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 /* Control VQ buffers: protected by the rtnl lock */
122 struct virtio_net_ctrl_hdr hdr
;
123 virtio_net_ctrl_ack status
;
124 struct virtio_net_ctrl_mq mq
;
131 struct virtnet_info
{
132 struct virtio_device
*vdev
;
133 struct virtqueue
*cvq
;
134 struct net_device
*dev
;
135 struct send_queue
*sq
;
136 struct receive_queue
*rq
;
139 /* Max # of queue pairs supported by the device */
142 /* # of queue pairs currently used by the driver */
143 u16 curr_queue_pairs
;
145 /* # of XDP queue pairs currently used by the driver */
148 /* I like... big packets and I cannot lie! */
151 /* Host will merge rx buffers for big packets (shake it! shake it!) */
152 bool mergeable_rx_bufs
;
154 /* Has control virtqueue */
157 /* Host can handle any s/g split between our header and packet data */
160 /* Packet virtio header size */
163 /* Active statistics */
164 struct virtnet_stats __percpu
*stats
;
166 /* Work struct for refilling if we run low on memory. */
167 struct delayed_work refill
;
169 /* Work struct for config space updates */
170 struct work_struct config_work
;
172 /* Does the affinity hint is set for virtqueues? */
173 bool affinity_hint_set
;
175 /* CPU hotplug instances for online & dead */
176 struct hlist_node node
;
177 struct hlist_node node_dead
;
179 struct control_buf
*ctrl
;
181 /* Ethtool settings */
185 unsigned long guest_offloads
;
188 struct padded_vnet_hdr
{
189 struct virtio_net_hdr_mrg_rxbuf hdr
;
191 * hdr is in a separate sg buffer, and data sg buffer shares same page
192 * with this header sg. This padding makes next sg 16 byte aligned
198 /* Converting between virtqueue no. and kernel tx/rx queue no.
199 * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
201 static int vq2txq(struct virtqueue
*vq
)
203 return (vq
->index
- 1) / 2;
206 static int txq2vq(int txq
)
211 static int vq2rxq(struct virtqueue
*vq
)
213 return vq
->index
/ 2;
216 static int rxq2vq(int rxq
)
221 static inline struct virtio_net_hdr_mrg_rxbuf
*skb_vnet_hdr(struct sk_buff
*skb
)
223 return (struct virtio_net_hdr_mrg_rxbuf
*)skb
->cb
;
227 * private is used to chain pages for big packets, put the whole
228 * most recent used list in the beginning for reuse
230 static void give_pages(struct receive_queue
*rq
, struct page
*page
)
234 /* Find end of list, sew whole thing into vi->rq.pages. */
235 for (end
= page
; end
->private; end
= (struct page
*)end
->private);
236 end
->private = (unsigned long)rq
->pages
;
240 static struct page
*get_a_page(struct receive_queue
*rq
, gfp_t gfp_mask
)
242 struct page
*p
= rq
->pages
;
245 rq
->pages
= (struct page
*)p
->private;
246 /* clear private here, it is used to chain pages */
249 p
= alloc_page(gfp_mask
);
253 static void virtqueue_napi_schedule(struct napi_struct
*napi
,
254 struct virtqueue
*vq
)
256 if (napi_schedule_prep(napi
)) {
257 virtqueue_disable_cb(vq
);
258 __napi_schedule(napi
);
262 static void virtqueue_napi_complete(struct napi_struct
*napi
,
263 struct virtqueue
*vq
, int processed
)
267 opaque
= virtqueue_enable_cb_prepare(vq
);
268 if (napi_complete_done(napi
, processed
)) {
269 if (unlikely(virtqueue_poll(vq
, opaque
)))
270 virtqueue_napi_schedule(napi
, vq
);
272 virtqueue_disable_cb(vq
);
276 static void skb_xmit_done(struct virtqueue
*vq
)
278 struct virtnet_info
*vi
= vq
->vdev
->priv
;
279 struct napi_struct
*napi
= &vi
->sq
[vq2txq(vq
)].napi
;
281 /* Suppress further interrupts. */
282 virtqueue_disable_cb(vq
);
285 virtqueue_napi_schedule(napi
, vq
);
287 /* We were probably waiting for more output buffers. */
288 netif_wake_subqueue(vi
->dev
, vq2txq(vq
));
291 #define MRG_CTX_HEADER_SHIFT 22
292 static void *mergeable_len_to_ctx(unsigned int truesize
,
293 unsigned int headroom
)
295 return (void *)(unsigned long)((headroom
<< MRG_CTX_HEADER_SHIFT
) | truesize
);
298 static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx
)
300 return (unsigned long)mrg_ctx
>> MRG_CTX_HEADER_SHIFT
;
303 static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx
)
305 return (unsigned long)mrg_ctx
& ((1 << MRG_CTX_HEADER_SHIFT
) - 1);
308 /* Called from bottom half context */
309 static struct sk_buff
*page_to_skb(struct virtnet_info
*vi
,
310 struct receive_queue
*rq
,
311 struct page
*page
, unsigned int offset
,
312 unsigned int len
, unsigned int truesize
)
315 struct virtio_net_hdr_mrg_rxbuf
*hdr
;
316 unsigned int copy
, hdr_len
, hdr_padded_len
;
319 p
= page_address(page
) + offset
;
321 /* copy small packet so we can reuse these pages for small data */
322 skb
= napi_alloc_skb(&rq
->napi
, GOOD_COPY_LEN
);
326 hdr
= skb_vnet_hdr(skb
);
328 hdr_len
= vi
->hdr_len
;
329 if (vi
->mergeable_rx_bufs
)
330 hdr_padded_len
= sizeof(*hdr
);
332 hdr_padded_len
= sizeof(struct padded_vnet_hdr
);
334 memcpy(hdr
, p
, hdr_len
);
337 offset
+= hdr_padded_len
;
341 if (copy
> skb_tailroom(skb
))
342 copy
= skb_tailroom(skb
);
343 skb_put_data(skb
, p
, copy
);
348 if (vi
->mergeable_rx_bufs
) {
350 skb_add_rx_frag(skb
, 0, page
, offset
, len
, truesize
);
357 * Verify that we can indeed put this data into a skb.
358 * This is here to handle cases when the device erroneously
359 * tries to receive more than is possible. This is usually
360 * the case of a broken device.
362 if (unlikely(len
> MAX_SKB_FRAGS
* PAGE_SIZE
)) {
363 net_dbg_ratelimited("%s: too much data\n", skb
->dev
->name
);
367 BUG_ON(offset
>= PAGE_SIZE
);
369 unsigned int frag_size
= min((unsigned)PAGE_SIZE
- offset
, len
);
370 skb_add_rx_frag(skb
, skb_shinfo(skb
)->nr_frags
, page
, offset
,
371 frag_size
, truesize
);
373 page
= (struct page
*)page
->private;
378 give_pages(rq
, page
);
383 static void virtnet_xdp_flush(struct net_device
*dev
)
385 struct virtnet_info
*vi
= netdev_priv(dev
);
386 struct send_queue
*sq
;
389 qp
= vi
->curr_queue_pairs
- vi
->xdp_queue_pairs
+ smp_processor_id();
392 virtqueue_kick(sq
->vq
);
395 static bool __virtnet_xdp_xmit(struct virtnet_info
*vi
,
396 struct xdp_buff
*xdp
)
398 struct virtio_net_hdr_mrg_rxbuf
*hdr
;
400 struct send_queue
*sq
;
405 qp
= vi
->curr_queue_pairs
- vi
->xdp_queue_pairs
+ smp_processor_id();
408 /* Free up any pending old buffers before queueing new ones. */
409 while ((xdp_sent
= virtqueue_get_buf(sq
->vq
, &len
)) != NULL
) {
410 struct page
*sent_page
= virt_to_head_page(xdp_sent
);
415 xdp
->data
-= vi
->hdr_len
;
416 /* Zero header and leave csum up to XDP layers */
418 memset(hdr
, 0, vi
->hdr_len
);
420 sg_init_one(sq
->sg
, xdp
->data
, xdp
->data_end
- xdp
->data
);
422 err
= virtqueue_add_outbuf(sq
->vq
, sq
->sg
, 1, xdp
->data
, GFP_ATOMIC
);
424 return false; /* Caller handle free/refcnt */
429 static int virtnet_xdp_xmit(struct net_device
*dev
, struct xdp_buff
*xdp
)
431 struct virtnet_info
*vi
= netdev_priv(dev
);
432 bool sent
= __virtnet_xdp_xmit(vi
, xdp
);
439 static unsigned int virtnet_get_headroom(struct virtnet_info
*vi
)
441 return vi
->xdp_queue_pairs
? VIRTIO_XDP_HEADROOM
: 0;
444 /* We copy the packet for XDP in the following cases:
446 * 1) Packet is scattered across multiple rx buffers.
447 * 2) Headroom space is insufficient.
449 * This is inefficient but it's a temporary condition that
450 * we hit right after XDP is enabled and until queue is refilled
451 * with large buffers with sufficient headroom - so it should affect
452 * at most queue size packets.
453 * Afterwards, the conditions to enable
454 * XDP should preclude the underlying device from sending packets
455 * across multiple buffers (num_buf > 1), and we make sure buffers
456 * have enough headroom.
458 static struct page
*xdp_linearize_page(struct receive_queue
*rq
,
465 struct page
*page
= alloc_page(GFP_ATOMIC
);
470 memcpy(page_address(page
) + page_off
, page_address(p
) + offset
, *len
);
478 buf
= virtqueue_get_buf(rq
->vq
, &buflen
);
482 p
= virt_to_head_page(buf
);
483 off
= buf
- page_address(p
);
485 /* guard against a misconfigured or uncooperative backend that
486 * is sending packet larger than the MTU.
488 if ((page_off
+ buflen
) > PAGE_SIZE
) {
493 memcpy(page_address(page
) + page_off
,
494 page_address(p
) + off
, buflen
);
499 /* Headroom does not contribute to packet length */
500 *len
= page_off
- VIRTIO_XDP_HEADROOM
;
503 __free_pages(page
, 0);
507 static struct sk_buff
*receive_small(struct net_device
*dev
,
508 struct virtnet_info
*vi
,
509 struct receive_queue
*rq
,
510 void *buf
, void *ctx
,
515 struct bpf_prog
*xdp_prog
;
516 unsigned int xdp_headroom
= (unsigned long)ctx
;
517 unsigned int header_offset
= VIRTNET_RX_PAD
+ xdp_headroom
;
518 unsigned int headroom
= vi
->hdr_len
+ header_offset
;
519 unsigned int buflen
= SKB_DATA_ALIGN(GOOD_PACKET_LEN
+ headroom
) +
520 SKB_DATA_ALIGN(sizeof(struct skb_shared_info
));
521 struct page
*page
= virt_to_head_page(buf
);
522 unsigned int delta
= 0;
523 struct page
*xdp_page
;
530 xdp_prog
= rcu_dereference(rq
->xdp_prog
);
532 struct virtio_net_hdr_mrg_rxbuf
*hdr
= buf
+ header_offset
;
537 if (unlikely(hdr
->hdr
.gso_type
))
540 if (unlikely(xdp_headroom
< virtnet_get_headroom(vi
))) {
541 int offset
= buf
- page_address(page
) + header_offset
;
542 unsigned int tlen
= len
+ vi
->hdr_len
;
545 xdp_headroom
= virtnet_get_headroom(vi
);
546 header_offset
= VIRTNET_RX_PAD
+ xdp_headroom
;
547 headroom
= vi
->hdr_len
+ header_offset
;
548 buflen
= SKB_DATA_ALIGN(GOOD_PACKET_LEN
+ headroom
) +
549 SKB_DATA_ALIGN(sizeof(struct skb_shared_info
));
550 xdp_page
= xdp_linearize_page(rq
, &num_buf
, page
,
551 offset
, header_offset
,
556 buf
= page_address(xdp_page
);
561 xdp
.data_hard_start
= buf
+ VIRTNET_RX_PAD
+ vi
->hdr_len
;
562 xdp
.data
= xdp
.data_hard_start
+ xdp_headroom
;
563 xdp_set_data_meta_invalid(&xdp
);
564 xdp
.data_end
= xdp
.data
+ len
;
565 orig_data
= xdp
.data
;
566 act
= bpf_prog_run_xdp(xdp_prog
, &xdp
);
570 /* Recalculate length in case bpf program changed it */
571 delta
= orig_data
- xdp
.data
;
574 sent
= __virtnet_xdp_xmit(vi
, &xdp
);
575 if (unlikely(!sent
)) {
576 trace_xdp_exception(vi
->dev
, xdp_prog
, act
);
583 err
= xdp_do_redirect(dev
, &xdp
, xdp_prog
);
590 bpf_warn_invalid_xdp_action(act
);
592 trace_xdp_exception(vi
->dev
, xdp_prog
, act
);
599 skb
= build_skb(buf
, buflen
);
604 skb_reserve(skb
, headroom
- delta
);
605 skb_put(skb
, len
+ delta
);
607 buf
+= header_offset
;
608 memcpy(skb_vnet_hdr(skb
), buf
, vi
->hdr_len
);
609 } /* keep zeroed vnet hdr since packet was changed by bpf */
616 dev
->stats
.rx_dropped
++;
622 static struct sk_buff
*receive_big(struct net_device
*dev
,
623 struct virtnet_info
*vi
,
624 struct receive_queue
*rq
,
628 struct page
*page
= buf
;
629 struct sk_buff
*skb
= page_to_skb(vi
, rq
, page
, 0, len
, PAGE_SIZE
);
637 dev
->stats
.rx_dropped
++;
638 give_pages(rq
, page
);
642 static struct sk_buff
*receive_mergeable(struct net_device
*dev
,
643 struct virtnet_info
*vi
,
644 struct receive_queue
*rq
,
650 struct virtio_net_hdr_mrg_rxbuf
*hdr
= buf
;
651 u16 num_buf
= virtio16_to_cpu(vi
->vdev
, hdr
->num_buffers
);
652 struct page
*page
= virt_to_head_page(buf
);
653 int offset
= buf
- page_address(page
);
654 struct sk_buff
*head_skb
, *curr_skb
;
655 struct bpf_prog
*xdp_prog
;
656 unsigned int truesize
;
657 unsigned int headroom
= mergeable_ctx_to_headroom(ctx
);
664 xdp_prog
= rcu_dereference(rq
->xdp_prog
);
666 struct page
*xdp_page
;
671 /* Transient failure which in theory could occur if
672 * in-flight packets from before XDP was enabled reach
673 * the receive path after XDP is loaded.
675 if (unlikely(hdr
->hdr
.gso_type
))
678 /* This happens when rx buffer size is underestimated */
679 if (unlikely(num_buf
> 1 ||
680 headroom
< virtnet_get_headroom(vi
))) {
681 /* linearize data for XDP */
682 xdp_page
= xdp_linearize_page(rq
, &num_buf
,
688 offset
= VIRTIO_XDP_HEADROOM
;
693 /* Allow consuming headroom but reserve enough space to push
694 * the descriptor on if we get an XDP_TX return code.
696 data
= page_address(xdp_page
) + offset
;
697 xdp
.data_hard_start
= data
- VIRTIO_XDP_HEADROOM
+ vi
->hdr_len
;
698 xdp
.data
= data
+ vi
->hdr_len
;
699 xdp_set_data_meta_invalid(&xdp
);
700 xdp
.data_end
= xdp
.data
+ (len
- vi
->hdr_len
);
701 act
= bpf_prog_run_xdp(xdp_prog
, &xdp
);
704 ewma_pkt_len_add(&rq
->mrg_avg_pkt_len
, len
);
708 /* recalculate offset to account for any header
709 * adjustments. Note other cases do not build an
710 * skb and avoid using offset
713 page_address(xdp_page
) - vi
->hdr_len
;
715 /* We can only create skb based on xdp_page. */
716 if (unlikely(xdp_page
!= page
)) {
719 head_skb
= page_to_skb(vi
, rq
, xdp_page
,
720 offset
, len
, PAGE_SIZE
);
725 sent
= __virtnet_xdp_xmit(vi
, &xdp
);
726 if (unlikely(!sent
)) {
727 trace_xdp_exception(vi
->dev
, xdp_prog
, act
);
728 if (unlikely(xdp_page
!= page
))
733 if (unlikely(xdp_page
!= page
))
738 err
= xdp_do_redirect(dev
, &xdp
, xdp_prog
);
744 bpf_warn_invalid_xdp_action(act
);
746 trace_xdp_exception(vi
->dev
, xdp_prog
, act
);
748 if (unlikely(xdp_page
!= page
))
749 __free_pages(xdp_page
, 0);
755 truesize
= mergeable_ctx_to_truesize(ctx
);
756 if (unlikely(len
> truesize
)) {
757 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
758 dev
->name
, len
, (unsigned long)ctx
);
759 dev
->stats
.rx_length_errors
++;
763 head_skb
= page_to_skb(vi
, rq
, page
, offset
, len
, truesize
);
766 if (unlikely(!curr_skb
))
771 buf
= virtqueue_get_buf_ctx(rq
->vq
, &len
, &ctx
);
772 if (unlikely(!buf
)) {
773 pr_debug("%s: rx error: %d buffers out of %d missing\n",
775 virtio16_to_cpu(vi
->vdev
,
777 dev
->stats
.rx_length_errors
++;
781 page
= virt_to_head_page(buf
);
783 truesize
= mergeable_ctx_to_truesize(ctx
);
784 if (unlikely(len
> truesize
)) {
785 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
786 dev
->name
, len
, (unsigned long)ctx
);
787 dev
->stats
.rx_length_errors
++;
791 num_skb_frags
= skb_shinfo(curr_skb
)->nr_frags
;
792 if (unlikely(num_skb_frags
== MAX_SKB_FRAGS
)) {
793 struct sk_buff
*nskb
= alloc_skb(0, GFP_ATOMIC
);
797 if (curr_skb
== head_skb
)
798 skb_shinfo(curr_skb
)->frag_list
= nskb
;
800 curr_skb
->next
= nskb
;
802 head_skb
->truesize
+= nskb
->truesize
;
805 if (curr_skb
!= head_skb
) {
806 head_skb
->data_len
+= len
;
807 head_skb
->len
+= len
;
808 head_skb
->truesize
+= truesize
;
810 offset
= buf
- page_address(page
);
811 if (skb_can_coalesce(curr_skb
, num_skb_frags
, page
, offset
)) {
813 skb_coalesce_rx_frag(curr_skb
, num_skb_frags
- 1,
816 skb_add_rx_frag(curr_skb
, num_skb_frags
, page
,
817 offset
, len
, truesize
);
821 ewma_pkt_len_add(&rq
->mrg_avg_pkt_len
, head_skb
->len
);
828 while (num_buf
-- > 1) {
829 buf
= virtqueue_get_buf(rq
->vq
, &len
);
830 if (unlikely(!buf
)) {
831 pr_debug("%s: rx error: %d buffers missing\n",
833 dev
->stats
.rx_length_errors
++;
836 page
= virt_to_head_page(buf
);
840 dev
->stats
.rx_dropped
++;
841 dev_kfree_skb(head_skb
);
846 static int receive_buf(struct virtnet_info
*vi
, struct receive_queue
*rq
,
847 void *buf
, unsigned int len
, void **ctx
, bool *xdp_xmit
)
849 struct net_device
*dev
= vi
->dev
;
851 struct virtio_net_hdr_mrg_rxbuf
*hdr
;
854 if (unlikely(len
< vi
->hdr_len
+ ETH_HLEN
)) {
855 pr_debug("%s: short packet %i\n", dev
->name
, len
);
856 dev
->stats
.rx_length_errors
++;
857 if (vi
->mergeable_rx_bufs
) {
858 put_page(virt_to_head_page(buf
));
859 } else if (vi
->big_packets
) {
862 put_page(virt_to_head_page(buf
));
867 if (vi
->mergeable_rx_bufs
)
868 skb
= receive_mergeable(dev
, vi
, rq
, buf
, ctx
, len
, xdp_xmit
);
869 else if (vi
->big_packets
)
870 skb
= receive_big(dev
, vi
, rq
, buf
, len
);
872 skb
= receive_small(dev
, vi
, rq
, buf
, ctx
, len
, xdp_xmit
);
877 hdr
= skb_vnet_hdr(skb
);
881 if (hdr
->hdr
.flags
& VIRTIO_NET_HDR_F_DATA_VALID
)
882 skb
->ip_summed
= CHECKSUM_UNNECESSARY
;
884 if (virtio_net_hdr_to_skb(skb
, &hdr
->hdr
,
885 virtio_is_little_endian(vi
->vdev
))) {
886 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
887 dev
->name
, hdr
->hdr
.gso_type
,
892 skb
->protocol
= eth_type_trans(skb
, dev
);
893 pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
894 ntohs(skb
->protocol
), skb
->len
, skb
->pkt_type
);
896 napi_gro_receive(&rq
->napi
, skb
);
900 dev
->stats
.rx_frame_errors
++;
905 /* Unlike mergeable buffers, all buffers are allocated to the
906 * same size, except for the headroom. For this reason we do
907 * not need to use mergeable_len_to_ctx here - it is enough
908 * to store the headroom as the context ignoring the truesize.
910 static int add_recvbuf_small(struct virtnet_info
*vi
, struct receive_queue
*rq
,
913 struct page_frag
*alloc_frag
= &rq
->alloc_frag
;
915 unsigned int xdp_headroom
= virtnet_get_headroom(vi
);
916 void *ctx
= (void *)(unsigned long)xdp_headroom
;
917 int len
= vi
->hdr_len
+ VIRTNET_RX_PAD
+ GOOD_PACKET_LEN
+ xdp_headroom
;
920 len
= SKB_DATA_ALIGN(len
) +
921 SKB_DATA_ALIGN(sizeof(struct skb_shared_info
));
922 if (unlikely(!skb_page_frag_refill(len
, alloc_frag
, gfp
)))
925 buf
= (char *)page_address(alloc_frag
->page
) + alloc_frag
->offset
;
926 get_page(alloc_frag
->page
);
927 alloc_frag
->offset
+= len
;
928 sg_init_one(rq
->sg
, buf
+ VIRTNET_RX_PAD
+ xdp_headroom
,
929 vi
->hdr_len
+ GOOD_PACKET_LEN
);
930 err
= virtqueue_add_inbuf_ctx(rq
->vq
, rq
->sg
, 1, buf
, ctx
, gfp
);
932 put_page(virt_to_head_page(buf
));
936 static int add_recvbuf_big(struct virtnet_info
*vi
, struct receive_queue
*rq
,
939 struct page
*first
, *list
= NULL
;
943 sg_init_table(rq
->sg
, MAX_SKB_FRAGS
+ 2);
945 /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
946 for (i
= MAX_SKB_FRAGS
+ 1; i
> 1; --i
) {
947 first
= get_a_page(rq
, gfp
);
950 give_pages(rq
, list
);
953 sg_set_buf(&rq
->sg
[i
], page_address(first
), PAGE_SIZE
);
955 /* chain new page in list head to match sg */
956 first
->private = (unsigned long)list
;
960 first
= get_a_page(rq
, gfp
);
962 give_pages(rq
, list
);
965 p
= page_address(first
);
967 /* rq->sg[0], rq->sg[1] share the same page */
968 /* a separated rq->sg[0] for header - required in case !any_header_sg */
969 sg_set_buf(&rq
->sg
[0], p
, vi
->hdr_len
);
971 /* rq->sg[1] for data packet, from offset */
972 offset
= sizeof(struct padded_vnet_hdr
);
973 sg_set_buf(&rq
->sg
[1], p
+ offset
, PAGE_SIZE
- offset
);
975 /* chain first in list head */
976 first
->private = (unsigned long)list
;
977 err
= virtqueue_add_inbuf(rq
->vq
, rq
->sg
, MAX_SKB_FRAGS
+ 2,
980 give_pages(rq
, first
);
985 static unsigned int get_mergeable_buf_len(struct receive_queue
*rq
,
986 struct ewma_pkt_len
*avg_pkt_len
)
988 const size_t hdr_len
= sizeof(struct virtio_net_hdr_mrg_rxbuf
);
991 len
= hdr_len
+ clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len
),
992 rq
->min_buf_len
, PAGE_SIZE
- hdr_len
);
993 return ALIGN(len
, L1_CACHE_BYTES
);
996 static int add_recvbuf_mergeable(struct virtnet_info
*vi
,
997 struct receive_queue
*rq
, gfp_t gfp
)
999 struct page_frag
*alloc_frag
= &rq
->alloc_frag
;
1000 unsigned int headroom
= virtnet_get_headroom(vi
);
1004 unsigned int len
, hole
;
1006 len
= get_mergeable_buf_len(rq
, &rq
->mrg_avg_pkt_len
);
1007 if (unlikely(!skb_page_frag_refill(len
+ headroom
, alloc_frag
, gfp
)))
1010 buf
= (char *)page_address(alloc_frag
->page
) + alloc_frag
->offset
;
1011 buf
+= headroom
; /* advance address leaving hole at front of pkt */
1012 get_page(alloc_frag
->page
);
1013 alloc_frag
->offset
+= len
+ headroom
;
1014 hole
= alloc_frag
->size
- alloc_frag
->offset
;
1015 if (hole
< len
+ headroom
) {
1016 /* To avoid internal fragmentation, if there is very likely not
1017 * enough space for another buffer, add the remaining space to
1018 * the current buffer.
1021 alloc_frag
->offset
+= hole
;
1024 sg_init_one(rq
->sg
, buf
, len
);
1025 ctx
= mergeable_len_to_ctx(len
, headroom
);
1026 err
= virtqueue_add_inbuf_ctx(rq
->vq
, rq
->sg
, 1, buf
, ctx
, gfp
);
1028 put_page(virt_to_head_page(buf
));
1034 * Returns false if we couldn't fill entirely (OOM).
1036 * Normally run in the receive path, but can also be run from ndo_open
1037 * before we're receiving packets, or from refill_work which is
1038 * careful to disable receiving (using napi_disable).
1040 static bool try_fill_recv(struct virtnet_info
*vi
, struct receive_queue
*rq
,
1047 if (vi
->mergeable_rx_bufs
)
1048 err
= add_recvbuf_mergeable(vi
, rq
, gfp
);
1049 else if (vi
->big_packets
)
1050 err
= add_recvbuf_big(vi
, rq
, gfp
);
1052 err
= add_recvbuf_small(vi
, rq
, gfp
);
1054 oom
= err
== -ENOMEM
;
1057 } while (rq
->vq
->num_free
);
1058 virtqueue_kick(rq
->vq
);
1062 static void skb_recv_done(struct virtqueue
*rvq
)
1064 struct virtnet_info
*vi
= rvq
->vdev
->priv
;
1065 struct receive_queue
*rq
= &vi
->rq
[vq2rxq(rvq
)];
1067 virtqueue_napi_schedule(&rq
->napi
, rvq
);
1070 static void virtnet_napi_enable(struct virtqueue
*vq
, struct napi_struct
*napi
)
1074 /* If all buffers were filled by other side before we napi_enabled, we
1075 * won't get another interrupt, so process any outstanding packets now.
1076 * Call local_bh_enable after to trigger softIRQ processing.
1079 virtqueue_napi_schedule(napi
, vq
);
1083 static void virtnet_napi_tx_enable(struct virtnet_info
*vi
,
1084 struct virtqueue
*vq
,
1085 struct napi_struct
*napi
)
1090 /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
1091 * enable the feature if this is likely affine with the transmit path.
1093 if (!vi
->affinity_hint_set
) {
1098 return virtnet_napi_enable(vq
, napi
);
1101 static void virtnet_napi_tx_disable(struct napi_struct
*napi
)
1107 static void refill_work(struct work_struct
*work
)
1109 struct virtnet_info
*vi
=
1110 container_of(work
, struct virtnet_info
, refill
.work
);
1114 for (i
= 0; i
< vi
->curr_queue_pairs
; i
++) {
1115 struct receive_queue
*rq
= &vi
->rq
[i
];
1117 napi_disable(&rq
->napi
);
1118 still_empty
= !try_fill_recv(vi
, rq
, GFP_KERNEL
);
1119 virtnet_napi_enable(rq
->vq
, &rq
->napi
);
1121 /* In theory, this can happen: if we don't get any buffers in
1122 * we will *never* try to fill again.
1125 schedule_delayed_work(&vi
->refill
, HZ
/2);
1129 static int virtnet_receive(struct receive_queue
*rq
, int budget
, bool *xdp_xmit
)
1131 struct virtnet_info
*vi
= rq
->vq
->vdev
->priv
;
1132 unsigned int len
, received
= 0, bytes
= 0;
1134 struct virtnet_stats
*stats
= this_cpu_ptr(vi
->stats
);
1136 if (!vi
->big_packets
|| vi
->mergeable_rx_bufs
) {
1139 while (received
< budget
&&
1140 (buf
= virtqueue_get_buf_ctx(rq
->vq
, &len
, &ctx
))) {
1141 bytes
+= receive_buf(vi
, rq
, buf
, len
, ctx
, xdp_xmit
);
1145 while (received
< budget
&&
1146 (buf
= virtqueue_get_buf(rq
->vq
, &len
)) != NULL
) {
1147 bytes
+= receive_buf(vi
, rq
, buf
, len
, NULL
, xdp_xmit
);
1152 if (rq
->vq
->num_free
> virtqueue_get_vring_size(rq
->vq
) / 2) {
1153 if (!try_fill_recv(vi
, rq
, GFP_ATOMIC
))
1154 schedule_delayed_work(&vi
->refill
, 0);
1157 u64_stats_update_begin(&stats
->rx_syncp
);
1158 stats
->rx_bytes
+= bytes
;
1159 stats
->rx_packets
+= received
;
1160 u64_stats_update_end(&stats
->rx_syncp
);
1165 static void free_old_xmit_skbs(struct send_queue
*sq
)
1167 struct sk_buff
*skb
;
1169 struct virtnet_info
*vi
= sq
->vq
->vdev
->priv
;
1170 struct virtnet_stats
*stats
= this_cpu_ptr(vi
->stats
);
1171 unsigned int packets
= 0;
1172 unsigned int bytes
= 0;
1174 while ((skb
= virtqueue_get_buf(sq
->vq
, &len
)) != NULL
) {
1175 pr_debug("Sent skb %p\n", skb
);
1180 dev_consume_skb_any(skb
);
1183 /* Avoid overhead when no packets have been processed
1184 * happens when called speculatively from start_xmit.
1189 u64_stats_update_begin(&stats
->tx_syncp
);
1190 stats
->tx_bytes
+= bytes
;
1191 stats
->tx_packets
+= packets
;
1192 u64_stats_update_end(&stats
->tx_syncp
);
1195 static void virtnet_poll_cleantx(struct receive_queue
*rq
)
1197 struct virtnet_info
*vi
= rq
->vq
->vdev
->priv
;
1198 unsigned int index
= vq2rxq(rq
->vq
);
1199 struct send_queue
*sq
= &vi
->sq
[index
];
1200 struct netdev_queue
*txq
= netdev_get_tx_queue(vi
->dev
, index
);
1202 if (!sq
->napi
.weight
)
1205 if (__netif_tx_trylock(txq
)) {
1206 free_old_xmit_skbs(sq
);
1207 __netif_tx_unlock(txq
);
1210 if (sq
->vq
->num_free
>= 2 + MAX_SKB_FRAGS
)
1211 netif_tx_wake_queue(txq
);
1214 static int virtnet_poll(struct napi_struct
*napi
, int budget
)
1216 struct receive_queue
*rq
=
1217 container_of(napi
, struct receive_queue
, napi
);
1218 struct virtnet_info
*vi
= rq
->vq
->vdev
->priv
;
1219 struct send_queue
*sq
;
1220 unsigned int received
, qp
;
1221 bool xdp_xmit
= false;
1223 virtnet_poll_cleantx(rq
);
1225 received
= virtnet_receive(rq
, budget
, &xdp_xmit
);
1227 /* Out of packets? */
1228 if (received
< budget
)
1229 virtqueue_napi_complete(napi
, rq
->vq
, received
);
1232 qp
= vi
->curr_queue_pairs
- vi
->xdp_queue_pairs
+
1235 virtqueue_kick(sq
->vq
);
1242 static int virtnet_open(struct net_device
*dev
)
1244 struct virtnet_info
*vi
= netdev_priv(dev
);
1247 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1248 if (i
< vi
->curr_queue_pairs
)
1249 /* Make sure we have some buffers: if oom use wq. */
1250 if (!try_fill_recv(vi
, &vi
->rq
[i
], GFP_KERNEL
))
1251 schedule_delayed_work(&vi
->refill
, 0);
1252 virtnet_napi_enable(vi
->rq
[i
].vq
, &vi
->rq
[i
].napi
);
1253 virtnet_napi_tx_enable(vi
, vi
->sq
[i
].vq
, &vi
->sq
[i
].napi
);
1259 static int virtnet_poll_tx(struct napi_struct
*napi
, int budget
)
1261 struct send_queue
*sq
= container_of(napi
, struct send_queue
, napi
);
1262 struct virtnet_info
*vi
= sq
->vq
->vdev
->priv
;
1263 struct netdev_queue
*txq
= netdev_get_tx_queue(vi
->dev
, vq2txq(sq
->vq
));
1265 __netif_tx_lock(txq
, raw_smp_processor_id());
1266 free_old_xmit_skbs(sq
);
1267 __netif_tx_unlock(txq
);
1269 virtqueue_napi_complete(napi
, sq
->vq
, 0);
1271 if (sq
->vq
->num_free
>= 2 + MAX_SKB_FRAGS
)
1272 netif_tx_wake_queue(txq
);
1277 static int xmit_skb(struct send_queue
*sq
, struct sk_buff
*skb
)
1279 struct virtio_net_hdr_mrg_rxbuf
*hdr
;
1280 const unsigned char *dest
= ((struct ethhdr
*)skb
->data
)->h_dest
;
1281 struct virtnet_info
*vi
= sq
->vq
->vdev
->priv
;
1283 unsigned hdr_len
= vi
->hdr_len
;
1286 pr_debug("%s: xmit %p %pM\n", vi
->dev
->name
, skb
, dest
);
1288 can_push
= vi
->any_header_sg
&&
1289 !((unsigned long)skb
->data
& (__alignof__(*hdr
) - 1)) &&
1290 !skb_header_cloned(skb
) && skb_headroom(skb
) >= hdr_len
;
1291 /* Even if we can, don't push here yet as this would skew
1292 * csum_start offset below. */
1294 hdr
= (struct virtio_net_hdr_mrg_rxbuf
*)(skb
->data
- hdr_len
);
1296 hdr
= skb_vnet_hdr(skb
);
1298 if (virtio_net_hdr_from_skb(skb
, &hdr
->hdr
,
1299 virtio_is_little_endian(vi
->vdev
), false,
1303 if (vi
->mergeable_rx_bufs
)
1304 hdr
->num_buffers
= 0;
1306 sg_init_table(sq
->sg
, skb_shinfo(skb
)->nr_frags
+ (can_push
? 1 : 2));
1308 __skb_push(skb
, hdr_len
);
1309 num_sg
= skb_to_sgvec(skb
, sq
->sg
, 0, skb
->len
);
1310 if (unlikely(num_sg
< 0))
1312 /* Pull header back to avoid skew in tx bytes calculations. */
1313 __skb_pull(skb
, hdr_len
);
1315 sg_set_buf(sq
->sg
, hdr
, hdr_len
);
1316 num_sg
= skb_to_sgvec(skb
, sq
->sg
+ 1, 0, skb
->len
);
1317 if (unlikely(num_sg
< 0))
1321 return virtqueue_add_outbuf(sq
->vq
, sq
->sg
, num_sg
, skb
, GFP_ATOMIC
);
1324 static netdev_tx_t
start_xmit(struct sk_buff
*skb
, struct net_device
*dev
)
1326 struct virtnet_info
*vi
= netdev_priv(dev
);
1327 int qnum
= skb_get_queue_mapping(skb
);
1328 struct send_queue
*sq
= &vi
->sq
[qnum
];
1330 struct netdev_queue
*txq
= netdev_get_tx_queue(dev
, qnum
);
1331 bool kick
= !skb
->xmit_more
;
1332 bool use_napi
= sq
->napi
.weight
;
1334 /* Free up any pending old buffers before queueing new ones. */
1335 free_old_xmit_skbs(sq
);
1337 if (use_napi
&& kick
)
1338 virtqueue_enable_cb_delayed(sq
->vq
);
1340 /* timestamp packet in software */
1341 skb_tx_timestamp(skb
);
1343 /* Try to transmit */
1344 err
= xmit_skb(sq
, skb
);
1346 /* This should not happen! */
1347 if (unlikely(err
)) {
1348 dev
->stats
.tx_fifo_errors
++;
1349 if (net_ratelimit())
1351 "Unexpected TXQ (%d) queue failure: %d\n", qnum
, err
);
1352 dev
->stats
.tx_dropped
++;
1353 dev_kfree_skb_any(skb
);
1354 return NETDEV_TX_OK
;
1357 /* Don't wait up for transmitted skbs to be freed. */
1363 /* If running out of space, stop queue to avoid getting packets that we
1364 * are then unable to transmit.
1365 * An alternative would be to force queuing layer to requeue the skb by
1366 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1367 * returned in a normal path of operation: it means that driver is not
1368 * maintaining the TX queue stop/start state properly, and causes
1369 * the stack to do a non-trivial amount of useless work.
1370 * Since most packets only take 1 or 2 ring slots, stopping the queue
1371 * early means 16 slots are typically wasted.
1373 if (sq
->vq
->num_free
< 2+MAX_SKB_FRAGS
) {
1374 netif_stop_subqueue(dev
, qnum
);
1376 unlikely(!virtqueue_enable_cb_delayed(sq
->vq
))) {
1377 /* More just got used, free them then recheck. */
1378 free_old_xmit_skbs(sq
);
1379 if (sq
->vq
->num_free
>= 2+MAX_SKB_FRAGS
) {
1380 netif_start_subqueue(dev
, qnum
);
1381 virtqueue_disable_cb(sq
->vq
);
1386 if (kick
|| netif_xmit_stopped(txq
))
1387 virtqueue_kick(sq
->vq
);
1389 return NETDEV_TX_OK
;
1393 * Send command via the control virtqueue and check status. Commands
1394 * supported by the hypervisor, as indicated by feature bits, should
1395 * never fail unless improperly formatted.
1397 static bool virtnet_send_command(struct virtnet_info
*vi
, u8
class, u8 cmd
,
1398 struct scatterlist
*out
)
1400 struct scatterlist
*sgs
[4], hdr
, stat
;
1401 unsigned out_num
= 0, tmp
;
1403 /* Caller should know better */
1404 BUG_ON(!virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_CTRL_VQ
));
1406 vi
->ctrl
->status
= ~0;
1407 vi
->ctrl
->hdr
.class = class;
1408 vi
->ctrl
->hdr
.cmd
= cmd
;
1410 sg_init_one(&hdr
, &vi
->ctrl
->hdr
, sizeof(vi
->ctrl
->hdr
));
1411 sgs
[out_num
++] = &hdr
;
1414 sgs
[out_num
++] = out
;
1416 /* Add return status. */
1417 sg_init_one(&stat
, &vi
->ctrl
->status
, sizeof(vi
->ctrl
->status
));
1418 sgs
[out_num
] = &stat
;
1420 BUG_ON(out_num
+ 1 > ARRAY_SIZE(sgs
));
1421 virtqueue_add_sgs(vi
->cvq
, sgs
, out_num
, 1, vi
, GFP_ATOMIC
);
1423 if (unlikely(!virtqueue_kick(vi
->cvq
)))
1424 return vi
->ctrl
->status
== VIRTIO_NET_OK
;
1426 /* Spin for a response, the kick causes an ioport write, trapping
1427 * into the hypervisor, so the request should be handled immediately.
1429 while (!virtqueue_get_buf(vi
->cvq
, &tmp
) &&
1430 !virtqueue_is_broken(vi
->cvq
))
1433 return vi
->ctrl
->status
== VIRTIO_NET_OK
;
1436 static int virtnet_set_mac_address(struct net_device
*dev
, void *p
)
1438 struct virtnet_info
*vi
= netdev_priv(dev
);
1439 struct virtio_device
*vdev
= vi
->vdev
;
1441 struct sockaddr
*addr
;
1442 struct scatterlist sg
;
1444 addr
= kmemdup(p
, sizeof(*addr
), GFP_KERNEL
);
1448 ret
= eth_prepare_mac_addr_change(dev
, addr
);
1452 if (virtio_has_feature(vdev
, VIRTIO_NET_F_CTRL_MAC_ADDR
)) {
1453 sg_init_one(&sg
, addr
->sa_data
, dev
->addr_len
);
1454 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_MAC
,
1455 VIRTIO_NET_CTRL_MAC_ADDR_SET
, &sg
)) {
1456 dev_warn(&vdev
->dev
,
1457 "Failed to set mac address by vq command.\n");
1461 } else if (virtio_has_feature(vdev
, VIRTIO_NET_F_MAC
) &&
1462 !virtio_has_feature(vdev
, VIRTIO_F_VERSION_1
)) {
1465 /* Naturally, this has an atomicity problem. */
1466 for (i
= 0; i
< dev
->addr_len
; i
++)
1467 virtio_cwrite8(vdev
,
1468 offsetof(struct virtio_net_config
, mac
) +
1469 i
, addr
->sa_data
[i
]);
1472 eth_commit_mac_addr_change(dev
, p
);
1480 static void virtnet_stats(struct net_device
*dev
,
1481 struct rtnl_link_stats64
*tot
)
1483 struct virtnet_info
*vi
= netdev_priv(dev
);
1487 for_each_possible_cpu(cpu
) {
1488 struct virtnet_stats
*stats
= per_cpu_ptr(vi
->stats
, cpu
);
1489 u64 tpackets
, tbytes
, rpackets
, rbytes
;
1492 start
= u64_stats_fetch_begin_irq(&stats
->tx_syncp
);
1493 tpackets
= stats
->tx_packets
;
1494 tbytes
= stats
->tx_bytes
;
1495 } while (u64_stats_fetch_retry_irq(&stats
->tx_syncp
, start
));
1498 start
= u64_stats_fetch_begin_irq(&stats
->rx_syncp
);
1499 rpackets
= stats
->rx_packets
;
1500 rbytes
= stats
->rx_bytes
;
1501 } while (u64_stats_fetch_retry_irq(&stats
->rx_syncp
, start
));
1503 tot
->rx_packets
+= rpackets
;
1504 tot
->tx_packets
+= tpackets
;
1505 tot
->rx_bytes
+= rbytes
;
1506 tot
->tx_bytes
+= tbytes
;
1509 tot
->tx_dropped
= dev
->stats
.tx_dropped
;
1510 tot
->tx_fifo_errors
= dev
->stats
.tx_fifo_errors
;
1511 tot
->rx_dropped
= dev
->stats
.rx_dropped
;
1512 tot
->rx_length_errors
= dev
->stats
.rx_length_errors
;
1513 tot
->rx_frame_errors
= dev
->stats
.rx_frame_errors
;
1516 #ifdef CONFIG_NET_POLL_CONTROLLER
1517 static void virtnet_netpoll(struct net_device
*dev
)
1519 struct virtnet_info
*vi
= netdev_priv(dev
);
1522 for (i
= 0; i
< vi
->curr_queue_pairs
; i
++)
1523 napi_schedule(&vi
->rq
[i
].napi
);
1527 static void virtnet_ack_link_announce(struct virtnet_info
*vi
)
1530 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_ANNOUNCE
,
1531 VIRTIO_NET_CTRL_ANNOUNCE_ACK
, NULL
))
1532 dev_warn(&vi
->dev
->dev
, "Failed to ack link announce.\n");
1536 static int _virtnet_set_queues(struct virtnet_info
*vi
, u16 queue_pairs
)
1538 struct scatterlist sg
;
1539 struct net_device
*dev
= vi
->dev
;
1541 if (!vi
->has_cvq
|| !virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_MQ
))
1544 vi
->ctrl
->mq
.virtqueue_pairs
= cpu_to_virtio16(vi
->vdev
, queue_pairs
);
1545 sg_init_one(&sg
, &vi
->ctrl
->mq
, sizeof(vi
->ctrl
->mq
));
1547 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_MQ
,
1548 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET
, &sg
)) {
1549 dev_warn(&dev
->dev
, "Fail to set num of queue pairs to %d\n",
1553 vi
->curr_queue_pairs
= queue_pairs
;
1554 /* virtnet_open() will refill when device is going to up. */
1555 if (dev
->flags
& IFF_UP
)
1556 schedule_delayed_work(&vi
->refill
, 0);
1562 static int virtnet_set_queues(struct virtnet_info
*vi
, u16 queue_pairs
)
1567 err
= _virtnet_set_queues(vi
, queue_pairs
);
1572 static int virtnet_close(struct net_device
*dev
)
1574 struct virtnet_info
*vi
= netdev_priv(dev
);
1577 /* Make sure refill_work doesn't re-enable napi! */
1578 cancel_delayed_work_sync(&vi
->refill
);
1580 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1581 napi_disable(&vi
->rq
[i
].napi
);
1582 virtnet_napi_tx_disable(&vi
->sq
[i
].napi
);
1588 static void virtnet_set_rx_mode(struct net_device
*dev
)
1590 struct virtnet_info
*vi
= netdev_priv(dev
);
1591 struct scatterlist sg
[2];
1592 struct virtio_net_ctrl_mac
*mac_data
;
1593 struct netdev_hw_addr
*ha
;
1599 /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1600 if (!virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_CTRL_RX
))
1603 vi
->ctrl
->promisc
= ((dev
->flags
& IFF_PROMISC
) != 0);
1604 vi
->ctrl
->allmulti
= ((dev
->flags
& IFF_ALLMULTI
) != 0);
1606 sg_init_one(sg
, &vi
->ctrl
->promisc
, sizeof(vi
->ctrl
->promisc
));
1608 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_RX
,
1609 VIRTIO_NET_CTRL_RX_PROMISC
, sg
))
1610 dev_warn(&dev
->dev
, "Failed to %sable promisc mode.\n",
1611 vi
->ctrl
->promisc
? "en" : "dis");
1613 sg_init_one(sg
, &vi
->ctrl
->allmulti
, sizeof(vi
->ctrl
->allmulti
));
1615 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_RX
,
1616 VIRTIO_NET_CTRL_RX_ALLMULTI
, sg
))
1617 dev_warn(&dev
->dev
, "Failed to %sable allmulti mode.\n",
1618 vi
->ctrl
->allmulti
? "en" : "dis");
1620 uc_count
= netdev_uc_count(dev
);
1621 mc_count
= netdev_mc_count(dev
);
1622 /* MAC filter - use one buffer for both lists */
1623 buf
= kzalloc(((uc_count
+ mc_count
) * ETH_ALEN
) +
1624 (2 * sizeof(mac_data
->entries
)), GFP_ATOMIC
);
1629 sg_init_table(sg
, 2);
1631 /* Store the unicast list and count in the front of the buffer */
1632 mac_data
->entries
= cpu_to_virtio32(vi
->vdev
, uc_count
);
1634 netdev_for_each_uc_addr(ha
, dev
)
1635 memcpy(&mac_data
->macs
[i
++][0], ha
->addr
, ETH_ALEN
);
1637 sg_set_buf(&sg
[0], mac_data
,
1638 sizeof(mac_data
->entries
) + (uc_count
* ETH_ALEN
));
1640 /* multicast list and count fill the end */
1641 mac_data
= (void *)&mac_data
->macs
[uc_count
][0];
1643 mac_data
->entries
= cpu_to_virtio32(vi
->vdev
, mc_count
);
1645 netdev_for_each_mc_addr(ha
, dev
)
1646 memcpy(&mac_data
->macs
[i
++][0], ha
->addr
, ETH_ALEN
);
1648 sg_set_buf(&sg
[1], mac_data
,
1649 sizeof(mac_data
->entries
) + (mc_count
* ETH_ALEN
));
1651 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_MAC
,
1652 VIRTIO_NET_CTRL_MAC_TABLE_SET
, sg
))
1653 dev_warn(&dev
->dev
, "Failed to set MAC filter table.\n");
1658 static int virtnet_vlan_rx_add_vid(struct net_device
*dev
,
1659 __be16 proto
, u16 vid
)
1661 struct virtnet_info
*vi
= netdev_priv(dev
);
1662 struct scatterlist sg
;
1664 vi
->ctrl
->vid
= cpu_to_virtio16(vi
->vdev
, vid
);
1665 sg_init_one(&sg
, &vi
->ctrl
->vid
, sizeof(vi
->ctrl
->vid
));
1667 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_VLAN
,
1668 VIRTIO_NET_CTRL_VLAN_ADD
, &sg
))
1669 dev_warn(&dev
->dev
, "Failed to add VLAN ID %d.\n", vid
);
1673 static int virtnet_vlan_rx_kill_vid(struct net_device
*dev
,
1674 __be16 proto
, u16 vid
)
1676 struct virtnet_info
*vi
= netdev_priv(dev
);
1677 struct scatterlist sg
;
1679 vi
->ctrl
->vid
= cpu_to_virtio16(vi
->vdev
, vid
);
1680 sg_init_one(&sg
, &vi
->ctrl
->vid
, sizeof(vi
->ctrl
->vid
));
1682 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_VLAN
,
1683 VIRTIO_NET_CTRL_VLAN_DEL
, &sg
))
1684 dev_warn(&dev
->dev
, "Failed to kill VLAN ID %d.\n", vid
);
1688 static void virtnet_clean_affinity(struct virtnet_info
*vi
, long hcpu
)
1692 if (vi
->affinity_hint_set
) {
1693 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1694 virtqueue_set_affinity(vi
->rq
[i
].vq
, -1);
1695 virtqueue_set_affinity(vi
->sq
[i
].vq
, -1);
1698 vi
->affinity_hint_set
= false;
1702 static void virtnet_set_affinity(struct virtnet_info
*vi
)
1707 /* In multiqueue mode, when the number of cpu is equal to the number of
1708 * queue pairs, we let the queue pairs to be private to one cpu by
1709 * setting the affinity hint to eliminate the contention.
1711 if (vi
->curr_queue_pairs
== 1 ||
1712 vi
->max_queue_pairs
!= num_online_cpus()) {
1713 virtnet_clean_affinity(vi
, -1);
1718 for_each_online_cpu(cpu
) {
1719 virtqueue_set_affinity(vi
->rq
[i
].vq
, cpu
);
1720 virtqueue_set_affinity(vi
->sq
[i
].vq
, cpu
);
1721 netif_set_xps_queue(vi
->dev
, cpumask_of(cpu
), i
);
1725 vi
->affinity_hint_set
= true;
1728 static int virtnet_cpu_online(unsigned int cpu
, struct hlist_node
*node
)
1730 struct virtnet_info
*vi
= hlist_entry_safe(node
, struct virtnet_info
,
1732 virtnet_set_affinity(vi
);
1736 static int virtnet_cpu_dead(unsigned int cpu
, struct hlist_node
*node
)
1738 struct virtnet_info
*vi
= hlist_entry_safe(node
, struct virtnet_info
,
1740 virtnet_set_affinity(vi
);
1744 static int virtnet_cpu_down_prep(unsigned int cpu
, struct hlist_node
*node
)
1746 struct virtnet_info
*vi
= hlist_entry_safe(node
, struct virtnet_info
,
1749 virtnet_clean_affinity(vi
, cpu
);
1753 static enum cpuhp_state virtionet_online
;
1755 static int virtnet_cpu_notif_add(struct virtnet_info
*vi
)
1759 ret
= cpuhp_state_add_instance_nocalls(virtionet_online
, &vi
->node
);
1762 ret
= cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD
,
1766 cpuhp_state_remove_instance_nocalls(virtionet_online
, &vi
->node
);
1770 static void virtnet_cpu_notif_remove(struct virtnet_info
*vi
)
1772 cpuhp_state_remove_instance_nocalls(virtionet_online
, &vi
->node
);
1773 cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD
,
1777 static void virtnet_get_ringparam(struct net_device
*dev
,
1778 struct ethtool_ringparam
*ring
)
1780 struct virtnet_info
*vi
= netdev_priv(dev
);
1782 ring
->rx_max_pending
= virtqueue_get_vring_size(vi
->rq
[0].vq
);
1783 ring
->tx_max_pending
= virtqueue_get_vring_size(vi
->sq
[0].vq
);
1784 ring
->rx_pending
= ring
->rx_max_pending
;
1785 ring
->tx_pending
= ring
->tx_max_pending
;
1789 static void virtnet_get_drvinfo(struct net_device
*dev
,
1790 struct ethtool_drvinfo
*info
)
1792 struct virtnet_info
*vi
= netdev_priv(dev
);
1793 struct virtio_device
*vdev
= vi
->vdev
;
1795 strlcpy(info
->driver
, KBUILD_MODNAME
, sizeof(info
->driver
));
1796 strlcpy(info
->version
, VIRTNET_DRIVER_VERSION
, sizeof(info
->version
));
1797 strlcpy(info
->bus_info
, virtio_bus_name(vdev
), sizeof(info
->bus_info
));
1801 /* TODO: Eliminate OOO packets during switching */
1802 static int virtnet_set_channels(struct net_device
*dev
,
1803 struct ethtool_channels
*channels
)
1805 struct virtnet_info
*vi
= netdev_priv(dev
);
1806 u16 queue_pairs
= channels
->combined_count
;
1809 /* We don't support separate rx/tx channels.
1810 * We don't allow setting 'other' channels.
1812 if (channels
->rx_count
|| channels
->tx_count
|| channels
->other_count
)
1815 if (queue_pairs
> vi
->max_queue_pairs
|| queue_pairs
== 0)
1818 /* For now we don't support modifying channels while XDP is loaded
1819 * also when XDP is loaded all RX queues have XDP programs so we only
1820 * need to check a single RX queue.
1822 if (vi
->rq
[0].xdp_prog
)
1826 err
= _virtnet_set_queues(vi
, queue_pairs
);
1828 netif_set_real_num_tx_queues(dev
, queue_pairs
);
1829 netif_set_real_num_rx_queues(dev
, queue_pairs
);
1831 virtnet_set_affinity(vi
);
1838 static void virtnet_get_channels(struct net_device
*dev
,
1839 struct ethtool_channels
*channels
)
1841 struct virtnet_info
*vi
= netdev_priv(dev
);
1843 channels
->combined_count
= vi
->curr_queue_pairs
;
1844 channels
->max_combined
= vi
->max_queue_pairs
;
1845 channels
->max_other
= 0;
1846 channels
->rx_count
= 0;
1847 channels
->tx_count
= 0;
1848 channels
->other_count
= 0;
1851 /* Check if the user is trying to change anything besides speed/duplex */
1853 virtnet_validate_ethtool_cmd(const struct ethtool_link_ksettings
*cmd
)
1855 struct ethtool_link_ksettings diff1
= *cmd
;
1856 struct ethtool_link_ksettings diff2
= {};
1858 /* cmd is always set so we need to clear it, validate the port type
1859 * and also without autonegotiation we can ignore advertising
1861 diff1
.base
.speed
= 0;
1862 diff2
.base
.port
= PORT_OTHER
;
1863 ethtool_link_ksettings_zero_link_mode(&diff1
, advertising
);
1864 diff1
.base
.duplex
= 0;
1866 diff1
.base
.link_mode_masks_nwords
= 0;
1868 return !memcmp(&diff1
.base
, &diff2
.base
, sizeof(diff1
.base
)) &&
1869 bitmap_empty(diff1
.link_modes
.supported
,
1870 __ETHTOOL_LINK_MODE_MASK_NBITS
) &&
1871 bitmap_empty(diff1
.link_modes
.advertising
,
1872 __ETHTOOL_LINK_MODE_MASK_NBITS
) &&
1873 bitmap_empty(diff1
.link_modes
.lp_advertising
,
1874 __ETHTOOL_LINK_MODE_MASK_NBITS
);
1877 static int virtnet_set_link_ksettings(struct net_device
*dev
,
1878 const struct ethtool_link_ksettings
*cmd
)
1880 struct virtnet_info
*vi
= netdev_priv(dev
);
1883 speed
= cmd
->base
.speed
;
1884 /* don't allow custom speed and duplex */
1885 if (!ethtool_validate_speed(speed
) ||
1886 !ethtool_validate_duplex(cmd
->base
.duplex
) ||
1887 !virtnet_validate_ethtool_cmd(cmd
))
1890 vi
->duplex
= cmd
->base
.duplex
;
1895 static int virtnet_get_link_ksettings(struct net_device
*dev
,
1896 struct ethtool_link_ksettings
*cmd
)
1898 struct virtnet_info
*vi
= netdev_priv(dev
);
1900 cmd
->base
.speed
= vi
->speed
;
1901 cmd
->base
.duplex
= vi
->duplex
;
1902 cmd
->base
.port
= PORT_OTHER
;
1907 static void virtnet_init_settings(struct net_device
*dev
)
1909 struct virtnet_info
*vi
= netdev_priv(dev
);
1911 vi
->speed
= SPEED_UNKNOWN
;
1912 vi
->duplex
= DUPLEX_UNKNOWN
;
1915 static const struct ethtool_ops virtnet_ethtool_ops
= {
1916 .get_drvinfo
= virtnet_get_drvinfo
,
1917 .get_link
= ethtool_op_get_link
,
1918 .get_ringparam
= virtnet_get_ringparam
,
1919 .set_channels
= virtnet_set_channels
,
1920 .get_channels
= virtnet_get_channels
,
1921 .get_ts_info
= ethtool_op_get_ts_info
,
1922 .get_link_ksettings
= virtnet_get_link_ksettings
,
1923 .set_link_ksettings
= virtnet_set_link_ksettings
,
1926 static void virtnet_freeze_down(struct virtio_device
*vdev
)
1928 struct virtnet_info
*vi
= vdev
->priv
;
1931 /* Make sure no work handler is accessing the device */
1932 flush_work(&vi
->config_work
);
1934 netif_device_detach(vi
->dev
);
1935 netif_tx_disable(vi
->dev
);
1936 cancel_delayed_work_sync(&vi
->refill
);
1938 if (netif_running(vi
->dev
)) {
1939 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1940 napi_disable(&vi
->rq
[i
].napi
);
1941 virtnet_napi_tx_disable(&vi
->sq
[i
].napi
);
1946 static int init_vqs(struct virtnet_info
*vi
);
1948 static int virtnet_restore_up(struct virtio_device
*vdev
)
1950 struct virtnet_info
*vi
= vdev
->priv
;
1957 virtio_device_ready(vdev
);
1959 if (netif_running(vi
->dev
)) {
1960 for (i
= 0; i
< vi
->curr_queue_pairs
; i
++)
1961 if (!try_fill_recv(vi
, &vi
->rq
[i
], GFP_KERNEL
))
1962 schedule_delayed_work(&vi
->refill
, 0);
1964 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1965 virtnet_napi_enable(vi
->rq
[i
].vq
, &vi
->rq
[i
].napi
);
1966 virtnet_napi_tx_enable(vi
, vi
->sq
[i
].vq
,
1971 netif_device_attach(vi
->dev
);
1975 static int virtnet_set_guest_offloads(struct virtnet_info
*vi
, u64 offloads
)
1977 struct scatterlist sg
;
1978 vi
->ctrl
->offloads
= cpu_to_virtio64(vi
->vdev
, offloads
);
1980 sg_init_one(&sg
, &vi
->ctrl
->offloads
, sizeof(vi
->ctrl
->offloads
));
1982 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_GUEST_OFFLOADS
,
1983 VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET
, &sg
)) {
1984 dev_warn(&vi
->dev
->dev
, "Fail to set guest offload. \n");
1991 static int virtnet_clear_guest_offloads(struct virtnet_info
*vi
)
1995 if (!vi
->guest_offloads
)
1998 if (virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_CSUM
))
1999 offloads
= 1ULL << VIRTIO_NET_F_GUEST_CSUM
;
2001 return virtnet_set_guest_offloads(vi
, offloads
);
2004 static int virtnet_restore_guest_offloads(struct virtnet_info
*vi
)
2006 u64 offloads
= vi
->guest_offloads
;
2008 if (!vi
->guest_offloads
)
2010 if (virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_CSUM
))
2011 offloads
|= 1ULL << VIRTIO_NET_F_GUEST_CSUM
;
2013 return virtnet_set_guest_offloads(vi
, offloads
);
2016 static int virtnet_xdp_set(struct net_device
*dev
, struct bpf_prog
*prog
,
2017 struct netlink_ext_ack
*extack
)
2019 unsigned long int max_sz
= PAGE_SIZE
- sizeof(struct padded_vnet_hdr
);
2020 struct virtnet_info
*vi
= netdev_priv(dev
);
2021 struct bpf_prog
*old_prog
;
2022 u16 xdp_qp
= 0, curr_qp
;
2025 if (!virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS
)
2026 && (virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_TSO4
) ||
2027 virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_TSO6
) ||
2028 virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_ECN
) ||
2029 virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_UFO
))) {
2030 NL_SET_ERR_MSG_MOD(extack
, "Can't set XDP while host is implementing LRO, disable LRO first");
2034 if (vi
->mergeable_rx_bufs
&& !vi
->any_header_sg
) {
2035 NL_SET_ERR_MSG_MOD(extack
, "XDP expects header/data in single page, any_header_sg required");
2039 if (dev
->mtu
> max_sz
) {
2040 NL_SET_ERR_MSG_MOD(extack
, "MTU too large to enable XDP");
2041 netdev_warn(dev
, "XDP requires MTU less than %lu\n", max_sz
);
2045 curr_qp
= vi
->curr_queue_pairs
- vi
->xdp_queue_pairs
;
2047 xdp_qp
= nr_cpu_ids
;
2049 /* XDP requires extra queues for XDP_TX */
2050 if (curr_qp
+ xdp_qp
> vi
->max_queue_pairs
) {
2051 NL_SET_ERR_MSG_MOD(extack
, "Too few free TX rings available");
2052 netdev_warn(dev
, "request %i queues but max is %i\n",
2053 curr_qp
+ xdp_qp
, vi
->max_queue_pairs
);
2058 prog
= bpf_prog_add(prog
, vi
->max_queue_pairs
- 1);
2060 return PTR_ERR(prog
);
2063 /* Make sure NAPI is not using any XDP TX queues for RX. */
2064 if (netif_running(dev
))
2065 for (i
= 0; i
< vi
->max_queue_pairs
; i
++)
2066 napi_disable(&vi
->rq
[i
].napi
);
2068 netif_set_real_num_rx_queues(dev
, curr_qp
+ xdp_qp
);
2069 err
= _virtnet_set_queues(vi
, curr_qp
+ xdp_qp
);
2072 vi
->xdp_queue_pairs
= xdp_qp
;
2074 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2075 old_prog
= rtnl_dereference(vi
->rq
[i
].xdp_prog
);
2076 rcu_assign_pointer(vi
->rq
[i
].xdp_prog
, prog
);
2079 virtnet_clear_guest_offloads(vi
);
2081 virtnet_restore_guest_offloads(vi
);
2084 bpf_prog_put(old_prog
);
2085 if (netif_running(dev
))
2086 virtnet_napi_enable(vi
->rq
[i
].vq
, &vi
->rq
[i
].napi
);
2092 for (i
= 0; i
< vi
->max_queue_pairs
; i
++)
2093 virtnet_napi_enable(vi
->rq
[i
].vq
, &vi
->rq
[i
].napi
);
2095 bpf_prog_sub(prog
, vi
->max_queue_pairs
- 1);
2099 static u32
virtnet_xdp_query(struct net_device
*dev
)
2101 struct virtnet_info
*vi
= netdev_priv(dev
);
2102 const struct bpf_prog
*xdp_prog
;
2105 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2106 xdp_prog
= rtnl_dereference(vi
->rq
[i
].xdp_prog
);
2108 return xdp_prog
->aux
->id
;
2113 static int virtnet_xdp(struct net_device
*dev
, struct netdev_bpf
*xdp
)
2115 switch (xdp
->command
) {
2116 case XDP_SETUP_PROG
:
2117 return virtnet_xdp_set(dev
, xdp
->prog
, xdp
->extack
);
2118 case XDP_QUERY_PROG
:
2119 xdp
->prog_id
= virtnet_xdp_query(dev
);
2120 xdp
->prog_attached
= !!xdp
->prog_id
;
2127 static const struct net_device_ops virtnet_netdev
= {
2128 .ndo_open
= virtnet_open
,
2129 .ndo_stop
= virtnet_close
,
2130 .ndo_start_xmit
= start_xmit
,
2131 .ndo_validate_addr
= eth_validate_addr
,
2132 .ndo_set_mac_address
= virtnet_set_mac_address
,
2133 .ndo_set_rx_mode
= virtnet_set_rx_mode
,
2134 .ndo_get_stats64
= virtnet_stats
,
2135 .ndo_vlan_rx_add_vid
= virtnet_vlan_rx_add_vid
,
2136 .ndo_vlan_rx_kill_vid
= virtnet_vlan_rx_kill_vid
,
2137 #ifdef CONFIG_NET_POLL_CONTROLLER
2138 .ndo_poll_controller
= virtnet_netpoll
,
2140 .ndo_bpf
= virtnet_xdp
,
2141 .ndo_xdp_xmit
= virtnet_xdp_xmit
,
2142 .ndo_xdp_flush
= virtnet_xdp_flush
,
2143 .ndo_features_check
= passthru_features_check
,
2146 static void virtnet_config_changed_work(struct work_struct
*work
)
2148 struct virtnet_info
*vi
=
2149 container_of(work
, struct virtnet_info
, config_work
);
2152 if (virtio_cread_feature(vi
->vdev
, VIRTIO_NET_F_STATUS
,
2153 struct virtio_net_config
, status
, &v
) < 0)
2156 if (v
& VIRTIO_NET_S_ANNOUNCE
) {
2157 netdev_notify_peers(vi
->dev
);
2158 virtnet_ack_link_announce(vi
);
2161 /* Ignore unknown (future) status bits */
2162 v
&= VIRTIO_NET_S_LINK_UP
;
2164 if (vi
->status
== v
)
2169 if (vi
->status
& VIRTIO_NET_S_LINK_UP
) {
2170 netif_carrier_on(vi
->dev
);
2171 netif_tx_wake_all_queues(vi
->dev
);
2173 netif_carrier_off(vi
->dev
);
2174 netif_tx_stop_all_queues(vi
->dev
);
2178 static void virtnet_config_changed(struct virtio_device
*vdev
)
2180 struct virtnet_info
*vi
= vdev
->priv
;
2182 schedule_work(&vi
->config_work
);
2185 static void virtnet_free_queues(struct virtnet_info
*vi
)
2189 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2190 napi_hash_del(&vi
->rq
[i
].napi
);
2191 netif_napi_del(&vi
->rq
[i
].napi
);
2192 netif_napi_del(&vi
->sq
[i
].napi
);
2195 /* We called napi_hash_del() before netif_napi_del(),
2196 * we need to respect an RCU grace period before freeing vi->rq
2205 static void _free_receive_bufs(struct virtnet_info
*vi
)
2207 struct bpf_prog
*old_prog
;
2210 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2211 while (vi
->rq
[i
].pages
)
2212 __free_pages(get_a_page(&vi
->rq
[i
], GFP_KERNEL
), 0);
2214 old_prog
= rtnl_dereference(vi
->rq
[i
].xdp_prog
);
2215 RCU_INIT_POINTER(vi
->rq
[i
].xdp_prog
, NULL
);
2217 bpf_prog_put(old_prog
);
2221 static void free_receive_bufs(struct virtnet_info
*vi
)
2224 _free_receive_bufs(vi
);
2228 static void free_receive_page_frags(struct virtnet_info
*vi
)
2231 for (i
= 0; i
< vi
->max_queue_pairs
; i
++)
2232 if (vi
->rq
[i
].alloc_frag
.page
)
2233 put_page(vi
->rq
[i
].alloc_frag
.page
);
2236 static bool is_xdp_raw_buffer_queue(struct virtnet_info
*vi
, int q
)
2238 if (q
< (vi
->curr_queue_pairs
- vi
->xdp_queue_pairs
))
2240 else if (q
< vi
->curr_queue_pairs
)
2246 static void free_unused_bufs(struct virtnet_info
*vi
)
2251 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2252 struct virtqueue
*vq
= vi
->sq
[i
].vq
;
2253 while ((buf
= virtqueue_detach_unused_buf(vq
)) != NULL
) {
2254 if (!is_xdp_raw_buffer_queue(vi
, i
))
2257 put_page(virt_to_head_page(buf
));
2261 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2262 struct virtqueue
*vq
= vi
->rq
[i
].vq
;
2264 while ((buf
= virtqueue_detach_unused_buf(vq
)) != NULL
) {
2265 if (vi
->mergeable_rx_bufs
) {
2266 put_page(virt_to_head_page(buf
));
2267 } else if (vi
->big_packets
) {
2268 give_pages(&vi
->rq
[i
], buf
);
2270 put_page(virt_to_head_page(buf
));
2276 static void virtnet_del_vqs(struct virtnet_info
*vi
)
2278 struct virtio_device
*vdev
= vi
->vdev
;
2280 virtnet_clean_affinity(vi
, -1);
2282 vdev
->config
->del_vqs(vdev
);
2284 virtnet_free_queues(vi
);
2287 /* How large should a single buffer be so a queue full of these can fit at
2288 * least one full packet?
2289 * Logic below assumes the mergeable buffer header is used.
2291 static unsigned int mergeable_min_buf_len(struct virtnet_info
*vi
, struct virtqueue
*vq
)
2293 const unsigned int hdr_len
= sizeof(struct virtio_net_hdr_mrg_rxbuf
);
2294 unsigned int rq_size
= virtqueue_get_vring_size(vq
);
2295 unsigned int packet_len
= vi
->big_packets
? IP_MAX_MTU
: vi
->dev
->max_mtu
;
2296 unsigned int buf_len
= hdr_len
+ ETH_HLEN
+ VLAN_HLEN
+ packet_len
;
2297 unsigned int min_buf_len
= DIV_ROUND_UP(buf_len
, rq_size
);
2299 return max(max(min_buf_len
, hdr_len
) - hdr_len
,
2300 (unsigned int)GOOD_PACKET_LEN
);
2303 static int virtnet_find_vqs(struct virtnet_info
*vi
)
2305 vq_callback_t
**callbacks
;
2306 struct virtqueue
**vqs
;
2312 /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
2313 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
2314 * possible control vq.
2316 total_vqs
= vi
->max_queue_pairs
* 2 +
2317 virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_CTRL_VQ
);
2319 /* Allocate space for find_vqs parameters */
2320 vqs
= kzalloc(total_vqs
* sizeof(*vqs
), GFP_KERNEL
);
2323 callbacks
= kmalloc(total_vqs
* sizeof(*callbacks
), GFP_KERNEL
);
2326 names
= kmalloc(total_vqs
* sizeof(*names
), GFP_KERNEL
);
2329 if (!vi
->big_packets
|| vi
->mergeable_rx_bufs
) {
2330 ctx
= kzalloc(total_vqs
* sizeof(*ctx
), GFP_KERNEL
);
2337 /* Parameters for control virtqueue, if any */
2339 callbacks
[total_vqs
- 1] = NULL
;
2340 names
[total_vqs
- 1] = "control";
2343 /* Allocate/initialize parameters for send/receive virtqueues */
2344 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2345 callbacks
[rxq2vq(i
)] = skb_recv_done
;
2346 callbacks
[txq2vq(i
)] = skb_xmit_done
;
2347 sprintf(vi
->rq
[i
].name
, "input.%d", i
);
2348 sprintf(vi
->sq
[i
].name
, "output.%d", i
);
2349 names
[rxq2vq(i
)] = vi
->rq
[i
].name
;
2350 names
[txq2vq(i
)] = vi
->sq
[i
].name
;
2352 ctx
[rxq2vq(i
)] = true;
2355 ret
= vi
->vdev
->config
->find_vqs(vi
->vdev
, total_vqs
, vqs
, callbacks
,
2361 vi
->cvq
= vqs
[total_vqs
- 1];
2362 if (virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_CTRL_VLAN
))
2363 vi
->dev
->features
|= NETIF_F_HW_VLAN_CTAG_FILTER
;
2366 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2367 vi
->rq
[i
].vq
= vqs
[rxq2vq(i
)];
2368 vi
->rq
[i
].min_buf_len
= mergeable_min_buf_len(vi
, vi
->rq
[i
].vq
);
2369 vi
->sq
[i
].vq
= vqs
[txq2vq(i
)];
2391 static int virtnet_alloc_queues(struct virtnet_info
*vi
)
2395 vi
->ctrl
= kzalloc(sizeof(*vi
->ctrl
), GFP_KERNEL
);
2398 vi
->sq
= kzalloc(sizeof(*vi
->sq
) * vi
->max_queue_pairs
, GFP_KERNEL
);
2401 vi
->rq
= kzalloc(sizeof(*vi
->rq
) * vi
->max_queue_pairs
, GFP_KERNEL
);
2405 INIT_DELAYED_WORK(&vi
->refill
, refill_work
);
2406 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2407 vi
->rq
[i
].pages
= NULL
;
2408 netif_napi_add(vi
->dev
, &vi
->rq
[i
].napi
, virtnet_poll
,
2410 netif_tx_napi_add(vi
->dev
, &vi
->sq
[i
].napi
, virtnet_poll_tx
,
2411 napi_tx
? napi_weight
: 0);
2413 sg_init_table(vi
->rq
[i
].sg
, ARRAY_SIZE(vi
->rq
[i
].sg
));
2414 ewma_pkt_len_init(&vi
->rq
[i
].mrg_avg_pkt_len
);
2415 sg_init_table(vi
->sq
[i
].sg
, ARRAY_SIZE(vi
->sq
[i
].sg
));
2428 static int init_vqs(struct virtnet_info
*vi
)
2432 /* Allocate send & receive queues */
2433 ret
= virtnet_alloc_queues(vi
);
2437 ret
= virtnet_find_vqs(vi
);
2442 virtnet_set_affinity(vi
);
2448 virtnet_free_queues(vi
);
2454 static ssize_t
mergeable_rx_buffer_size_show(struct netdev_rx_queue
*queue
,
2457 struct virtnet_info
*vi
= netdev_priv(queue
->dev
);
2458 unsigned int queue_index
= get_netdev_rx_queue_index(queue
);
2459 struct ewma_pkt_len
*avg
;
2461 BUG_ON(queue_index
>= vi
->max_queue_pairs
);
2462 avg
= &vi
->rq
[queue_index
].mrg_avg_pkt_len
;
2463 return sprintf(buf
, "%u\n",
2464 get_mergeable_buf_len(&vi
->rq
[queue_index
], avg
));
2467 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute
=
2468 __ATTR_RO(mergeable_rx_buffer_size
);
2470 static struct attribute
*virtio_net_mrg_rx_attrs
[] = {
2471 &mergeable_rx_buffer_size_attribute
.attr
,
2475 static const struct attribute_group virtio_net_mrg_rx_group
= {
2476 .name
= "virtio_net",
2477 .attrs
= virtio_net_mrg_rx_attrs
2481 static bool virtnet_fail_on_feature(struct virtio_device
*vdev
,
2483 const char *fname
, const char *dname
)
2485 if (!virtio_has_feature(vdev
, fbit
))
2488 dev_err(&vdev
->dev
, "device advertises feature %s but not %s",
2494 #define VIRTNET_FAIL_ON(vdev, fbit, dbit) \
2495 virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
2497 static bool virtnet_validate_features(struct virtio_device
*vdev
)
2499 if (!virtio_has_feature(vdev
, VIRTIO_NET_F_CTRL_VQ
) &&
2500 (VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_CTRL_RX
,
2501 "VIRTIO_NET_F_CTRL_VQ") ||
2502 VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_CTRL_VLAN
,
2503 "VIRTIO_NET_F_CTRL_VQ") ||
2504 VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_GUEST_ANNOUNCE
,
2505 "VIRTIO_NET_F_CTRL_VQ") ||
2506 VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_MQ
, "VIRTIO_NET_F_CTRL_VQ") ||
2507 VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_CTRL_MAC_ADDR
,
2508 "VIRTIO_NET_F_CTRL_VQ"))) {
2515 #define MIN_MTU ETH_MIN_MTU
2516 #define MAX_MTU ETH_MAX_MTU
2518 static int virtnet_validate(struct virtio_device
*vdev
)
2520 if (!vdev
->config
->get
) {
2521 dev_err(&vdev
->dev
, "%s failure: config access disabled\n",
2526 if (!virtnet_validate_features(vdev
))
2529 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MTU
)) {
2530 int mtu
= virtio_cread16(vdev
,
2531 offsetof(struct virtio_net_config
,
2534 __virtio_clear_bit(vdev
, VIRTIO_NET_F_MTU
);
2540 static int virtnet_probe(struct virtio_device
*vdev
)
2543 struct net_device
*dev
;
2544 struct virtnet_info
*vi
;
2545 u16 max_queue_pairs
;
2548 /* Find if host supports multiqueue virtio_net device */
2549 err
= virtio_cread_feature(vdev
, VIRTIO_NET_F_MQ
,
2550 struct virtio_net_config
,
2551 max_virtqueue_pairs
, &max_queue_pairs
);
2553 /* We need at least 2 queue's */
2554 if (err
|| max_queue_pairs
< VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN
||
2555 max_queue_pairs
> VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX
||
2556 !virtio_has_feature(vdev
, VIRTIO_NET_F_CTRL_VQ
))
2557 max_queue_pairs
= 1;
2559 /* Allocate ourselves a network device with room for our info */
2560 dev
= alloc_etherdev_mq(sizeof(struct virtnet_info
), max_queue_pairs
);
2564 /* Set up network device as normal. */
2565 dev
->priv_flags
|= IFF_UNICAST_FLT
| IFF_LIVE_ADDR_CHANGE
;
2566 dev
->netdev_ops
= &virtnet_netdev
;
2567 dev
->features
= NETIF_F_HIGHDMA
;
2569 dev
->ethtool_ops
= &virtnet_ethtool_ops
;
2570 SET_NETDEV_DEV(dev
, &vdev
->dev
);
2572 /* Do we support "hardware" checksums? */
2573 if (virtio_has_feature(vdev
, VIRTIO_NET_F_CSUM
)) {
2574 /* This opens up the world of extra features. */
2575 dev
->hw_features
|= NETIF_F_HW_CSUM
| NETIF_F_SG
;
2577 dev
->features
|= NETIF_F_HW_CSUM
| NETIF_F_SG
;
2579 if (virtio_has_feature(vdev
, VIRTIO_NET_F_GSO
)) {
2580 dev
->hw_features
|= NETIF_F_TSO
2581 | NETIF_F_TSO_ECN
| NETIF_F_TSO6
;
2583 /* Individual feature bits: what can host handle? */
2584 if (virtio_has_feature(vdev
, VIRTIO_NET_F_HOST_TSO4
))
2585 dev
->hw_features
|= NETIF_F_TSO
;
2586 if (virtio_has_feature(vdev
, VIRTIO_NET_F_HOST_TSO6
))
2587 dev
->hw_features
|= NETIF_F_TSO6
;
2588 if (virtio_has_feature(vdev
, VIRTIO_NET_F_HOST_ECN
))
2589 dev
->hw_features
|= NETIF_F_TSO_ECN
;
2591 dev
->features
|= NETIF_F_GSO_ROBUST
;
2594 dev
->features
|= dev
->hw_features
& NETIF_F_ALL_TSO
;
2595 /* (!csum && gso) case will be fixed by register_netdev() */
2597 if (virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_CSUM
))
2598 dev
->features
|= NETIF_F_RXCSUM
;
2600 dev
->vlan_features
= dev
->features
;
2602 /* MTU range: 68 - 65535 */
2603 dev
->min_mtu
= MIN_MTU
;
2604 dev
->max_mtu
= MAX_MTU
;
2606 /* Configuration may specify what MAC to use. Otherwise random. */
2607 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MAC
))
2608 virtio_cread_bytes(vdev
,
2609 offsetof(struct virtio_net_config
, mac
),
2610 dev
->dev_addr
, dev
->addr_len
);
2612 eth_hw_addr_random(dev
);
2614 /* Set up our device-specific information */
2615 vi
= netdev_priv(dev
);
2619 vi
->stats
= alloc_percpu(struct virtnet_stats
);
2621 if (vi
->stats
== NULL
)
2624 for_each_possible_cpu(i
) {
2625 struct virtnet_stats
*virtnet_stats
;
2626 virtnet_stats
= per_cpu_ptr(vi
->stats
, i
);
2627 u64_stats_init(&virtnet_stats
->tx_syncp
);
2628 u64_stats_init(&virtnet_stats
->rx_syncp
);
2631 INIT_WORK(&vi
->config_work
, virtnet_config_changed_work
);
2633 /* If we can receive ANY GSO packets, we must allocate large ones. */
2634 if (virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_TSO4
) ||
2635 virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_TSO6
) ||
2636 virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_ECN
) ||
2637 virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_UFO
))
2638 vi
->big_packets
= true;
2640 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MRG_RXBUF
))
2641 vi
->mergeable_rx_bufs
= true;
2643 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MRG_RXBUF
) ||
2644 virtio_has_feature(vdev
, VIRTIO_F_VERSION_1
))
2645 vi
->hdr_len
= sizeof(struct virtio_net_hdr_mrg_rxbuf
);
2647 vi
->hdr_len
= sizeof(struct virtio_net_hdr
);
2649 if (virtio_has_feature(vdev
, VIRTIO_F_ANY_LAYOUT
) ||
2650 virtio_has_feature(vdev
, VIRTIO_F_VERSION_1
))
2651 vi
->any_header_sg
= true;
2653 if (virtio_has_feature(vdev
, VIRTIO_NET_F_CTRL_VQ
))
2656 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MTU
)) {
2657 mtu
= virtio_cread16(vdev
,
2658 offsetof(struct virtio_net_config
,
2660 if (mtu
< dev
->min_mtu
) {
2661 /* Should never trigger: MTU was previously validated
2662 * in virtnet_validate.
2664 dev_err(&vdev
->dev
, "device MTU appears to have changed "
2665 "it is now %d < %d", mtu
, dev
->min_mtu
);
2672 /* TODO: size buffers correctly in this case. */
2673 if (dev
->mtu
> ETH_DATA_LEN
)
2674 vi
->big_packets
= true;
2677 if (vi
->any_header_sg
)
2678 dev
->needed_headroom
= vi
->hdr_len
;
2680 /* Enable multiqueue by default */
2681 if (num_online_cpus() >= max_queue_pairs
)
2682 vi
->curr_queue_pairs
= max_queue_pairs
;
2684 vi
->curr_queue_pairs
= num_online_cpus();
2685 vi
->max_queue_pairs
= max_queue_pairs
;
2687 /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
2693 if (vi
->mergeable_rx_bufs
)
2694 dev
->sysfs_rx_queue_group
= &virtio_net_mrg_rx_group
;
2696 netif_set_real_num_tx_queues(dev
, vi
->curr_queue_pairs
);
2697 netif_set_real_num_rx_queues(dev
, vi
->curr_queue_pairs
);
2699 virtnet_init_settings(dev
);
2701 err
= register_netdev(dev
);
2703 pr_debug("virtio_net: registering device failed\n");
2707 virtio_device_ready(vdev
);
2709 err
= virtnet_cpu_notif_add(vi
);
2711 pr_debug("virtio_net: registering cpu notifier failed\n");
2712 goto free_unregister_netdev
;
2715 virtnet_set_queues(vi
, vi
->curr_queue_pairs
);
2717 /* Assume link up if device can't report link status,
2718 otherwise get link status from config. */
2719 netif_carrier_off(dev
);
2720 if (virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_STATUS
)) {
2721 schedule_work(&vi
->config_work
);
2723 vi
->status
= VIRTIO_NET_S_LINK_UP
;
2724 netif_carrier_on(dev
);
2727 for (i
= 0; i
< ARRAY_SIZE(guest_offloads
); i
++)
2728 if (virtio_has_feature(vi
->vdev
, guest_offloads
[i
]))
2729 set_bit(guest_offloads
[i
], &vi
->guest_offloads
);
2731 pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
2732 dev
->name
, max_queue_pairs
);
2736 free_unregister_netdev
:
2737 vi
->vdev
->config
->reset(vdev
);
2739 unregister_netdev(dev
);
2741 cancel_delayed_work_sync(&vi
->refill
);
2742 free_receive_page_frags(vi
);
2743 virtnet_del_vqs(vi
);
2745 free_percpu(vi
->stats
);
2751 static void remove_vq_common(struct virtnet_info
*vi
)
2753 vi
->vdev
->config
->reset(vi
->vdev
);
2755 /* Free unused buffers in both send and recv, if any. */
2756 free_unused_bufs(vi
);
2758 free_receive_bufs(vi
);
2760 free_receive_page_frags(vi
);
2762 virtnet_del_vqs(vi
);
2765 static void virtnet_remove(struct virtio_device
*vdev
)
2767 struct virtnet_info
*vi
= vdev
->priv
;
2769 virtnet_cpu_notif_remove(vi
);
2771 /* Make sure no work handler is accessing the device. */
2772 flush_work(&vi
->config_work
);
2774 unregister_netdev(vi
->dev
);
2776 remove_vq_common(vi
);
2778 free_percpu(vi
->stats
);
2779 free_netdev(vi
->dev
);
2782 static __maybe_unused
int virtnet_freeze(struct virtio_device
*vdev
)
2784 struct virtnet_info
*vi
= vdev
->priv
;
2786 virtnet_cpu_notif_remove(vi
);
2787 virtnet_freeze_down(vdev
);
2788 remove_vq_common(vi
);
2793 static __maybe_unused
int virtnet_restore(struct virtio_device
*vdev
)
2795 struct virtnet_info
*vi
= vdev
->priv
;
2798 err
= virtnet_restore_up(vdev
);
2801 virtnet_set_queues(vi
, vi
->curr_queue_pairs
);
2803 err
= virtnet_cpu_notif_add(vi
);
2810 static struct virtio_device_id id_table
[] = {
2811 { VIRTIO_ID_NET
, VIRTIO_DEV_ANY_ID
},
2815 #define VIRTNET_FEATURES \
2816 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
2818 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
2819 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
2820 VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
2821 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
2822 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
2823 VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
2824 VIRTIO_NET_F_CTRL_MAC_ADDR, \
2825 VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS
2827 static unsigned int features
[] = {
2831 static unsigned int features_legacy
[] = {
2834 VIRTIO_F_ANY_LAYOUT
,
2837 static struct virtio_driver virtio_net_driver
= {
2838 .feature_table
= features
,
2839 .feature_table_size
= ARRAY_SIZE(features
),
2840 .feature_table_legacy
= features_legacy
,
2841 .feature_table_size_legacy
= ARRAY_SIZE(features_legacy
),
2842 .driver
.name
= KBUILD_MODNAME
,
2843 .driver
.owner
= THIS_MODULE
,
2844 .id_table
= id_table
,
2845 .validate
= virtnet_validate
,
2846 .probe
= virtnet_probe
,
2847 .remove
= virtnet_remove
,
2848 .config_changed
= virtnet_config_changed
,
2849 #ifdef CONFIG_PM_SLEEP
2850 .freeze
= virtnet_freeze
,
2851 .restore
= virtnet_restore
,
2855 static __init
int virtio_net_driver_init(void)
2859 ret
= cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN
, "virtio/net:online",
2861 virtnet_cpu_down_prep
);
2864 virtionet_online
= ret
;
2865 ret
= cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD
, "virtio/net:dead",
2866 NULL
, virtnet_cpu_dead
);
2870 ret
= register_virtio_driver(&virtio_net_driver
);
2875 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD
);
2877 cpuhp_remove_multi_state(virtionet_online
);
2881 module_init(virtio_net_driver_init
);
2883 static __exit
void virtio_net_driver_exit(void)
2885 unregister_virtio_driver(&virtio_net_driver
);
2886 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD
);
2887 cpuhp_remove_multi_state(virtionet_online
);
2889 module_exit(virtio_net_driver_exit
);
2891 MODULE_DEVICE_TABLE(virtio
, id_table
);
2892 MODULE_DESCRIPTION("Virtio network driver");
2893 MODULE_LICENSE("GPL");