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>
33 static int napi_weight
= NAPI_POLL_WEIGHT
;
34 module_param(napi_weight
, int, 0444);
36 static bool csum
= true, gso
= true;
37 module_param(csum
, bool, 0444);
38 module_param(gso
, bool, 0444);
40 /* FIXME: MTU in config. */
41 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
42 #define GOOD_COPY_LEN 128
44 /* Amount of XDP headroom to prepend to packets for use by xdp_adjust_head */
45 #define VIRTIO_XDP_HEADROOM 256
47 /* RX packet size EWMA. The average packet size is used to determine the packet
48 * buffer size when refilling RX rings. As the entire RX ring may be refilled
49 * at once, the weight is chosen so that the EWMA will be insensitive to short-
50 * term, transient changes in packet size.
52 DECLARE_EWMA(pkt_len
, 1, 64)
54 /* With mergeable buffers we align buffer address and use the low bits to
55 * encode its true size. Buffer size is up to 1 page so we need to align to
56 * square root of page size to ensure we reserve enough bits to encode the true
59 #define MERGEABLE_BUFFER_MIN_ALIGN_SHIFT ((PAGE_SHIFT + 1) / 2)
61 /* Minimum alignment for mergeable packet buffers. */
62 #define MERGEABLE_BUFFER_ALIGN max(L1_CACHE_BYTES, \
63 1 << MERGEABLE_BUFFER_MIN_ALIGN_SHIFT)
65 #define VIRTNET_DRIVER_VERSION "1.0.0"
67 struct virtnet_stats
{
68 struct u64_stats_sync tx_syncp
;
69 struct u64_stats_sync rx_syncp
;
77 /* Internal representation of a send virtqueue */
79 /* Virtqueue associated with this send _queue */
82 /* TX: fragments + linear part + virtio header */
83 struct scatterlist sg
[MAX_SKB_FRAGS
+ 2];
85 /* Name of the send queue: output.$index */
89 /* Internal representation of a receive virtqueue */
90 struct receive_queue
{
91 /* Virtqueue associated with this receive_queue */
94 struct napi_struct napi
;
96 struct bpf_prog __rcu
*xdp_prog
;
98 /* Chain pages by the private ptr. */
101 /* Average packet length for mergeable receive buffers. */
102 struct ewma_pkt_len mrg_avg_pkt_len
;
104 /* Page frag for packet buffer allocation. */
105 struct page_frag alloc_frag
;
107 /* RX: fragments + linear part + virtio header */
108 struct scatterlist sg
[MAX_SKB_FRAGS
+ 2];
110 /* Name of this receive queue: input.$index */
114 struct virtnet_info
{
115 struct virtio_device
*vdev
;
116 struct virtqueue
*cvq
;
117 struct net_device
*dev
;
118 struct send_queue
*sq
;
119 struct receive_queue
*rq
;
122 /* Max # of queue pairs supported by the device */
125 /* # of queue pairs currently used by the driver */
126 u16 curr_queue_pairs
;
128 /* # of XDP queue pairs currently used by the driver */
131 /* I like... big packets and I cannot lie! */
134 /* Host will merge rx buffers for big packets (shake it! shake it!) */
135 bool mergeable_rx_bufs
;
137 /* Has control virtqueue */
140 /* Host can handle any s/g split between our header and packet data */
143 /* Packet virtio header size */
146 /* Active statistics */
147 struct virtnet_stats __percpu
*stats
;
149 /* Work struct for refilling if we run low on memory. */
150 struct delayed_work refill
;
152 /* Work struct for config space updates */
153 struct work_struct config_work
;
155 /* Does the affinity hint is set for virtqueues? */
156 bool affinity_hint_set
;
158 /* CPU hotplug instances for online & dead */
159 struct hlist_node node
;
160 struct hlist_node node_dead
;
162 /* Control VQ buffers: protected by the rtnl lock */
163 struct virtio_net_ctrl_hdr ctrl_hdr
;
164 virtio_net_ctrl_ack ctrl_status
;
165 struct virtio_net_ctrl_mq ctrl_mq
;
170 /* Ethtool settings */
175 struct padded_vnet_hdr
{
176 struct virtio_net_hdr_mrg_rxbuf hdr
;
178 * hdr is in a separate sg buffer, and data sg buffer shares same page
179 * with this header sg. This padding makes next sg 16 byte aligned
185 /* Converting between virtqueue no. and kernel tx/rx queue no.
186 * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
188 static int vq2txq(struct virtqueue
*vq
)
190 return (vq
->index
- 1) / 2;
193 static int txq2vq(int txq
)
198 static int vq2rxq(struct virtqueue
*vq
)
200 return vq
->index
/ 2;
203 static int rxq2vq(int rxq
)
208 static inline struct virtio_net_hdr_mrg_rxbuf
*skb_vnet_hdr(struct sk_buff
*skb
)
210 return (struct virtio_net_hdr_mrg_rxbuf
*)skb
->cb
;
214 * private is used to chain pages for big packets, put the whole
215 * most recent used list in the beginning for reuse
217 static void give_pages(struct receive_queue
*rq
, struct page
*page
)
221 /* Find end of list, sew whole thing into vi->rq.pages. */
222 for (end
= page
; end
->private; end
= (struct page
*)end
->private);
223 end
->private = (unsigned long)rq
->pages
;
227 static struct page
*get_a_page(struct receive_queue
*rq
, gfp_t gfp_mask
)
229 struct page
*p
= rq
->pages
;
232 rq
->pages
= (struct page
*)p
->private;
233 /* clear private here, it is used to chain pages */
236 p
= alloc_page(gfp_mask
);
240 static void skb_xmit_done(struct virtqueue
*vq
)
242 struct virtnet_info
*vi
= vq
->vdev
->priv
;
244 /* Suppress further interrupts. */
245 virtqueue_disable_cb(vq
);
247 /* We were probably waiting for more output buffers. */
248 netif_wake_subqueue(vi
->dev
, vq2txq(vq
));
251 static unsigned int mergeable_ctx_to_buf_truesize(unsigned long mrg_ctx
)
253 unsigned int truesize
= mrg_ctx
& (MERGEABLE_BUFFER_ALIGN
- 1);
254 return (truesize
+ 1) * MERGEABLE_BUFFER_ALIGN
;
257 static void *mergeable_ctx_to_buf_address(unsigned long mrg_ctx
)
259 return (void *)(mrg_ctx
& -MERGEABLE_BUFFER_ALIGN
);
263 static unsigned long mergeable_buf_to_ctx(void *buf
, unsigned int truesize
)
265 unsigned int size
= truesize
/ MERGEABLE_BUFFER_ALIGN
;
266 return (unsigned long)buf
| (size
- 1);
269 /* Called from bottom half context */
270 static struct sk_buff
*page_to_skb(struct virtnet_info
*vi
,
271 struct receive_queue
*rq
,
272 struct page
*page
, unsigned int offset
,
273 unsigned int len
, unsigned int truesize
)
276 struct virtio_net_hdr_mrg_rxbuf
*hdr
;
277 unsigned int copy
, hdr_len
, hdr_padded_len
;
280 p
= page_address(page
) + offset
;
282 /* copy small packet so we can reuse these pages for small data */
283 skb
= napi_alloc_skb(&rq
->napi
, GOOD_COPY_LEN
);
287 hdr
= skb_vnet_hdr(skb
);
289 hdr_len
= vi
->hdr_len
;
290 if (vi
->mergeable_rx_bufs
)
291 hdr_padded_len
= sizeof *hdr
;
293 hdr_padded_len
= sizeof(struct padded_vnet_hdr
);
295 memcpy(hdr
, p
, hdr_len
);
298 offset
+= hdr_padded_len
;
302 if (copy
> skb_tailroom(skb
))
303 copy
= skb_tailroom(skb
);
304 memcpy(skb_put(skb
, copy
), p
, copy
);
309 if (vi
->mergeable_rx_bufs
) {
311 skb_add_rx_frag(skb
, 0, page
, offset
, len
, truesize
);
318 * Verify that we can indeed put this data into a skb.
319 * This is here to handle cases when the device erroneously
320 * tries to receive more than is possible. This is usually
321 * the case of a broken device.
323 if (unlikely(len
> MAX_SKB_FRAGS
* PAGE_SIZE
)) {
324 net_dbg_ratelimited("%s: too much data\n", skb
->dev
->name
);
328 BUG_ON(offset
>= PAGE_SIZE
);
330 unsigned int frag_size
= min((unsigned)PAGE_SIZE
- offset
, len
);
331 skb_add_rx_frag(skb
, skb_shinfo(skb
)->nr_frags
, page
, offset
,
332 frag_size
, truesize
);
334 page
= (struct page
*)page
->private;
339 give_pages(rq
, page
);
344 static bool virtnet_xdp_xmit(struct virtnet_info
*vi
,
345 struct receive_queue
*rq
,
346 struct xdp_buff
*xdp
,
349 struct virtio_net_hdr_mrg_rxbuf
*hdr
;
350 unsigned int num_sg
, len
;
351 struct send_queue
*sq
;
356 qp
= vi
->curr_queue_pairs
- vi
->xdp_queue_pairs
+ smp_processor_id();
359 /* Free up any pending old buffers before queueing new ones. */
360 while ((xdp_sent
= virtqueue_get_buf(sq
->vq
, &len
)) != NULL
) {
361 if (vi
->mergeable_rx_bufs
) {
362 struct page
*sent_page
= virt_to_head_page(xdp_sent
);
365 } else { /* small buffer */
366 struct sk_buff
*skb
= xdp_sent
;
372 if (vi
->mergeable_rx_bufs
) {
373 xdp
->data
-= sizeof(struct virtio_net_hdr_mrg_rxbuf
);
374 /* Zero header and leave csum up to XDP layers */
376 memset(hdr
, 0, vi
->hdr_len
);
379 sg_init_one(sq
->sg
, xdp
->data
, xdp
->data_end
- xdp
->data
);
380 } else { /* small buffer */
381 struct sk_buff
*skb
= data
;
383 /* Zero header and leave csum up to XDP layers */
384 hdr
= skb_vnet_hdr(skb
);
385 memset(hdr
, 0, vi
->hdr_len
);
388 sg_init_table(sq
->sg
, 2);
389 sg_set_buf(sq
->sg
, hdr
, vi
->hdr_len
);
390 skb_to_sgvec(skb
, sq
->sg
+ 1,
391 xdp
->data
- xdp
->data_hard_start
,
392 xdp
->data_end
- xdp
->data
);
394 err
= virtqueue_add_outbuf(sq
->vq
, sq
->sg
, num_sg
,
397 if (vi
->mergeable_rx_bufs
) {
398 struct page
*page
= virt_to_head_page(xdp
->data
);
401 } else /* small buffer */
403 /* On error abort to avoid unnecessary kick */
407 virtqueue_kick(sq
->vq
);
411 static struct sk_buff
*receive_small(struct net_device
*dev
,
412 struct virtnet_info
*vi
,
413 struct receive_queue
*rq
,
414 void *buf
, unsigned int len
)
416 struct sk_buff
* skb
= buf
;
417 struct bpf_prog
*xdp_prog
;
422 xdp_prog
= rcu_dereference(rq
->xdp_prog
);
424 struct virtio_net_hdr_mrg_rxbuf
*hdr
= buf
;
428 if (unlikely(hdr
->hdr
.gso_type
|| hdr
->hdr
.flags
))
431 xdp
.data_hard_start
= skb
->data
;
432 xdp
.data
= skb
->data
+ VIRTIO_XDP_HEADROOM
;
433 xdp
.data_end
= xdp
.data
+ len
;
434 act
= bpf_prog_run_xdp(xdp_prog
, &xdp
);
438 /* Recalculate length in case bpf program changed it */
439 __skb_pull(skb
, xdp
.data
- xdp
.data_hard_start
);
440 len
= xdp
.data_end
- xdp
.data
;
443 if (unlikely(!virtnet_xdp_xmit(vi
, rq
, &xdp
, skb
)))
444 trace_xdp_exception(vi
->dev
, xdp_prog
, act
);
448 bpf_warn_invalid_xdp_action(act
);
450 trace_xdp_exception(vi
->dev
, xdp_prog
, act
);
462 dev
->stats
.rx_dropped
++;
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
);
520 ctx
= (unsigned long)virtqueue_get_buf(rq
->vq
, &buflen
);
524 buf
= mergeable_ctx_to_buf_address(ctx
);
525 p
= virt_to_head_page(buf
);
526 off
= buf
- page_address(p
);
528 /* guard against a misconfigured or uncooperative backend that
529 * is sending packet larger than the MTU.
531 if ((page_off
+ buflen
) > PAGE_SIZE
) {
536 memcpy(page_address(page
) + page_off
,
537 page_address(p
) + off
, buflen
);
542 /* Headroom does not contribute to packet length */
543 *len
= page_off
- VIRTIO_XDP_HEADROOM
;
546 __free_pages(page
, 0);
550 static struct sk_buff
*receive_mergeable(struct net_device
*dev
,
551 struct virtnet_info
*vi
,
552 struct receive_queue
*rq
,
556 void *buf
= mergeable_ctx_to_buf_address(ctx
);
557 struct virtio_net_hdr_mrg_rxbuf
*hdr
= buf
;
558 u16 num_buf
= virtio16_to_cpu(vi
->vdev
, hdr
->num_buffers
);
559 struct page
*page
= virt_to_head_page(buf
);
560 int offset
= buf
- page_address(page
);
561 struct sk_buff
*head_skb
, *curr_skb
;
562 struct bpf_prog
*xdp_prog
;
563 unsigned int truesize
;
568 xdp_prog
= rcu_dereference(rq
->xdp_prog
);
570 struct page
*xdp_page
;
575 /* This happens when rx buffer size is underestimated */
576 if (unlikely(num_buf
> 1)) {
577 /* linearize data for XDP */
578 xdp_page
= xdp_linearize_page(rq
, &num_buf
,
582 offset
= VIRTIO_XDP_HEADROOM
;
587 /* Transient failure which in theory could occur if
588 * in-flight packets from before XDP was enabled reach
589 * the receive path after XDP is loaded. In practice I
590 * was not able to create this condition.
592 if (unlikely(hdr
->hdr
.gso_type
))
595 /* Allow consuming headroom but reserve enough space to push
596 * the descriptor on if we get an XDP_TX return code.
598 data
= page_address(xdp_page
) + offset
;
599 xdp
.data_hard_start
= data
- VIRTIO_XDP_HEADROOM
+ vi
->hdr_len
;
600 xdp
.data
= data
+ vi
->hdr_len
;
601 xdp
.data_end
= xdp
.data
+ (len
- vi
->hdr_len
);
602 act
= bpf_prog_run_xdp(xdp_prog
, &xdp
);
606 /* recalculate offset to account for any header
607 * adjustments. Note other cases do not build an
608 * skb and avoid using offset
611 page_address(xdp_page
) - vi
->hdr_len
;
613 /* We can only create skb based on xdp_page. */
614 if (unlikely(xdp_page
!= page
)) {
617 head_skb
= page_to_skb(vi
, rq
, xdp_page
,
618 offset
, len
, PAGE_SIZE
);
619 ewma_pkt_len_add(&rq
->mrg_avg_pkt_len
, len
);
624 if (unlikely(!virtnet_xdp_xmit(vi
, rq
, &xdp
, data
)))
625 trace_xdp_exception(vi
->dev
, xdp_prog
, act
);
626 ewma_pkt_len_add(&rq
->mrg_avg_pkt_len
, len
);
627 if (unlikely(xdp_page
!= page
))
632 bpf_warn_invalid_xdp_action(act
);
634 trace_xdp_exception(vi
->dev
, xdp_prog
, act
);
636 if (unlikely(xdp_page
!= page
))
637 __free_pages(xdp_page
, 0);
638 ewma_pkt_len_add(&rq
->mrg_avg_pkt_len
, len
);
644 truesize
= max(len
, mergeable_ctx_to_buf_truesize(ctx
));
645 head_skb
= page_to_skb(vi
, rq
, page
, offset
, len
, truesize
);
648 if (unlikely(!curr_skb
))
653 ctx
= (unsigned long)virtqueue_get_buf(rq
->vq
, &len
);
654 if (unlikely(!ctx
)) {
655 pr_debug("%s: rx error: %d buffers out of %d missing\n",
657 virtio16_to_cpu(vi
->vdev
,
659 dev
->stats
.rx_length_errors
++;
663 buf
= mergeable_ctx_to_buf_address(ctx
);
664 page
= virt_to_head_page(buf
);
666 num_skb_frags
= skb_shinfo(curr_skb
)->nr_frags
;
667 if (unlikely(num_skb_frags
== MAX_SKB_FRAGS
)) {
668 struct sk_buff
*nskb
= alloc_skb(0, GFP_ATOMIC
);
672 if (curr_skb
== head_skb
)
673 skb_shinfo(curr_skb
)->frag_list
= nskb
;
675 curr_skb
->next
= nskb
;
677 head_skb
->truesize
+= nskb
->truesize
;
680 truesize
= max(len
, mergeable_ctx_to_buf_truesize(ctx
));
681 if (curr_skb
!= head_skb
) {
682 head_skb
->data_len
+= len
;
683 head_skb
->len
+= len
;
684 head_skb
->truesize
+= truesize
;
686 offset
= buf
- page_address(page
);
687 if (skb_can_coalesce(curr_skb
, num_skb_frags
, page
, offset
)) {
689 skb_coalesce_rx_frag(curr_skb
, num_skb_frags
- 1,
692 skb_add_rx_frag(curr_skb
, num_skb_frags
, page
,
693 offset
, len
, truesize
);
697 ewma_pkt_len_add(&rq
->mrg_avg_pkt_len
, head_skb
->len
);
705 ctx
= (unsigned long)virtqueue_get_buf(rq
->vq
, &len
);
706 if (unlikely(!ctx
)) {
707 pr_debug("%s: rx error: %d buffers missing\n",
709 dev
->stats
.rx_length_errors
++;
712 page
= virt_to_head_page(mergeable_ctx_to_buf_address(ctx
));
716 dev
->stats
.rx_dropped
++;
717 dev_kfree_skb(head_skb
);
722 static void receive_buf(struct virtnet_info
*vi
, struct receive_queue
*rq
,
723 void *buf
, unsigned int len
)
725 struct net_device
*dev
= vi
->dev
;
726 struct virtnet_stats
*stats
= this_cpu_ptr(vi
->stats
);
728 struct virtio_net_hdr_mrg_rxbuf
*hdr
;
730 if (unlikely(len
< vi
->hdr_len
+ ETH_HLEN
)) {
731 pr_debug("%s: short packet %i\n", dev
->name
, len
);
732 dev
->stats
.rx_length_errors
++;
733 if (vi
->mergeable_rx_bufs
) {
734 unsigned long ctx
= (unsigned long)buf
;
735 void *base
= mergeable_ctx_to_buf_address(ctx
);
736 put_page(virt_to_head_page(base
));
737 } else if (vi
->big_packets
) {
745 if (vi
->mergeable_rx_bufs
)
746 skb
= receive_mergeable(dev
, vi
, rq
, (unsigned long)buf
, len
);
747 else if (vi
->big_packets
)
748 skb
= receive_big(dev
, vi
, rq
, buf
, len
);
750 skb
= receive_small(dev
, vi
, rq
, buf
, len
);
755 hdr
= skb_vnet_hdr(skb
);
757 u64_stats_update_begin(&stats
->rx_syncp
);
758 stats
->rx_bytes
+= skb
->len
;
760 u64_stats_update_end(&stats
->rx_syncp
);
762 if (hdr
->hdr
.flags
& VIRTIO_NET_HDR_F_DATA_VALID
)
763 skb
->ip_summed
= CHECKSUM_UNNECESSARY
;
765 if (virtio_net_hdr_to_skb(skb
, &hdr
->hdr
,
766 virtio_is_little_endian(vi
->vdev
))) {
767 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
768 dev
->name
, hdr
->hdr
.gso_type
,
773 skb
->protocol
= eth_type_trans(skb
, dev
);
774 pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
775 ntohs(skb
->protocol
), skb
->len
, skb
->pkt_type
);
777 napi_gro_receive(&rq
->napi
, skb
);
781 dev
->stats
.rx_frame_errors
++;
785 static unsigned int virtnet_get_headroom(struct virtnet_info
*vi
)
787 return vi
->xdp_queue_pairs
? VIRTIO_XDP_HEADROOM
: 0;
790 static int add_recvbuf_small(struct virtnet_info
*vi
, struct receive_queue
*rq
,
793 int headroom
= GOOD_PACKET_LEN
+ virtnet_get_headroom(vi
);
794 unsigned int xdp_headroom
= virtnet_get_headroom(vi
);
796 struct virtio_net_hdr_mrg_rxbuf
*hdr
;
799 skb
= __netdev_alloc_skb_ip_align(vi
->dev
, headroom
, gfp
);
803 skb_put(skb
, headroom
);
805 hdr
= skb_vnet_hdr(skb
);
806 sg_init_table(rq
->sg
, 2);
807 sg_set_buf(rq
->sg
, hdr
, vi
->hdr_len
);
808 skb_to_sgvec(skb
, rq
->sg
+ 1, xdp_headroom
, skb
->len
- xdp_headroom
);
810 err
= virtqueue_add_inbuf(rq
->vq
, rq
->sg
, 2, skb
, gfp
);
817 static int add_recvbuf_big(struct virtnet_info
*vi
, struct receive_queue
*rq
,
820 struct page
*first
, *list
= NULL
;
824 sg_init_table(rq
->sg
, MAX_SKB_FRAGS
+ 2);
826 /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
827 for (i
= MAX_SKB_FRAGS
+ 1; i
> 1; --i
) {
828 first
= get_a_page(rq
, gfp
);
831 give_pages(rq
, list
);
834 sg_set_buf(&rq
->sg
[i
], page_address(first
), PAGE_SIZE
);
836 /* chain new page in list head to match sg */
837 first
->private = (unsigned long)list
;
841 first
= get_a_page(rq
, gfp
);
843 give_pages(rq
, list
);
846 p
= page_address(first
);
848 /* rq->sg[0], rq->sg[1] share the same page */
849 /* a separated rq->sg[0] for header - required in case !any_header_sg */
850 sg_set_buf(&rq
->sg
[0], p
, vi
->hdr_len
);
852 /* rq->sg[1] for data packet, from offset */
853 offset
= sizeof(struct padded_vnet_hdr
);
854 sg_set_buf(&rq
->sg
[1], p
+ offset
, PAGE_SIZE
- offset
);
856 /* chain first in list head */
857 first
->private = (unsigned long)list
;
858 err
= virtqueue_add_inbuf(rq
->vq
, rq
->sg
, MAX_SKB_FRAGS
+ 2,
861 give_pages(rq
, first
);
866 static unsigned int get_mergeable_buf_len(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 GOOD_PACKET_LEN
, PAGE_SIZE
- hdr_len
);
873 return ALIGN(len
, MERGEABLE_BUFFER_ALIGN
);
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
->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
= mergeable_buf_to_ctx(buf
, 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(rq
->vq
, rq
->sg
, 1, (void *)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 /* Schedule NAPI, Suppress further interrupts if successful. */
950 if (napi_schedule_prep(&rq
->napi
)) {
951 virtqueue_disable_cb(rvq
);
952 __napi_schedule(&rq
->napi
);
956 static void virtnet_napi_enable(struct receive_queue
*rq
)
958 napi_enable(&rq
->napi
);
960 /* If all buffers were filled by other side before we napi_enabled, we
961 * won't get another interrupt, so process any outstanding packets
962 * now. virtnet_poll wants re-enable the queue, so we disable here.
963 * We synchronize against interrupts via NAPI_STATE_SCHED */
964 if (napi_schedule_prep(&rq
->napi
)) {
965 virtqueue_disable_cb(rq
->vq
);
967 __napi_schedule(&rq
->napi
);
972 static void refill_work(struct work_struct
*work
)
974 struct virtnet_info
*vi
=
975 container_of(work
, struct virtnet_info
, refill
.work
);
979 for (i
= 0; i
< vi
->curr_queue_pairs
; i
++) {
980 struct receive_queue
*rq
= &vi
->rq
[i
];
982 napi_disable(&rq
->napi
);
983 still_empty
= !try_fill_recv(vi
, rq
, GFP_KERNEL
);
984 virtnet_napi_enable(rq
);
986 /* In theory, this can happen: if we don't get any buffers in
987 * we will *never* try to fill again.
990 schedule_delayed_work(&vi
->refill
, HZ
/2);
994 static int virtnet_receive(struct receive_queue
*rq
, int budget
)
996 struct virtnet_info
*vi
= rq
->vq
->vdev
->priv
;
997 unsigned int len
, received
= 0;
1000 while (received
< budget
&&
1001 (buf
= virtqueue_get_buf(rq
->vq
, &len
)) != NULL
) {
1002 receive_buf(vi
, rq
, buf
, len
);
1006 if (rq
->vq
->num_free
> virtqueue_get_vring_size(rq
->vq
) / 2) {
1007 if (!try_fill_recv(vi
, rq
, GFP_ATOMIC
))
1008 schedule_delayed_work(&vi
->refill
, 0);
1014 static int virtnet_poll(struct napi_struct
*napi
, int budget
)
1016 struct receive_queue
*rq
=
1017 container_of(napi
, struct receive_queue
, napi
);
1018 unsigned int r
, received
;
1020 received
= virtnet_receive(rq
, budget
);
1022 /* Out of packets? */
1023 if (received
< budget
) {
1024 r
= virtqueue_enable_cb_prepare(rq
->vq
);
1025 if (napi_complete_done(napi
, received
)) {
1026 if (unlikely(virtqueue_poll(rq
->vq
, r
)) &&
1027 napi_schedule_prep(napi
)) {
1028 virtqueue_disable_cb(rq
->vq
);
1029 __napi_schedule(napi
);
1037 static int virtnet_open(struct net_device
*dev
)
1039 struct virtnet_info
*vi
= netdev_priv(dev
);
1042 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1043 if (i
< vi
->curr_queue_pairs
)
1044 /* Make sure we have some buffers: if oom use wq. */
1045 if (!try_fill_recv(vi
, &vi
->rq
[i
], GFP_KERNEL
))
1046 schedule_delayed_work(&vi
->refill
, 0);
1047 virtnet_napi_enable(&vi
->rq
[i
]);
1053 static void free_old_xmit_skbs(struct send_queue
*sq
)
1055 struct sk_buff
*skb
;
1057 struct virtnet_info
*vi
= sq
->vq
->vdev
->priv
;
1058 struct virtnet_stats
*stats
= this_cpu_ptr(vi
->stats
);
1060 while ((skb
= virtqueue_get_buf(sq
->vq
, &len
)) != NULL
) {
1061 pr_debug("Sent skb %p\n", skb
);
1063 u64_stats_update_begin(&stats
->tx_syncp
);
1064 stats
->tx_bytes
+= skb
->len
;
1065 stats
->tx_packets
++;
1066 u64_stats_update_end(&stats
->tx_syncp
);
1068 dev_kfree_skb_any(skb
);
1072 static int xmit_skb(struct send_queue
*sq
, struct sk_buff
*skb
)
1074 struct virtio_net_hdr_mrg_rxbuf
*hdr
;
1075 const unsigned char *dest
= ((struct ethhdr
*)skb
->data
)->h_dest
;
1076 struct virtnet_info
*vi
= sq
->vq
->vdev
->priv
;
1078 unsigned hdr_len
= vi
->hdr_len
;
1081 pr_debug("%s: xmit %p %pM\n", vi
->dev
->name
, skb
, dest
);
1083 can_push
= vi
->any_header_sg
&&
1084 !((unsigned long)skb
->data
& (__alignof__(*hdr
) - 1)) &&
1085 !skb_header_cloned(skb
) && skb_headroom(skb
) >= hdr_len
;
1086 /* Even if we can, don't push here yet as this would skew
1087 * csum_start offset below. */
1089 hdr
= (struct virtio_net_hdr_mrg_rxbuf
*)(skb
->data
- hdr_len
);
1091 hdr
= skb_vnet_hdr(skb
);
1093 if (virtio_net_hdr_from_skb(skb
, &hdr
->hdr
,
1094 virtio_is_little_endian(vi
->vdev
), false))
1097 if (vi
->mergeable_rx_bufs
)
1098 hdr
->num_buffers
= 0;
1100 sg_init_table(sq
->sg
, skb_shinfo(skb
)->nr_frags
+ (can_push
? 1 : 2));
1102 __skb_push(skb
, hdr_len
);
1103 num_sg
= skb_to_sgvec(skb
, sq
->sg
, 0, skb
->len
);
1104 /* Pull header back to avoid skew in tx bytes calculations. */
1105 __skb_pull(skb
, hdr_len
);
1107 sg_set_buf(sq
->sg
, hdr
, hdr_len
);
1108 num_sg
= skb_to_sgvec(skb
, sq
->sg
+ 1, 0, skb
->len
) + 1;
1110 return virtqueue_add_outbuf(sq
->vq
, sq
->sg
, num_sg
, skb
, GFP_ATOMIC
);
1113 static netdev_tx_t
start_xmit(struct sk_buff
*skb
, struct net_device
*dev
)
1115 struct virtnet_info
*vi
= netdev_priv(dev
);
1116 int qnum
= skb_get_queue_mapping(skb
);
1117 struct send_queue
*sq
= &vi
->sq
[qnum
];
1119 struct netdev_queue
*txq
= netdev_get_tx_queue(dev
, qnum
);
1120 bool kick
= !skb
->xmit_more
;
1122 /* Free up any pending old buffers before queueing new ones. */
1123 free_old_xmit_skbs(sq
);
1125 /* timestamp packet in software */
1126 skb_tx_timestamp(skb
);
1128 /* Try to transmit */
1129 err
= xmit_skb(sq
, skb
);
1131 /* This should not happen! */
1132 if (unlikely(err
)) {
1133 dev
->stats
.tx_fifo_errors
++;
1134 if (net_ratelimit())
1136 "Unexpected TXQ (%d) queue failure: %d\n", qnum
, err
);
1137 dev
->stats
.tx_dropped
++;
1138 dev_kfree_skb_any(skb
);
1139 return NETDEV_TX_OK
;
1142 /* Don't wait up for transmitted skbs to be freed. */
1146 /* If running out of space, stop queue to avoid getting packets that we
1147 * are then unable to transmit.
1148 * An alternative would be to force queuing layer to requeue the skb by
1149 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1150 * returned in a normal path of operation: it means that driver is not
1151 * maintaining the TX queue stop/start state properly, and causes
1152 * the stack to do a non-trivial amount of useless work.
1153 * Since most packets only take 1 or 2 ring slots, stopping the queue
1154 * early means 16 slots are typically wasted.
1156 if (sq
->vq
->num_free
< 2+MAX_SKB_FRAGS
) {
1157 netif_stop_subqueue(dev
, qnum
);
1158 if (unlikely(!virtqueue_enable_cb_delayed(sq
->vq
))) {
1159 /* More just got used, free them then recheck. */
1160 free_old_xmit_skbs(sq
);
1161 if (sq
->vq
->num_free
>= 2+MAX_SKB_FRAGS
) {
1162 netif_start_subqueue(dev
, qnum
);
1163 virtqueue_disable_cb(sq
->vq
);
1168 if (kick
|| netif_xmit_stopped(txq
))
1169 virtqueue_kick(sq
->vq
);
1171 return NETDEV_TX_OK
;
1175 * Send command via the control virtqueue and check status. Commands
1176 * supported by the hypervisor, as indicated by feature bits, should
1177 * never fail unless improperly formatted.
1179 static bool virtnet_send_command(struct virtnet_info
*vi
, u8
class, u8 cmd
,
1180 struct scatterlist
*out
)
1182 struct scatterlist
*sgs
[4], hdr
, stat
;
1183 unsigned out_num
= 0, tmp
;
1185 /* Caller should know better */
1186 BUG_ON(!virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_CTRL_VQ
));
1188 vi
->ctrl_status
= ~0;
1189 vi
->ctrl_hdr
.class = class;
1190 vi
->ctrl_hdr
.cmd
= cmd
;
1192 sg_init_one(&hdr
, &vi
->ctrl_hdr
, sizeof(vi
->ctrl_hdr
));
1193 sgs
[out_num
++] = &hdr
;
1196 sgs
[out_num
++] = out
;
1198 /* Add return status. */
1199 sg_init_one(&stat
, &vi
->ctrl_status
, sizeof(vi
->ctrl_status
));
1200 sgs
[out_num
] = &stat
;
1202 BUG_ON(out_num
+ 1 > ARRAY_SIZE(sgs
));
1203 virtqueue_add_sgs(vi
->cvq
, sgs
, out_num
, 1, vi
, GFP_ATOMIC
);
1205 if (unlikely(!virtqueue_kick(vi
->cvq
)))
1206 return vi
->ctrl_status
== VIRTIO_NET_OK
;
1208 /* Spin for a response, the kick causes an ioport write, trapping
1209 * into the hypervisor, so the request should be handled immediately.
1211 while (!virtqueue_get_buf(vi
->cvq
, &tmp
) &&
1212 !virtqueue_is_broken(vi
->cvq
))
1215 return vi
->ctrl_status
== VIRTIO_NET_OK
;
1218 static int virtnet_set_mac_address(struct net_device
*dev
, void *p
)
1220 struct virtnet_info
*vi
= netdev_priv(dev
);
1221 struct virtio_device
*vdev
= vi
->vdev
;
1223 struct sockaddr
*addr
;
1224 struct scatterlist sg
;
1226 addr
= kmemdup(p
, sizeof(*addr
), GFP_KERNEL
);
1230 ret
= eth_prepare_mac_addr_change(dev
, addr
);
1234 if (virtio_has_feature(vdev
, VIRTIO_NET_F_CTRL_MAC_ADDR
)) {
1235 sg_init_one(&sg
, addr
->sa_data
, dev
->addr_len
);
1236 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_MAC
,
1237 VIRTIO_NET_CTRL_MAC_ADDR_SET
, &sg
)) {
1238 dev_warn(&vdev
->dev
,
1239 "Failed to set mac address by vq command.\n");
1243 } else if (virtio_has_feature(vdev
, VIRTIO_NET_F_MAC
) &&
1244 !virtio_has_feature(vdev
, VIRTIO_F_VERSION_1
)) {
1247 /* Naturally, this has an atomicity problem. */
1248 for (i
= 0; i
< dev
->addr_len
; i
++)
1249 virtio_cwrite8(vdev
,
1250 offsetof(struct virtio_net_config
, mac
) +
1251 i
, addr
->sa_data
[i
]);
1254 eth_commit_mac_addr_change(dev
, p
);
1262 static void virtnet_stats(struct net_device
*dev
,
1263 struct rtnl_link_stats64
*tot
)
1265 struct virtnet_info
*vi
= netdev_priv(dev
);
1269 for_each_possible_cpu(cpu
) {
1270 struct virtnet_stats
*stats
= per_cpu_ptr(vi
->stats
, cpu
);
1271 u64 tpackets
, tbytes
, rpackets
, rbytes
;
1274 start
= u64_stats_fetch_begin_irq(&stats
->tx_syncp
);
1275 tpackets
= stats
->tx_packets
;
1276 tbytes
= stats
->tx_bytes
;
1277 } while (u64_stats_fetch_retry_irq(&stats
->tx_syncp
, start
));
1280 start
= u64_stats_fetch_begin_irq(&stats
->rx_syncp
);
1281 rpackets
= stats
->rx_packets
;
1282 rbytes
= stats
->rx_bytes
;
1283 } while (u64_stats_fetch_retry_irq(&stats
->rx_syncp
, start
));
1285 tot
->rx_packets
+= rpackets
;
1286 tot
->tx_packets
+= tpackets
;
1287 tot
->rx_bytes
+= rbytes
;
1288 tot
->tx_bytes
+= tbytes
;
1291 tot
->tx_dropped
= dev
->stats
.tx_dropped
;
1292 tot
->tx_fifo_errors
= dev
->stats
.tx_fifo_errors
;
1293 tot
->rx_dropped
= dev
->stats
.rx_dropped
;
1294 tot
->rx_length_errors
= dev
->stats
.rx_length_errors
;
1295 tot
->rx_frame_errors
= dev
->stats
.rx_frame_errors
;
1298 #ifdef CONFIG_NET_POLL_CONTROLLER
1299 static void virtnet_netpoll(struct net_device
*dev
)
1301 struct virtnet_info
*vi
= netdev_priv(dev
);
1304 for (i
= 0; i
< vi
->curr_queue_pairs
; i
++)
1305 napi_schedule(&vi
->rq
[i
].napi
);
1309 static void virtnet_ack_link_announce(struct virtnet_info
*vi
)
1312 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_ANNOUNCE
,
1313 VIRTIO_NET_CTRL_ANNOUNCE_ACK
, NULL
))
1314 dev_warn(&vi
->dev
->dev
, "Failed to ack link announce.\n");
1318 static int _virtnet_set_queues(struct virtnet_info
*vi
, u16 queue_pairs
)
1320 struct scatterlist sg
;
1321 struct net_device
*dev
= vi
->dev
;
1323 if (!vi
->has_cvq
|| !virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_MQ
))
1326 vi
->ctrl_mq
.virtqueue_pairs
= cpu_to_virtio16(vi
->vdev
, queue_pairs
);
1327 sg_init_one(&sg
, &vi
->ctrl_mq
, sizeof(vi
->ctrl_mq
));
1329 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_MQ
,
1330 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET
, &sg
)) {
1331 dev_warn(&dev
->dev
, "Fail to set num of queue pairs to %d\n",
1335 vi
->curr_queue_pairs
= queue_pairs
;
1336 /* virtnet_open() will refill when device is going to up. */
1337 if (dev
->flags
& IFF_UP
)
1338 schedule_delayed_work(&vi
->refill
, 0);
1344 static int virtnet_set_queues(struct virtnet_info
*vi
, u16 queue_pairs
)
1349 err
= _virtnet_set_queues(vi
, queue_pairs
);
1354 static int virtnet_close(struct net_device
*dev
)
1356 struct virtnet_info
*vi
= netdev_priv(dev
);
1359 /* Make sure refill_work doesn't re-enable napi! */
1360 cancel_delayed_work_sync(&vi
->refill
);
1362 for (i
= 0; i
< vi
->max_queue_pairs
; i
++)
1363 napi_disable(&vi
->rq
[i
].napi
);
1368 static void virtnet_set_rx_mode(struct net_device
*dev
)
1370 struct virtnet_info
*vi
= netdev_priv(dev
);
1371 struct scatterlist sg
[2];
1372 struct virtio_net_ctrl_mac
*mac_data
;
1373 struct netdev_hw_addr
*ha
;
1379 /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1380 if (!virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_CTRL_RX
))
1383 vi
->ctrl_promisc
= ((dev
->flags
& IFF_PROMISC
) != 0);
1384 vi
->ctrl_allmulti
= ((dev
->flags
& IFF_ALLMULTI
) != 0);
1386 sg_init_one(sg
, &vi
->ctrl_promisc
, sizeof(vi
->ctrl_promisc
));
1388 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_RX
,
1389 VIRTIO_NET_CTRL_RX_PROMISC
, sg
))
1390 dev_warn(&dev
->dev
, "Failed to %sable promisc mode.\n",
1391 vi
->ctrl_promisc
? "en" : "dis");
1393 sg_init_one(sg
, &vi
->ctrl_allmulti
, sizeof(vi
->ctrl_allmulti
));
1395 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_RX
,
1396 VIRTIO_NET_CTRL_RX_ALLMULTI
, sg
))
1397 dev_warn(&dev
->dev
, "Failed to %sable allmulti mode.\n",
1398 vi
->ctrl_allmulti
? "en" : "dis");
1400 uc_count
= netdev_uc_count(dev
);
1401 mc_count
= netdev_mc_count(dev
);
1402 /* MAC filter - use one buffer for both lists */
1403 buf
= kzalloc(((uc_count
+ mc_count
) * ETH_ALEN
) +
1404 (2 * sizeof(mac_data
->entries
)), GFP_ATOMIC
);
1409 sg_init_table(sg
, 2);
1411 /* Store the unicast list and count in the front of the buffer */
1412 mac_data
->entries
= cpu_to_virtio32(vi
->vdev
, uc_count
);
1414 netdev_for_each_uc_addr(ha
, dev
)
1415 memcpy(&mac_data
->macs
[i
++][0], ha
->addr
, ETH_ALEN
);
1417 sg_set_buf(&sg
[0], mac_data
,
1418 sizeof(mac_data
->entries
) + (uc_count
* ETH_ALEN
));
1420 /* multicast list and count fill the end */
1421 mac_data
= (void *)&mac_data
->macs
[uc_count
][0];
1423 mac_data
->entries
= cpu_to_virtio32(vi
->vdev
, mc_count
);
1425 netdev_for_each_mc_addr(ha
, dev
)
1426 memcpy(&mac_data
->macs
[i
++][0], ha
->addr
, ETH_ALEN
);
1428 sg_set_buf(&sg
[1], mac_data
,
1429 sizeof(mac_data
->entries
) + (mc_count
* ETH_ALEN
));
1431 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_MAC
,
1432 VIRTIO_NET_CTRL_MAC_TABLE_SET
, sg
))
1433 dev_warn(&dev
->dev
, "Failed to set MAC filter table.\n");
1438 static int virtnet_vlan_rx_add_vid(struct net_device
*dev
,
1439 __be16 proto
, u16 vid
)
1441 struct virtnet_info
*vi
= netdev_priv(dev
);
1442 struct scatterlist sg
;
1445 sg_init_one(&sg
, &vi
->ctrl_vid
, sizeof(vi
->ctrl_vid
));
1447 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_VLAN
,
1448 VIRTIO_NET_CTRL_VLAN_ADD
, &sg
))
1449 dev_warn(&dev
->dev
, "Failed to add VLAN ID %d.\n", vid
);
1453 static int virtnet_vlan_rx_kill_vid(struct net_device
*dev
,
1454 __be16 proto
, u16 vid
)
1456 struct virtnet_info
*vi
= netdev_priv(dev
);
1457 struct scatterlist sg
;
1460 sg_init_one(&sg
, &vi
->ctrl_vid
, sizeof(vi
->ctrl_vid
));
1462 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_VLAN
,
1463 VIRTIO_NET_CTRL_VLAN_DEL
, &sg
))
1464 dev_warn(&dev
->dev
, "Failed to kill VLAN ID %d.\n", vid
);
1468 static void virtnet_clean_affinity(struct virtnet_info
*vi
, long hcpu
)
1472 if (vi
->affinity_hint_set
) {
1473 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1474 virtqueue_set_affinity(vi
->rq
[i
].vq
, -1);
1475 virtqueue_set_affinity(vi
->sq
[i
].vq
, -1);
1478 vi
->affinity_hint_set
= false;
1482 static void virtnet_set_affinity(struct virtnet_info
*vi
)
1487 /* In multiqueue mode, when the number of cpu is equal to the number of
1488 * queue pairs, we let the queue pairs to be private to one cpu by
1489 * setting the affinity hint to eliminate the contention.
1491 if (vi
->curr_queue_pairs
== 1 ||
1492 vi
->max_queue_pairs
!= num_online_cpus()) {
1493 virtnet_clean_affinity(vi
, -1);
1498 for_each_online_cpu(cpu
) {
1499 virtqueue_set_affinity(vi
->rq
[i
].vq
, cpu
);
1500 virtqueue_set_affinity(vi
->sq
[i
].vq
, cpu
);
1501 netif_set_xps_queue(vi
->dev
, cpumask_of(cpu
), i
);
1505 vi
->affinity_hint_set
= true;
1508 static int virtnet_cpu_online(unsigned int cpu
, struct hlist_node
*node
)
1510 struct virtnet_info
*vi
= hlist_entry_safe(node
, struct virtnet_info
,
1512 virtnet_set_affinity(vi
);
1516 static int virtnet_cpu_dead(unsigned int cpu
, struct hlist_node
*node
)
1518 struct virtnet_info
*vi
= hlist_entry_safe(node
, struct virtnet_info
,
1520 virtnet_set_affinity(vi
);
1524 static int virtnet_cpu_down_prep(unsigned int cpu
, struct hlist_node
*node
)
1526 struct virtnet_info
*vi
= hlist_entry_safe(node
, struct virtnet_info
,
1529 virtnet_clean_affinity(vi
, cpu
);
1533 static enum cpuhp_state virtionet_online
;
1535 static int virtnet_cpu_notif_add(struct virtnet_info
*vi
)
1539 ret
= cpuhp_state_add_instance_nocalls(virtionet_online
, &vi
->node
);
1542 ret
= cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD
,
1546 cpuhp_state_remove_instance_nocalls(virtionet_online
, &vi
->node
);
1550 static void virtnet_cpu_notif_remove(struct virtnet_info
*vi
)
1552 cpuhp_state_remove_instance_nocalls(virtionet_online
, &vi
->node
);
1553 cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD
,
1557 static void virtnet_get_ringparam(struct net_device
*dev
,
1558 struct ethtool_ringparam
*ring
)
1560 struct virtnet_info
*vi
= netdev_priv(dev
);
1562 ring
->rx_max_pending
= virtqueue_get_vring_size(vi
->rq
[0].vq
);
1563 ring
->tx_max_pending
= virtqueue_get_vring_size(vi
->sq
[0].vq
);
1564 ring
->rx_pending
= ring
->rx_max_pending
;
1565 ring
->tx_pending
= ring
->tx_max_pending
;
1569 static void virtnet_get_drvinfo(struct net_device
*dev
,
1570 struct ethtool_drvinfo
*info
)
1572 struct virtnet_info
*vi
= netdev_priv(dev
);
1573 struct virtio_device
*vdev
= vi
->vdev
;
1575 strlcpy(info
->driver
, KBUILD_MODNAME
, sizeof(info
->driver
));
1576 strlcpy(info
->version
, VIRTNET_DRIVER_VERSION
, sizeof(info
->version
));
1577 strlcpy(info
->bus_info
, virtio_bus_name(vdev
), sizeof(info
->bus_info
));
1581 /* TODO: Eliminate OOO packets during switching */
1582 static int virtnet_set_channels(struct net_device
*dev
,
1583 struct ethtool_channels
*channels
)
1585 struct virtnet_info
*vi
= netdev_priv(dev
);
1586 u16 queue_pairs
= channels
->combined_count
;
1589 /* We don't support separate rx/tx channels.
1590 * We don't allow setting 'other' channels.
1592 if (channels
->rx_count
|| channels
->tx_count
|| channels
->other_count
)
1595 if (queue_pairs
> vi
->max_queue_pairs
|| queue_pairs
== 0)
1598 /* For now we don't support modifying channels while XDP is loaded
1599 * also when XDP is loaded all RX queues have XDP programs so we only
1600 * need to check a single RX queue.
1602 if (vi
->rq
[0].xdp_prog
)
1606 err
= _virtnet_set_queues(vi
, queue_pairs
);
1608 netif_set_real_num_tx_queues(dev
, queue_pairs
);
1609 netif_set_real_num_rx_queues(dev
, queue_pairs
);
1611 virtnet_set_affinity(vi
);
1618 static void virtnet_get_channels(struct net_device
*dev
,
1619 struct ethtool_channels
*channels
)
1621 struct virtnet_info
*vi
= netdev_priv(dev
);
1623 channels
->combined_count
= vi
->curr_queue_pairs
;
1624 channels
->max_combined
= vi
->max_queue_pairs
;
1625 channels
->max_other
= 0;
1626 channels
->rx_count
= 0;
1627 channels
->tx_count
= 0;
1628 channels
->other_count
= 0;
1631 /* Check if the user is trying to change anything besides speed/duplex */
1632 static bool virtnet_validate_ethtool_cmd(const struct ethtool_cmd
*cmd
)
1634 struct ethtool_cmd diff1
= *cmd
;
1635 struct ethtool_cmd diff2
= {};
1637 /* cmd is always set so we need to clear it, validate the port type
1638 * and also without autonegotiation we can ignore advertising
1640 ethtool_cmd_speed_set(&diff1
, 0);
1641 diff2
.port
= PORT_OTHER
;
1642 diff1
.advertising
= 0;
1646 return !memcmp(&diff1
, &diff2
, sizeof(diff1
));
1649 static int virtnet_set_settings(struct net_device
*dev
, struct ethtool_cmd
*cmd
)
1651 struct virtnet_info
*vi
= netdev_priv(dev
);
1654 speed
= ethtool_cmd_speed(cmd
);
1655 /* don't allow custom speed and duplex */
1656 if (!ethtool_validate_speed(speed
) ||
1657 !ethtool_validate_duplex(cmd
->duplex
) ||
1658 !virtnet_validate_ethtool_cmd(cmd
))
1661 vi
->duplex
= cmd
->duplex
;
1666 static int virtnet_get_settings(struct net_device
*dev
, struct ethtool_cmd
*cmd
)
1668 struct virtnet_info
*vi
= netdev_priv(dev
);
1670 ethtool_cmd_speed_set(cmd
, vi
->speed
);
1671 cmd
->duplex
= vi
->duplex
;
1672 cmd
->port
= PORT_OTHER
;
1677 static void virtnet_init_settings(struct net_device
*dev
)
1679 struct virtnet_info
*vi
= netdev_priv(dev
);
1681 vi
->speed
= SPEED_UNKNOWN
;
1682 vi
->duplex
= DUPLEX_UNKNOWN
;
1685 static const struct ethtool_ops virtnet_ethtool_ops
= {
1686 .get_drvinfo
= virtnet_get_drvinfo
,
1687 .get_link
= ethtool_op_get_link
,
1688 .get_ringparam
= virtnet_get_ringparam
,
1689 .set_channels
= virtnet_set_channels
,
1690 .get_channels
= virtnet_get_channels
,
1691 .get_ts_info
= ethtool_op_get_ts_info
,
1692 .get_settings
= virtnet_get_settings
,
1693 .set_settings
= virtnet_set_settings
,
1696 static void virtnet_freeze_down(struct virtio_device
*vdev
)
1698 struct virtnet_info
*vi
= vdev
->priv
;
1701 /* Make sure no work handler is accessing the device */
1702 flush_work(&vi
->config_work
);
1704 netif_device_detach(vi
->dev
);
1705 cancel_delayed_work_sync(&vi
->refill
);
1707 if (netif_running(vi
->dev
)) {
1708 for (i
= 0; i
< vi
->max_queue_pairs
; i
++)
1709 napi_disable(&vi
->rq
[i
].napi
);
1713 static int init_vqs(struct virtnet_info
*vi
);
1714 static void _remove_vq_common(struct virtnet_info
*vi
);
1716 static int virtnet_restore_up(struct virtio_device
*vdev
)
1718 struct virtnet_info
*vi
= vdev
->priv
;
1725 virtio_device_ready(vdev
);
1727 if (netif_running(vi
->dev
)) {
1728 for (i
= 0; i
< vi
->curr_queue_pairs
; i
++)
1729 if (!try_fill_recv(vi
, &vi
->rq
[i
], GFP_KERNEL
))
1730 schedule_delayed_work(&vi
->refill
, 0);
1732 for (i
= 0; i
< vi
->max_queue_pairs
; i
++)
1733 virtnet_napi_enable(&vi
->rq
[i
]);
1736 netif_device_attach(vi
->dev
);
1740 static int virtnet_reset(struct virtnet_info
*vi
)
1742 struct virtio_device
*dev
= vi
->vdev
;
1745 virtio_config_disable(dev
);
1746 dev
->failed
= dev
->config
->get_status(dev
) & VIRTIO_CONFIG_S_FAILED
;
1747 virtnet_freeze_down(dev
);
1748 _remove_vq_common(vi
);
1750 dev
->config
->reset(dev
);
1751 virtio_add_status(dev
, VIRTIO_CONFIG_S_ACKNOWLEDGE
);
1752 virtio_add_status(dev
, VIRTIO_CONFIG_S_DRIVER
);
1754 ret
= virtio_finalize_features(dev
);
1758 ret
= virtnet_restore_up(dev
);
1761 ret
= _virtnet_set_queues(vi
, vi
->curr_queue_pairs
);
1765 virtio_add_status(dev
, VIRTIO_CONFIG_S_DRIVER_OK
);
1766 virtio_config_enable(dev
);
1769 virtio_add_status(dev
, VIRTIO_CONFIG_S_FAILED
);
1773 static int virtnet_xdp_set(struct net_device
*dev
, struct bpf_prog
*prog
)
1775 unsigned long int max_sz
= PAGE_SIZE
- sizeof(struct padded_vnet_hdr
);
1776 struct virtnet_info
*vi
= netdev_priv(dev
);
1777 struct bpf_prog
*old_prog
;
1778 u16 oxdp_qp
, xdp_qp
= 0, curr_qp
;
1781 if (virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_TSO4
) ||
1782 virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_TSO6
) ||
1783 virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_ECN
) ||
1784 virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_UFO
)) {
1785 netdev_warn(dev
, "can't set XDP while host is implementing LRO, disable LRO first\n");
1789 if (vi
->mergeable_rx_bufs
&& !vi
->any_header_sg
) {
1790 netdev_warn(dev
, "XDP expects header/data in single page, any_header_sg required\n");
1794 if (dev
->mtu
> max_sz
) {
1795 netdev_warn(dev
, "XDP requires MTU less than %lu\n", max_sz
);
1799 curr_qp
= vi
->curr_queue_pairs
- vi
->xdp_queue_pairs
;
1801 xdp_qp
= nr_cpu_ids
;
1803 /* XDP requires extra queues for XDP_TX */
1804 if (curr_qp
+ xdp_qp
> vi
->max_queue_pairs
) {
1805 netdev_warn(dev
, "request %i queues but max is %i\n",
1806 curr_qp
+ xdp_qp
, vi
->max_queue_pairs
);
1811 prog
= bpf_prog_add(prog
, vi
->max_queue_pairs
- 1);
1813 return PTR_ERR(prog
);
1816 err
= _virtnet_set_queues(vi
, curr_qp
+ xdp_qp
);
1818 dev_warn(&dev
->dev
, "XDP Device queue allocation failure.\n");
1819 goto virtio_queue_err
;
1822 oxdp_qp
= vi
->xdp_queue_pairs
;
1824 /* Changing the headroom in buffers is a disruptive operation because
1825 * existing buffers must be flushed and reallocated. This will happen
1826 * when a xdp program is initially added or xdp is disabled by removing
1827 * the xdp program resulting in number of XDP queues changing.
1829 if (vi
->xdp_queue_pairs
!= xdp_qp
) {
1830 vi
->xdp_queue_pairs
= xdp_qp
;
1831 err
= virtnet_reset(vi
);
1833 goto virtio_reset_err
;
1836 netif_set_real_num_rx_queues(dev
, curr_qp
+ xdp_qp
);
1838 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1839 old_prog
= rtnl_dereference(vi
->rq
[i
].xdp_prog
);
1840 rcu_assign_pointer(vi
->rq
[i
].xdp_prog
, prog
);
1842 bpf_prog_put(old_prog
);
1848 /* On reset error do our best to unwind XDP changes inflight and return
1849 * error up to user space for resolution. The underlying reset hung on
1850 * us so not much we can do here.
1852 dev_warn(&dev
->dev
, "XDP reset failure and queues unstable\n");
1853 vi
->xdp_queue_pairs
= oxdp_qp
;
1855 /* On queue set error we can unwind bpf ref count and user space can
1856 * retry this is most likely an allocation failure.
1859 bpf_prog_sub(prog
, vi
->max_queue_pairs
- 1);
1863 static bool virtnet_xdp_query(struct net_device
*dev
)
1865 struct virtnet_info
*vi
= netdev_priv(dev
);
1868 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1869 if (vi
->rq
[i
].xdp_prog
)
1875 static int virtnet_xdp(struct net_device
*dev
, struct netdev_xdp
*xdp
)
1877 switch (xdp
->command
) {
1878 case XDP_SETUP_PROG
:
1879 return virtnet_xdp_set(dev
, xdp
->prog
);
1880 case XDP_QUERY_PROG
:
1881 xdp
->prog_attached
= virtnet_xdp_query(dev
);
1888 static const struct net_device_ops virtnet_netdev
= {
1889 .ndo_open
= virtnet_open
,
1890 .ndo_stop
= virtnet_close
,
1891 .ndo_start_xmit
= start_xmit
,
1892 .ndo_validate_addr
= eth_validate_addr
,
1893 .ndo_set_mac_address
= virtnet_set_mac_address
,
1894 .ndo_set_rx_mode
= virtnet_set_rx_mode
,
1895 .ndo_get_stats64
= virtnet_stats
,
1896 .ndo_vlan_rx_add_vid
= virtnet_vlan_rx_add_vid
,
1897 .ndo_vlan_rx_kill_vid
= virtnet_vlan_rx_kill_vid
,
1898 #ifdef CONFIG_NET_POLL_CONTROLLER
1899 .ndo_poll_controller
= virtnet_netpoll
,
1901 .ndo_xdp
= virtnet_xdp
,
1904 static void virtnet_config_changed_work(struct work_struct
*work
)
1906 struct virtnet_info
*vi
=
1907 container_of(work
, struct virtnet_info
, config_work
);
1910 if (virtio_cread_feature(vi
->vdev
, VIRTIO_NET_F_STATUS
,
1911 struct virtio_net_config
, status
, &v
) < 0)
1914 if (v
& VIRTIO_NET_S_ANNOUNCE
) {
1915 netdev_notify_peers(vi
->dev
);
1916 virtnet_ack_link_announce(vi
);
1919 /* Ignore unknown (future) status bits */
1920 v
&= VIRTIO_NET_S_LINK_UP
;
1922 if (vi
->status
== v
)
1927 if (vi
->status
& VIRTIO_NET_S_LINK_UP
) {
1928 netif_carrier_on(vi
->dev
);
1929 netif_tx_wake_all_queues(vi
->dev
);
1931 netif_carrier_off(vi
->dev
);
1932 netif_tx_stop_all_queues(vi
->dev
);
1936 static void virtnet_config_changed(struct virtio_device
*vdev
)
1938 struct virtnet_info
*vi
= vdev
->priv
;
1940 schedule_work(&vi
->config_work
);
1943 static void virtnet_free_queues(struct virtnet_info
*vi
)
1947 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1948 napi_hash_del(&vi
->rq
[i
].napi
);
1949 netif_napi_del(&vi
->rq
[i
].napi
);
1952 /* We called napi_hash_del() before netif_napi_del(),
1953 * we need to respect an RCU grace period before freeing vi->rq
1961 static void _free_receive_bufs(struct virtnet_info
*vi
)
1963 struct bpf_prog
*old_prog
;
1966 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1967 while (vi
->rq
[i
].pages
)
1968 __free_pages(get_a_page(&vi
->rq
[i
], GFP_KERNEL
), 0);
1970 old_prog
= rtnl_dereference(vi
->rq
[i
].xdp_prog
);
1971 RCU_INIT_POINTER(vi
->rq
[i
].xdp_prog
, NULL
);
1973 bpf_prog_put(old_prog
);
1977 static void free_receive_bufs(struct virtnet_info
*vi
)
1980 _free_receive_bufs(vi
);
1984 static void free_receive_page_frags(struct virtnet_info
*vi
)
1987 for (i
= 0; i
< vi
->max_queue_pairs
; i
++)
1988 if (vi
->rq
[i
].alloc_frag
.page
)
1989 put_page(vi
->rq
[i
].alloc_frag
.page
);
1992 static bool is_xdp_raw_buffer_queue(struct virtnet_info
*vi
, int q
)
1994 /* For small receive mode always use kfree_skb variants */
1995 if (!vi
->mergeable_rx_bufs
)
1998 if (q
< (vi
->curr_queue_pairs
- vi
->xdp_queue_pairs
))
2000 else if (q
< vi
->curr_queue_pairs
)
2006 static void free_unused_bufs(struct virtnet_info
*vi
)
2011 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2012 struct virtqueue
*vq
= vi
->sq
[i
].vq
;
2013 while ((buf
= virtqueue_detach_unused_buf(vq
)) != NULL
) {
2014 if (!is_xdp_raw_buffer_queue(vi
, i
))
2017 put_page(virt_to_head_page(buf
));
2021 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2022 struct virtqueue
*vq
= vi
->rq
[i
].vq
;
2024 while ((buf
= virtqueue_detach_unused_buf(vq
)) != NULL
) {
2025 if (vi
->mergeable_rx_bufs
) {
2026 unsigned long ctx
= (unsigned long)buf
;
2027 void *base
= mergeable_ctx_to_buf_address(ctx
);
2028 put_page(virt_to_head_page(base
));
2029 } else if (vi
->big_packets
) {
2030 give_pages(&vi
->rq
[i
], buf
);
2038 static void virtnet_del_vqs(struct virtnet_info
*vi
)
2040 struct virtio_device
*vdev
= vi
->vdev
;
2042 virtnet_clean_affinity(vi
, -1);
2044 vdev
->config
->del_vqs(vdev
);
2046 virtnet_free_queues(vi
);
2049 static int virtnet_find_vqs(struct virtnet_info
*vi
)
2051 vq_callback_t
**callbacks
;
2052 struct virtqueue
**vqs
;
2057 /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
2058 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
2059 * possible control vq.
2061 total_vqs
= vi
->max_queue_pairs
* 2 +
2062 virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_CTRL_VQ
);
2064 /* Allocate space for find_vqs parameters */
2065 vqs
= kzalloc(total_vqs
* sizeof(*vqs
), GFP_KERNEL
);
2068 callbacks
= kmalloc(total_vqs
* sizeof(*callbacks
), GFP_KERNEL
);
2071 names
= kmalloc(total_vqs
* sizeof(*names
), GFP_KERNEL
);
2075 /* Parameters for control virtqueue, if any */
2077 callbacks
[total_vqs
- 1] = NULL
;
2078 names
[total_vqs
- 1] = "control";
2081 /* Allocate/initialize parameters for send/receive virtqueues */
2082 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2083 callbacks
[rxq2vq(i
)] = skb_recv_done
;
2084 callbacks
[txq2vq(i
)] = skb_xmit_done
;
2085 sprintf(vi
->rq
[i
].name
, "input.%d", i
);
2086 sprintf(vi
->sq
[i
].name
, "output.%d", i
);
2087 names
[rxq2vq(i
)] = vi
->rq
[i
].name
;
2088 names
[txq2vq(i
)] = vi
->sq
[i
].name
;
2091 ret
= vi
->vdev
->config
->find_vqs(vi
->vdev
, total_vqs
, vqs
, callbacks
,
2097 vi
->cvq
= vqs
[total_vqs
- 1];
2098 if (virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_CTRL_VLAN
))
2099 vi
->dev
->features
|= NETIF_F_HW_VLAN_CTAG_FILTER
;
2102 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2103 vi
->rq
[i
].vq
= vqs
[rxq2vq(i
)];
2104 vi
->sq
[i
].vq
= vqs
[txq2vq(i
)];
2123 static int virtnet_alloc_queues(struct virtnet_info
*vi
)
2127 vi
->sq
= kzalloc(sizeof(*vi
->sq
) * vi
->max_queue_pairs
, GFP_KERNEL
);
2130 vi
->rq
= kzalloc(sizeof(*vi
->rq
) * vi
->max_queue_pairs
, GFP_KERNEL
);
2134 INIT_DELAYED_WORK(&vi
->refill
, refill_work
);
2135 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2136 vi
->rq
[i
].pages
= NULL
;
2137 netif_napi_add(vi
->dev
, &vi
->rq
[i
].napi
, virtnet_poll
,
2140 sg_init_table(vi
->rq
[i
].sg
, ARRAY_SIZE(vi
->rq
[i
].sg
));
2141 ewma_pkt_len_init(&vi
->rq
[i
].mrg_avg_pkt_len
);
2142 sg_init_table(vi
->sq
[i
].sg
, ARRAY_SIZE(vi
->sq
[i
].sg
));
2153 static int init_vqs(struct virtnet_info
*vi
)
2157 /* Allocate send & receive queues */
2158 ret
= virtnet_alloc_queues(vi
);
2162 ret
= virtnet_find_vqs(vi
);
2167 virtnet_set_affinity(vi
);
2173 virtnet_free_queues(vi
);
2179 static ssize_t
mergeable_rx_buffer_size_show(struct netdev_rx_queue
*queue
,
2180 struct rx_queue_attribute
*attribute
, char *buf
)
2182 struct virtnet_info
*vi
= netdev_priv(queue
->dev
);
2183 unsigned int queue_index
= get_netdev_rx_queue_index(queue
);
2184 struct ewma_pkt_len
*avg
;
2186 BUG_ON(queue_index
>= vi
->max_queue_pairs
);
2187 avg
= &vi
->rq
[queue_index
].mrg_avg_pkt_len
;
2188 return sprintf(buf
, "%u\n", get_mergeable_buf_len(avg
));
2191 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute
=
2192 __ATTR_RO(mergeable_rx_buffer_size
);
2194 static struct attribute
*virtio_net_mrg_rx_attrs
[] = {
2195 &mergeable_rx_buffer_size_attribute
.attr
,
2199 static const struct attribute_group virtio_net_mrg_rx_group
= {
2200 .name
= "virtio_net",
2201 .attrs
= virtio_net_mrg_rx_attrs
2205 static bool virtnet_fail_on_feature(struct virtio_device
*vdev
,
2207 const char *fname
, const char *dname
)
2209 if (!virtio_has_feature(vdev
, fbit
))
2212 dev_err(&vdev
->dev
, "device advertises feature %s but not %s",
2218 #define VIRTNET_FAIL_ON(vdev, fbit, dbit) \
2219 virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
2221 static bool virtnet_validate_features(struct virtio_device
*vdev
)
2223 if (!virtio_has_feature(vdev
, VIRTIO_NET_F_CTRL_VQ
) &&
2224 (VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_CTRL_RX
,
2225 "VIRTIO_NET_F_CTRL_VQ") ||
2226 VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_CTRL_VLAN
,
2227 "VIRTIO_NET_F_CTRL_VQ") ||
2228 VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_GUEST_ANNOUNCE
,
2229 "VIRTIO_NET_F_CTRL_VQ") ||
2230 VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_MQ
, "VIRTIO_NET_F_CTRL_VQ") ||
2231 VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_CTRL_MAC_ADDR
,
2232 "VIRTIO_NET_F_CTRL_VQ"))) {
2239 #define MIN_MTU ETH_MIN_MTU
2240 #define MAX_MTU ETH_MAX_MTU
2242 static int virtnet_probe(struct virtio_device
*vdev
)
2245 struct net_device
*dev
;
2246 struct virtnet_info
*vi
;
2247 u16 max_queue_pairs
;
2250 if (!vdev
->config
->get
) {
2251 dev_err(&vdev
->dev
, "%s failure: config access disabled\n",
2256 if (!virtnet_validate_features(vdev
))
2259 /* Find if host supports multiqueue virtio_net device */
2260 err
= virtio_cread_feature(vdev
, VIRTIO_NET_F_MQ
,
2261 struct virtio_net_config
,
2262 max_virtqueue_pairs
, &max_queue_pairs
);
2264 /* We need at least 2 queue's */
2265 if (err
|| max_queue_pairs
< VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN
||
2266 max_queue_pairs
> VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX
||
2267 !virtio_has_feature(vdev
, VIRTIO_NET_F_CTRL_VQ
))
2268 max_queue_pairs
= 1;
2270 /* Allocate ourselves a network device with room for our info */
2271 dev
= alloc_etherdev_mq(sizeof(struct virtnet_info
), max_queue_pairs
);
2275 /* Set up network device as normal. */
2276 dev
->priv_flags
|= IFF_UNICAST_FLT
| IFF_LIVE_ADDR_CHANGE
;
2277 dev
->netdev_ops
= &virtnet_netdev
;
2278 dev
->features
= NETIF_F_HIGHDMA
;
2280 dev
->ethtool_ops
= &virtnet_ethtool_ops
;
2281 SET_NETDEV_DEV(dev
, &vdev
->dev
);
2283 /* Do we support "hardware" checksums? */
2284 if (virtio_has_feature(vdev
, VIRTIO_NET_F_CSUM
)) {
2285 /* This opens up the world of extra features. */
2286 dev
->hw_features
|= NETIF_F_HW_CSUM
| NETIF_F_SG
;
2288 dev
->features
|= NETIF_F_HW_CSUM
| NETIF_F_SG
;
2290 if (virtio_has_feature(vdev
, VIRTIO_NET_F_GSO
)) {
2291 dev
->hw_features
|= NETIF_F_TSO
| NETIF_F_UFO
2292 | NETIF_F_TSO_ECN
| NETIF_F_TSO6
;
2294 /* Individual feature bits: what can host handle? */
2295 if (virtio_has_feature(vdev
, VIRTIO_NET_F_HOST_TSO4
))
2296 dev
->hw_features
|= NETIF_F_TSO
;
2297 if (virtio_has_feature(vdev
, VIRTIO_NET_F_HOST_TSO6
))
2298 dev
->hw_features
|= NETIF_F_TSO6
;
2299 if (virtio_has_feature(vdev
, VIRTIO_NET_F_HOST_ECN
))
2300 dev
->hw_features
|= NETIF_F_TSO_ECN
;
2301 if (virtio_has_feature(vdev
, VIRTIO_NET_F_HOST_UFO
))
2302 dev
->hw_features
|= NETIF_F_UFO
;
2304 dev
->features
|= NETIF_F_GSO_ROBUST
;
2307 dev
->features
|= dev
->hw_features
& (NETIF_F_ALL_TSO
|NETIF_F_UFO
);
2308 /* (!csum && gso) case will be fixed by register_netdev() */
2310 if (virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_CSUM
))
2311 dev
->features
|= NETIF_F_RXCSUM
;
2313 dev
->vlan_features
= dev
->features
;
2315 /* MTU range: 68 - 65535 */
2316 dev
->min_mtu
= MIN_MTU
;
2317 dev
->max_mtu
= MAX_MTU
;
2319 /* Configuration may specify what MAC to use. Otherwise random. */
2320 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MAC
))
2321 virtio_cread_bytes(vdev
,
2322 offsetof(struct virtio_net_config
, mac
),
2323 dev
->dev_addr
, dev
->addr_len
);
2325 eth_hw_addr_random(dev
);
2327 /* Set up our device-specific information */
2328 vi
= netdev_priv(dev
);
2332 vi
->stats
= alloc_percpu(struct virtnet_stats
);
2334 if (vi
->stats
== NULL
)
2337 for_each_possible_cpu(i
) {
2338 struct virtnet_stats
*virtnet_stats
;
2339 virtnet_stats
= per_cpu_ptr(vi
->stats
, i
);
2340 u64_stats_init(&virtnet_stats
->tx_syncp
);
2341 u64_stats_init(&virtnet_stats
->rx_syncp
);
2344 INIT_WORK(&vi
->config_work
, virtnet_config_changed_work
);
2346 /* If we can receive ANY GSO packets, we must allocate large ones. */
2347 if (virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_TSO4
) ||
2348 virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_TSO6
) ||
2349 virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_ECN
) ||
2350 virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_UFO
))
2351 vi
->big_packets
= true;
2353 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MRG_RXBUF
))
2354 vi
->mergeable_rx_bufs
= true;
2356 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MRG_RXBUF
) ||
2357 virtio_has_feature(vdev
, VIRTIO_F_VERSION_1
))
2358 vi
->hdr_len
= sizeof(struct virtio_net_hdr_mrg_rxbuf
);
2360 vi
->hdr_len
= sizeof(struct virtio_net_hdr
);
2362 if (virtio_has_feature(vdev
, VIRTIO_F_ANY_LAYOUT
) ||
2363 virtio_has_feature(vdev
, VIRTIO_F_VERSION_1
))
2364 vi
->any_header_sg
= true;
2366 if (virtio_has_feature(vdev
, VIRTIO_NET_F_CTRL_VQ
))
2369 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MTU
)) {
2370 mtu
= virtio_cread16(vdev
,
2371 offsetof(struct virtio_net_config
,
2373 if (mtu
< dev
->min_mtu
) {
2374 __virtio_clear_bit(vdev
, VIRTIO_NET_F_MTU
);
2381 if (vi
->any_header_sg
)
2382 dev
->needed_headroom
= vi
->hdr_len
;
2384 /* Enable multiqueue by default */
2385 if (num_online_cpus() >= max_queue_pairs
)
2386 vi
->curr_queue_pairs
= max_queue_pairs
;
2388 vi
->curr_queue_pairs
= num_online_cpus();
2389 vi
->max_queue_pairs
= max_queue_pairs
;
2391 /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
2397 if (vi
->mergeable_rx_bufs
)
2398 dev
->sysfs_rx_queue_group
= &virtio_net_mrg_rx_group
;
2400 netif_set_real_num_tx_queues(dev
, vi
->curr_queue_pairs
);
2401 netif_set_real_num_rx_queues(dev
, vi
->curr_queue_pairs
);
2403 virtnet_init_settings(dev
);
2405 err
= register_netdev(dev
);
2407 pr_debug("virtio_net: registering device failed\n");
2411 virtio_device_ready(vdev
);
2413 err
= virtnet_cpu_notif_add(vi
);
2415 pr_debug("virtio_net: registering cpu notifier failed\n");
2416 goto free_unregister_netdev
;
2419 virtnet_set_queues(vi
, vi
->curr_queue_pairs
);
2421 /* Assume link up if device can't report link status,
2422 otherwise get link status from config. */
2423 if (virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_STATUS
)) {
2424 netif_carrier_off(dev
);
2425 schedule_work(&vi
->config_work
);
2427 vi
->status
= VIRTIO_NET_S_LINK_UP
;
2428 netif_carrier_on(dev
);
2431 pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
2432 dev
->name
, max_queue_pairs
);
2436 free_unregister_netdev
:
2437 vi
->vdev
->config
->reset(vdev
);
2439 unregister_netdev(dev
);
2441 cancel_delayed_work_sync(&vi
->refill
);
2442 free_receive_page_frags(vi
);
2443 virtnet_del_vqs(vi
);
2445 free_percpu(vi
->stats
);
2451 static void _remove_vq_common(struct virtnet_info
*vi
)
2453 vi
->vdev
->config
->reset(vi
->vdev
);
2454 free_unused_bufs(vi
);
2455 _free_receive_bufs(vi
);
2456 free_receive_page_frags(vi
);
2457 virtnet_del_vqs(vi
);
2460 static void remove_vq_common(struct virtnet_info
*vi
)
2462 vi
->vdev
->config
->reset(vi
->vdev
);
2464 /* Free unused buffers in both send and recv, if any. */
2465 free_unused_bufs(vi
);
2467 free_receive_bufs(vi
);
2469 free_receive_page_frags(vi
);
2471 virtnet_del_vqs(vi
);
2474 static void virtnet_remove(struct virtio_device
*vdev
)
2476 struct virtnet_info
*vi
= vdev
->priv
;
2478 virtnet_cpu_notif_remove(vi
);
2480 /* Make sure no work handler is accessing the device. */
2481 flush_work(&vi
->config_work
);
2483 unregister_netdev(vi
->dev
);
2485 remove_vq_common(vi
);
2487 free_percpu(vi
->stats
);
2488 free_netdev(vi
->dev
);
2491 #ifdef CONFIG_PM_SLEEP
2492 static int virtnet_freeze(struct virtio_device
*vdev
)
2494 struct virtnet_info
*vi
= vdev
->priv
;
2496 virtnet_cpu_notif_remove(vi
);
2497 virtnet_freeze_down(vdev
);
2498 remove_vq_common(vi
);
2503 static int virtnet_restore(struct virtio_device
*vdev
)
2505 struct virtnet_info
*vi
= vdev
->priv
;
2508 err
= virtnet_restore_up(vdev
);
2511 virtnet_set_queues(vi
, vi
->curr_queue_pairs
);
2513 err
= virtnet_cpu_notif_add(vi
);
2521 static struct virtio_device_id id_table
[] = {
2522 { VIRTIO_ID_NET
, VIRTIO_DEV_ANY_ID
},
2526 #define VIRTNET_FEATURES \
2527 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
2529 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
2530 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
2531 VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
2532 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
2533 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
2534 VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
2535 VIRTIO_NET_F_CTRL_MAC_ADDR, \
2538 static unsigned int features
[] = {
2542 static unsigned int features_legacy
[] = {
2545 VIRTIO_F_ANY_LAYOUT
,
2548 static struct virtio_driver virtio_net_driver
= {
2549 .feature_table
= features
,
2550 .feature_table_size
= ARRAY_SIZE(features
),
2551 .feature_table_legacy
= features_legacy
,
2552 .feature_table_size_legacy
= ARRAY_SIZE(features_legacy
),
2553 .driver
.name
= KBUILD_MODNAME
,
2554 .driver
.owner
= THIS_MODULE
,
2555 .id_table
= id_table
,
2556 .probe
= virtnet_probe
,
2557 .remove
= virtnet_remove
,
2558 .config_changed
= virtnet_config_changed
,
2559 #ifdef CONFIG_PM_SLEEP
2560 .freeze
= virtnet_freeze
,
2561 .restore
= virtnet_restore
,
2565 static __init
int virtio_net_driver_init(void)
2569 ret
= cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN
, "virtio/net:online",
2571 virtnet_cpu_down_prep
);
2574 virtionet_online
= ret
;
2575 ret
= cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD
, "virtio/net:dead",
2576 NULL
, virtnet_cpu_dead
);
2580 ret
= register_virtio_driver(&virtio_net_driver
);
2585 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD
);
2587 cpuhp_remove_multi_state(virtionet_online
);
2591 module_init(virtio_net_driver_init
);
2593 static __exit
void virtio_net_driver_exit(void)
2595 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD
);
2596 cpuhp_remove_multi_state(virtionet_online
);
2597 unregister_virtio_driver(&virtio_net_driver
);
2599 module_exit(virtio_net_driver_exit
);
2601 MODULE_DEVICE_TABLE(virtio
, id_table
);
2602 MODULE_DESCRIPTION("Virtio network driver");
2603 MODULE_LICENSE("GPL");