if (copy
> skb_tailroom(skb
))
335 copy
= skb_tailroom(skb
);
336 skb_put_data(skb
, p
, copy
);
341 if (vi
->mergeable_rx_bufs
) {
343 skb_add_rx_frag(skb
, 0, page
, offset
, len
, truesize
);
350 * Verify that we can indeed put this data into a skb.
351 * This is here to handle cases when the device erroneously
352 * tries to receive more than is possible. This is usually
353 * the case of a broken device.
355 if (unlikely(len
> MAX_SKB_FRAGS
* PAGE_SIZE
)) {
356 net_dbg_ratelimited("%s: too much data\n", skb
->dev
->name
);
360 BUG_ON(offset
>= PAGE_SIZE
);
362 unsigned int frag_size
= min((unsigned)PAGE_SIZE
- offset
, len
);
363 skb_add_rx_frag(skb
, skb_shinfo(skb
)->nr_frags
, page
, offset
,
364 frag_size
, truesize
);
366 page
= (struct page
*)page
->private;
371 give_pages(rq
, page
);
376 static void virtnet_xdp_flush(struct net_device
*dev
)
378 struct virtnet_info
*vi
= netdev_priv(dev
);
379 struct send_queue
*sq
;
382 qp
= vi
->curr_queue_pairs
- vi
->xdp_queue_pairs
+ smp_processor_id();
385 virtqueue_kick(sq
->vq
);
388 static bool __virtnet_xdp_xmit(struct virtnet_info
*vi
,
389 struct xdp_buff
*xdp
)
391 struct virtio_net_hdr_mrg_rxbuf
*hdr
;
393 struct send_queue
*sq
;
398 qp
= vi
->curr_queue_pairs
- vi
->xdp_queue_pairs
+ smp_processor_id();
401 /* Free up any pending old buffers before queueing new ones. */
402 while ((xdp_sent
= virtqueue_get_buf(sq
->vq
, &len
)) != NULL
) {
403 struct page
*sent_page
= virt_to_head_page(xdp_sent
);
408 xdp
->data
-= vi
->hdr_len
;
409 /* Zero header and leave csum up to XDP layers */
411 memset(hdr
, 0, vi
->hdr_len
);
413 sg_init_one(sq
->sg
, xdp
->data
, xdp
->data_end
- xdp
->data
);
415 err
= virtqueue_add_outbuf(sq
->vq
, sq
->sg
, 1, xdp
->data
, GFP_ATOMIC
);
417 struct page
*page
= virt_to_head_page(xdp
->data
);
426 static int virtnet_xdp_xmit(struct net_device
*dev
, struct xdp_buff
*xdp
)
428 struct virtnet_info
*vi
= netdev_priv(dev
);
429 bool sent
= __virtnet_xdp_xmit(vi
, xdp
);
436 static unsigned int virtnet_get_headroom(struct virtnet_info
*vi
)
438 return vi
->xdp_queue_pairs
? VIRTIO_XDP_HEADROOM
: 0;
441 /* We copy the packet for XDP in the following cases:
443 * 1) Packet is scattered across multiple rx buffers.
444 * 2) Headroom space is insufficient.
446 * This is inefficient but it's a temporary condition that
447 * we hit right after XDP is enabled and until queue is refilled
448 * with large buffers with sufficient headroom - so it should affect
449 * at most queue size packets.
450 * Afterwards, the conditions to enable
451 * XDP should preclude the underlying device from sending packets
452 * across multiple buffers (num_buf > 1), and we make sure buffers
453 * have enough headroom.
455 static struct page
*xdp_linearize_page(struct receive_queue
*rq
,
462 struct page
*page
= alloc_page(GFP_ATOMIC
);
467 memcpy(page_address(page
) + page_off
, page_address(p
) + offset
, *len
);
475 buf
= virtqueue_get_buf(rq
->vq
, &buflen
);
479 p
= virt_to_head_page(buf
);
480 off
= buf
- page_address(p
);
482 /* guard against a misconfigured or uncooperative backend that
483 * is sending packet larger than the MTU.
485 if ((page_off
+ buflen
) > PAGE_SIZE
) {
490 memcpy(page_address(page
) + page_off
,
491 page_address(p
) + off
, buflen
);
496 /* Headroom does not contribute to packet length */
497 *len
= page_off
- VIRTIO_XDP_HEADROOM
;
500 __free_pages(page
, 0);
504 static struct sk_buff
*receive_small(struct net_device
*dev
,
505 struct virtnet_info
*vi
,
506 struct receive_queue
*rq
,
507 void *buf
, void *ctx
,
512 struct bpf_prog
*xdp_prog
;
513 unsigned int xdp_headroom
= (unsigned long)ctx
;
514 unsigned int header_offset
= VIRTNET_RX_PAD
+ xdp_headroom
;
515 unsigned int headroom
= vi
->hdr_len
+ header_offset
;
516 unsigned int buflen
= SKB_DATA_ALIGN(GOOD_PACKET_LEN
+ headroom
) +
517 SKB_DATA_ALIGN(sizeof(struct skb_shared_info
));
518 struct page
*page
= virt_to_head_page(buf
);
519 unsigned int delta
= 0, err
;
520 struct page
*xdp_page
;
524 xdp_prog
= rcu_dereference(rq
->xdp_prog
);
526 struct virtio_net_hdr_mrg_rxbuf
*hdr
= buf
+ header_offset
;
531 if (unlikely(hdr
->hdr
.gso_type
|| hdr
->hdr
.flags
))
534 if (unlikely(xdp_headroom
< virtnet_get_headroom(vi
))) {
535 int offset
= buf
- page_address(page
) + header_offset
;
536 unsigned int tlen
= len
+ vi
->hdr_len
;
539 xdp_headroom
= virtnet_get_headroom(vi
);
540 header_offset
= VIRTNET_RX_PAD
+ xdp_headroom
;
541 headroom
= vi
->hdr_len
+ header_offset
;
542 buflen
= SKB_DATA_ALIGN(GOOD_PACKET_LEN
+ headroom
) +
543 SKB_DATA_ALIGN(sizeof(struct skb_shared_info
));
544 xdp_page
= xdp_linearize_page(rq
, &num_buf
, page
,
545 offset
, header_offset
,
550 buf
= page_address(xdp_page
);
555 xdp
.data_hard_start
= buf
+ VIRTNET_RX_PAD
+ vi
->hdr_len
;
556 xdp
.data
= xdp
.data_hard_start
+ xdp_headroom
;
557 xdp_set_data_meta_invalid(&xdp
);
558 xdp
.data_end
= xdp
.data
+ len
;
559 orig_data
= xdp
.data
;
560 act
= bpf_prog_run_xdp(xdp_prog
, &xdp
);
564 /* Recalculate length in case bpf program changed it */
565 delta
= orig_data
- xdp
.data
;
568 if (unlikely(!__virtnet_xdp_xmit(vi
, &xdp
)))
569 trace_xdp_exception(vi
->dev
, xdp_prog
, act
);
575 err
= xdp_do_redirect(dev
, &xdp
, xdp_prog
);
581 bpf_warn_invalid_xdp_action(act
);
583 trace_xdp_exception(vi
->dev
, xdp_prog
, act
);
590 skb
= build_skb(buf
, buflen
);
595 skb_reserve(skb
, headroom
- delta
);
596 skb_put(skb
, len
+ delta
);
598 buf
+= header_offset
;
599 memcpy(skb_vnet_hdr(skb
), buf
, vi
->hdr_len
);
600 } /* keep zeroed vnet hdr since packet was changed by bpf */
607 dev
->stats
.rx_dropped
++;
613 static struct sk_buff
*receive_big(struct net_device
*dev
,
614 struct virtnet_info
*vi
,
615 struct receive_queue
*rq
,
619 struct page
*page
= buf
;
620 struct sk_buff
*skb
= page_to_skb(vi
, rq
, page
, 0, len
, PAGE_SIZE
);
628 dev
->stats
.rx_dropped
++;
629 give_pages(rq
, page
);
633 static struct sk_buff
*receive_mergeable(struct net_device
*dev
,
634 struct virtnet_info
*vi
,
635 struct receive_queue
*rq
,
641 struct virtio_net_hdr_mrg_rxbuf
*hdr
= buf
;
642 u16 num_buf
= virtio16_to_cpu(vi
->vdev
, hdr
->num_buffers
);
643 struct page
*page
= virt_to_head_page(buf
);
644 int offset
= buf
- page_address(page
);
645 struct sk_buff
*head_skb
, *curr_skb
;
646 struct bpf_prog
*xdp_prog
;
647 unsigned int truesize
;
648 unsigned int headroom
= mergeable_ctx_to_headroom(ctx
);
654 xdp_prog
= rcu_dereference(rq
->xdp_prog
);
656 struct page
*xdp_page
;
661 /* This happens when rx buffer size is underestimated */
662 if (unlikely(num_buf
> 1 ||
663 headroom
< virtnet_get_headroom(vi
))) {
664 /* linearize data for XDP */
665 xdp_page
= xdp_linearize_page(rq
, &num_buf
,
671 offset
= VIRTIO_XDP_HEADROOM
;
676 /* Transient failure which in theory could occur if
677 * in-flight packets from before XDP was enabled reach
678 * the receive path after XDP is loaded. In practice I
679 * was not able to create this condition.
681 if (unlikely(hdr
->hdr
.gso_type
))
684 /* Allow consuming headroom but reserve enough space to push
685 * the descriptor on if we get an XDP_TX return code.
687 data
= page_address(xdp_page
) + offset
;
688 xdp
.data_hard_start
= data
- VIRTIO_XDP_HEADROOM
+ vi
->hdr_len
;
689 xdp
.data
= data
+ vi
->hdr_len
;
690 xdp_set_data_meta_invalid(&xdp
);
691 xdp
.data_end
= xdp
.data
+ (len
- vi
->hdr_len
);
692 act
= bpf_prog_run_xdp(xdp_prog
, &xdp
);
695 ewma_pkt_len_add(&rq
->mrg_avg_pkt_len
, len
);
699 /* recalculate offset to account for any header
700 * adjustments. Note other cases do not build an
701 * skb and avoid using offset
704 page_address(xdp_page
) - vi
->hdr_len
;
706 /* We can only create skb based on xdp_page. */
707 if (unlikely(xdp_page
!= page
)) {
710 head_skb
= page_to_skb(vi
, rq
, xdp_page
,
711 offset
, len
, PAGE_SIZE
);
716 if (unlikely(!__virtnet_xdp_xmit(vi
, &xdp
)))
717 trace_xdp_exception(vi
->dev
, xdp_prog
, act
);
720 if (unlikely(xdp_page
!= page
))
725 err
= xdp_do_redirect(dev
, &xdp
, xdp_prog
);
731 bpf_warn_invalid_xdp_action(act
);
733 trace_xdp_exception(vi
->dev
, xdp_prog
, act
);
735 if (unlikely(xdp_page
!= page
))
736 __free_pages(xdp_page
, 0);
742 truesize
= mergeable_ctx_to_truesize(ctx
);
743 if (unlikely(len
> truesize
)) {
744 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
745 dev
->name
, len
, (unsigned long)ctx
);
746 dev
->stats
.rx_length_errors
++;
750 head_skb
= page_to_skb(vi
, rq
, page
, offset
, len
, truesize
);
753 if (unlikely(!curr_skb
))
758 buf
= virtqueue_get_buf_ctx(rq
->vq
, &len
, &ctx
);
759 if (unlikely(!buf
)) {
760 pr_debug("%s: rx error: %d buffers out of %d missing\n",
762 virtio16_to_cpu(vi
->vdev
,
764 dev
->stats
.rx_length_errors
++;
768 page
= virt_to_head_page(buf
);
770 truesize
= mergeable_ctx_to_truesize(ctx
);
771 if (unlikely(len
> truesize
)) {
772 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
773 dev
->name
, len
, (unsigned long)ctx
);
774 dev
->stats
.rx_length_errors
++;
778 num_skb_frags
= skb_shinfo(curr_skb
)->nr_frags
;
779 if (unlikely(num_skb_frags
== MAX_SKB_FRAGS
)) {
780 struct sk_buff
*nskb
= alloc_skb(0, GFP_ATOMIC
);
784 if (curr_skb
== head_skb
)
785 skb_shinfo(curr_skb
)->frag_list
= nskb
;
787 curr_skb
->next
= nskb
;
789 head_skb
->truesize
+= nskb
->truesize
;
792 if (curr_skb
!= head_skb
) {
793 head_skb
->data_len
+= len
;
794 head_skb
->len
+= len
;
795 head_skb
->truesize
+= truesize
;
797 offset
= buf
- page_address(page
);
798 if (skb_can_coalesce(curr_skb
, num_skb_frags
, page
, offset
)) {
800 skb_coalesce_rx_frag(curr_skb
, num_skb_frags
- 1,
803 skb_add_rx_frag(curr_skb
, num_skb_frags
, page
,
804 offset
, len
, truesize
);
808 ewma_pkt_len_add(&rq
->mrg_avg_pkt_len
, head_skb
->len
);
816 buf
= virtqueue_get_buf(rq
->vq
, &len
);
817 if (unlikely(!buf
)) {
818 pr_debug("%s: rx error: %d buffers missing\n",
820 dev
->stats
.rx_length_errors
++;
823 page
= virt_to_head_page(buf
);
827 dev
->stats
.rx_dropped
++;
828 dev_kfree_skb(head_skb
);
833 static int receive_buf(struct virtnet_info
*vi
, struct receive_queue
*rq
,
834 void *buf
, unsigned int len
, void **ctx
, bool *xdp_xmit
)
836 struct net_device
*dev
= vi
->dev
;
838 struct virtio_net_hdr_mrg_rxbuf
*hdr
;
841 if (unlikely(len
< vi
->hdr_len
+ ETH_HLEN
)) {
842 pr_debug("%s: short packet %i\n", dev
->name
, len
);
843 dev
->stats
.rx_length_errors
++;
844 if (vi
->mergeable_rx_bufs
) {
845 put_page(virt_to_head_page(buf
));
846 } else if (vi
->big_packets
) {
849 put_page(virt_to_head_page(buf
));
854 if (vi
->mergeable_rx_bufs
)
855 skb
= receive_mergeable(dev
, vi
, rq
, buf
, ctx
, len
, xdp_xmit
);
856 else if (vi
->big_packets
)
857 skb
= receive_big(dev
, vi
, rq
, buf
, len
);
859 skb
= receive_small(dev
, vi
, rq
, buf
, ctx
, len
, xdp_xmit
);
864 hdr
= skb_vnet_hdr(skb
);
868 if (hdr
->hdr
.flags
& VIRTIO_NET_HDR_F_DATA_VALID
)
869 skb
->ip_summed
= CHECKSUM_UNNECESSARY
;
871 if (virtio_net_hdr_to_skb(skb
, &hdr
->hdr
,
872 virtio_is_little_endian(vi
->vdev
))) {
873 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
874 dev
->name
, hdr
->hdr
.gso_type
,
879 skb
->protocol
= eth_type_trans(skb
, dev
);
880 pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
881 ntohs(skb
->protocol
), skb
->len
, skb
->pkt_type
);
883 napi_gro_receive(&rq
->napi
, skb
);
887 dev
->stats
.rx_frame_errors
++;
892 /* Unlike mergeable buffers, all buffers are allocated to the
893 * same size, except for the headroom. For this reason we do
894 * not need to use mergeable_len_to_ctx here - it is enough
895 * to store the headroom as the context ignoring the truesize.
897 static int add_recvbuf_small(struct virtnet_info
*vi
, struct receive_queue
*rq
,
900 struct page_frag
*alloc_frag
= &rq
->alloc_frag
;
902 unsigned int xdp_headroom
= virtnet_get_headroom(vi
);
903 void *ctx
= (void *)(unsigned long)xdp_headroom
;
904 int len
= vi
->hdr_len
+ VIRTNET_RX_PAD
+ GOOD_PACKET_LEN
+ xdp_headroom
;
907 len
= SKB_DATA_ALIGN(len
) +
908 SKB_DATA_ALIGN(sizeof(struct skb_shared_info
));
909 if (unlikely(!skb_page_frag_refill(len
, alloc_frag
, gfp
)))
912 buf
= (char *)page_address(alloc_frag
->page
) + alloc_frag
->offset
;
913 get_page(alloc_frag
->page
);
914 alloc_frag
->offset
+= len
;
915 sg_init_one(rq
->sg
, buf
+ VIRTNET_RX_PAD
+ xdp_headroom
,
916 vi
->hdr_len
+ GOOD_PACKET_LEN
);
917 err
= virtqueue_add_inbuf_ctx(rq
->vq
, rq
->sg
, 1, buf
, ctx
, gfp
);
919 put_page(virt_to_head_page(buf
));
923 static int add_recvbuf_big(struct virtnet_info
*vi
, struct receive_queue
*rq
,
926 struct page
*first
, *list
= NULL
;
930 sg_init_table(rq
->sg
, MAX_SKB_FRAGS
+ 2);
932 /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
933 for (i
= MAX_SKB_FRAGS
+ 1; i
> 1; --i
) {
934 first
= get_a_page(rq
, gfp
);
937 give_pages(rq
, list
);
940 sg_set_buf(&rq
->sg
[i
], page_address(first
), PAGE_SIZE
);
942 /* chain new page in list head to match sg */
943 first
->private = (unsigned long)list
;
947 first
= get_a_page(rq
, gfp
);
949 give_pages(rq
, list
);
952 p
= page_address(first
);
954 /* rq->sg[0], rq->sg[1] share the same page */
955 /* a separated rq->sg[0] for header - required in case !any_header_sg */
956 sg_set_buf(&rq
->sg
[0], p
, vi
->hdr_len
);
958 /* rq->sg[1] for data packet, from offset */
959 offset
= sizeof(struct padded_vnet_hdr
);
960 sg_set_buf(&rq
->sg
[1], p
+ offset
, PAGE_SIZE
- offset
);
962 /* chain first in list head */
963 first
->private = (unsigned long)list
;
964 err
= virtqueue_add_inbuf(rq
->vq
, rq
->sg
, MAX_SKB_FRAGS
+ 2,
967 give_pages(rq
, first
);
972 static unsigned int get_mergeable_buf_len(struct receive_queue
*rq
,
973 struct ewma_pkt_len
*avg_pkt_len
)
975 const size_t hdr_len
= sizeof(struct virtio_net_hdr_mrg_rxbuf
);
978 len
= hdr_len
+ clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len
),
979 rq
->min_buf_len
, PAGE_SIZE
- hdr_len
);
980 return ALIGN(len
, L1_CACHE_BYTES
);
983 static int add_recvbuf_mergeable(struct virtnet_info
*vi
,
984 struct receive_queue
*rq
, gfp_t gfp
)
986 struct page_frag
*alloc_frag
= &rq
->alloc_frag
;
987 unsigned int headroom
= virtnet_get_headroom(vi
);
991 unsigned int len
, hole
;
993 len
= get_mergeable_buf_len(rq
, &rq
->mrg_avg_pkt_len
);
994 if (unlikely(!skb_page_frag_refill(len
+ headroom
, alloc_frag
, gfp
)))
997 buf
= (char *)page_address(alloc_frag
->page
) + alloc_frag
->offset
;
998 buf
+= headroom
; /* advance address leaving hole at front of pkt */
999 get_page(alloc_frag
->page
);
1000 alloc_frag
->offset
+= len
+ headroom
;
1001 hole
= alloc_frag
->size
- alloc_frag
->offset
;
1002 if (hole
< len
+ headroom
) {
1003 /* To avoid internal fragmentation, if there is very likely not
1004 * enough space for another buffer, add the remaining space to
1005 * the current buffer.
1008 alloc_frag
->offset
+= hole
;
1011 sg_init_one(rq
->sg
, buf
, len
);
1012 ctx
= mergeable_len_to_ctx(len
, headroom
);
1013 err
= virtqueue_add_inbuf_ctx(rq
->vq
, rq
->sg
, 1, buf
, ctx
, gfp
);
1015 put_page(virt_to_head_page(buf
));
1021 * Returns false if we couldn't fill entirely (OOM).
1023 * Normally run in the receive path, but can also be run from ndo_open
1024 * before we're receiving packets, or from refill_work which is
1025 * careful to disable receiving (using napi_disable).
1027 static bool try_fill_recv(struct virtnet_info
*vi
, struct receive_queue
*rq
,
1034 if (vi
->mergeable_rx_bufs
)
1035 err
= add_recvbuf_mergeable(vi
, rq
, gfp
);
1036 else if (vi
->big_packets
)
1037 err
= add_recvbuf_big(vi
, rq
, gfp
);
1039 err
= add_recvbuf_small(vi
, rq
, gfp
);
1041 oom
= err
== -ENOMEM
;
1044 } while (rq
->vq
->num_free
);
1045 virtqueue_kick(rq
->vq
);
1049 static void skb_recv_done(struct virtqueue
*rvq
)
1051 struct virtnet_info
*vi
= rvq
->vdev
->priv
;
1052 struct receive_queue
*rq
= &vi
->rq
[vq2rxq(rvq
)];
1054 virtqueue_napi_schedule(&rq
->napi
, rvq
);
1057 static void virtnet_napi_enable(struct virtqueue
*vq
, struct napi_struct
*napi
)
1061 /* If all buffers were filled by other side before we napi_enabled, we
1062 * won't get another interrupt, so process any outstanding packets now.
1063 * Call local_bh_enable after to trigger softIRQ processing.
1066 virtqueue_napi_schedule(napi
, vq
);
1070 static void virtnet_napi_tx_enable(struct virtnet_info
*vi
,
1071 struct virtqueue
*vq
,
1072 struct napi_struct
*napi
)
1077 /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
1078 * enable the feature if this is likely affine with the transmit path.
1080 if (!vi
->affinity_hint_set
) {
1085 return virtnet_napi_enable(vq
, napi
);
1088 static void virtnet_napi_tx_disable(struct napi_struct
*napi
)
1094 static void refill_work(struct work_struct
*work
)
1096 struct virtnet_info
*vi
=
1097 container_of(work
, struct virtnet_info
, refill
.work
);
1101 for (i
= 0; i
< vi
->curr_queue_pairs
; i
++) {
1102 struct receive_queue
*rq
= &vi
->rq
[i
];
1104 napi_disable(&rq
->napi
);
1105 still_empty
= !try_fill_recv(vi
, rq
, GFP_KERNEL
);
1106 virtnet_napi_enable(rq
->vq
, &rq
->napi
);
1108 /* In theory, this can happen: if we don't get any buffers in
1109 * we will *never* try to fill again.
1112 schedule_delayed_work(&vi
->refill
, HZ
/2);
1116 static int virtnet_receive(struct receive_queue
*rq
, int budget
, bool *xdp_xmit
)
1118 struct virtnet_info
*vi
= rq
->vq
->vdev
->priv
;
1119 unsigned int len
, received
= 0, bytes
= 0;
1121 struct virtnet_stats
*stats
= this_cpu_ptr(vi
->stats
);
1123 if (!vi
->big_packets
|| vi
->mergeable_rx_bufs
) {
1126 while (received
< budget
&&
1127 (buf
= virtqueue_get_buf_ctx(rq
->vq
, &len
, &ctx
))) {
1128 bytes
+= receive_buf(vi
, rq
, buf
, len
, ctx
, xdp_xmit
);
1132 while (received
< budget
&&
1133 (buf
= virtqueue_get_buf(rq
->vq
, &len
)) != NULL
) {
1134 bytes
+= receive_buf(vi
, rq
, buf
, len
, NULL
, xdp_xmit
);
1139 if (rq
->vq
->num_free
> virtqueue_get_vring_size(rq
->vq
) / 2) {
1140 if (!try_fill_recv(vi
, rq
, GFP_ATOMIC
))
1141 schedule_delayed_work(&vi
->refill
, 0);
1144 u64_stats_update_begin(&stats
->rx_syncp
);
1145 stats
->rx_bytes
+= bytes
;
1146 stats
->rx_packets
+= received
;
1147 u64_stats_update_end(&stats
->rx_syncp
);
1152 static void free_old_xmit_skbs(struct send_queue
*sq
)
1154 struct sk_buff
*skb
;
1156 struct virtnet_info
*vi
= sq
->vq
->vdev
->priv
;
1157 struct virtnet_stats
*stats
= this_cpu_ptr(vi
->stats
);
1158 unsigned int packets
= 0;
1159 unsigned int bytes
= 0;
1161 while ((skb
= virtqueue_get_buf(sq
->vq
, &len
)) != NULL
) {
1162 pr_debug("Sent skb %p\n", skb
);
1167 dev_consume_skb_any(skb
);
1170 /* Avoid overhead when no packets have been processed
1171 * happens when called speculatively from start_xmit.
1176 u64_stats_update_begin(&stats
->tx_syncp
);
1177 stats
->tx_bytes
+= bytes
;
1178 stats
->tx_packets
+= packets
;
1179 u64_stats_update_end(&stats
->tx_syncp
);
1182 static void virtnet_poll_cleantx(struct receive_queue
*rq
)
1184 struct virtnet_info
*vi
= rq
->vq
->vdev
->priv
;
1185 unsigned int index
= vq2rxq(rq
->vq
);
1186 struct send_queue
*sq
= &vi
->sq
[index
];
1187 struct netdev_queue
*txq
= netdev_get_tx_queue(vi
->dev
, index
);
1189 if (!sq
->napi
.weight
)
1192 if (__netif_tx_trylock(txq
)) {
1193 free_old_xmit_skbs(sq
);
1194 __netif_tx_unlock(txq
);
1197 if (sq
->vq
->num_free
>= 2 + MAX_SKB_FRAGS
)
1198 netif_tx_wake_queue(txq
);
1201 static int virtnet_poll(struct napi_struct
*napi
, int budget
)
1203 struct receive_queue
*rq
=
1204 container_of(napi
, struct receive_queue
, napi
);
1205 unsigned int received
;
1206 bool xdp_xmit
= false;
1208 virtnet_poll_cleantx(rq
);
1210 received
= virtnet_receive(rq
, budget
, &xdp_xmit
);
1212 /* Out of packets? */
1213 if (received
< budget
)
1214 virtqueue_napi_complete(napi
, rq
->vq
, received
);
1222 static int virtnet_open(struct net_device
*dev
)
1224 struct virtnet_info
*vi
= netdev_priv(dev
);
1227 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1228 if (i
< vi
->curr_queue_pairs
)
1229 /* Make sure we have some buffers: if oom use wq. */
1230 if (!try_fill_recv(vi
, &vi
->rq
[i
], GFP_KERNEL
))
1231 schedule_delayed_work(&vi
->refill
, 0);
1232 virtnet_napi_enable(vi
->rq
[i
].vq
, &vi
->rq
[i
].napi
);
1233 virtnet_napi_tx_enable(vi
, vi
->sq
[i
].vq
, &vi
->sq
[i
].napi
);
1239 static int virtnet_poll_tx(struct napi_struct
*napi
, int budget
)
1241 struct send_queue
*sq
= container_of(napi
, struct send_queue
, napi
);
1242 struct virtnet_info
*vi
= sq
->vq
->vdev
->priv
;
1243 struct netdev_queue
*txq
= netdev_get_tx_queue(vi
->dev
, vq2txq(sq
->vq
));
1245 __netif_tx_lock(txq
, raw_smp_processor_id());
1246 free_old_xmit_skbs(sq
);
1247 __netif_tx_unlock(txq
);
1249 virtqueue_napi_complete(napi
, sq
->vq
, 0);
1251 if (sq
->vq
->num_free
>= 2 + MAX_SKB_FRAGS
)
1252 netif_tx_wake_queue(txq
);
1257 static int xmit_skb(struct send_queue
*sq
, struct sk_buff
*skb
)
1259 struct virtio_net_hdr_mrg_rxbuf
*hdr
;
1260 const unsigned char *dest
= ((struct ethhdr
*)skb
->data
)->h_dest
;
1261 struct virtnet_info
*vi
= sq
->vq
->vdev
->priv
;
1263 unsigned hdr_len
= vi
->hdr_len
;
1266 pr_debug("%s: xmit %p %pM\n", vi
->dev
->name
, skb
, dest
);
1268 can_push
= vi
->any_header_sg
&&
1269 !((unsigned long)skb
->data
& (__alignof__(*hdr
) - 1)) &&
1270 !skb_header_cloned(skb
) && skb_headroom(skb
) >= hdr_len
;
1271 /* Even if we can, don't push here yet as this would skew
1272 * csum_start offset below. */
1274 hdr
= (struct virtio_net_hdr_mrg_rxbuf
*)(skb
->data
- hdr_len
);
1276 hdr
= skb_vnet_hdr(skb
);
1278 if (virtio_net_hdr_from_skb(skb
, &hdr
->hdr
,
1279 virtio_is_little_endian(vi
->vdev
), false))
1282 if (vi
->mergeable_rx_bufs
)
1283 hdr
->num_buffers
= 0;
1285 sg_init_table(sq
->sg
, skb_shinfo(skb
)->nr_frags
+ (can_push
? 1 : 2));
1287 __skb_push(skb
, hdr_len
);
1288 num_sg
= skb_to_sgvec(skb
, sq
->sg
, 0, skb
->len
);
1289 if (unlikely(num_sg
< 0))
1291 /* Pull header back to avoid skew in tx bytes calculations. */
1292 __skb_pull(skb
, hdr_len
);
1294 sg_set_buf(sq
->sg
, hdr
, hdr_len
);
1295 num_sg
= skb_to_sgvec(skb
, sq
->sg
+ 1, 0, skb
->len
);
1296 if (unlikely(num_sg
< 0))
1300 return virtqueue_add_outbuf(sq
->vq
, sq
->sg
, num_sg
, skb
, GFP_ATOMIC
);
1303 static netdev_tx_t
start_xmit(struct sk_buff
*skb
, struct net_device
*dev
)
1305 struct virtnet_info
*vi
= netdev_priv(dev
);
1306 int qnum
= skb_get_queue_mapping(skb
);
1307 struct send_queue
*sq
= &vi
->sq
[qnum
];
1309 struct netdev_queue
*txq
= netdev_get_tx_queue(dev
, qnum
);
1310 bool kick
= !skb
->xmit_more
;
1311 bool use_napi
= sq
->napi
.weight
;
1313 /* Free up any pending old buffers before queueing new ones. */
1314 free_old_xmit_skbs(sq
);
1316 if (use_napi
&& kick
)
1317 virtqueue_enable_cb_delayed(sq
->vq
);
1319 /* timestamp packet in software */
1320 skb_tx_timestamp(skb
);
1322 /* Try to transmit */
1323 err
= xmit_skb(sq
, skb
);
1325 /* This should not happen! */
1326 if (unlikely(err
)) {
1327 dev
->stats
.tx_fifo_errors
++;
1328 if (net_ratelimit())
1330 "Unexpected TXQ (%d) queue failure: %d\n", qnum
, err
);
1331 dev
->stats
.tx_dropped
++;
1332 dev_kfree_skb_any(skb
);
1333 return NETDEV_TX_OK
;
1336 /* Don't wait up for transmitted skbs to be freed. */
1342 /* If running out of space, stop queue to avoid getting packets that we
1343 * are then unable to transmit.
1344 * An alternative would be to force queuing layer to requeue the skb by
1345 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1346 * returned in a normal path of operation: it means that driver is not
1347 * maintaining the TX queue stop/start state properly, and causes
1348 * the stack to do a non-trivial amount of useless work.
1349 * Since most packets only take 1 or 2 ring slots, stopping the queue
1350 * early means 16 slots are typically wasted.
1352 if (sq
->vq
->num_free
< 2+MAX_SKB_FRAGS
) {
1353 netif_stop_subqueue(dev
, qnum
);
1355 unlikely(!virtqueue_enable_cb_delayed(sq
->vq
))) {
1356 /* More just got used, free them then recheck. */
1357 free_old_xmit_skbs(sq
);
1358 if (sq
->vq
->num_free
>= 2+MAX_SKB_FRAGS
) {
1359 netif_start_subqueue(dev
, qnum
);
1360 virtqueue_disable_cb(sq
->vq
);
1365 if (kick
|| netif_xmit_stopped(txq
))
1366 virtqueue_kick(sq
->vq
);
1368 return NETDEV_TX_OK
;
1372 * Send command via the control virtqueue and check status. Commands
1373 * supported by the hypervisor, as indicated by feature bits, should
1374 * never fail unless improperly formatted.
1376 static bool virtnet_send_command(struct virtnet_info
*vi
, u8
class, u8 cmd
,
1377 struct scatterlist
*out
)
1379 struct scatterlist
*sgs
[4], hdr
, stat
;
1380 unsigned out_num
= 0, tmp
;
1382 /* Caller should know better */
1383 BUG_ON(!virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_CTRL_VQ
));
1385 vi
->ctrl_status
= ~0;
1386 vi
->ctrl_hdr
.class = class;
1387 vi
->ctrl_hdr
.cmd
= cmd
;
1389 sg_init_one(&hdr
, &vi
->ctrl_hdr
, sizeof(vi
->ctrl_hdr
));
1390 sgs
[out_num
++] = &hdr
;
1393 sgs
[out_num
++] = out
;
1395 /* Add return status. */
1396 sg_init_one(&stat
, &vi
->ctrl_status
, sizeof(vi
->ctrl_status
));
1397 sgs
[out_num
] = &stat
;
1399 BUG_ON(out_num
+ 1 > ARRAY_SIZE(sgs
));
1400 virtqueue_add_sgs(vi
->cvq
, sgs
, out_num
, 1, vi
, GFP_ATOMIC
);
1402 if (unlikely(!virtqueue_kick(vi
->cvq
)))
1403 return vi
->ctrl_status
== VIRTIO_NET_OK
;
1405 /* Spin for a response, the kick causes an ioport write, trapping
1406 * into the hypervisor, so the request should be handled immediately.
1408 while (!virtqueue_get_buf(vi
->cvq
, &tmp
) &&
1409 !virtqueue_is_broken(vi
->cvq
))
1412 return vi
->ctrl_status
== VIRTIO_NET_OK
;
1415 static int virtnet_set_mac_address(struct net_device
*dev
, void *p
)
1417 struct virtnet_info
*vi
= netdev_priv(dev
);
1418 struct virtio_device
*vdev
= vi
->vdev
;
1420 struct sockaddr
*addr
;
1421 struct scatterlist sg
;
1423 addr
= kmemdup(p
, sizeof(*addr
), GFP_KERNEL
);
1427 ret
= eth_prepare_mac_addr_change(dev
, addr
);
1431 if (virtio_has_feature(vdev
, VIRTIO_NET_F_CTRL_MAC_ADDR
)) {
1432 sg_init_one(&sg
, addr
->sa_data
, dev
->addr_len
);
1433 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_MAC
,
1434 VIRTIO_NET_CTRL_MAC_ADDR_SET
, &sg
)) {
1435 dev_warn(&vdev
->dev
,
1436 "Failed to set mac address by vq command.\n");
1440 } else if (virtio_has_feature(vdev
, VIRTIO_NET_F_MAC
) &&
1441 !virtio_has_feature(vdev
, VIRTIO_F_VERSION_1
)) {
1444 /* Naturally, this has an atomicity problem. */
1445 for (i
= 0; i
< dev
->addr_len
; i
++)
1446 virtio_cwrite8(vdev
,
1447 offsetof(struct virtio_net_config
, mac
) +
1448 i
, addr
->sa_data
[i
]);
1451 eth_commit_mac_addr_change(dev
, p
);
1459 static void virtnet_stats(struct net_device
*dev
,
1460 struct rtnl_link_stats64
*tot
)
1462 struct virtnet_info
*vi
= netdev_priv(dev
);
1466 for_each_possible_cpu(cpu
) {
1467 struct virtnet_stats
*stats
= per_cpu_ptr(vi
->stats
, cpu
);
1468 u64 tpackets
, tbytes
, rpackets
, rbytes
;
1471 start
= u64_stats_fetch_begin_irq(&stats
->tx_syncp
);
1472 tpackets
= stats
->tx_packets
;
1473 tbytes
= stats
->tx_bytes
;
1474 } while (u64_stats_fetch_retry_irq(&stats
->tx_syncp
, start
));
1477 start
= u64_stats_fetch_begin_irq(&stats
->rx_syncp
);
1478 rpackets
= stats
->rx_packets
;
1479 rbytes
= stats
->rx_bytes
;
1480 } while (u64_stats_fetch_retry_irq(&stats
->rx_syncp
, start
));
1482 tot
->rx_packets
+= rpackets
;
1483 tot
->tx_packets
+= tpackets
;
1484 tot
->rx_bytes
+= rbytes
;
1485 tot
->tx_bytes
+= tbytes
;
1488 tot
->tx_dropped
= dev
->stats
.tx_dropped
;
1489 tot
->tx_fifo_errors
= dev
->stats
.tx_fifo_errors
;
1490 tot
->rx_dropped
= dev
->stats
.rx_dropped
;
1491 tot
->rx_length_errors
= dev
->stats
.rx_length_errors
;
1492 tot
->rx_frame_errors
= dev
->stats
.rx_frame_errors
;
1495 #ifdef CONFIG_NET_POLL_CONTROLLER
1496 static void virtnet_netpoll(struct net_device
*dev
)
1498 struct virtnet_info
*vi
= netdev_priv(dev
);
1501 for (i
= 0; i
< vi
->curr_queue_pairs
; i
++)
1502 napi_schedule(&vi
->rq
[i
].napi
);
1506 static void virtnet_ack_link_announce(struct virtnet_info
*vi
)
1509 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_ANNOUNCE
,
1510 VIRTIO_NET_CTRL_ANNOUNCE_ACK
, NULL
))
1511 dev_warn(&vi
->dev
->dev
, "Failed to ack link announce.\n");
1515 static int _virtnet_set_queues(struct virtnet_info
*vi
, u16 queue_pairs
)
1517 struct scatterlist sg
;
1518 struct net_device
*dev
= vi
->dev
;
1520 if (!vi
->has_cvq
|| !virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_MQ
))
1523 vi
->ctrl_mq
.virtqueue_pairs
= cpu_to_virtio16(vi
->vdev
, queue_pairs
);
1524 sg_init_one(&sg
, &vi
->ctrl_mq
, sizeof(vi
->ctrl_mq
));
1526 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_MQ
,
1527 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET
, &sg
)) {
1528 dev_warn(&dev
->dev
, "Fail to set num of queue pairs to %d\n",
1532 vi
->curr_queue_pairs
= queue_pairs
;
1533 /* virtnet_open() will refill when device is going to up. */
1534 if (dev
->flags
& IFF_UP
)
1535 schedule_delayed_work(&vi
->refill
, 0);
1541 static int virtnet_set_queues(struct virtnet_info
*vi
, u16 queue_pairs
)
1546 err
= _virtnet_set_queues(vi
, queue_pairs
);
1551 static int virtnet_close(struct net_device
*dev
)
1553 struct virtnet_info
*vi
= netdev_priv(dev
);
1556 /* Make sure refill_work doesn't re-enable napi! */
1557 cancel_delayed_work_sync(&vi
->refill
);
1559 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1560 napi_disable(&vi
->rq
[i
].napi
);
1561 virtnet_napi_tx_disable(&vi
->sq
[i
].napi
);
1567 static void virtnet_set_rx_mode(struct net_device
*dev
)
1569 struct virtnet_info
*vi
= netdev_priv(dev
);
1570 struct scatterlist sg
[2];
1571 struct virtio_net_ctrl_mac
*mac_data
;
1572 struct netdev_hw_addr
*ha
;
1578 /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1579 if (!virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_CTRL_RX
))
1582 vi
->ctrl_promisc
= ((dev
->flags
& IFF_PROMISC
) != 0);
1583 vi
->ctrl_allmulti
= ((dev
->flags
& IFF_ALLMULTI
) != 0);
1585 sg_init_one(sg
, &vi
->ctrl_promisc
, sizeof(vi
->ctrl_promisc
));
1587 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_RX
,
1588 VIRTIO_NET_CTRL_RX_PROMISC
, sg
))
1589 dev_warn(&dev
->dev
, "Failed to %sable promisc mode.\n",
1590 vi
->ctrl_promisc
? "en" : "dis");
1592 sg_init_one(sg
, &vi
->ctrl_allmulti
, sizeof(vi
->ctrl_allmulti
));
1594 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_RX
,
1595 VIRTIO_NET_CTRL_RX_ALLMULTI
, sg
))
1596 dev_warn(&dev
->dev
, "Failed to %sable allmulti mode.\n",
1597 vi
->ctrl_allmulti
? "en" : "dis");
1599 uc_count
= netdev_uc_count(dev
);
1600 mc_count
= netdev_mc_count(dev
);
1601 /* MAC filter - use one buffer for both lists */
1602 buf
= kzalloc(((uc_count
+ mc_count
) * ETH_ALEN
) +
1603 (2 * sizeof(mac_data
->entries
)), GFP_ATOMIC
);
1608 sg_init_table(sg
, 2);
1610 /* Store the unicast list and count in the front of the buffer */
1611 mac_data
->entries
= cpu_to_virtio32(vi
->vdev
, uc_count
);
1613 netdev_for_each_uc_addr(ha
, dev
)
1614 memcpy(&mac_data
->macs
[i
++][0], ha
->addr
, ETH_ALEN
);
1616 sg_set_buf(&sg
[0], mac_data
,
1617 sizeof(mac_data
->entries
) + (uc_count
* ETH_ALEN
));
1619 /* multicast list and count fill the end */
1620 mac_data
= (void *)&mac_data
->macs
[uc_count
][0];
1622 mac_data
->entries
= cpu_to_virtio32(vi
->vdev
, mc_count
);
1624 netdev_for_each_mc_addr(ha
, dev
)
1625 memcpy(&mac_data
->macs
[i
++][0], ha
->addr
, ETH_ALEN
);
1627 sg_set_buf(&sg
[1], mac_data
,
1628 sizeof(mac_data
->entries
) + (mc_count
* ETH_ALEN
));
1630 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_MAC
,
1631 VIRTIO_NET_CTRL_MAC_TABLE_SET
, sg
))
1632 dev_warn(&dev
->dev
, "Failed to set MAC filter table.\n");
1637 static int virtnet_vlan_rx_add_vid(struct net_device
*dev
,
1638 __be16 proto
, u16 vid
)
1640 struct virtnet_info
*vi
= netdev_priv(dev
);
1641 struct scatterlist sg
;
1644 sg_init_one(&sg
, &vi
->ctrl_vid
, sizeof(vi
->ctrl_vid
));
1646 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_VLAN
,
1647 VIRTIO_NET_CTRL_VLAN_ADD
, &sg
))
1648 dev_warn(&dev
->dev
, "Failed to add VLAN ID %d.\n", vid
);
1652 static int virtnet_vlan_rx_kill_vid(struct net_device
*dev
,
1653 __be16 proto
, u16 vid
)
1655 struct virtnet_info
*vi
= netdev_priv(dev
);
1656 struct scatterlist sg
;
1659 sg_init_one(&sg
, &vi
->ctrl_vid
, sizeof(vi
->ctrl_vid
));
1661 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_VLAN
,
1662 VIRTIO_NET_CTRL_VLAN_DEL
, &sg
))
1663 dev_warn(&dev
->dev
, "Failed to kill VLAN ID %d.\n", vid
);
1667 static void virtnet_clean_affinity(struct virtnet_info
*vi
, long hcpu
)
1671 if (vi
->affinity_hint_set
) {
1672 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1673 virtqueue_set_affinity(vi
->rq
[i
].vq
, -1);
1674 virtqueue_set_affinity(vi
->sq
[i
].vq
, -1);
1677 vi
->affinity_hint_set
= false;
1681 static void virtnet_set_affinity(struct virtnet_info
*vi
)
1686 /* In multiqueue mode, when the number of cpu is equal to the number of
1687 * queue pairs, we let the queue pairs to be private to one cpu by
1688 * setting the affinity hint to eliminate the contention.
1690 if (vi
->curr_queue_pairs
== 1 ||
1691 vi
->max_queue_pairs
!= num_online_cpus()) {
1692 virtnet_clean_affinity(vi
, -1);
1697 for_each_online_cpu(cpu
) {
1698 virtqueue_set_affinity(vi
->rq
[i
].vq
, cpu
);
1699 virtqueue_set_affinity(vi
->sq
[i
].vq
, cpu
);
1700 netif_set_xps_queue(vi
->dev
, cpumask_of(cpu
), i
);
1704 vi
->affinity_hint_set
= true;
1707 static int virtnet_cpu_online(unsigned int cpu
, struct hlist_node
*node
)
1709 struct virtnet_info
*vi
= hlist_entry_safe(node
, struct virtnet_info
,
1711 virtnet_set_affinity(vi
);
1715 static int virtnet_cpu_dead(unsigned int cpu
, struct hlist_node
*node
)
1717 struct virtnet_info
*vi
= hlist_entry_safe(node
, struct virtnet_info
,
1719 virtnet_set_affinity(vi
);
1723 static int virtnet_cpu_down_prep(unsigned int cpu
, struct hlist_node
*node
)
1725 struct virtnet_info
*vi
= hlist_entry_safe(node
, struct virtnet_info
,
1728 virtnet_clean_affinity(vi
, cpu
);
1732 static enum cpuhp_state virtionet_online
;
1734 static int virtnet_cpu_notif_add(struct virtnet_info
*vi
)
1738 ret
= cpuhp_state_add_instance_nocalls(virtionet_online
, &vi
->node
);
1741 ret
= cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD
,
1745 cpuhp_state_remove_instance_nocalls(virtionet_online
, &vi
->node
);
1749 static void virtnet_cpu_notif_remove(struct virtnet_info
*vi
)
1751 cpuhp_state_remove_instance_nocalls(virtionet_online
, &vi
->node
);
1752 cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD
,
1756 static void virtnet_get_ringparam(struct net_device
*dev
,
1757 struct ethtool_ringparam
*ring
)
1759 struct virtnet_info
*vi
= netdev_priv(dev
);
1761 ring
->rx_max_pending
= virtqueue_get_vring_size(vi
->rq
[0].vq
);
1762 ring
->tx_max_pending
= virtqueue_get_vring_size(vi
->sq
[0].vq
);
1763 ring
->rx_pending
= ring
->rx_max_pending
;
1764 ring
->tx_pending
= ring
->tx_max_pending
;
1768 static void virtnet_get_drvinfo(struct net_device
*dev
,
1769 struct ethtool_drvinfo
*info
)
1771 struct virtnet_info
*vi
= netdev_priv(dev
);
1772 struct virtio_device
*vdev
= vi
->vdev
;
1774 strlcpy(info
->driver
, KBUILD_MODNAME
, sizeof(info
->driver
));
1775 strlcpy(info
->version
, VIRTNET_DRIVER_VERSION
, sizeof(info
->version
));
1776 strlcpy(info
->bus_info
, virtio_bus_name(vdev
), sizeof(info
->bus_info
));
1780 /* TODO: Eliminate OOO packets during switching */
1781 static int virtnet_set_channels(struct net_device
*dev
,
1782 struct ethtool_channels
*channels
)
1784 struct virtnet_info
*vi
= netdev_priv(dev
);
1785 u16 queue_pairs
= channels
->combined_count
;
1788 /* We don't support separate rx/tx channels.
1789 * We don't allow setting 'other' channels.
1791 if (channels
->rx_count
|| channels
->tx_count
|| channels
->other_count
)
1794 if (queue_pairs
> vi
->max_queue_pairs
|| queue_pairs
== 0)
1797 /* For now we don't support modifying channels while XDP is loaded
1798 * also when XDP is loaded all RX queues have XDP programs so we only
1799 * need to check a single RX queue.
1801 if (vi
->rq
[0].xdp_prog
)
1805 err
= _virtnet_set_queues(vi
, queue_pairs
);
1807 netif_set_real_num_tx_queues(dev
, queue_pairs
);
1808 netif_set_real_num_rx_queues(dev
, queue_pairs
);
1810 virtnet_set_affinity(vi
);
1817 static void virtnet_get_channels(struct net_device
*dev
,
1818 struct ethtool_channels
*channels
)
1820 struct virtnet_info
*vi
= netdev_priv(dev
);
1822 channels
->combined_count
= vi
->curr_queue_pairs
;
1823 channels
->max_combined
= vi
->max_queue_pairs
;
1824 channels
->max_other
= 0;
1825 channels
->rx_count
= 0;
1826 channels
->tx_count
= 0;
1827 channels
->other_count
= 0;
1830 /* Check if the user is trying to change anything besides speed/duplex */
1832 virtnet_validate_ethtool_cmd(const struct ethtool_link_ksettings
*cmd
)
1834 struct ethtool_link_ksettings diff1
= *cmd
;
1835 struct ethtool_link_ksettings diff2
= {};
1837 /* cmd is always set so we need to clear it, validate the port type
1838 * and also without autonegotiation we can ignore advertising
1840 diff1
.base
.speed
= 0;
1841 diff2
.base
.port
= PORT_OTHER
;
1842 ethtool_link_ksettings_zero_link_mode(&diff1
, advertising
);
1843 diff1
.base
.duplex
= 0;
1845 diff1
.base
.link_mode_masks_nwords
= 0;
1847 return !memcmp(&diff1
.base
, &diff2
.base
, sizeof(diff1
.base
)) &&
1848 bitmap_empty(diff1
.link_modes
.supported
,
1849 __ETHTOOL_LINK_MODE_MASK_NBITS
) &&
1850 bitmap_empty(diff1
.link_modes
.advertising
,
1851 __ETHTOOL_LINK_MODE_MASK_NBITS
) &&
1852 bitmap_empty(diff1
.link_modes
.lp_advertising
,
1853 __ETHTOOL_LINK_MODE_MASK_NBITS
);
1856 static int virtnet_set_link_ksettings(struct net_device
*dev
,
1857 const struct ethtool_link_ksettings
*cmd
)
1859 struct virtnet_info
*vi
= netdev_priv(dev
);
1862 speed
= cmd
->base
.speed
;
1863 /* don't allow custom speed and duplex */
1864 if (!ethtool_validate_speed(speed
) ||
1865 !ethtool_validate_duplex(cmd
->base
.duplex
) ||
1866 !virtnet_validate_ethtool_cmd(cmd
))
1869 vi
->duplex
= cmd
->base
.duplex
;
1874 static int virtnet_get_link_ksettings(struct net_device
*dev
,
1875 struct ethtool_link_ksettings
*cmd
)
1877 struct virtnet_info
*vi
= netdev_priv(dev
);
1879 cmd
->base
.speed
= vi
->speed
;
1880 cmd
->base
.duplex
= vi
->duplex
;
1881 cmd
->base
.port
= PORT_OTHER
;
1886 static void virtnet_init_settings(struct net_device
*dev
)
1888 struct virtnet_info
*vi
= netdev_priv(dev
);
1890 vi
->speed
= SPEED_UNKNOWN
;
1891 vi
->duplex
= DUPLEX_UNKNOWN
;
1894 static const struct ethtool_ops virtnet_ethtool_ops
= {
1895 .get_drvinfo
= virtnet_get_drvinfo
,
1896 .get_link
= ethtool_op_get_link
,
1897 .get_ringparam
= virtnet_get_ringparam
,
1898 .set_channels
= virtnet_set_channels
,
1899 .get_channels
= virtnet_get_channels
,
1900 .get_ts_info
= ethtool_op_get_ts_info
,
1901 .get_link_ksettings
= virtnet_get_link_ksettings
,
1902 .set_link_ksettings
= virtnet_set_link_ksettings
,
1905 static void virtnet_freeze_down(struct virtio_device
*vdev
)
1907 struct virtnet_info
*vi
= vdev
->priv
;
1910 /* Make sure no work handler is accessing the device */
1911 flush_work(&vi
->config_work
);
1913 netif_device_detach(vi
->dev
);
1914 netif_tx_disable(vi
->dev
);
1915 cancel_delayed_work_sync(&vi
->refill
);
1917 if (netif_running(vi
->dev
)) {
1918 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1919 napi_disable(&vi
->rq
[i
].napi
);
1920 virtnet_napi_tx_disable(&vi
->sq
[i
].napi
);
1925 static int init_vqs(struct virtnet_info
*vi
);
1927 static int virtnet_restore_up(struct virtio_device
*vdev
)
1929 struct virtnet_info
*vi
= vdev
->priv
;
1936 virtio_device_ready(vdev
);
1938 if (netif_running(vi
->dev
)) {
1939 for (i
= 0; i
< vi
->curr_queue_pairs
; i
++)
1940 if (!try_fill_recv(vi
, &vi
->rq
[i
], GFP_KERNEL
))
1941 schedule_delayed_work(&vi
->refill
, 0);
1943 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1944 virtnet_napi_enable(vi
->rq
[i
].vq
, &vi
->rq
[i
].napi
);
1945 virtnet_napi_tx_enable(vi
, vi
->sq
[i
].vq
,
1950 netif_device_attach(vi
->dev
);
1954 static int virtnet_set_guest_offloads(struct virtnet_info
*vi
, u64 offloads
)
1956 struct scatterlist sg
;
1957 vi
->ctrl_offloads
= cpu_to_virtio64(vi
->vdev
, offloads
);
1959 sg_init_one(&sg
, &vi
->ctrl_offloads
, sizeof(vi
->ctrl_offloads
));
1961 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_GUEST_OFFLOADS
,
1962 VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET
, &sg
)) {
1963 dev_warn(&vi
->dev
->dev
, "Fail to set guest offload. \n");
1970 static int virtnet_clear_guest_offloads(struct virtnet_info
*vi
)
1974 if (!vi
->guest_offloads
)
1977 if (virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_CSUM
))
1978 offloads
= 1ULL << VIRTIO_NET_F_GUEST_CSUM
;
1980 return virtnet_set_guest_offloads(vi
, offloads
);
1983 static int virtnet_restore_guest_offloads(struct virtnet_info
*vi
)
1985 u64 offloads
= vi
->guest_offloads
;
1987 if (!vi
->guest_offloads
)
1989 if (virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_CSUM
))
1990 offloads
|= 1ULL << VIRTIO_NET_F_GUEST_CSUM
;
1992 return virtnet_set_guest_offloads(vi
, offloads
);
1995 static int virtnet_xdp_set(struct net_device
*dev
, struct bpf_prog
*prog
,
1996 struct netlink_ext_ack
*extack
)
1998 unsigned long int max_sz
= PAGE_SIZE
- sizeof(struct padded_vnet_hdr
);
1999 struct virtnet_info
*vi
= netdev_priv(dev
);
2000 struct bpf_prog
*old_prog
;
2001 u16 xdp_qp
= 0, curr_qp
;
2004 if (!virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS
)
2005 && (virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_TSO4
) ||
2006 virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_TSO6
) ||
2007 virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_ECN
) ||
2008 virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_UFO
))) {
2009 NL_SET_ERR_MSG_MOD(extack
, "Can't set XDP while host is implementing LRO, disable LRO first");
2013 if (vi
->mergeable_rx_bufs
&& !vi
->any_header_sg
) {
2014 NL_SET_ERR_MSG_MOD(extack
, "XDP expects header/data in single page, any_header_sg required");
2018 if (dev
->mtu
> max_sz
) {
2019 NL_SET_ERR_MSG_MOD(extack
, "MTU too large to enable XDP");
2020 netdev_warn(dev
, "XDP requires MTU less than %lu\n", max_sz
);
2024 curr_qp
= vi
->curr_queue_pairs
- vi
->xdp_queue_pairs
;
2026 xdp_qp
= nr_cpu_ids
;
2028 /* XDP requires extra queues for XDP_TX */
2029 if (curr_qp
+ xdp_qp
> vi
->max_queue_pairs
) {
2030 NL_SET_ERR_MSG_MOD(extack
, "Too few free TX rings available");
2031 netdev_warn(dev
, "request %i queues but max is %i\n",
2032 curr_qp
+ xdp_qp
, vi
->max_queue_pairs
);
2037 prog
= bpf_prog_add(prog
, vi
->max_queue_pairs
- 1);
2039 return PTR_ERR(prog
);
2042 /* Make sure NAPI is not using any XDP TX queues for RX. */
2043 for (i
= 0; i
< vi
->max_queue_pairs
; i
++)
2044 napi_disable(&vi
->rq
[i
].napi
);
2046 netif_set_real_num_rx_queues(dev
, curr_qp
+ xdp_qp
);
2047 err
= _virtnet_set_queues(vi
, curr_qp
+ xdp_qp
);
2050 vi
->xdp_queue_pairs
= xdp_qp
;
2052 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2053 old_prog
= rtnl_dereference(vi
->rq
[i
].xdp_prog
);
2054 rcu_assign_pointer(vi
->rq
[i
].xdp_prog
, prog
);
2057 virtnet_clear_guest_offloads(vi
);
2059 virtnet_restore_guest_offloads(vi
);
2062 bpf_prog_put(old_prog
);
2063 virtnet_napi_enable(vi
->rq
[i
].vq
, &vi
->rq
[i
].napi
);
2069 for (i
= 0; i
< vi
->max_queue_pairs
; i
++)
2070 virtnet_napi_enable(vi
->rq
[i
].vq
, &vi
->rq
[i
].napi
);
2072 bpf_prog_sub(prog
, vi
->max_queue_pairs
- 1);
2076 static u32
virtnet_xdp_query(struct net_device
*dev
)
2078 struct virtnet_info
*vi
= netdev_priv(dev
);
2079 const struct bpf_prog
*xdp_prog
;
2082 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2083 xdp_prog
= rtnl_dereference(vi
->rq
[i
].xdp_prog
);
2085 return xdp_prog
->aux
->id
;
2090 static int virtnet_xdp(struct net_device
*dev
, struct netdev_bpf
*xdp
)
2092 switch (xdp
->command
) {
2093 case XDP_SETUP_PROG
:
2094 return virtnet_xdp_set(dev
, xdp
->prog
, xdp
->extack
);
2095 case XDP_QUERY_PROG
:
2096 xdp
->prog_id
= virtnet_xdp_query(dev
);
2097 xdp
->prog_attached
= !!xdp
->prog_id
;
2104 static const struct net_device_ops virtnet_netdev
= {
2105 .ndo_open
= virtnet_open
,
2106 .ndo_stop
= virtnet_close
,
2107 .ndo_start_xmit
= start_xmit
,
2108 .ndo_validate_addr
= eth_validate_addr
,
2109 .ndo_set_mac_address
= virtnet_set_mac_address
,
2110 .ndo_set_rx_mode
= virtnet_set_rx_mode
,
2111 .ndo_get_stats64
= virtnet_stats
,
2112 .ndo_vlan_rx_add_vid
= virtnet_vlan_rx_add_vid
,
2113 .ndo_vlan_rx_kill_vid
= virtnet_vlan_rx_kill_vid
,
2114 #ifdef CONFIG_NET_POLL_CONTROLLER
2115 .ndo_poll_controller
= virtnet_netpoll
,
2117 .ndo_bpf
= virtnet_xdp
,
2118 .ndo_xdp_xmit
= virtnet_xdp_xmit
,
2119 .ndo_xdp_flush
= virtnet_xdp_flush
,
2120 .ndo_features_check
= passthru_features_check
,
2123 static void virtnet_config_changed_work(struct work_struct
*work
)
2125 struct virtnet_info
*vi
=
2126 container_of(work
, struct virtnet_info
, config_work
);
2129 if (virtio_cread_feature(vi
->vdev
, VIRTIO_NET_F_STATUS
,
2130 struct virtio_net_config
, status
, &v
) < 0)
2133 if (v
& VIRTIO_NET_S_ANNOUNCE
) {
2134 netdev_notify_peers(vi
->dev
);
2135 virtnet_ack_link_announce(vi
);
2138 /* Ignore unknown (future) status bits */
2139 v
&= VIRTIO_NET_S_LINK_UP
;
2141 if (vi
->status
== v
)
2146 if (vi
->status
& VIRTIO_NET_S_LINK_UP
) {
2147 netif_carrier_on(vi
->dev
);
2148 netif_tx_wake_all_queues(vi
->dev
);
2150 netif_carrier_off(vi
->dev
);
2151 netif_tx_stop_all_queues(vi
->dev
);
2155 static void virtnet_config_changed(struct virtio_device
*vdev
)
2157 struct virtnet_info
*vi
= vdev
->priv
;
2159 schedule_work(&vi
->config_work
);
2162 static void virtnet_free_queues(struct virtnet_info
*vi
)
2166 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2167 napi_hash_del(&vi
->rq
[i
].napi
);
2168 netif_napi_del(&vi
->rq
[i
].napi
);
2169 netif_napi_del(&vi
->sq
[i
].napi
);
2172 /* We called napi_hash_del() before netif_napi_del(),
2173 * we need to respect an RCU grace period before freeing vi->rq
2181 static void _free_receive_bufs(struct virtnet_info
*vi
)
2183 struct bpf_prog
*old_prog
;
2186 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2187 while (vi
->rq
[i
].pages
)
2188 __free_pages(get_a_page(&vi
->rq
[i
], GFP_KERNEL
), 0);
2190 old_prog
= rtnl_dereference(vi
->rq
[i
].xdp_prog
);
2191 RCU_INIT_POINTER(vi
->rq
[i
].xdp_prog
, NULL
);
2193 bpf_prog_put(old_prog
);
2197 static void free_receive_bufs(struct virtnet_info
*vi
)
2200 _free_receive_bufs(vi
);
2204 static void free_receive_page_frags(struct virtnet_info
*vi
)
2207 for (i
= 0; i
< vi
->max_queue_pairs
; i
++)
2208 if (vi
->rq
[i
].alloc_frag
.page
)
2209 put_page(vi
->rq
[i
].alloc_frag
.page
);
2212 static bool is_xdp_raw_buffer_queue(struct virtnet_info
*vi
, int q
)
2214 if (q
< (vi
->curr_queue_pairs
- vi
->xdp_queue_pairs
))
2216 else if (q
< vi
->curr_queue_pairs
)
2222 static void free_unused_bufs(struct virtnet_info
*vi
)
2227 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2228 struct virtqueue
*vq
= vi
->sq
[i
].vq
;
2229 while ((buf
= virtqueue_detach_unused_buf(vq
)) != NULL
) {
2230 if (!is_xdp_raw_buffer_queue(vi
, i
))
2233 put_page(virt_to_head_page(buf
));
2237 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2238 struct virtqueue
*vq
= vi
->rq
[i
].vq
;
2240 while ((buf
= virtqueue_detach_unused_buf(vq
)) != NULL
) {
2241 if (vi
->mergeable_rx_bufs
) {
2242 put_page(virt_to_head_page(buf
));
2243 } else if (vi
->big_packets
) {
2244 give_pages(&vi
->rq
[i
], buf
);
2246 put_page(virt_to_head_page(buf
));
2252 static void virtnet_del_vqs(struct virtnet_info
*vi
)
2254 struct virtio_device
*vdev
= vi
->vdev
;
2256 virtnet_clean_affinity(vi
, -1);
2258 vdev
->config
->del_vqs(vdev
);
2260 virtnet_free_queues(vi
);
2263 /* How large should a single buffer be so a queue full of these can fit at
2264 * least one full packet?
2265 * Logic below assumes the mergeable buffer header is used.
2267 static unsigned int mergeable_min_buf_len(struct virtnet_info
*vi
, struct virtqueue
*vq
)
2269 const unsigned int hdr_len
= sizeof(struct virtio_net_hdr_mrg_rxbuf
);
2270 unsigned int rq_size
= virtqueue_get_vring_size(vq
);
2271 unsigned int packet_len
= vi
->big_packets
? IP_MAX_MTU
: vi
->dev
->max_mtu
;
2272 unsigned int buf_len
= hdr_len
+ ETH_HLEN
+ VLAN_HLEN
+ packet_len
;
2273 unsigned int min_buf_len
= DIV_ROUND_UP(buf_len
, rq_size
);
2275 return max(max(min_buf_len
, hdr_len
) - hdr_len
,
2276 (unsigned int)GOOD_PACKET_LEN
);
2279 static int virtnet_find_vqs(struct virtnet_info
*vi
)
2281 vq_callback_t
**callbacks
;
2282 struct virtqueue
**vqs
;
2288 /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
2289 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
2290 * possible control vq.
2292 total_vqs
= vi
->max_queue_pairs
* 2 +
2293 virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_CTRL_VQ
);
2295 /* Allocate space for find_vqs parameters */
2296 vqs
= kzalloc(total_vqs
* sizeof(*vqs
), GFP_KERNEL
);
2299 callbacks
= kmalloc(total_vqs
* sizeof(*callbacks
), GFP_KERNEL
);
2302 names
= kmalloc(total_vqs
* sizeof(*names
), GFP_KERNEL
);
2305 if (!vi
->big_packets
|| vi
->mergeable_rx_bufs
) {
2306 ctx
= kzalloc(total_vqs
* sizeof(*ctx
), GFP_KERNEL
);
2313 /* Parameters for control virtqueue, if any */
2315 callbacks
[total_vqs
- 1] = NULL
;
2316 names
[total_vqs
- 1] = "control";
2319 /* Allocate/initialize parameters for send/receive virtqueues */
2320 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2321 callbacks
[rxq2vq(i
)] = skb_recv_done
;
2322 callbacks
[txq2vq(i
)] = skb_xmit_done
;
2323 sprintf(vi
->rq
[i
].name
, "input.%d", i
);
2324 sprintf(vi
->sq
[i
].name
, "output.%d", i
);
2325 names
[rxq2vq(i
)] = vi
->rq
[i
].name
;
2326 names
[txq2vq(i
)] = vi
->sq
[i
].name
;
2328 ctx
[rxq2vq(i
)] = true;
2331 ret
= vi
->vdev
->config
->find_vqs(vi
->vdev
, total_vqs
, vqs
, callbacks
,
2337 vi
->cvq
= vqs
[total_vqs
- 1];
2338 if (virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_CTRL_VLAN
))
2339 vi
->dev
->features
|= NETIF_F_HW_VLAN_CTAG_FILTER
;
2342 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2343 vi
->rq
[i
].vq
= vqs
[rxq2vq(i
)];
2344 vi
->rq
[i
].min_buf_len
= mergeable_min_buf_len(vi
, vi
->rq
[i
].vq
);
2345 vi
->sq
[i
].vq
= vqs
[txq2vq(i
)];
2367 static int virtnet_alloc_queues(struct virtnet_info
*vi
)
2371 vi
->sq
= kzalloc(sizeof(*vi
->sq
) * vi
->max_queue_pairs
, GFP_KERNEL
);
2374 vi
->rq
= kzalloc(sizeof(*vi
->rq
) * vi
->max_queue_pairs
, GFP_KERNEL
);
2378 INIT_DELAYED_WORK(&vi
->refill
, refill_work
);
2379 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2380 vi
->rq
[i
].pages
= NULL
;
2381 netif_napi_add(vi
->dev
, &vi
->rq
[i
].napi
, virtnet_poll
,
2383 netif_tx_napi_add(vi
->dev
, &vi
->sq
[i
].napi
, virtnet_poll_tx
,
2384 napi_tx
? napi_weight
: 0);
2386 sg_init_table(vi
->rq
[i
].sg
, ARRAY_SIZE(vi
->rq
[i
].sg
));
2387 ewma_pkt_len_init(&vi
->rq
[i
].mrg_avg_pkt_len
);
2388 sg_init_table(vi
->sq
[i
].sg
, ARRAY_SIZE(vi
->sq
[i
].sg
));
2399 static int init_vqs(struct virtnet_info
*vi
)
2403 /* Allocate send & receive queues */
2404 ret
= virtnet_alloc_queues(vi
);
2408 ret
= virtnet_find_vqs(vi
);
2413 virtnet_set_affinity(vi
);
2419 virtnet_free_queues(vi
);
2425 static ssize_t
mergeable_rx_buffer_size_show(struct netdev_rx_queue
*queue
,
2428 struct virtnet_info
*vi
= netdev_priv(queue
->dev
);
2429 unsigned int queue_index
= get_netdev_rx_queue_index(queue
);
2430 struct ewma_pkt_len
*avg
;
2432 BUG_ON(queue_index
>= vi
->max_queue_pairs
);
2433 avg
= &vi
->rq
[queue_index
].mrg_avg_pkt_len
;
2434 return sprintf(buf
, "%u\n",
2435 get_mergeable_buf_len(&vi
->rq
[queue_index
], avg
));
2438 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute
=
2439 __ATTR_RO(mergeable_rx_buffer_size
);
2441 static struct attribute
*virtio_net_mrg_rx_attrs
[] = {
2442 &mergeable_rx_buffer_size_attribute
.attr
,
2446 static const struct attribute_group virtio_net_mrg_rx_group
= {
2447 .name
= "virtio_net",
2448 .attrs
= virtio_net_mrg_rx_attrs
2452 static bool virtnet_fail_on_feature(struct virtio_device
*vdev
,
2454 const char *fname
, const char *dname
)
2456 if (!virtio_has_feature(vdev
, fbit
))
2459 dev_err(&vdev
->dev
, "device advertises feature %s but not %s",
2465 #define VIRTNET_FAIL_ON(vdev, fbit, dbit) \
2466 virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
2468 static bool virtnet_validate_features(struct virtio_device
*vdev
)
2470 if (!virtio_has_feature(vdev
, VIRTIO_NET_F_CTRL_VQ
) &&
2471 (VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_CTRL_RX
,
2472 "VIRTIO_NET_F_CTRL_VQ") ||
2473 VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_CTRL_VLAN
,
2474 "VIRTIO_NET_F_CTRL_VQ") ||
2475 VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_GUEST_ANNOUNCE
,
2476 "VIRTIO_NET_F_CTRL_VQ") ||
2477 VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_MQ
, "VIRTIO_NET_F_CTRL_VQ") ||
2478 VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_CTRL_MAC_ADDR
,
2479 "VIRTIO_NET_F_CTRL_VQ"))) {
2486 #define MIN_MTU ETH_MIN_MTU
2487 #define MAX_MTU ETH_MAX_MTU
2489 static int virtnet_validate(struct virtio_device
*vdev
)
2491 if (!vdev
->config
->get
) {
2492 dev_err(&vdev
->dev
, "%s failure: config access disabled\n",
2497 if (!virtnet_validate_features(vdev
))
2500 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MTU
)) {
2501 int mtu
= virtio_cread16(vdev
,
2502 offsetof(struct virtio_net_config
,
2505 __virtio_clear_bit(vdev
, VIRTIO_NET_F_MTU
);
2511 static int virtnet_probe(struct virtio_device
*vdev
)
2514 struct net_device
*dev
;
2515 struct virtnet_info
*vi
;
2516 u16 max_queue_pairs
;
2519 /* Find if host supports multiqueue virtio_net device */
2520 err
= virtio_cread_feature(vdev
, VIRTIO_NET_F_MQ
,
2521 struct virtio_net_config
,
2522 max_virtqueue_pairs
, &max_queue_pairs
);
2524 /* We need at least 2 queue's */
2525 if (err
|| max_queue_pairs
< VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN
||
2526 max_queue_pairs
> VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX
||
2527 !virtio_has_feature(vdev
, VIRTIO_NET_F_CTRL_VQ
))
2528 max_queue_pairs
= 1;
2530 /* Allocate ourselves a network device with room for our info */
2531 dev
= alloc_etherdev_mq(sizeof(struct virtnet_info
), max_queue_pairs
);
2535 /* Set up network device as normal. */
2536 dev
->priv_flags
|= IFF_UNICAST_FLT
| IFF_LIVE_ADDR_CHANGE
;
2537 dev
->netdev_ops
= &virtnet_netdev
;
2538 dev
->features
= NETIF_F_HIGHDMA
;
2540 dev
->ethtool_ops
= &virtnet_ethtool_ops
;
2541 SET_NETDEV_DEV(dev
, &vdev
->dev
);
2543 /* Do we support "hardware" checksums? */
2544 if (virtio_has_feature(vdev
, VIRTIO_NET_F_CSUM
)) {
2545 /* This opens up the world of extra features. */
2546 dev
->hw_features
|= NETIF_F_HW_CSUM
| NETIF_F_SG
;
2548 dev
->features
|= NETIF_F_HW_CSUM
| NETIF_F_SG
;
2550 if (virtio_has_feature(vdev
, VIRTIO_NET_F_GSO
)) {
2551 dev
->hw_features
|= NETIF_F_TSO
2552 | NETIF_F_TSO_ECN
| NETIF_F_TSO6
;
2554 /* Individual feature bits: what can host handle? */
2555 if (virtio_has_feature(vdev
, VIRTIO_NET_F_HOST_TSO4
))
2556 dev
->hw_features
|= NETIF_F_TSO
;
2557 if (virtio_has_feature(vdev
, VIRTIO_NET_F_HOST_TSO6
))
2558 dev
->hw_features
|= NETIF_F_TSO6
;
2559 if (virtio_has_feature(vdev
, VIRTIO_NET_F_HOST_ECN
))
2560 dev
->hw_features
|= NETIF_F_TSO_ECN
;
2562 dev
->features
|= NETIF_F_GSO_ROBUST
;
2565 dev
->features
|= dev
->hw_features
& NETIF_F_ALL_TSO
;
2566 /* (!csum && gso) case will be fixed by register_netdev() */
2568 if (virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_CSUM
))
2569 dev
->features
|= NETIF_F_RXCSUM
;
2571 dev
->vlan_features
= dev
->features
;
2573 /* MTU range: 68 - 65535 */
2574 dev
->min_mtu
= MIN_MTU
;
2575 dev
->max_mtu
= MAX_MTU
;
2577 /* Configuration may specify what MAC to use. Otherwise random. */
2578 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MAC
))
2579 virtio_cread_bytes(vdev
,
2580 offsetof(struct virtio_net_config
, mac
),
2581 dev
->dev_addr
, dev
->addr_len
);
2583 eth_hw_addr_random(dev
);
2585 /* Set up our device-specific information */
2586 vi
= netdev_priv(dev
);
2590 vi
->stats
= alloc_percpu(struct virtnet_stats
);
2592 if (vi
->stats
== NULL
)
2595 for_each_possible_cpu(i
) {
2596 struct virtnet_stats
*virtnet_stats
;
2597 virtnet_stats
= per_cpu_ptr(vi
->stats
, i
);
2598 u64_stats_init(&virtnet_stats
->tx_syncp
);
2599 u64_stats_init(&virtnet_stats
->rx_syncp
);
2602 INIT_WORK(&vi
->config_work
, virtnet_config_changed_work
);
2604 /* If we can receive ANY GSO packets, we must allocate large ones. */
2605 if (virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_TSO4
) ||
2606 virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_TSO6
) ||
2607 virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_ECN
) ||
2608 virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_UFO
))
2609 vi
->big_packets
= true;
2611 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MRG_RXBUF
))
2612 vi
->mergeable_rx_bufs
= true;
2614 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MRG_RXBUF
) ||
2615 virtio_has_feature(vdev
, VIRTIO_F_VERSION_1
))
2616 vi
->hdr_len
= sizeof(struct virtio_net_hdr_mrg_rxbuf
);
2618 vi
->hdr_len
= sizeof(struct virtio_net_hdr
);
2620 if (virtio_has_feature(vdev
, VIRTIO_F_ANY_LAYOUT
) ||
2621 virtio_has_feature(vdev
, VIRTIO_F_VERSION_1
))
2622 vi
->any_header_sg
= true;
2624 if (virtio_has_feature(vdev
, VIRTIO_NET_F_CTRL_VQ
))
2627 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MTU
)) {
2628 mtu
= virtio_cread16(vdev
,
2629 offsetof(struct virtio_net_config
,
2631 if (mtu
< dev
->min_mtu
) {
2632 /* Should never trigger: MTU was previously validated
2633 * in virtnet_validate.
2635 dev_err(&vdev
->dev
, "device MTU appears to have changed "
2636 "it is now %d < %d", mtu
, dev
->min_mtu
);
2643 /* TODO: size buffers correctly in this case. */
2644 if (dev
->mtu
> ETH_DATA_LEN
)
2645 vi
->big_packets
= true;
2648 if (vi
->any_header_sg
)
2649 dev
->needed_headroom
= vi
->hdr_len
;
2651 /* Enable multiqueue by default */
2652 if (num_online_cpus() >= max_queue_pairs
)
2653 vi
->curr_queue_pairs
= max_queue_pairs
;
2655 vi
->curr_queue_pairs
= num_online_cpus();
2656 vi
->max_queue_pairs
= max_queue_pairs
;
2658 /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
2664 if (vi
->mergeable_rx_bufs
)
2665 dev
->sysfs_rx_queue_group
= &virtio_net_mrg_rx_group
;
2667 netif_set_real_num_tx_queues(dev
, vi
->curr_queue_pairs
);
2668 netif_set_real_num_rx_queues(dev
, vi
->curr_queue_pairs
);
2670 virtnet_init_settings(dev
);
2672 err
= register_netdev(dev
);
2674 pr_debug("virtio_net: registering device failed\n");
2678 virtio_device_ready(vdev
);
2680 err
= virtnet_cpu_notif_add(vi
);
2682 pr_debug("virtio_net: registering cpu notifier failed\n");
2683 goto free_unregister_netdev
;
2686 virtnet_set_queues(vi
, vi
->curr_queue_pairs
);
2688 /* Assume link up if device can't report link status,
2689 otherwise get link status from config. */
2690 if (virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_STATUS
)) {
2691 netif_carrier_off(dev
);
2692 schedule_work(&vi
->config_work
);
2694 vi
->status
= VIRTIO_NET_S_LINK_UP
;
2695 netif_carrier_on(dev
);
2698 for (i
= 0; i
< ARRAY_SIZE(guest_offloads
); i
++)
2699 if (virtio_has_feature(vi
->vdev
, guest_offloads
[i
]))
2700 set_bit(guest_offloads
[i
], &vi
->guest_offloads
);
2702 pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
2703 dev
->name
, max_queue_pairs
);
2707 free_unregister_netdev
:
2708 vi
->vdev
->config
->reset(vdev
);
2710 unregister_netdev(dev
);
2712 cancel_delayed_work_sync(&vi
->refill
);
2713 free_receive_page_frags(vi
);
2714 virtnet_del_vqs(vi
);
2716 free_percpu(vi
->stats
);
2722 static void remove_vq_common(struct virtnet_info
*vi
)
2724 vi
->vdev
->config
->reset(vi
->vdev
);
2726 /* Free unused buffers in both send and recv, if any. */
2727 free_unused_bufs(vi
);
2729 free_receive_bufs(vi
);
2731 free_receive_page_frags(vi
);
2733 virtnet_del_vqs(vi
);
2736 static void virtnet_remove(struct virtio_device
*vdev
)
2738 struct virtnet_info
*vi
= vdev
->priv
;
2740 virtnet_cpu_notif_remove(vi
);
2742 /* Make sure no work handler is accessing the device. */
2743 flush_work(&vi
->config_work
);
2745 unregister_netdev(vi
->dev
);
2747 remove_vq_common(vi
);
2749 free_percpu(vi
->stats
);
2750 free_netdev(vi
->dev
);
2753 static __maybe_unused
int virtnet_freeze(struct virtio_device
*vdev
)
2755 struct virtnet_info
*vi
= vdev
->priv
;
2757 virtnet_cpu_notif_remove(vi
);
2758 virtnet_freeze_down(vdev
);
2759 remove_vq_common(vi
);
2764 static __maybe_unused
int virtnet_restore(struct virtio_device
*vdev
)
2766 struct virtnet_info
*vi
= vdev
->priv
;
2769 err
= virtnet_restore_up(vdev
);
2772 virtnet_set_queues(vi
, vi
->curr_queue_pairs
);
2774 err
= virtnet_cpu_notif_add(vi
);
2781 static struct virtio_device_id id_table
[] = {
2782 { VIRTIO_ID_NET
, VIRTIO_DEV_ANY_ID
},
2786 #define VIRTNET_FEATURES \
2787 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
2789 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
2790 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
2791 VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
2792 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
2793 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
2794 VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
2795 VIRTIO_NET_F_CTRL_MAC_ADDR, \
2796 VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS
2798 static unsigned int features
[] = {
2802 static unsigned int features_legacy
[] = {
2805 VIRTIO_F_ANY_LAYOUT
,
2808 static struct virtio_driver virtio_net_driver
= {
2809 .feature_table
= features
,
2810 .feature_table_size
= ARRAY_SIZE(features
),
2811 .feature_table_legacy
= features_legacy
,
2812 .feature_table_size_legacy
= ARRAY_SIZE(features_legacy
),
2813 .driver
.name
= KBUILD_MODNAME
,
2814 .driver
.owner
= THIS_MODULE
,
2815 .id_table
= id_table
,
2816 .validate
= virtnet_validate
,
2817 .probe
= virtnet_probe
,
2818 .remove
= virtnet_remove
,
2819 .config_changed
= virtnet_config_changed
,
2820 #ifdef CONFIG_PM_SLEEP
2821 .freeze
= virtnet_freeze
,
2822 .restore
= virtnet_restore
,
2826 static __init
int virtio_net_driver_init(void)
2830 ret
= cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN
, "virtio/net:online",
2832 virtnet_cpu_down_prep
);
2835 virtionet_online
= ret
;
2836 ret
= cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD
, "virtio/net:dead",
2837 NULL
, virtnet_cpu_dead
);
2841 ret
= register_virtio_driver(&virtio_net_driver
);
2846 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD
);
2848 cpuhp_remove_multi_state(virtionet_online
);
2852 module_init(virtio_net_driver_init
);
2854 static __exit
void virtio_net_driver_exit(void)
2856 unregister_virtio_driver(&virtio_net_driver
);
2857 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD
);
2858 cpuhp_remove_multi_state(virtionet_online
);
2860 module_exit(virtio_net_driver_exit
);
2862 MODULE_DEVICE_TABLE(virtio
, id_table
);
2863 MODULE_DESCRIPTION("Virtio network driver");
2864 MODULE_LICENSE("GPL");