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 <net/route.h>
34 static int napi_weight
= NAPI_POLL_WEIGHT
;
35 module_param(napi_weight
, int, 0444);
37 static bool csum
= true, gso
= true, napi_tx
;
38 module_param(csum
, bool, 0444);
39 module_param(gso
, bool, 0444);
40 module_param(napi_tx
, bool, 0644);
42 /* FIXME: MTU in config. */
43 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
44 #define GOOD_COPY_LEN 128
46 #define VIRTNET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
48 /* Amount of XDP headroom to prepend to packets for use by xdp_adjust_head */
49 #define VIRTIO_XDP_HEADROOM 256
51 /* RX packet size EWMA. The average packet size is used to determine the packet
52 * buffer size when refilling RX rings. As the entire RX ring may be refilled
53 * at once, the weight is chosen so that the EWMA will be insensitive to short-
54 * term, transient changes in packet size.
56 DECLARE_EWMA(pkt_len
, 0, 64)
58 #define VIRTNET_DRIVER_VERSION "1.0.0"
60 struct virtnet_stats
{
61 struct u64_stats_sync tx_syncp
;
62 struct u64_stats_sync rx_syncp
;
70 /* Internal representation of a send virtqueue */
72 /* Virtqueue associated with this send _queue */
75 /* TX: fragments + linear part + virtio header */
76 struct scatterlist sg
[MAX_SKB_FRAGS
+ 2];
78 /* Name of the send queue: output.$index */
81 struct napi_struct napi
;
84 /* Internal representation of a receive virtqueue */
85 struct receive_queue
{
86 /* Virtqueue associated with this receive_queue */
89 struct napi_struct napi
;
91 struct bpf_prog __rcu
*xdp_prog
;
93 /* Chain pages by the private ptr. */
96 /* Average packet length for mergeable receive buffers. */
97 struct ewma_pkt_len mrg_avg_pkt_len
;
99 /* Page frag for packet buffer allocation. */
100 struct page_frag alloc_frag
;
102 /* RX: fragments + linear part + virtio header */
103 struct scatterlist sg
[MAX_SKB_FRAGS
+ 2];
105 /* Min single buffer size for mergeable buffers case. */
106 unsigned int min_buf_len
;
108 /* Name of this receive queue: input.$index */
112 struct virtnet_info
{
113 struct virtio_device
*vdev
;
114 struct virtqueue
*cvq
;
115 struct net_device
*dev
;
116 struct send_queue
*sq
;
117 struct receive_queue
*rq
;
120 /* Max # of queue pairs supported by the device */
123 /* # of queue pairs currently used by the driver */
124 u16 curr_queue_pairs
;
126 /* # of XDP queue pairs currently used by the driver */
129 /* I like... big packets and I cannot lie! */
132 /* Host will merge rx buffers for big packets (shake it! shake it!) */
133 bool mergeable_rx_bufs
;
135 /* Has control virtqueue */
138 /* Host can handle any s/g split between our header and packet data */
141 /* Packet virtio header size */
144 /* Active statistics */
145 struct virtnet_stats __percpu
*stats
;
147 /* Work struct for refilling if we run low on memory. */
148 struct delayed_work refill
;
150 /* Work struct for config space updates */
151 struct work_struct config_work
;
153 /* Does the affinity hint is set for virtqueues? */
154 bool affinity_hint_set
;
156 /* CPU hotplug instances for online & dead */
157 struct hlist_node node
;
158 struct hlist_node node_dead
;
160 /* Control VQ buffers: protected by the rtnl lock */
161 struct virtio_net_ctrl_hdr ctrl_hdr
;
162 virtio_net_ctrl_ack ctrl_status
;
163 struct virtio_net_ctrl_mq ctrl_mq
;
168 /* Ethtool settings */
173 struct padded_vnet_hdr
{
174 struct virtio_net_hdr_mrg_rxbuf hdr
;
176 * hdr is in a separate sg buffer, and data sg buffer shares same page
177 * with this header sg. This padding makes next sg 16 byte aligned
183 /* Converting between virtqueue no. and kernel tx/rx queue no.
184 * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
186 static int vq2txq(struct virtqueue
*vq
)
188 return (vq
->index
- 1) / 2;
191 static int txq2vq(int txq
)
196 static int vq2rxq(struct virtqueue
*vq
)
198 return vq
->index
/ 2;
201 static int rxq2vq(int rxq
)
206 static inline struct virtio_net_hdr_mrg_rxbuf
*skb_vnet_hdr(struct sk_buff
*skb
)
208 return (struct virtio_net_hdr_mrg_rxbuf
*)skb
->cb
;
212 * private is used to chain pages for big packets, put the whole
213 * most recent used list in the beginning for reuse
215 static void give_pages(struct receive_queue
*rq
, struct page
*page
)
219 /* Find end of list, sew whole thing into vi->rq.pages. */
220 for (end
= page
; end
->private; end
= (struct page
*)end
->private);
221 end
->private = (unsigned long)rq
->pages
;
225 static struct page
*get_a_page(struct receive_queue
*rq
, gfp_t gfp_mask
)
227 struct page
*p
= rq
->pages
;
230 rq
->pages
= (struct page
*)p
->private;
231 /* clear private here, it is used to chain pages */
234 p
= alloc_page(gfp_mask
);
238 static void virtqueue_napi_schedule(struct napi_struct
*napi
,
239 struct virtqueue
*vq
)
241 if (napi_schedule_prep(napi
)) {
242 virtqueue_disable_cb(vq
);
243 __napi_schedule(napi
);
247 static void virtqueue_napi_complete(struct napi_struct
*napi
,
248 struct virtqueue
*vq
, int processed
)
252 opaque
= virtqueue_enable_cb_prepare(vq
);
253 if (napi_complete_done(napi
, processed
) &&
254 unlikely(virtqueue_poll(vq
, opaque
)))
255 virtqueue_napi_schedule(napi
, vq
);
258 static void skb_xmit_done(struct virtqueue
*vq
)
260 struct virtnet_info
*vi
= vq
->vdev
->priv
;
261 struct napi_struct
*napi
= &vi
->sq
[vq2txq(vq
)].napi
;
263 /* Suppress further interrupts. */
264 virtqueue_disable_cb(vq
);
267 virtqueue_napi_schedule(napi
, vq
);
269 /* We were probably waiting for more output buffers. */
270 netif_wake_subqueue(vi
->dev
, vq2txq(vq
));
273 /* Called from bottom half context */
274 static struct sk_buff
*page_to_skb(struct virtnet_info
*vi
,
275 struct receive_queue
*rq
,
276 struct page
*page
, unsigned int offset
,
277 unsigned int len
, unsigned int truesize
)
280 struct virtio_net_hdr_mrg_rxbuf
*hdr
;
281 unsigned int copy
, hdr_len
, hdr_padded_len
;
284 p
= page_address(page
) + offset
;
286 /* copy small packet so we can reuse these pages for small data */
287 skb
= napi_alloc_skb(&rq
->napi
, GOOD_COPY_LEN
);
291 hdr
= skb_vnet_hdr(skb
);
293 hdr_len
= vi
->hdr_len
;
294 if (vi
->mergeable_rx_bufs
)
295 hdr_padded_len
= sizeof *hdr
;
297 hdr_padded_len
= sizeof(struct padded_vnet_hdr
);
299 memcpy(hdr
, p
, hdr_len
);
302 offset
+= hdr_padded_len
;
306 if (copy
> skb_tailroom(skb
))
307 copy
= skb_tailroom(skb
);
308 skb_put_data(skb
, p
, copy
);
313 if (vi
->mergeable_rx_bufs
) {
315 skb_add_rx_frag(skb
, 0, page
, offset
, len
, truesize
);
322 * Verify that we can indeed put this data into a skb.
323 * This is here to handle cases when the device erroneously
324 * tries to receive more than is possible. This is usually
325 * the case of a broken device.
327 if (unlikely(len
> MAX_SKB_FRAGS
* PAGE_SIZE
)) {
328 net_dbg_ratelimited("%s: too much data\n", skb
->dev
->name
);
332 BUG_ON(offset
>= PAGE_SIZE
);
334 unsigned int frag_size
= min((unsigned)PAGE_SIZE
- offset
, len
);
335 skb_add_rx_frag(skb
, skb_shinfo(skb
)->nr_frags
, page
, offset
,
336 frag_size
, truesize
);
338 page
= (struct page
*)page
->private;
343 give_pages(rq
, page
);
348 static bool virtnet_xdp_xmit(struct virtnet_info
*vi
,
349 struct receive_queue
*rq
,
350 struct xdp_buff
*xdp
)
352 struct virtio_net_hdr_mrg_rxbuf
*hdr
;
354 struct send_queue
*sq
;
359 qp
= vi
->curr_queue_pairs
- vi
->xdp_queue_pairs
+ smp_processor_id();
362 /* Free up any pending old buffers before queueing new ones. */
363 while ((xdp_sent
= virtqueue_get_buf(sq
->vq
, &len
)) != NULL
) {
364 struct page
*sent_page
= virt_to_head_page(xdp_sent
);
369 xdp
->data
-= vi
->hdr_len
;
370 /* Zero header and leave csum up to XDP layers */
372 memset(hdr
, 0, vi
->hdr_len
);
374 sg_init_one(sq
->sg
, xdp
->data
, xdp
->data_end
- xdp
->data
);
376 err
= virtqueue_add_outbuf(sq
->vq
, sq
->sg
, 1, xdp
->data
, GFP_ATOMIC
);
378 struct page
*page
= virt_to_head_page(xdp
->data
);
384 virtqueue_kick(sq
->vq
);
388 static unsigned int virtnet_get_headroom(struct virtnet_info
*vi
)
390 return vi
->xdp_queue_pairs
? VIRTIO_XDP_HEADROOM
: 0;
393 static struct sk_buff
*receive_small(struct net_device
*dev
,
394 struct virtnet_info
*vi
,
395 struct receive_queue
*rq
,
396 void *buf
, unsigned int len
)
399 struct bpf_prog
*xdp_prog
;
400 unsigned int xdp_headroom
= virtnet_get_headroom(vi
);
401 unsigned int header_offset
= VIRTNET_RX_PAD
+ xdp_headroom
;
402 unsigned int headroom
= vi
->hdr_len
+ header_offset
;
403 unsigned int buflen
= SKB_DATA_ALIGN(GOOD_PACKET_LEN
+ headroom
) +
404 SKB_DATA_ALIGN(sizeof(struct skb_shared_info
));
405 unsigned int delta
= 0;
409 xdp_prog
= rcu_dereference(rq
->xdp_prog
);
411 struct virtio_net_hdr_mrg_rxbuf
*hdr
= buf
+ header_offset
;
416 if (unlikely(hdr
->hdr
.gso_type
|| hdr
->hdr
.flags
))
419 xdp
.data_hard_start
= buf
+ VIRTNET_RX_PAD
+ vi
->hdr_len
;
420 xdp
.data
= xdp
.data_hard_start
+ xdp_headroom
;
421 xdp
.data_end
= xdp
.data
+ len
;
422 orig_data
= xdp
.data
;
423 act
= bpf_prog_run_xdp(xdp_prog
, &xdp
);
427 /* Recalculate length in case bpf program changed it */
428 delta
= orig_data
- xdp
.data
;
431 if (unlikely(!virtnet_xdp_xmit(vi
, rq
, &xdp
)))
432 trace_xdp_exception(vi
->dev
, xdp_prog
, act
);
436 bpf_warn_invalid_xdp_action(act
);
438 trace_xdp_exception(vi
->dev
, xdp_prog
, act
);
445 skb
= build_skb(buf
, buflen
);
447 put_page(virt_to_head_page(buf
));
450 skb_reserve(skb
, headroom
- delta
);
451 skb_put(skb
, len
+ delta
);
453 buf
+= header_offset
;
454 memcpy(skb_vnet_hdr(skb
), buf
, vi
->hdr_len
);
455 } /* keep zeroed vnet hdr since packet was changed by bpf */
462 dev
->stats
.rx_dropped
++;
463 put_page(virt_to_head_page(buf
));
468 static struct sk_buff
*receive_big(struct net_device
*dev
,
469 struct virtnet_info
*vi
,
470 struct receive_queue
*rq
,
474 struct page
*page
= buf
;
475 struct sk_buff
*skb
= page_to_skb(vi
, rq
, page
, 0, len
, PAGE_SIZE
);
483 dev
->stats
.rx_dropped
++;
484 give_pages(rq
, page
);
488 /* The conditions to enable XDP should preclude the underlying device from
489 * sending packets across multiple buffers (num_buf > 1). However per spec
490 * it does not appear to be illegal to do so but rather just against convention.
491 * So in order to avoid making a system unresponsive the packets are pushed
492 * into a page and the XDP program is run. This will be extremely slow and we
493 * push a warning to the user to fix this as soon as possible. Fixing this may
494 * require resolving the underlying hardware to determine why multiple buffers
495 * are being received or simply loading the XDP program in the ingress stack
496 * after the skb is built because there is no advantage to running it here
499 static struct page
*xdp_linearize_page(struct receive_queue
*rq
,
505 struct page
*page
= alloc_page(GFP_ATOMIC
);
506 unsigned int page_off
= VIRTIO_XDP_HEADROOM
;
511 memcpy(page_address(page
) + page_off
, page_address(p
) + offset
, *len
);
519 buf
= virtqueue_get_buf(rq
->vq
, &buflen
);
523 p
= virt_to_head_page(buf
);
524 off
= buf
- page_address(p
);
526 /* guard against a misconfigured or uncooperative backend that
527 * is sending packet larger than the MTU.
529 if ((page_off
+ buflen
) > PAGE_SIZE
) {
534 memcpy(page_address(page
) + page_off
,
535 page_address(p
) + off
, buflen
);
540 /* Headroom does not contribute to packet length */
541 *len
= page_off
- VIRTIO_XDP_HEADROOM
;
544 __free_pages(page
, 0);
548 static struct sk_buff
*receive_mergeable(struct net_device
*dev
,
549 struct virtnet_info
*vi
,
550 struct receive_queue
*rq
,
555 struct virtio_net_hdr_mrg_rxbuf
*hdr
= buf
;
556 u16 num_buf
= virtio16_to_cpu(vi
->vdev
, hdr
->num_buffers
);
557 struct page
*page
= virt_to_head_page(buf
);
558 int offset
= buf
- page_address(page
);
559 struct sk_buff
*head_skb
, *curr_skb
;
560 struct bpf_prog
*xdp_prog
;
561 unsigned int truesize
;
566 xdp_prog
= rcu_dereference(rq
->xdp_prog
);
568 struct page
*xdp_page
;
573 /* This happens when rx buffer size is underestimated */
574 if (unlikely(num_buf
> 1)) {
575 /* linearize data for XDP */
576 xdp_page
= xdp_linearize_page(rq
, &num_buf
,
580 offset
= VIRTIO_XDP_HEADROOM
;
585 /* Transient failure which in theory could occur if
586 * in-flight packets from before XDP was enabled reach
587 * the receive path after XDP is loaded. In practice I
588 * was not able to create this condition.
590 if (unlikely(hdr
->hdr
.gso_type
))
593 /* Allow consuming headroom but reserve enough space to push
594 * the descriptor on if we get an XDP_TX return code.
596 data
= page_address(xdp_page
) + offset
;
597 xdp
.data_hard_start
= data
- VIRTIO_XDP_HEADROOM
+ vi
->hdr_len
;
598 xdp
.data
= data
+ vi
->hdr_len
;
599 xdp
.data_end
= xdp
.data
+ (len
- vi
->hdr_len
);
600 act
= bpf_prog_run_xdp(xdp_prog
, &xdp
);
604 /* recalculate offset to account for any header
605 * adjustments. Note other cases do not build an
606 * skb and avoid using offset
609 page_address(xdp_page
) - vi
->hdr_len
;
611 /* We can only create skb based on xdp_page. */
612 if (unlikely(xdp_page
!= page
)) {
615 head_skb
= page_to_skb(vi
, rq
, xdp_page
,
616 offset
, len
, PAGE_SIZE
);
617 ewma_pkt_len_add(&rq
->mrg_avg_pkt_len
, len
);
622 if (unlikely(!virtnet_xdp_xmit(vi
, rq
, &xdp
)))
623 trace_xdp_exception(vi
->dev
, xdp_prog
, act
);
624 ewma_pkt_len_add(&rq
->mrg_avg_pkt_len
, len
);
625 if (unlikely(xdp_page
!= page
))
630 bpf_warn_invalid_xdp_action(act
);
632 trace_xdp_exception(vi
->dev
, xdp_prog
, act
);
634 if (unlikely(xdp_page
!= page
))
635 __free_pages(xdp_page
, 0);
636 ewma_pkt_len_add(&rq
->mrg_avg_pkt_len
, len
);
642 if (unlikely(len
> (unsigned long)ctx
)) {
643 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
644 dev
->name
, len
, (unsigned long)ctx
);
645 dev
->stats
.rx_length_errors
++;
648 truesize
= (unsigned long)ctx
;
649 head_skb
= page_to_skb(vi
, rq
, page
, offset
, len
, truesize
);
652 if (unlikely(!curr_skb
))
657 buf
= virtqueue_get_buf_ctx(rq
->vq
, &len
, &ctx
);
658 if (unlikely(!ctx
)) {
659 pr_debug("%s: rx error: %d buffers out of %d missing\n",
661 virtio16_to_cpu(vi
->vdev
,
663 dev
->stats
.rx_length_errors
++;
667 page
= virt_to_head_page(buf
);
668 if (unlikely(len
> (unsigned long)ctx
)) {
669 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
670 dev
->name
, len
, (unsigned long)ctx
);
671 dev
->stats
.rx_length_errors
++;
674 truesize
= (unsigned long)ctx
;
676 num_skb_frags
= skb_shinfo(curr_skb
)->nr_frags
;
677 if (unlikely(num_skb_frags
== MAX_SKB_FRAGS
)) {
678 struct sk_buff
*nskb
= alloc_skb(0, GFP_ATOMIC
);
682 if (curr_skb
== head_skb
)
683 skb_shinfo(curr_skb
)->frag_list
= nskb
;
685 curr_skb
->next
= nskb
;
687 head_skb
->truesize
+= nskb
->truesize
;
690 if (curr_skb
!= head_skb
) {
691 head_skb
->data_len
+= len
;
692 head_skb
->len
+= len
;
693 head_skb
->truesize
+= truesize
;
695 offset
= buf
- page_address(page
);
696 if (skb_can_coalesce(curr_skb
, num_skb_frags
, page
, offset
)) {
698 skb_coalesce_rx_frag(curr_skb
, num_skb_frags
- 1,
701 skb_add_rx_frag(curr_skb
, num_skb_frags
, page
,
702 offset
, len
, truesize
);
706 ewma_pkt_len_add(&rq
->mrg_avg_pkt_len
, head_skb
->len
);
714 buf
= virtqueue_get_buf(rq
->vq
, &len
);
715 if (unlikely(!buf
)) {
716 pr_debug("%s: rx error: %d buffers missing\n",
718 dev
->stats
.rx_length_errors
++;
721 page
= virt_to_head_page(buf
);
725 dev
->stats
.rx_dropped
++;
726 dev_kfree_skb(head_skb
);
731 static int receive_buf(struct virtnet_info
*vi
, struct receive_queue
*rq
,
732 void *buf
, unsigned int len
, void **ctx
)
734 struct net_device
*dev
= vi
->dev
;
736 struct virtio_net_hdr_mrg_rxbuf
*hdr
;
739 if (unlikely(len
< vi
->hdr_len
+ ETH_HLEN
)) {
740 pr_debug("%s: short packet %i\n", dev
->name
, len
);
741 dev
->stats
.rx_length_errors
++;
742 if (vi
->mergeable_rx_bufs
) {
743 put_page(virt_to_head_page(buf
));
744 } else if (vi
->big_packets
) {
747 put_page(virt_to_head_page(buf
));
752 if (vi
->mergeable_rx_bufs
)
753 skb
= receive_mergeable(dev
, vi
, rq
, buf
, ctx
, len
);
754 else if (vi
->big_packets
)
755 skb
= receive_big(dev
, vi
, rq
, buf
, len
);
757 skb
= receive_small(dev
, vi
, rq
, buf
, len
);
762 hdr
= skb_vnet_hdr(skb
);
766 if (hdr
->hdr
.flags
& VIRTIO_NET_HDR_F_DATA_VALID
)
767 skb
->ip_summed
= CHECKSUM_UNNECESSARY
;
769 if (virtio_net_hdr_to_skb(skb
, &hdr
->hdr
,
770 virtio_is_little_endian(vi
->vdev
))) {
771 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
772 dev
->name
, hdr
->hdr
.gso_type
,
777 skb
->protocol
= eth_type_trans(skb
, dev
);
778 pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
779 ntohs(skb
->protocol
), skb
->len
, skb
->pkt_type
);
781 napi_gro_receive(&rq
->napi
, skb
);
785 dev
->stats
.rx_frame_errors
++;
790 static int add_recvbuf_small(struct virtnet_info
*vi
, struct receive_queue
*rq
,
793 struct page_frag
*alloc_frag
= &rq
->alloc_frag
;
795 unsigned int xdp_headroom
= virtnet_get_headroom(vi
);
796 int len
= vi
->hdr_len
+ VIRTNET_RX_PAD
+ GOOD_PACKET_LEN
+ xdp_headroom
;
799 len
= SKB_DATA_ALIGN(len
) +
800 SKB_DATA_ALIGN(sizeof(struct skb_shared_info
));
801 if (unlikely(!skb_page_frag_refill(len
, alloc_frag
, gfp
)))
804 buf
= (char *)page_address(alloc_frag
->page
) + alloc_frag
->offset
;
805 get_page(alloc_frag
->page
);
806 alloc_frag
->offset
+= len
;
807 sg_init_one(rq
->sg
, buf
+ VIRTNET_RX_PAD
+ xdp_headroom
,
808 vi
->hdr_len
+ GOOD_PACKET_LEN
);
809 err
= virtqueue_add_inbuf(rq
->vq
, rq
->sg
, 1, buf
, gfp
);
811 put_page(virt_to_head_page(buf
));
816 static int add_recvbuf_big(struct virtnet_info
*vi
, struct receive_queue
*rq
,
819 struct page
*first
, *list
= NULL
;
823 sg_init_table(rq
->sg
, MAX_SKB_FRAGS
+ 2);
825 /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
826 for (i
= MAX_SKB_FRAGS
+ 1; i
> 1; --i
) {
827 first
= get_a_page(rq
, gfp
);
830 give_pages(rq
, list
);
833 sg_set_buf(&rq
->sg
[i
], page_address(first
), PAGE_SIZE
);
835 /* chain new page in list head to match sg */
836 first
->private = (unsigned long)list
;
840 first
= get_a_page(rq
, gfp
);
842 give_pages(rq
, list
);
845 p
= page_address(first
);
847 /* rq->sg[0], rq->sg[1] share the same page */
848 /* a separated rq->sg[0] for header - required in case !any_header_sg */
849 sg_set_buf(&rq
->sg
[0], p
, vi
->hdr_len
);
851 /* rq->sg[1] for data packet, from offset */
852 offset
= sizeof(struct padded_vnet_hdr
);
853 sg_set_buf(&rq
->sg
[1], p
+ offset
, PAGE_SIZE
- offset
);
855 /* chain first in list head */
856 first
->private = (unsigned long)list
;
857 err
= virtqueue_add_inbuf(rq
->vq
, rq
->sg
, MAX_SKB_FRAGS
+ 2,
860 give_pages(rq
, first
);
865 static unsigned int get_mergeable_buf_len(struct receive_queue
*rq
,
866 struct ewma_pkt_len
*avg_pkt_len
)
868 const size_t hdr_len
= sizeof(struct virtio_net_hdr_mrg_rxbuf
);
871 len
= hdr_len
+ clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len
),
872 rq
->min_buf_len
, PAGE_SIZE
- hdr_len
);
873 return ALIGN(len
, L1_CACHE_BYTES
);
876 static int add_recvbuf_mergeable(struct virtnet_info
*vi
,
877 struct receive_queue
*rq
, gfp_t gfp
)
879 struct page_frag
*alloc_frag
= &rq
->alloc_frag
;
880 unsigned int headroom
= virtnet_get_headroom(vi
);
884 unsigned int len
, hole
;
886 len
= get_mergeable_buf_len(rq
, &rq
->mrg_avg_pkt_len
);
887 if (unlikely(!skb_page_frag_refill(len
+ headroom
, alloc_frag
, gfp
)))
890 buf
= (char *)page_address(alloc_frag
->page
) + alloc_frag
->offset
;
891 buf
+= headroom
; /* advance address leaving hole at front of pkt */
892 ctx
= (void *)(unsigned long)len
;
893 get_page(alloc_frag
->page
);
894 alloc_frag
->offset
+= len
+ headroom
;
895 hole
= alloc_frag
->size
- alloc_frag
->offset
;
896 if (hole
< len
+ headroom
) {
897 /* To avoid internal fragmentation, if there is very likely not
898 * enough space for another buffer, add the remaining space to
899 * the current buffer. This extra space is not included in
900 * the truesize stored in ctx.
903 alloc_frag
->offset
+= hole
;
906 sg_init_one(rq
->sg
, buf
, len
);
907 err
= virtqueue_add_inbuf_ctx(rq
->vq
, rq
->sg
, 1, buf
, ctx
, gfp
);
909 put_page(virt_to_head_page(buf
));
915 * Returns false if we couldn't fill entirely (OOM).
917 * Normally run in the receive path, but can also be run from ndo_open
918 * before we're receiving packets, or from refill_work which is
919 * careful to disable receiving (using napi_disable).
921 static bool try_fill_recv(struct virtnet_info
*vi
, struct receive_queue
*rq
,
929 if (vi
->mergeable_rx_bufs
)
930 err
= add_recvbuf_mergeable(vi
, rq
, gfp
);
931 else if (vi
->big_packets
)
932 err
= add_recvbuf_big(vi
, rq
, gfp
);
934 err
= add_recvbuf_small(vi
, rq
, gfp
);
936 oom
= err
== -ENOMEM
;
939 } while (rq
->vq
->num_free
);
940 virtqueue_kick(rq
->vq
);
944 static void skb_recv_done(struct virtqueue
*rvq
)
946 struct virtnet_info
*vi
= rvq
->vdev
->priv
;
947 struct receive_queue
*rq
= &vi
->rq
[vq2rxq(rvq
)];
949 virtqueue_napi_schedule(&rq
->napi
, rvq
);
952 static void virtnet_napi_enable(struct virtqueue
*vq
, struct napi_struct
*napi
)
956 /* If all buffers were filled by other side before we napi_enabled, we
957 * won't get another interrupt, so process any outstanding packets now.
958 * Call local_bh_enable after to trigger softIRQ processing.
961 virtqueue_napi_schedule(napi
, vq
);
965 static void virtnet_napi_tx_enable(struct virtnet_info
*vi
,
966 struct virtqueue
*vq
,
967 struct napi_struct
*napi
)
972 /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
973 * enable the feature if this is likely affine with the transmit path.
975 if (!vi
->affinity_hint_set
) {
980 return virtnet_napi_enable(vq
, napi
);
983 static void virtnet_napi_tx_disable(struct napi_struct
*napi
)
989 static void refill_work(struct work_struct
*work
)
991 struct virtnet_info
*vi
=
992 container_of(work
, struct virtnet_info
, refill
.work
);
996 for (i
= 0; i
< vi
->curr_queue_pairs
; i
++) {
997 struct receive_queue
*rq
= &vi
->rq
[i
];
999 napi_disable(&rq
->napi
);
1000 still_empty
= !try_fill_recv(vi
, rq
, GFP_KERNEL
);
1001 virtnet_napi_enable(rq
->vq
, &rq
->napi
);
1003 /* In theory, this can happen: if we don't get any buffers in
1004 * we will *never* try to fill again.
1007 schedule_delayed_work(&vi
->refill
, HZ
/2);
1011 static int virtnet_receive(struct receive_queue
*rq
, int budget
)
1013 struct virtnet_info
*vi
= rq
->vq
->vdev
->priv
;
1014 unsigned int len
, received
= 0, bytes
= 0;
1016 struct virtnet_stats
*stats
= this_cpu_ptr(vi
->stats
);
1018 if (vi
->mergeable_rx_bufs
) {
1021 while (received
< budget
&&
1022 (buf
= virtqueue_get_buf_ctx(rq
->vq
, &len
, &ctx
))) {
1023 bytes
+= receive_buf(vi
, rq
, buf
, len
, ctx
);
1027 while (received
< budget
&&
1028 (buf
= virtqueue_get_buf(rq
->vq
, &len
)) != NULL
) {
1029 bytes
+= receive_buf(vi
, rq
, buf
, len
, NULL
);
1034 if (rq
->vq
->num_free
> virtqueue_get_vring_size(rq
->vq
) / 2) {
1035 if (!try_fill_recv(vi
, rq
, GFP_ATOMIC
))
1036 schedule_delayed_work(&vi
->refill
, 0);
1039 u64_stats_update_begin(&stats
->rx_syncp
);
1040 stats
->rx_bytes
+= bytes
;
1041 stats
->rx_packets
+= received
;
1042 u64_stats_update_end(&stats
->rx_syncp
);
1047 static void free_old_xmit_skbs(struct send_queue
*sq
)
1049 struct sk_buff
*skb
;
1051 struct virtnet_info
*vi
= sq
->vq
->vdev
->priv
;
1052 struct virtnet_stats
*stats
= this_cpu_ptr(vi
->stats
);
1053 unsigned int packets
= 0;
1054 unsigned int bytes
= 0;
1056 while ((skb
= virtqueue_get_buf(sq
->vq
, &len
)) != NULL
) {
1057 pr_debug("Sent skb %p\n", skb
);
1062 dev_kfree_skb_any(skb
);
1065 /* Avoid overhead when no packets have been processed
1066 * happens when called speculatively from start_xmit.
1071 u64_stats_update_begin(&stats
->tx_syncp
);
1072 stats
->tx_bytes
+= bytes
;
1073 stats
->tx_packets
+= packets
;
1074 u64_stats_update_end(&stats
->tx_syncp
);
1077 static void virtnet_poll_cleantx(struct receive_queue
*rq
)
1079 struct virtnet_info
*vi
= rq
->vq
->vdev
->priv
;
1080 unsigned int index
= vq2rxq(rq
->vq
);
1081 struct send_queue
*sq
= &vi
->sq
[index
];
1082 struct netdev_queue
*txq
= netdev_get_tx_queue(vi
->dev
, index
);
1084 if (!sq
->napi
.weight
)
1087 if (__netif_tx_trylock(txq
)) {
1088 free_old_xmit_skbs(sq
);
1089 __netif_tx_unlock(txq
);
1092 if (sq
->vq
->num_free
>= 2 + MAX_SKB_FRAGS
)
1093 netif_tx_wake_queue(txq
);
1096 static int virtnet_poll(struct napi_struct
*napi
, int budget
)
1098 struct receive_queue
*rq
=
1099 container_of(napi
, struct receive_queue
, napi
);
1100 unsigned int received
;
1102 virtnet_poll_cleantx(rq
);
1104 received
= virtnet_receive(rq
, budget
);
1106 /* Out of packets? */
1107 if (received
< budget
)
1108 virtqueue_napi_complete(napi
, rq
->vq
, received
);
1113 static int virtnet_open(struct net_device
*dev
)
1115 struct virtnet_info
*vi
= netdev_priv(dev
);
1118 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1119 if (i
< vi
->curr_queue_pairs
)
1120 /* Make sure we have some buffers: if oom use wq. */
1121 if (!try_fill_recv(vi
, &vi
->rq
[i
], GFP_KERNEL
))
1122 schedule_delayed_work(&vi
->refill
, 0);
1123 virtnet_napi_enable(vi
->rq
[i
].vq
, &vi
->rq
[i
].napi
);
1124 virtnet_napi_tx_enable(vi
, vi
->sq
[i
].vq
, &vi
->sq
[i
].napi
);
1130 static int virtnet_poll_tx(struct napi_struct
*napi
, int budget
)
1132 struct send_queue
*sq
= container_of(napi
, struct send_queue
, napi
);
1133 struct virtnet_info
*vi
= sq
->vq
->vdev
->priv
;
1134 struct netdev_queue
*txq
= netdev_get_tx_queue(vi
->dev
, vq2txq(sq
->vq
));
1136 __netif_tx_lock(txq
, raw_smp_processor_id());
1137 free_old_xmit_skbs(sq
);
1138 __netif_tx_unlock(txq
);
1140 virtqueue_napi_complete(napi
, sq
->vq
, 0);
1142 if (sq
->vq
->num_free
>= 2 + MAX_SKB_FRAGS
)
1143 netif_tx_wake_queue(txq
);
1148 static int xmit_skb(struct send_queue
*sq
, struct sk_buff
*skb
)
1150 struct virtio_net_hdr_mrg_rxbuf
*hdr
;
1151 const unsigned char *dest
= ((struct ethhdr
*)skb
->data
)->h_dest
;
1152 struct virtnet_info
*vi
= sq
->vq
->vdev
->priv
;
1154 unsigned hdr_len
= vi
->hdr_len
;
1157 pr_debug("%s: xmit %p %pM\n", vi
->dev
->name
, skb
, dest
);
1159 can_push
= vi
->any_header_sg
&&
1160 !((unsigned long)skb
->data
& (__alignof__(*hdr
) - 1)) &&
1161 !skb_header_cloned(skb
) && skb_headroom(skb
) >= hdr_len
;
1162 /* Even if we can, don't push here yet as this would skew
1163 * csum_start offset below. */
1165 hdr
= (struct virtio_net_hdr_mrg_rxbuf
*)(skb
->data
- hdr_len
);
1167 hdr
= skb_vnet_hdr(skb
);
1169 if (virtio_net_hdr_from_skb(skb
, &hdr
->hdr
,
1170 virtio_is_little_endian(vi
->vdev
), false))
1173 if (vi
->mergeable_rx_bufs
)
1174 hdr
->num_buffers
= 0;
1176 sg_init_table(sq
->sg
, skb_shinfo(skb
)->nr_frags
+ (can_push
? 1 : 2));
1178 __skb_push(skb
, hdr_len
);
1179 num_sg
= skb_to_sgvec(skb
, sq
->sg
, 0, skb
->len
);
1180 if (unlikely(num_sg
< 0))
1182 /* Pull header back to avoid skew in tx bytes calculations. */
1183 __skb_pull(skb
, hdr_len
);
1185 sg_set_buf(sq
->sg
, hdr
, hdr_len
);
1186 num_sg
= skb_to_sgvec(skb
, sq
->sg
+ 1, 0, skb
->len
);
1187 if (unlikely(num_sg
< 0))
1191 return virtqueue_add_outbuf(sq
->vq
, sq
->sg
, num_sg
, skb
, GFP_ATOMIC
);
1194 static netdev_tx_t
start_xmit(struct sk_buff
*skb
, struct net_device
*dev
)
1196 struct virtnet_info
*vi
= netdev_priv(dev
);
1197 int qnum
= skb_get_queue_mapping(skb
);
1198 struct send_queue
*sq
= &vi
->sq
[qnum
];
1200 struct netdev_queue
*txq
= netdev_get_tx_queue(dev
, qnum
);
1201 bool kick
= !skb
->xmit_more
;
1202 bool use_napi
= sq
->napi
.weight
;
1204 /* Free up any pending old buffers before queueing new ones. */
1205 free_old_xmit_skbs(sq
);
1207 if (use_napi
&& kick
)
1208 virtqueue_enable_cb_delayed(sq
->vq
);
1210 /* timestamp packet in software */
1211 skb_tx_timestamp(skb
);
1213 /* Try to transmit */
1214 err
= xmit_skb(sq
, skb
);
1216 /* This should not happen! */
1217 if (unlikely(err
)) {
1218 dev
->stats
.tx_fifo_errors
++;
1219 if (net_ratelimit())
1221 "Unexpected TXQ (%d) queue failure: %d\n", qnum
, err
);
1222 dev
->stats
.tx_dropped
++;
1223 dev_kfree_skb_any(skb
);
1224 return NETDEV_TX_OK
;
1227 /* Don't wait up for transmitted skbs to be freed. */
1233 /* If running out of space, stop queue to avoid getting packets that we
1234 * are then unable to transmit.
1235 * An alternative would be to force queuing layer to requeue the skb by
1236 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1237 * returned in a normal path of operation: it means that driver is not
1238 * maintaining the TX queue stop/start state properly, and causes
1239 * the stack to do a non-trivial amount of useless work.
1240 * Since most packets only take 1 or 2 ring slots, stopping the queue
1241 * early means 16 slots are typically wasted.
1243 if (sq
->vq
->num_free
< 2+MAX_SKB_FRAGS
) {
1244 netif_stop_subqueue(dev
, qnum
);
1246 unlikely(!virtqueue_enable_cb_delayed(sq
->vq
))) {
1247 /* More just got used, free them then recheck. */
1248 free_old_xmit_skbs(sq
);
1249 if (sq
->vq
->num_free
>= 2+MAX_SKB_FRAGS
) {
1250 netif_start_subqueue(dev
, qnum
);
1251 virtqueue_disable_cb(sq
->vq
);
1256 if (kick
|| netif_xmit_stopped(txq
))
1257 virtqueue_kick(sq
->vq
);
1259 return NETDEV_TX_OK
;
1263 * Send command via the control virtqueue and check status. Commands
1264 * supported by the hypervisor, as indicated by feature bits, should
1265 * never fail unless improperly formatted.
1267 static bool virtnet_send_command(struct virtnet_info
*vi
, u8
class, u8 cmd
,
1268 struct scatterlist
*out
)
1270 struct scatterlist
*sgs
[4], hdr
, stat
;
1271 unsigned out_num
= 0, tmp
;
1273 /* Caller should know better */
1274 BUG_ON(!virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_CTRL_VQ
));
1276 vi
->ctrl_status
= ~0;
1277 vi
->ctrl_hdr
.class = class;
1278 vi
->ctrl_hdr
.cmd
= cmd
;
1280 sg_init_one(&hdr
, &vi
->ctrl_hdr
, sizeof(vi
->ctrl_hdr
));
1281 sgs
[out_num
++] = &hdr
;
1284 sgs
[out_num
++] = out
;
1286 /* Add return status. */
1287 sg_init_one(&stat
, &vi
->ctrl_status
, sizeof(vi
->ctrl_status
));
1288 sgs
[out_num
] = &stat
;
1290 BUG_ON(out_num
+ 1 > ARRAY_SIZE(sgs
));
1291 virtqueue_add_sgs(vi
->cvq
, sgs
, out_num
, 1, vi
, GFP_ATOMIC
);
1293 if (unlikely(!virtqueue_kick(vi
->cvq
)))
1294 return vi
->ctrl_status
== VIRTIO_NET_OK
;
1296 /* Spin for a response, the kick causes an ioport write, trapping
1297 * into the hypervisor, so the request should be handled immediately.
1299 while (!virtqueue_get_buf(vi
->cvq
, &tmp
) &&
1300 !virtqueue_is_broken(vi
->cvq
))
1303 return vi
->ctrl_status
== VIRTIO_NET_OK
;
1306 static int virtnet_set_mac_address(struct net_device
*dev
, void *p
)
1308 struct virtnet_info
*vi
= netdev_priv(dev
);
1309 struct virtio_device
*vdev
= vi
->vdev
;
1311 struct sockaddr
*addr
;
1312 struct scatterlist sg
;
1314 addr
= kmemdup(p
, sizeof(*addr
), GFP_KERNEL
);
1318 ret
= eth_prepare_mac_addr_change(dev
, addr
);
1322 if (virtio_has_feature(vdev
, VIRTIO_NET_F_CTRL_MAC_ADDR
)) {
1323 sg_init_one(&sg
, addr
->sa_data
, dev
->addr_len
);
1324 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_MAC
,
1325 VIRTIO_NET_CTRL_MAC_ADDR_SET
, &sg
)) {
1326 dev_warn(&vdev
->dev
,
1327 "Failed to set mac address by vq command.\n");
1331 } else if (virtio_has_feature(vdev
, VIRTIO_NET_F_MAC
) &&
1332 !virtio_has_feature(vdev
, VIRTIO_F_VERSION_1
)) {
1335 /* Naturally, this has an atomicity problem. */
1336 for (i
= 0; i
< dev
->addr_len
; i
++)
1337 virtio_cwrite8(vdev
,
1338 offsetof(struct virtio_net_config
, mac
) +
1339 i
, addr
->sa_data
[i
]);
1342 eth_commit_mac_addr_change(dev
, p
);
1350 static void virtnet_stats(struct net_device
*dev
,
1351 struct rtnl_link_stats64
*tot
)
1353 struct virtnet_info
*vi
= netdev_priv(dev
);
1357 for_each_possible_cpu(cpu
) {
1358 struct virtnet_stats
*stats
= per_cpu_ptr(vi
->stats
, cpu
);
1359 u64 tpackets
, tbytes
, rpackets
, rbytes
;
1362 start
= u64_stats_fetch_begin_irq(&stats
->tx_syncp
);
1363 tpackets
= stats
->tx_packets
;
1364 tbytes
= stats
->tx_bytes
;
1365 } while (u64_stats_fetch_retry_irq(&stats
->tx_syncp
, start
));
1368 start
= u64_stats_fetch_begin_irq(&stats
->rx_syncp
);
1369 rpackets
= stats
->rx_packets
;
1370 rbytes
= stats
->rx_bytes
;
1371 } while (u64_stats_fetch_retry_irq(&stats
->rx_syncp
, start
));
1373 tot
->rx_packets
+= rpackets
;
1374 tot
->tx_packets
+= tpackets
;
1375 tot
->rx_bytes
+= rbytes
;
1376 tot
->tx_bytes
+= tbytes
;
1379 tot
->tx_dropped
= dev
->stats
.tx_dropped
;
1380 tot
->tx_fifo_errors
= dev
->stats
.tx_fifo_errors
;
1381 tot
->rx_dropped
= dev
->stats
.rx_dropped
;
1382 tot
->rx_length_errors
= dev
->stats
.rx_length_errors
;
1383 tot
->rx_frame_errors
= dev
->stats
.rx_frame_errors
;
1386 #ifdef CONFIG_NET_POLL_CONTROLLER
1387 static void virtnet_netpoll(struct net_device
*dev
)
1389 struct virtnet_info
*vi
= netdev_priv(dev
);
1392 for (i
= 0; i
< vi
->curr_queue_pairs
; i
++)
1393 napi_schedule(&vi
->rq
[i
].napi
);
1397 static void virtnet_ack_link_announce(struct virtnet_info
*vi
)
1400 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_ANNOUNCE
,
1401 VIRTIO_NET_CTRL_ANNOUNCE_ACK
, NULL
))
1402 dev_warn(&vi
->dev
->dev
, "Failed to ack link announce.\n");
1406 static int _virtnet_set_queues(struct virtnet_info
*vi
, u16 queue_pairs
)
1408 struct scatterlist sg
;
1409 struct net_device
*dev
= vi
->dev
;
1411 if (!vi
->has_cvq
|| !virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_MQ
))
1414 vi
->ctrl_mq
.virtqueue_pairs
= cpu_to_virtio16(vi
->vdev
, queue_pairs
);
1415 sg_init_one(&sg
, &vi
->ctrl_mq
, sizeof(vi
->ctrl_mq
));
1417 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_MQ
,
1418 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET
, &sg
)) {
1419 dev_warn(&dev
->dev
, "Fail to set num of queue pairs to %d\n",
1423 vi
->curr_queue_pairs
= queue_pairs
;
1424 /* virtnet_open() will refill when device is going to up. */
1425 if (dev
->flags
& IFF_UP
)
1426 schedule_delayed_work(&vi
->refill
, 0);
1432 static int virtnet_set_queues(struct virtnet_info
*vi
, u16 queue_pairs
)
1437 err
= _virtnet_set_queues(vi
, queue_pairs
);
1442 static int virtnet_close(struct net_device
*dev
)
1444 struct virtnet_info
*vi
= netdev_priv(dev
);
1447 /* Make sure refill_work doesn't re-enable napi! */
1448 cancel_delayed_work_sync(&vi
->refill
);
1450 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1451 napi_disable(&vi
->rq
[i
].napi
);
1452 virtnet_napi_tx_disable(&vi
->sq
[i
].napi
);
1458 static void virtnet_set_rx_mode(struct net_device
*dev
)
1460 struct virtnet_info
*vi
= netdev_priv(dev
);
1461 struct scatterlist sg
[2];
1462 struct virtio_net_ctrl_mac
*mac_data
;
1463 struct netdev_hw_addr
*ha
;
1469 /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1470 if (!virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_CTRL_RX
))
1473 vi
->ctrl_promisc
= ((dev
->flags
& IFF_PROMISC
) != 0);
1474 vi
->ctrl_allmulti
= ((dev
->flags
& IFF_ALLMULTI
) != 0);
1476 sg_init_one(sg
, &vi
->ctrl_promisc
, sizeof(vi
->ctrl_promisc
));
1478 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_RX
,
1479 VIRTIO_NET_CTRL_RX_PROMISC
, sg
))
1480 dev_warn(&dev
->dev
, "Failed to %sable promisc mode.\n",
1481 vi
->ctrl_promisc
? "en" : "dis");
1483 sg_init_one(sg
, &vi
->ctrl_allmulti
, sizeof(vi
->ctrl_allmulti
));
1485 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_RX
,
1486 VIRTIO_NET_CTRL_RX_ALLMULTI
, sg
))
1487 dev_warn(&dev
->dev
, "Failed to %sable allmulti mode.\n",
1488 vi
->ctrl_allmulti
? "en" : "dis");
1490 uc_count
= netdev_uc_count(dev
);
1491 mc_count
= netdev_mc_count(dev
);
1492 /* MAC filter - use one buffer for both lists */
1493 buf
= kzalloc(((uc_count
+ mc_count
) * ETH_ALEN
) +
1494 (2 * sizeof(mac_data
->entries
)), GFP_ATOMIC
);
1499 sg_init_table(sg
, 2);
1501 /* Store the unicast list and count in the front of the buffer */
1502 mac_data
->entries
= cpu_to_virtio32(vi
->vdev
, uc_count
);
1504 netdev_for_each_uc_addr(ha
, dev
)
1505 memcpy(&mac_data
->macs
[i
++][0], ha
->addr
, ETH_ALEN
);
1507 sg_set_buf(&sg
[0], mac_data
,
1508 sizeof(mac_data
->entries
) + (uc_count
* ETH_ALEN
));
1510 /* multicast list and count fill the end */
1511 mac_data
= (void *)&mac_data
->macs
[uc_count
][0];
1513 mac_data
->entries
= cpu_to_virtio32(vi
->vdev
, mc_count
);
1515 netdev_for_each_mc_addr(ha
, dev
)
1516 memcpy(&mac_data
->macs
[i
++][0], ha
->addr
, ETH_ALEN
);
1518 sg_set_buf(&sg
[1], mac_data
,
1519 sizeof(mac_data
->entries
) + (mc_count
* ETH_ALEN
));
1521 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_MAC
,
1522 VIRTIO_NET_CTRL_MAC_TABLE_SET
, sg
))
1523 dev_warn(&dev
->dev
, "Failed to set MAC filter table.\n");
1528 static int virtnet_vlan_rx_add_vid(struct net_device
*dev
,
1529 __be16 proto
, u16 vid
)
1531 struct virtnet_info
*vi
= netdev_priv(dev
);
1532 struct scatterlist sg
;
1535 sg_init_one(&sg
, &vi
->ctrl_vid
, sizeof(vi
->ctrl_vid
));
1537 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_VLAN
,
1538 VIRTIO_NET_CTRL_VLAN_ADD
, &sg
))
1539 dev_warn(&dev
->dev
, "Failed to add VLAN ID %d.\n", vid
);
1543 static int virtnet_vlan_rx_kill_vid(struct net_device
*dev
,
1544 __be16 proto
, u16 vid
)
1546 struct virtnet_info
*vi
= netdev_priv(dev
);
1547 struct scatterlist sg
;
1550 sg_init_one(&sg
, &vi
->ctrl_vid
, sizeof(vi
->ctrl_vid
));
1552 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_VLAN
,
1553 VIRTIO_NET_CTRL_VLAN_DEL
, &sg
))
1554 dev_warn(&dev
->dev
, "Failed to kill VLAN ID %d.\n", vid
);
1558 static void virtnet_clean_affinity(struct virtnet_info
*vi
, long hcpu
)
1562 if (vi
->affinity_hint_set
) {
1563 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1564 virtqueue_set_affinity(vi
->rq
[i
].vq
, -1);
1565 virtqueue_set_affinity(vi
->sq
[i
].vq
, -1);
1568 vi
->affinity_hint_set
= false;
1572 static void virtnet_set_affinity(struct virtnet_info
*vi
)
1577 /* In multiqueue mode, when the number of cpu is equal to the number of
1578 * queue pairs, we let the queue pairs to be private to one cpu by
1579 * setting the affinity hint to eliminate the contention.
1581 if (vi
->curr_queue_pairs
== 1 ||
1582 vi
->max_queue_pairs
!= num_online_cpus()) {
1583 virtnet_clean_affinity(vi
, -1);
1588 for_each_online_cpu(cpu
) {
1589 virtqueue_set_affinity(vi
->rq
[i
].vq
, cpu
);
1590 virtqueue_set_affinity(vi
->sq
[i
].vq
, cpu
);
1591 netif_set_xps_queue(vi
->dev
, cpumask_of(cpu
), i
);
1595 vi
->affinity_hint_set
= true;
1598 static int virtnet_cpu_online(unsigned int cpu
, struct hlist_node
*node
)
1600 struct virtnet_info
*vi
= hlist_entry_safe(node
, struct virtnet_info
,
1602 virtnet_set_affinity(vi
);
1606 static int virtnet_cpu_dead(unsigned int cpu
, struct hlist_node
*node
)
1608 struct virtnet_info
*vi
= hlist_entry_safe(node
, struct virtnet_info
,
1610 virtnet_set_affinity(vi
);
1614 static int virtnet_cpu_down_prep(unsigned int cpu
, struct hlist_node
*node
)
1616 struct virtnet_info
*vi
= hlist_entry_safe(node
, struct virtnet_info
,
1619 virtnet_clean_affinity(vi
, cpu
);
1623 static enum cpuhp_state virtionet_online
;
1625 static int virtnet_cpu_notif_add(struct virtnet_info
*vi
)
1629 ret
= cpuhp_state_add_instance_nocalls(virtionet_online
, &vi
->node
);
1632 ret
= cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD
,
1636 cpuhp_state_remove_instance_nocalls(virtionet_online
, &vi
->node
);
1640 static void virtnet_cpu_notif_remove(struct virtnet_info
*vi
)
1642 cpuhp_state_remove_instance_nocalls(virtionet_online
, &vi
->node
);
1643 cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD
,
1647 static void virtnet_get_ringparam(struct net_device
*dev
,
1648 struct ethtool_ringparam
*ring
)
1650 struct virtnet_info
*vi
= netdev_priv(dev
);
1652 ring
->rx_max_pending
= virtqueue_get_vring_size(vi
->rq
[0].vq
);
1653 ring
->tx_max_pending
= virtqueue_get_vring_size(vi
->sq
[0].vq
);
1654 ring
->rx_pending
= ring
->rx_max_pending
;
1655 ring
->tx_pending
= ring
->tx_max_pending
;
1659 static void virtnet_get_drvinfo(struct net_device
*dev
,
1660 struct ethtool_drvinfo
*info
)
1662 struct virtnet_info
*vi
= netdev_priv(dev
);
1663 struct virtio_device
*vdev
= vi
->vdev
;
1665 strlcpy(info
->driver
, KBUILD_MODNAME
, sizeof(info
->driver
));
1666 strlcpy(info
->version
, VIRTNET_DRIVER_VERSION
, sizeof(info
->version
));
1667 strlcpy(info
->bus_info
, virtio_bus_name(vdev
), sizeof(info
->bus_info
));
1671 /* TODO: Eliminate OOO packets during switching */
1672 static int virtnet_set_channels(struct net_device
*dev
,
1673 struct ethtool_channels
*channels
)
1675 struct virtnet_info
*vi
= netdev_priv(dev
);
1676 u16 queue_pairs
= channels
->combined_count
;
1679 /* We don't support separate rx/tx channels.
1680 * We don't allow setting 'other' channels.
1682 if (channels
->rx_count
|| channels
->tx_count
|| channels
->other_count
)
1685 if (queue_pairs
> vi
->max_queue_pairs
|| queue_pairs
== 0)
1688 /* For now we don't support modifying channels while XDP is loaded
1689 * also when XDP is loaded all RX queues have XDP programs so we only
1690 * need to check a single RX queue.
1692 if (vi
->rq
[0].xdp_prog
)
1696 err
= _virtnet_set_queues(vi
, queue_pairs
);
1698 netif_set_real_num_tx_queues(dev
, queue_pairs
);
1699 netif_set_real_num_rx_queues(dev
, queue_pairs
);
1701 virtnet_set_affinity(vi
);
1708 static void virtnet_get_channels(struct net_device
*dev
,
1709 struct ethtool_channels
*channels
)
1711 struct virtnet_info
*vi
= netdev_priv(dev
);
1713 channels
->combined_count
= vi
->curr_queue_pairs
;
1714 channels
->max_combined
= vi
->max_queue_pairs
;
1715 channels
->max_other
= 0;
1716 channels
->rx_count
= 0;
1717 channels
->tx_count
= 0;
1718 channels
->other_count
= 0;
1721 /* Check if the user is trying to change anything besides speed/duplex */
1723 virtnet_validate_ethtool_cmd(const struct ethtool_link_ksettings
*cmd
)
1725 struct ethtool_link_ksettings diff1
= *cmd
;
1726 struct ethtool_link_ksettings diff2
= {};
1728 /* cmd is always set so we need to clear it, validate the port type
1729 * and also without autonegotiation we can ignore advertising
1731 diff1
.base
.speed
= 0;
1732 diff2
.base
.port
= PORT_OTHER
;
1733 ethtool_link_ksettings_zero_link_mode(&diff1
, advertising
);
1734 diff1
.base
.duplex
= 0;
1736 diff1
.base
.link_mode_masks_nwords
= 0;
1738 return !memcmp(&diff1
.base
, &diff2
.base
, sizeof(diff1
.base
)) &&
1739 bitmap_empty(diff1
.link_modes
.supported
,
1740 __ETHTOOL_LINK_MODE_MASK_NBITS
) &&
1741 bitmap_empty(diff1
.link_modes
.advertising
,
1742 __ETHTOOL_LINK_MODE_MASK_NBITS
) &&
1743 bitmap_empty(diff1
.link_modes
.lp_advertising
,
1744 __ETHTOOL_LINK_MODE_MASK_NBITS
);
1747 static int virtnet_set_link_ksettings(struct net_device
*dev
,
1748 const struct ethtool_link_ksettings
*cmd
)
1750 struct virtnet_info
*vi
= netdev_priv(dev
);
1753 speed
= cmd
->base
.speed
;
1754 /* don't allow custom speed and duplex */
1755 if (!ethtool_validate_speed(speed
) ||
1756 !ethtool_validate_duplex(cmd
->base
.duplex
) ||
1757 !virtnet_validate_ethtool_cmd(cmd
))
1760 vi
->duplex
= cmd
->base
.duplex
;
1765 static int virtnet_get_link_ksettings(struct net_device
*dev
,
1766 struct ethtool_link_ksettings
*cmd
)
1768 struct virtnet_info
*vi
= netdev_priv(dev
);
1770 cmd
->base
.speed
= vi
->speed
;
1771 cmd
->base
.duplex
= vi
->duplex
;
1772 cmd
->base
.port
= PORT_OTHER
;
1777 static void virtnet_init_settings(struct net_device
*dev
)
1779 struct virtnet_info
*vi
= netdev_priv(dev
);
1781 vi
->speed
= SPEED_UNKNOWN
;
1782 vi
->duplex
= DUPLEX_UNKNOWN
;
1785 static const struct ethtool_ops virtnet_ethtool_ops
= {
1786 .get_drvinfo
= virtnet_get_drvinfo
,
1787 .get_link
= ethtool_op_get_link
,
1788 .get_ringparam
= virtnet_get_ringparam
,
1789 .set_channels
= virtnet_set_channels
,
1790 .get_channels
= virtnet_get_channels
,
1791 .get_ts_info
= ethtool_op_get_ts_info
,
1792 .get_link_ksettings
= virtnet_get_link_ksettings
,
1793 .set_link_ksettings
= virtnet_set_link_ksettings
,
1796 static void virtnet_freeze_down(struct virtio_device
*vdev
)
1798 struct virtnet_info
*vi
= vdev
->priv
;
1801 /* Make sure no work handler is accessing the device */
1802 flush_work(&vi
->config_work
);
1804 netif_device_detach(vi
->dev
);
1805 cancel_delayed_work_sync(&vi
->refill
);
1807 if (netif_running(vi
->dev
)) {
1808 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1809 napi_disable(&vi
->rq
[i
].napi
);
1810 virtnet_napi_tx_disable(&vi
->sq
[i
].napi
);
1815 static int init_vqs(struct virtnet_info
*vi
);
1816 static void _remove_vq_common(struct virtnet_info
*vi
);
1818 static int virtnet_restore_up(struct virtio_device
*vdev
)
1820 struct virtnet_info
*vi
= vdev
->priv
;
1827 virtio_device_ready(vdev
);
1829 if (netif_running(vi
->dev
)) {
1830 for (i
= 0; i
< vi
->curr_queue_pairs
; i
++)
1831 if (!try_fill_recv(vi
, &vi
->rq
[i
], GFP_KERNEL
))
1832 schedule_delayed_work(&vi
->refill
, 0);
1834 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1835 virtnet_napi_enable(vi
->rq
[i
].vq
, &vi
->rq
[i
].napi
);
1836 virtnet_napi_tx_enable(vi
, vi
->sq
[i
].vq
,
1841 netif_device_attach(vi
->dev
);
1845 static int virtnet_reset(struct virtnet_info
*vi
, int curr_qp
, int xdp_qp
)
1847 struct virtio_device
*dev
= vi
->vdev
;
1850 virtio_config_disable(dev
);
1851 dev
->failed
= dev
->config
->get_status(dev
) & VIRTIO_CONFIG_S_FAILED
;
1852 virtnet_freeze_down(dev
);
1853 _remove_vq_common(vi
);
1855 virtio_add_status(dev
, VIRTIO_CONFIG_S_ACKNOWLEDGE
);
1856 virtio_add_status(dev
, VIRTIO_CONFIG_S_DRIVER
);
1858 ret
= virtio_finalize_features(dev
);
1862 vi
->xdp_queue_pairs
= xdp_qp
;
1863 ret
= virtnet_restore_up(dev
);
1866 ret
= _virtnet_set_queues(vi
, curr_qp
);
1870 virtio_add_status(dev
, VIRTIO_CONFIG_S_DRIVER_OK
);
1871 virtio_config_enable(dev
);
1874 virtio_add_status(dev
, VIRTIO_CONFIG_S_FAILED
);
1878 static int virtnet_xdp_set(struct net_device
*dev
, struct bpf_prog
*prog
,
1879 struct netlink_ext_ack
*extack
)
1881 unsigned long int max_sz
= PAGE_SIZE
- sizeof(struct padded_vnet_hdr
);
1882 struct virtnet_info
*vi
= netdev_priv(dev
);
1883 struct bpf_prog
*old_prog
;
1884 u16 xdp_qp
= 0, curr_qp
;
1887 if (virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_TSO4
) ||
1888 virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_TSO6
) ||
1889 virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_ECN
) ||
1890 virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_UFO
)) {
1891 NL_SET_ERR_MSG_MOD(extack
, "Can't set XDP while host is implementing LRO, disable LRO first");
1895 if (vi
->mergeable_rx_bufs
&& !vi
->any_header_sg
) {
1896 NL_SET_ERR_MSG_MOD(extack
, "XDP expects header/data in single page, any_header_sg required");
1900 if (dev
->mtu
> max_sz
) {
1901 NL_SET_ERR_MSG_MOD(extack
, "MTU too large to enable XDP");
1902 netdev_warn(dev
, "XDP requires MTU less than %lu\n", max_sz
);
1906 curr_qp
= vi
->curr_queue_pairs
- vi
->xdp_queue_pairs
;
1908 xdp_qp
= nr_cpu_ids
;
1910 /* XDP requires extra queues for XDP_TX */
1911 if (curr_qp
+ xdp_qp
> vi
->max_queue_pairs
) {
1912 NL_SET_ERR_MSG_MOD(extack
, "Too few free TX rings available");
1913 netdev_warn(dev
, "request %i queues but max is %i\n",
1914 curr_qp
+ xdp_qp
, vi
->max_queue_pairs
);
1919 prog
= bpf_prog_add(prog
, vi
->max_queue_pairs
- 1);
1921 return PTR_ERR(prog
);
1924 /* Changing the headroom in buffers is a disruptive operation because
1925 * existing buffers must be flushed and reallocated. This will happen
1926 * when a xdp program is initially added or xdp is disabled by removing
1927 * the xdp program resulting in number of XDP queues changing.
1929 if (vi
->xdp_queue_pairs
!= xdp_qp
) {
1930 err
= virtnet_reset(vi
, curr_qp
+ xdp_qp
, xdp_qp
);
1932 dev_warn(&dev
->dev
, "XDP reset failure.\n");
1933 goto virtio_reset_err
;
1937 netif_set_real_num_rx_queues(dev
, curr_qp
+ xdp_qp
);
1939 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1940 old_prog
= rtnl_dereference(vi
->rq
[i
].xdp_prog
);
1941 rcu_assign_pointer(vi
->rq
[i
].xdp_prog
, prog
);
1943 bpf_prog_put(old_prog
);
1949 /* On reset error do our best to unwind XDP changes inflight and return
1950 * error up to user space for resolution. The underlying reset hung on
1951 * us so not much we can do here.
1954 bpf_prog_sub(prog
, vi
->max_queue_pairs
- 1);
1958 static bool virtnet_xdp_query(struct net_device
*dev
)
1960 struct virtnet_info
*vi
= netdev_priv(dev
);
1963 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1964 if (vi
->rq
[i
].xdp_prog
)
1970 static int virtnet_xdp(struct net_device
*dev
, struct netdev_xdp
*xdp
)
1972 switch (xdp
->command
) {
1973 case XDP_SETUP_PROG
:
1974 return virtnet_xdp_set(dev
, xdp
->prog
, xdp
->extack
);
1975 case XDP_QUERY_PROG
:
1976 xdp
->prog_attached
= virtnet_xdp_query(dev
);
1983 static const struct net_device_ops virtnet_netdev
= {
1984 .ndo_open
= virtnet_open
,
1985 .ndo_stop
= virtnet_close
,
1986 .ndo_start_xmit
= start_xmit
,
1987 .ndo_validate_addr
= eth_validate_addr
,
1988 .ndo_set_mac_address
= virtnet_set_mac_address
,
1989 .ndo_set_rx_mode
= virtnet_set_rx_mode
,
1990 .ndo_get_stats64
= virtnet_stats
,
1991 .ndo_vlan_rx_add_vid
= virtnet_vlan_rx_add_vid
,
1992 .ndo_vlan_rx_kill_vid
= virtnet_vlan_rx_kill_vid
,
1993 #ifdef CONFIG_NET_POLL_CONTROLLER
1994 .ndo_poll_controller
= virtnet_netpoll
,
1996 .ndo_xdp
= virtnet_xdp
,
1997 .ndo_features_check
= passthru_features_check
,
2000 static void virtnet_config_changed_work(struct work_struct
*work
)
2002 struct virtnet_info
*vi
=
2003 container_of(work
, struct virtnet_info
, config_work
);
2006 if (virtio_cread_feature(vi
->vdev
, VIRTIO_NET_F_STATUS
,
2007 struct virtio_net_config
, status
, &v
) < 0)
2010 if (v
& VIRTIO_NET_S_ANNOUNCE
) {
2011 netdev_notify_peers(vi
->dev
);
2012 virtnet_ack_link_announce(vi
);
2015 /* Ignore unknown (future) status bits */
2016 v
&= VIRTIO_NET_S_LINK_UP
;
2018 if (vi
->status
== v
)
2023 if (vi
->status
& VIRTIO_NET_S_LINK_UP
) {
2024 netif_carrier_on(vi
->dev
);
2025 netif_tx_wake_all_queues(vi
->dev
);
2027 netif_carrier_off(vi
->dev
);
2028 netif_tx_stop_all_queues(vi
->dev
);
2032 static void virtnet_config_changed(struct virtio_device
*vdev
)
2034 struct virtnet_info
*vi
= vdev
->priv
;
2036 schedule_work(&vi
->config_work
);
2039 static void virtnet_free_queues(struct virtnet_info
*vi
)
2043 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2044 napi_hash_del(&vi
->rq
[i
].napi
);
2045 netif_napi_del(&vi
->rq
[i
].napi
);
2046 netif_napi_del(&vi
->sq
[i
].napi
);
2049 /* We called napi_hash_del() before netif_napi_del(),
2050 * we need to respect an RCU grace period before freeing vi->rq
2058 static void _free_receive_bufs(struct virtnet_info
*vi
)
2060 struct bpf_prog
*old_prog
;
2063 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2064 while (vi
->rq
[i
].pages
)
2065 __free_pages(get_a_page(&vi
->rq
[i
], GFP_KERNEL
), 0);
2067 old_prog
= rtnl_dereference(vi
->rq
[i
].xdp_prog
);
2068 RCU_INIT_POINTER(vi
->rq
[i
].xdp_prog
, NULL
);
2070 bpf_prog_put(old_prog
);
2074 static void free_receive_bufs(struct virtnet_info
*vi
)
2077 _free_receive_bufs(vi
);
2081 static void free_receive_page_frags(struct virtnet_info
*vi
)
2084 for (i
= 0; i
< vi
->max_queue_pairs
; i
++)
2085 if (vi
->rq
[i
].alloc_frag
.page
)
2086 put_page(vi
->rq
[i
].alloc_frag
.page
);
2089 static bool is_xdp_raw_buffer_queue(struct virtnet_info
*vi
, int q
)
2091 if (q
< (vi
->curr_queue_pairs
- vi
->xdp_queue_pairs
))
2093 else if (q
< vi
->curr_queue_pairs
)
2099 static void free_unused_bufs(struct virtnet_info
*vi
)
2104 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2105 struct virtqueue
*vq
= vi
->sq
[i
].vq
;
2106 while ((buf
= virtqueue_detach_unused_buf(vq
)) != NULL
) {
2107 if (!is_xdp_raw_buffer_queue(vi
, i
))
2110 put_page(virt_to_head_page(buf
));
2114 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2115 struct virtqueue
*vq
= vi
->rq
[i
].vq
;
2117 while ((buf
= virtqueue_detach_unused_buf(vq
)) != NULL
) {
2118 if (vi
->mergeable_rx_bufs
) {
2119 put_page(virt_to_head_page(buf
));
2120 } else if (vi
->big_packets
) {
2121 give_pages(&vi
->rq
[i
], buf
);
2123 put_page(virt_to_head_page(buf
));
2129 static void virtnet_del_vqs(struct virtnet_info
*vi
)
2131 struct virtio_device
*vdev
= vi
->vdev
;
2133 virtnet_clean_affinity(vi
, -1);
2135 vdev
->config
->del_vqs(vdev
);
2137 virtnet_free_queues(vi
);
2140 /* How large should a single buffer be so a queue full of these can fit at
2141 * least one full packet?
2142 * Logic below assumes the mergeable buffer header is used.
2144 static unsigned int mergeable_min_buf_len(struct virtnet_info
*vi
, struct virtqueue
*vq
)
2146 const unsigned int hdr_len
= sizeof(struct virtio_net_hdr_mrg_rxbuf
);
2147 unsigned int rq_size
= virtqueue_get_vring_size(vq
);
2148 unsigned int packet_len
= vi
->big_packets
? IP_MAX_MTU
: vi
->dev
->max_mtu
;
2149 unsigned int buf_len
= hdr_len
+ ETH_HLEN
+ VLAN_HLEN
+ packet_len
;
2150 unsigned int min_buf_len
= DIV_ROUND_UP(buf_len
, rq_size
);
2152 return max(max(min_buf_len
, hdr_len
) - hdr_len
,
2153 (unsigned int)GOOD_PACKET_LEN
);
2156 static int virtnet_find_vqs(struct virtnet_info
*vi
)
2158 vq_callback_t
**callbacks
;
2159 struct virtqueue
**vqs
;
2165 /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
2166 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
2167 * possible control vq.
2169 total_vqs
= vi
->max_queue_pairs
* 2 +
2170 virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_CTRL_VQ
);
2172 /* Allocate space for find_vqs parameters */
2173 vqs
= kzalloc(total_vqs
* sizeof(*vqs
), GFP_KERNEL
);
2176 callbacks
= kmalloc(total_vqs
* sizeof(*callbacks
), GFP_KERNEL
);
2179 names
= kmalloc(total_vqs
* sizeof(*names
), GFP_KERNEL
);
2182 if (vi
->mergeable_rx_bufs
) {
2183 ctx
= kzalloc(total_vqs
* sizeof(*ctx
), GFP_KERNEL
);
2190 /* Parameters for control virtqueue, if any */
2192 callbacks
[total_vqs
- 1] = NULL
;
2193 names
[total_vqs
- 1] = "control";
2196 /* Allocate/initialize parameters for send/receive virtqueues */
2197 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2198 callbacks
[rxq2vq(i
)] = skb_recv_done
;
2199 callbacks
[txq2vq(i
)] = skb_xmit_done
;
2200 sprintf(vi
->rq
[i
].name
, "input.%d", i
);
2201 sprintf(vi
->sq
[i
].name
, "output.%d", i
);
2202 names
[rxq2vq(i
)] = vi
->rq
[i
].name
;
2203 names
[txq2vq(i
)] = vi
->sq
[i
].name
;
2205 ctx
[rxq2vq(i
)] = true;
2208 ret
= vi
->vdev
->config
->find_vqs(vi
->vdev
, total_vqs
, vqs
, callbacks
,
2214 vi
->cvq
= vqs
[total_vqs
- 1];
2215 if (virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_CTRL_VLAN
))
2216 vi
->dev
->features
|= NETIF_F_HW_VLAN_CTAG_FILTER
;
2219 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2220 vi
->rq
[i
].vq
= vqs
[rxq2vq(i
)];
2221 vi
->rq
[i
].min_buf_len
= mergeable_min_buf_len(vi
, vi
->rq
[i
].vq
);
2222 vi
->sq
[i
].vq
= vqs
[txq2vq(i
)];
2243 static int virtnet_alloc_queues(struct virtnet_info
*vi
)
2247 vi
->sq
= kzalloc(sizeof(*vi
->sq
) * vi
->max_queue_pairs
, GFP_KERNEL
);
2250 vi
->rq
= kzalloc(sizeof(*vi
->rq
) * vi
->max_queue_pairs
, GFP_KERNEL
);
2254 INIT_DELAYED_WORK(&vi
->refill
, refill_work
);
2255 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2256 vi
->rq
[i
].pages
= NULL
;
2257 netif_napi_add(vi
->dev
, &vi
->rq
[i
].napi
, virtnet_poll
,
2259 netif_tx_napi_add(vi
->dev
, &vi
->sq
[i
].napi
, virtnet_poll_tx
,
2260 napi_tx
? napi_weight
: 0);
2262 sg_init_table(vi
->rq
[i
].sg
, ARRAY_SIZE(vi
->rq
[i
].sg
));
2263 ewma_pkt_len_init(&vi
->rq
[i
].mrg_avg_pkt_len
);
2264 sg_init_table(vi
->sq
[i
].sg
, ARRAY_SIZE(vi
->sq
[i
].sg
));
2275 static int init_vqs(struct virtnet_info
*vi
)
2279 /* Allocate send & receive queues */
2280 ret
= virtnet_alloc_queues(vi
);
2284 ret
= virtnet_find_vqs(vi
);
2289 virtnet_set_affinity(vi
);
2295 virtnet_free_queues(vi
);
2301 static ssize_t
mergeable_rx_buffer_size_show(struct netdev_rx_queue
*queue
,
2302 struct rx_queue_attribute
*attribute
, char *buf
)
2304 struct virtnet_info
*vi
= netdev_priv(queue
->dev
);
2305 unsigned int queue_index
= get_netdev_rx_queue_index(queue
);
2306 struct ewma_pkt_len
*avg
;
2308 BUG_ON(queue_index
>= vi
->max_queue_pairs
);
2309 avg
= &vi
->rq
[queue_index
].mrg_avg_pkt_len
;
2310 return sprintf(buf
, "%u\n",
2311 get_mergeable_buf_len(&vi
->rq
[queue_index
], avg
));
2314 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute
=
2315 __ATTR_RO(mergeable_rx_buffer_size
);
2317 static struct attribute
*virtio_net_mrg_rx_attrs
[] = {
2318 &mergeable_rx_buffer_size_attribute
.attr
,
2322 static const struct attribute_group virtio_net_mrg_rx_group
= {
2323 .name
= "virtio_net",
2324 .attrs
= virtio_net_mrg_rx_attrs
2328 static bool virtnet_fail_on_feature(struct virtio_device
*vdev
,
2330 const char *fname
, const char *dname
)
2332 if (!virtio_has_feature(vdev
, fbit
))
2335 dev_err(&vdev
->dev
, "device advertises feature %s but not %s",
2341 #define VIRTNET_FAIL_ON(vdev, fbit, dbit) \
2342 virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
2344 static bool virtnet_validate_features(struct virtio_device
*vdev
)
2346 if (!virtio_has_feature(vdev
, VIRTIO_NET_F_CTRL_VQ
) &&
2347 (VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_CTRL_RX
,
2348 "VIRTIO_NET_F_CTRL_VQ") ||
2349 VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_CTRL_VLAN
,
2350 "VIRTIO_NET_F_CTRL_VQ") ||
2351 VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_GUEST_ANNOUNCE
,
2352 "VIRTIO_NET_F_CTRL_VQ") ||
2353 VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_MQ
, "VIRTIO_NET_F_CTRL_VQ") ||
2354 VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_CTRL_MAC_ADDR
,
2355 "VIRTIO_NET_F_CTRL_VQ"))) {
2362 #define MIN_MTU ETH_MIN_MTU
2363 #define MAX_MTU ETH_MAX_MTU
2365 static int virtnet_validate(struct virtio_device
*vdev
)
2367 if (!vdev
->config
->get
) {
2368 dev_err(&vdev
->dev
, "%s failure: config access disabled\n",
2373 if (!virtnet_validate_features(vdev
))
2376 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MTU
)) {
2377 int mtu
= virtio_cread16(vdev
,
2378 offsetof(struct virtio_net_config
,
2381 __virtio_clear_bit(vdev
, VIRTIO_NET_F_MTU
);
2387 static int virtnet_probe(struct virtio_device
*vdev
)
2390 struct net_device
*dev
;
2391 struct virtnet_info
*vi
;
2392 u16 max_queue_pairs
;
2395 /* Find if host supports multiqueue virtio_net device */
2396 err
= virtio_cread_feature(vdev
, VIRTIO_NET_F_MQ
,
2397 struct virtio_net_config
,
2398 max_virtqueue_pairs
, &max_queue_pairs
);
2400 /* We need at least 2 queue's */
2401 if (err
|| max_queue_pairs
< VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN
||
2402 max_queue_pairs
> VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX
||
2403 !virtio_has_feature(vdev
, VIRTIO_NET_F_CTRL_VQ
))
2404 max_queue_pairs
= 1;
2406 /* Allocate ourselves a network device with room for our info */
2407 dev
= alloc_etherdev_mq(sizeof(struct virtnet_info
), max_queue_pairs
);
2411 /* Set up network device as normal. */
2412 dev
->priv_flags
|= IFF_UNICAST_FLT
| IFF_LIVE_ADDR_CHANGE
;
2413 dev
->netdev_ops
= &virtnet_netdev
;
2414 dev
->features
= NETIF_F_HIGHDMA
;
2416 dev
->ethtool_ops
= &virtnet_ethtool_ops
;
2417 SET_NETDEV_DEV(dev
, &vdev
->dev
);
2419 /* Do we support "hardware" checksums? */
2420 if (virtio_has_feature(vdev
, VIRTIO_NET_F_CSUM
)) {
2421 /* This opens up the world of extra features. */
2422 dev
->hw_features
|= NETIF_F_HW_CSUM
| NETIF_F_SG
;
2424 dev
->features
|= NETIF_F_HW_CSUM
| NETIF_F_SG
;
2426 if (virtio_has_feature(vdev
, VIRTIO_NET_F_GSO
)) {
2427 dev
->hw_features
|= NETIF_F_TSO
| NETIF_F_UFO
2428 | NETIF_F_TSO_ECN
| NETIF_F_TSO6
;
2430 /* Individual feature bits: what can host handle? */
2431 if (virtio_has_feature(vdev
, VIRTIO_NET_F_HOST_TSO4
))
2432 dev
->hw_features
|= NETIF_F_TSO
;
2433 if (virtio_has_feature(vdev
, VIRTIO_NET_F_HOST_TSO6
))
2434 dev
->hw_features
|= NETIF_F_TSO6
;
2435 if (virtio_has_feature(vdev
, VIRTIO_NET_F_HOST_ECN
))
2436 dev
->hw_features
|= NETIF_F_TSO_ECN
;
2437 if (virtio_has_feature(vdev
, VIRTIO_NET_F_HOST_UFO
))
2438 dev
->hw_features
|= NETIF_F_UFO
;
2440 dev
->features
|= NETIF_F_GSO_ROBUST
;
2443 dev
->features
|= dev
->hw_features
& (NETIF_F_ALL_TSO
|NETIF_F_UFO
);
2444 /* (!csum && gso) case will be fixed by register_netdev() */
2446 if (virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_CSUM
))
2447 dev
->features
|= NETIF_F_RXCSUM
;
2449 dev
->vlan_features
= dev
->features
;
2451 /* MTU range: 68 - 65535 */
2452 dev
->min_mtu
= MIN_MTU
;
2453 dev
->max_mtu
= MAX_MTU
;
2455 /* Configuration may specify what MAC to use. Otherwise random. */
2456 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MAC
))
2457 virtio_cread_bytes(vdev
,
2458 offsetof(struct virtio_net_config
, mac
),
2459 dev
->dev_addr
, dev
->addr_len
);
2461 eth_hw_addr_random(dev
);
2463 /* Set up our device-specific information */
2464 vi
= netdev_priv(dev
);
2468 vi
->stats
= alloc_percpu(struct virtnet_stats
);
2470 if (vi
->stats
== NULL
)
2473 for_each_possible_cpu(i
) {
2474 struct virtnet_stats
*virtnet_stats
;
2475 virtnet_stats
= per_cpu_ptr(vi
->stats
, i
);
2476 u64_stats_init(&virtnet_stats
->tx_syncp
);
2477 u64_stats_init(&virtnet_stats
->rx_syncp
);
2480 INIT_WORK(&vi
->config_work
, virtnet_config_changed_work
);
2482 /* If we can receive ANY GSO packets, we must allocate large ones. */
2483 if (virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_TSO4
) ||
2484 virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_TSO6
) ||
2485 virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_ECN
) ||
2486 virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_UFO
))
2487 vi
->big_packets
= true;
2489 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MRG_RXBUF
))
2490 vi
->mergeable_rx_bufs
= true;
2492 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MRG_RXBUF
) ||
2493 virtio_has_feature(vdev
, VIRTIO_F_VERSION_1
))
2494 vi
->hdr_len
= sizeof(struct virtio_net_hdr_mrg_rxbuf
);
2496 vi
->hdr_len
= sizeof(struct virtio_net_hdr
);
2498 if (virtio_has_feature(vdev
, VIRTIO_F_ANY_LAYOUT
) ||
2499 virtio_has_feature(vdev
, VIRTIO_F_VERSION_1
))
2500 vi
->any_header_sg
= true;
2502 if (virtio_has_feature(vdev
, VIRTIO_NET_F_CTRL_VQ
))
2505 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MTU
)) {
2506 mtu
= virtio_cread16(vdev
,
2507 offsetof(struct virtio_net_config
,
2509 if (mtu
< dev
->min_mtu
) {
2510 /* Should never trigger: MTU was previously validated
2511 * in virtnet_validate.
2513 dev_err(&vdev
->dev
, "device MTU appears to have changed "
2514 "it is now %d < %d", mtu
, dev
->min_mtu
);
2521 /* TODO: size buffers correctly in this case. */
2522 if (dev
->mtu
> ETH_DATA_LEN
)
2523 vi
->big_packets
= true;
2526 if (vi
->any_header_sg
)
2527 dev
->needed_headroom
= vi
->hdr_len
;
2529 /* Enable multiqueue by default */
2530 if (num_online_cpus() >= max_queue_pairs
)
2531 vi
->curr_queue_pairs
= max_queue_pairs
;
2533 vi
->curr_queue_pairs
= num_online_cpus();
2534 vi
->max_queue_pairs
= max_queue_pairs
;
2536 /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
2542 if (vi
->mergeable_rx_bufs
)
2543 dev
->sysfs_rx_queue_group
= &virtio_net_mrg_rx_group
;
2545 netif_set_real_num_tx_queues(dev
, vi
->curr_queue_pairs
);
2546 netif_set_real_num_rx_queues(dev
, vi
->curr_queue_pairs
);
2548 virtnet_init_settings(dev
);
2550 err
= register_netdev(dev
);
2552 pr_debug("virtio_net: registering device failed\n");
2556 virtio_device_ready(vdev
);
2558 err
= virtnet_cpu_notif_add(vi
);
2560 pr_debug("virtio_net: registering cpu notifier failed\n");
2561 goto free_unregister_netdev
;
2564 virtnet_set_queues(vi
, vi
->curr_queue_pairs
);
2566 /* Assume link up if device can't report link status,
2567 otherwise get link status from config. */
2568 if (virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_STATUS
)) {
2569 netif_carrier_off(dev
);
2570 schedule_work(&vi
->config_work
);
2572 vi
->status
= VIRTIO_NET_S_LINK_UP
;
2573 netif_carrier_on(dev
);
2576 pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
2577 dev
->name
, max_queue_pairs
);
2581 free_unregister_netdev
:
2582 vi
->vdev
->config
->reset(vdev
);
2584 unregister_netdev(dev
);
2586 cancel_delayed_work_sync(&vi
->refill
);
2587 free_receive_page_frags(vi
);
2588 virtnet_del_vqs(vi
);
2590 free_percpu(vi
->stats
);
2596 static void _remove_vq_common(struct virtnet_info
*vi
)
2598 vi
->vdev
->config
->reset(vi
->vdev
);
2599 free_unused_bufs(vi
);
2600 _free_receive_bufs(vi
);
2601 free_receive_page_frags(vi
);
2602 virtnet_del_vqs(vi
);
2605 static void remove_vq_common(struct virtnet_info
*vi
)
2607 vi
->vdev
->config
->reset(vi
->vdev
);
2609 /* Free unused buffers in both send and recv, if any. */
2610 free_unused_bufs(vi
);
2612 free_receive_bufs(vi
);
2614 free_receive_page_frags(vi
);
2616 virtnet_del_vqs(vi
);
2619 static void virtnet_remove(struct virtio_device
*vdev
)
2621 struct virtnet_info
*vi
= vdev
->priv
;
2623 virtnet_cpu_notif_remove(vi
);
2625 /* Make sure no work handler is accessing the device. */
2626 flush_work(&vi
->config_work
);
2628 unregister_netdev(vi
->dev
);
2630 remove_vq_common(vi
);
2632 free_percpu(vi
->stats
);
2633 free_netdev(vi
->dev
);
2636 #ifdef CONFIG_PM_SLEEP
2637 static int virtnet_freeze(struct virtio_device
*vdev
)
2639 struct virtnet_info
*vi
= vdev
->priv
;
2641 virtnet_cpu_notif_remove(vi
);
2642 virtnet_freeze_down(vdev
);
2643 remove_vq_common(vi
);
2648 static int virtnet_restore(struct virtio_device
*vdev
)
2650 struct virtnet_info
*vi
= vdev
->priv
;
2653 err
= virtnet_restore_up(vdev
);
2656 virtnet_set_queues(vi
, vi
->curr_queue_pairs
);
2658 err
= virtnet_cpu_notif_add(vi
);
2666 static struct virtio_device_id id_table
[] = {
2667 { VIRTIO_ID_NET
, VIRTIO_DEV_ANY_ID
},
2671 #define VIRTNET_FEATURES \
2672 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
2674 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
2675 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
2676 VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
2677 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
2678 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
2679 VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
2680 VIRTIO_NET_F_CTRL_MAC_ADDR, \
2683 static unsigned int features
[] = {
2687 static unsigned int features_legacy
[] = {
2690 VIRTIO_F_ANY_LAYOUT
,
2693 static struct virtio_driver virtio_net_driver
= {
2694 .feature_table
= features
,
2695 .feature_table_size
= ARRAY_SIZE(features
),
2696 .feature_table_legacy
= features_legacy
,
2697 .feature_table_size_legacy
= ARRAY_SIZE(features_legacy
),
2698 .driver
.name
= KBUILD_MODNAME
,
2699 .driver
.owner
= THIS_MODULE
,
2700 .id_table
= id_table
,
2701 .validate
= virtnet_validate
,
2702 .probe
= virtnet_probe
,
2703 .remove
= virtnet_remove
,
2704 .config_changed
= virtnet_config_changed
,
2705 #ifdef CONFIG_PM_SLEEP
2706 .freeze
= virtnet_freeze
,
2707 .restore
= virtnet_restore
,
2711 static __init
int virtio_net_driver_init(void)
2715 ret
= cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN
, "virtio/net:online",
2717 virtnet_cpu_down_prep
);
2720 virtionet_online
= ret
;
2721 ret
= cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD
, "virtio/net:dead",
2722 NULL
, virtnet_cpu_dead
);
2726 ret
= register_virtio_driver(&virtio_net_driver
);
2731 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD
);
2733 cpuhp_remove_multi_state(virtionet_online
);
2737 module_init(virtio_net_driver_init
);
2739 static __exit
void virtio_net_driver_exit(void)
2741 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD
);
2742 cpuhp_remove_multi_state(virtionet_online
);
2743 unregister_virtio_driver(&virtio_net_driver
);
2745 module_exit(virtio_net_driver_exit
);
2747 MODULE_DEVICE_TABLE(virtio
, id_table
);
2748 MODULE_DESCRIPTION("Virtio network driver");
2749 MODULE_LICENSE("GPL");