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 static const unsigned long guest_offloads
[] = {
61 VIRTIO_NET_F_GUEST_TSO4
,
62 VIRTIO_NET_F_GUEST_TSO6
,
63 VIRTIO_NET_F_GUEST_ECN
,
64 VIRTIO_NET_F_GUEST_UFO
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.$index */
88 struct napi_struct napi
;
91 /* Internal representation of a receive virtqueue */
92 struct receive_queue
{
93 /* Virtqueue associated with this receive_queue */
96 struct napi_struct napi
;
98 struct bpf_prog __rcu
*xdp_prog
;
100 /* Chain pages by the private ptr. */
103 /* Average packet length for mergeable receive buffers. */
104 struct ewma_pkt_len mrg_avg_pkt_len
;
106 /* Page frag for packet buffer allocation. */
107 struct page_frag alloc_frag
;
109 /* RX: fragments + linear part + virtio header */
110 struct scatterlist sg
[MAX_SKB_FRAGS
+ 2];
112 /* Min single buffer size for mergeable buffers case. */
113 unsigned int min_buf_len
;
115 /* Name of this receive queue: input.$index */
119 struct virtnet_info
{
120 struct virtio_device
*vdev
;
121 struct virtqueue
*cvq
;
122 struct net_device
*dev
;
123 struct send_queue
*sq
;
124 struct receive_queue
*rq
;
127 /* Max # of queue pairs supported by the device */
130 /* # of queue pairs currently used by the driver */
131 u16 curr_queue_pairs
;
133 /* # of XDP queue pairs currently used by the driver */
136 /* I like... big packets and I cannot lie! */
139 /* Host will merge rx buffers for big packets (shake it! shake it!) */
140 bool mergeable_rx_bufs
;
142 /* Has control virtqueue */
145 /* Host can handle any s/g split between our header and packet data */
148 /* Packet virtio header size */
151 /* Active statistics */
152 struct virtnet_stats __percpu
*stats
;
154 /* Work struct for refilling if we run low on memory. */
155 struct delayed_work refill
;
157 /* Work struct for config space updates */
158 struct work_struct config_work
;
160 /* Does the affinity hint is set for virtqueues? */
161 bool affinity_hint_set
;
163 /* CPU hotplug instances for online & dead */
164 struct hlist_node node
;
165 struct hlist_node node_dead
;
167 /* Control VQ buffers: protected by the rtnl lock */
168 struct virtio_net_ctrl_hdr ctrl_hdr
;
169 virtio_net_ctrl_ack ctrl_status
;
170 struct virtio_net_ctrl_mq ctrl_mq
;
176 /* Ethtool settings */
180 unsigned long guest_offloads
;
183 struct padded_vnet_hdr
{
184 struct virtio_net_hdr_mrg_rxbuf hdr
;
186 * hdr is in a separate sg buffer, and data sg buffer shares same page
187 * with this header sg. This padding makes next sg 16 byte aligned
193 /* Converting between virtqueue no. and kernel tx/rx queue no.
194 * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
196 static int vq2txq(struct virtqueue
*vq
)
198 return (vq
->index
- 1) / 2;
201 static int txq2vq(int txq
)
206 static int vq2rxq(struct virtqueue
*vq
)
208 return vq
->index
/ 2;
211 static int rxq2vq(int rxq
)
216 static inline struct virtio_net_hdr_mrg_rxbuf
*skb_vnet_hdr(struct sk_buff
*skb
)
218 return (struct virtio_net_hdr_mrg_rxbuf
*)skb
->cb
;
222 * private is used to chain pages for big packets, put the whole
223 * most recent used list in the beginning for reuse
225 static void give_pages(struct receive_queue
*rq
, struct page
*page
)
229 /* Find end of list, sew whole thing into vi->rq.pages. */
230 for (end
= page
; end
->private; end
= (struct page
*)end
->private);
231 end
->private = (unsigned long)rq
->pages
;
235 static struct page
*get_a_page(struct receive_queue
*rq
, gfp_t gfp_mask
)
237 struct page
*p
= rq
->pages
;
240 rq
->pages
= (struct page
*)p
->private;
241 /* clear private here, it is used to chain pages */
244 p
= alloc_page(gfp_mask
);
248 static void virtqueue_napi_schedule(struct napi_struct
*napi
,
249 struct virtqueue
*vq
)
251 if (napi_schedule_prep(napi
)) {
252 virtqueue_disable_cb(vq
);
253 __napi_schedule(napi
);
257 static void virtqueue_napi_complete(struct napi_struct
*napi
,
258 struct virtqueue
*vq
, int processed
)
262 opaque
= virtqueue_enable_cb_prepare(vq
);
263 if (napi_complete_done(napi
, processed
) &&
264 unlikely(virtqueue_poll(vq
, opaque
)))
265 virtqueue_napi_schedule(napi
, vq
);
268 static void skb_xmit_done(struct virtqueue
*vq
)
270 struct virtnet_info
*vi
= vq
->vdev
->priv
;
271 struct napi_struct
*napi
= &vi
->sq
[vq2txq(vq
)].napi
;
273 /* Suppress further interrupts. */
274 virtqueue_disable_cb(vq
);
277 virtqueue_napi_schedule(napi
, vq
);
279 /* We were probably waiting for more output buffers. */
280 netif_wake_subqueue(vi
->dev
, vq2txq(vq
));
283 #define MRG_CTX_HEADER_SHIFT 22
284 static void *mergeable_len_to_ctx(unsigned int truesize
,
285 unsigned int headroom
)
287 return (void *)(unsigned long)((headroom
<< MRG_CTX_HEADER_SHIFT
) | truesize
);
290 static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx
)
292 return (unsigned long)mrg_ctx
>> MRG_CTX_HEADER_SHIFT
;
295 static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx
)
297 return (unsigned long)mrg_ctx
& ((1 << MRG_CTX_HEADER_SHIFT
) - 1);
300 /* Called from bottom half context */
301 static struct sk_buff
*page_to_skb(struct virtnet_info
*vi
,
302 struct receive_queue
*rq
,
303 struct page
*page
, unsigned int offset
,
304 unsigned int len
, unsigned int truesize
)
307 struct virtio_net_hdr_mrg_rxbuf
*hdr
;
308 unsigned int copy
, hdr_len
, hdr_padded_len
;
311 p
= page_address(page
) + offset
;
313 /* copy small packet so we can reuse these pages for small data */
314 skb
= napi_alloc_skb(&rq
->napi
, GOOD_COPY_LEN
);
318 hdr
= skb_vnet_hdr(skb
);
320 hdr_len
= vi
->hdr_len
;
321 if (vi
->mergeable_rx_bufs
)
322 hdr_padded_len
= sizeof(*hdr
);
324 hdr_padded_len
= sizeof(struct padded_vnet_hdr
);
326 memcpy(hdr
, p
, hdr_len
);
329 offset
+= hdr_padded_len
;
333 if (copy
> skb_tailroom(skb
))
334 copy
= skb_tailroom(skb
);
335 skb_put_data(skb
, p
, copy
);
340 if (vi
->mergeable_rx_bufs
) {
342 skb_add_rx_frag(skb
, 0, page
, offset
, len
, truesize
);
349 * Verify that we can indeed put this data into a skb.
350 * This is here to handle cases when the device erroneously
351 * tries to receive more than is possible. This is usually
352 * the case of a broken device.
354 if (unlikely(len
> MAX_SKB_FRAGS
* PAGE_SIZE
)) {
355 net_dbg_ratelimited("%s: too much data\n", skb
->dev
->name
);
359 BUG_ON(offset
>= PAGE_SIZE
);
361 unsigned int frag_size
= min((unsigned)PAGE_SIZE
- offset
, len
);
362 skb_add_rx_frag(skb
, skb_shinfo(skb
)->nr_frags
, page
, offset
,
363 frag_size
, truesize
);
365 page
= (struct page
*)page
->private;
370 give_pages(rq
, page
);
375 static bool virtnet_xdp_xmit(struct virtnet_info
*vi
,
376 struct xdp_buff
*xdp
)
378 struct virtio_net_hdr_mrg_rxbuf
*hdr
;
380 struct send_queue
*sq
;
385 qp
= vi
->curr_queue_pairs
- vi
->xdp_queue_pairs
+ smp_processor_id();
388 /* Free up any pending old buffers before queueing new ones. */
389 while ((xdp_sent
= virtqueue_get_buf(sq
->vq
, &len
)) != NULL
) {
390 struct page
*sent_page
= virt_to_head_page(xdp_sent
);
395 xdp
->data
-= vi
->hdr_len
;
396 /* Zero header and leave csum up to XDP layers */
398 memset(hdr
, 0, vi
->hdr_len
);
400 sg_init_one(sq
->sg
, xdp
->data
, xdp
->data_end
- xdp
->data
);
402 err
= virtqueue_add_outbuf(sq
->vq
, sq
->sg
, 1, xdp
->data
, GFP_ATOMIC
);
404 struct page
*page
= virt_to_head_page(xdp
->data
);
410 virtqueue_kick(sq
->vq
);
414 static unsigned int virtnet_get_headroom(struct virtnet_info
*vi
)
416 return vi
->xdp_queue_pairs
? VIRTIO_XDP_HEADROOM
: 0;
419 /* We copy the packet for XDP in the following cases:
421 * 1) Packet is scattered across multiple rx buffers.
422 * 2) Headroom space is insufficient.
424 * This is inefficient but it's a temporary condition that
425 * we hit right after XDP is enabled and until queue is refilled
426 * with large buffers with sufficient headroom - so it should affect
427 * at most queue size packets.
428 * Afterwards, the conditions to enable
429 * XDP should preclude the underlying device from sending packets
430 * across multiple buffers (num_buf > 1), and we make sure buffers
431 * have enough headroom.
433 static struct page
*xdp_linearize_page(struct receive_queue
*rq
,
440 struct page
*page
= alloc_page(GFP_ATOMIC
);
445 memcpy(page_address(page
) + page_off
, page_address(p
) + offset
, *len
);
453 buf
= virtqueue_get_buf(rq
->vq
, &buflen
);
457 p
= virt_to_head_page(buf
);
458 off
= buf
- page_address(p
);
460 /* guard against a misconfigured or uncooperative backend that
461 * is sending packet larger than the MTU.
463 if ((page_off
+ buflen
) > PAGE_SIZE
) {
468 memcpy(page_address(page
) + page_off
,
469 page_address(p
) + off
, buflen
);
474 /* Headroom does not contribute to packet length */
475 *len
= page_off
- VIRTIO_XDP_HEADROOM
;
478 __free_pages(page
, 0);
482 static struct sk_buff
*receive_small(struct net_device
*dev
,
483 struct virtnet_info
*vi
,
484 struct receive_queue
*rq
,
485 void *buf
, void *ctx
,
489 struct bpf_prog
*xdp_prog
;
490 unsigned int xdp_headroom
= (unsigned long)ctx
;
491 unsigned int header_offset
= VIRTNET_RX_PAD
+ xdp_headroom
;
492 unsigned int headroom
= vi
->hdr_len
+ header_offset
;
493 unsigned int buflen
= SKB_DATA_ALIGN(GOOD_PACKET_LEN
+ headroom
) +
494 SKB_DATA_ALIGN(sizeof(struct skb_shared_info
));
495 struct page
*page
= virt_to_head_page(buf
);
496 unsigned int delta
= 0;
497 struct page
*xdp_page
;
501 xdp_prog
= rcu_dereference(rq
->xdp_prog
);
503 struct virtio_net_hdr_mrg_rxbuf
*hdr
= buf
+ header_offset
;
508 if (unlikely(hdr
->hdr
.gso_type
|| hdr
->hdr
.flags
))
511 if (unlikely(xdp_headroom
< virtnet_get_headroom(vi
))) {
512 int offset
= buf
- page_address(page
) + header_offset
;
513 unsigned int tlen
= len
+ vi
->hdr_len
;
516 xdp_headroom
= virtnet_get_headroom(vi
);
517 header_offset
= VIRTNET_RX_PAD
+ xdp_headroom
;
518 headroom
= vi
->hdr_len
+ header_offset
;
519 buflen
= SKB_DATA_ALIGN(GOOD_PACKET_LEN
+ headroom
) +
520 SKB_DATA_ALIGN(sizeof(struct skb_shared_info
));
521 xdp_page
= xdp_linearize_page(rq
, &num_buf
, page
,
522 offset
, header_offset
,
527 buf
= page_address(xdp_page
);
532 xdp
.data_hard_start
= buf
+ VIRTNET_RX_PAD
+ vi
->hdr_len
;
533 xdp
.data
= xdp
.data_hard_start
+ xdp_headroom
;
534 xdp
.data_end
= xdp
.data
+ len
;
535 orig_data
= xdp
.data
;
536 act
= bpf_prog_run_xdp(xdp_prog
, &xdp
);
540 /* Recalculate length in case bpf program changed it */
541 delta
= orig_data
- xdp
.data
;
544 if (unlikely(!virtnet_xdp_xmit(vi
, &xdp
)))
545 trace_xdp_exception(vi
->dev
, xdp_prog
, act
);
549 bpf_warn_invalid_xdp_action(act
);
551 trace_xdp_exception(vi
->dev
, xdp_prog
, act
);
558 skb
= build_skb(buf
, buflen
);
563 skb_reserve(skb
, headroom
- delta
);
564 skb_put(skb
, len
+ delta
);
566 buf
+= header_offset
;
567 memcpy(skb_vnet_hdr(skb
), buf
, vi
->hdr_len
);
568 } /* keep zeroed vnet hdr since packet was changed by bpf */
575 dev
->stats
.rx_dropped
++;
581 static struct sk_buff
*receive_big(struct net_device
*dev
,
582 struct virtnet_info
*vi
,
583 struct receive_queue
*rq
,
587 struct page
*page
= buf
;
588 struct sk_buff
*skb
= page_to_skb(vi
, rq
, page
, 0, len
, PAGE_SIZE
);
596 dev
->stats
.rx_dropped
++;
597 give_pages(rq
, page
);
601 static struct sk_buff
*receive_mergeable(struct net_device
*dev
,
602 struct virtnet_info
*vi
,
603 struct receive_queue
*rq
,
608 struct virtio_net_hdr_mrg_rxbuf
*hdr
= buf
;
609 u16 num_buf
= virtio16_to_cpu(vi
->vdev
, hdr
->num_buffers
);
610 struct page
*page
= virt_to_head_page(buf
);
611 int offset
= buf
- page_address(page
);
612 struct sk_buff
*head_skb
, *curr_skb
;
613 struct bpf_prog
*xdp_prog
;
614 unsigned int truesize
;
615 unsigned int headroom
= mergeable_ctx_to_headroom(ctx
);
620 xdp_prog
= rcu_dereference(rq
->xdp_prog
);
622 struct page
*xdp_page
;
627 /* This happens when rx buffer size is underestimated */
628 if (unlikely(num_buf
> 1 ||
629 headroom
< virtnet_get_headroom(vi
))) {
630 /* linearize data for XDP */
631 xdp_page
= xdp_linearize_page(rq
, &num_buf
,
637 offset
= VIRTIO_XDP_HEADROOM
;
642 /* Transient failure which in theory could occur if
643 * in-flight packets from before XDP was enabled reach
644 * the receive path after XDP is loaded. In practice I
645 * was not able to create this condition.
647 if (unlikely(hdr
->hdr
.gso_type
))
650 /* Allow consuming headroom but reserve enough space to push
651 * the descriptor on if we get an XDP_TX return code.
653 data
= page_address(xdp_page
) + offset
;
654 xdp
.data_hard_start
= data
- VIRTIO_XDP_HEADROOM
+ vi
->hdr_len
;
655 xdp
.data
= data
+ vi
->hdr_len
;
656 xdp
.data_end
= xdp
.data
+ (len
- vi
->hdr_len
);
657 act
= bpf_prog_run_xdp(xdp_prog
, &xdp
);
660 ewma_pkt_len_add(&rq
->mrg_avg_pkt_len
, len
);
664 /* recalculate offset to account for any header
665 * adjustments. Note other cases do not build an
666 * skb and avoid using offset
669 page_address(xdp_page
) - vi
->hdr_len
;
671 /* We can only create skb based on xdp_page. */
672 if (unlikely(xdp_page
!= page
)) {
675 head_skb
= page_to_skb(vi
, rq
, xdp_page
,
676 offset
, len
, PAGE_SIZE
);
681 if (unlikely(!virtnet_xdp_xmit(vi
, &xdp
)))
682 trace_xdp_exception(vi
->dev
, xdp_prog
, act
);
683 if (unlikely(xdp_page
!= page
))
688 bpf_warn_invalid_xdp_action(act
);
690 trace_xdp_exception(vi
->dev
, xdp_prog
, act
);
692 if (unlikely(xdp_page
!= page
))
693 __free_pages(xdp_page
, 0);
699 truesize
= mergeable_ctx_to_truesize(ctx
);
700 if (unlikely(len
> truesize
)) {
701 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
702 dev
->name
, len
, (unsigned long)ctx
);
703 dev
->stats
.rx_length_errors
++;
707 head_skb
= page_to_skb(vi
, rq
, page
, offset
, len
, truesize
);
710 if (unlikely(!curr_skb
))
715 buf
= virtqueue_get_buf_ctx(rq
->vq
, &len
, &ctx
);
716 if (unlikely(!ctx
)) {
717 pr_debug("%s: rx error: %d buffers out of %d missing\n",
719 virtio16_to_cpu(vi
->vdev
,
721 dev
->stats
.rx_length_errors
++;
725 page
= virt_to_head_page(buf
);
727 truesize
= mergeable_ctx_to_truesize(ctx
);
728 if (unlikely(len
> truesize
)) {
729 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
730 dev
->name
, len
, (unsigned long)ctx
);
731 dev
->stats
.rx_length_errors
++;
735 num_skb_frags
= skb_shinfo(curr_skb
)->nr_frags
;
736 if (unlikely(num_skb_frags
== MAX_SKB_FRAGS
)) {
737 struct sk_buff
*nskb
= alloc_skb(0, GFP_ATOMIC
);
741 if (curr_skb
== head_skb
)
742 skb_shinfo(curr_skb
)->frag_list
= nskb
;
744 curr_skb
->next
= nskb
;
746 head_skb
->truesize
+= nskb
->truesize
;
749 if (curr_skb
!= head_skb
) {
750 head_skb
->data_len
+= len
;
751 head_skb
->len
+= len
;
752 head_skb
->truesize
+= truesize
;
754 offset
= buf
- page_address(page
);
755 if (skb_can_coalesce(curr_skb
, num_skb_frags
, page
, offset
)) {
757 skb_coalesce_rx_frag(curr_skb
, num_skb_frags
- 1,
760 skb_add_rx_frag(curr_skb
, num_skb_frags
, page
,
761 offset
, len
, truesize
);
765 ewma_pkt_len_add(&rq
->mrg_avg_pkt_len
, head_skb
->len
);
773 buf
= virtqueue_get_buf(rq
->vq
, &len
);
774 if (unlikely(!buf
)) {
775 pr_debug("%s: rx error: %d buffers missing\n",
777 dev
->stats
.rx_length_errors
++;
780 page
= virt_to_head_page(buf
);
784 dev
->stats
.rx_dropped
++;
785 dev_kfree_skb(head_skb
);
790 static int receive_buf(struct virtnet_info
*vi
, struct receive_queue
*rq
,
791 void *buf
, unsigned int len
, void **ctx
)
793 struct net_device
*dev
= vi
->dev
;
795 struct virtio_net_hdr_mrg_rxbuf
*hdr
;
798 if (unlikely(len
< vi
->hdr_len
+ ETH_HLEN
)) {
799 pr_debug("%s: short packet %i\n", dev
->name
, len
);
800 dev
->stats
.rx_length_errors
++;
801 if (vi
->mergeable_rx_bufs
) {
802 put_page(virt_to_head_page(buf
));
803 } else if (vi
->big_packets
) {
806 put_page(virt_to_head_page(buf
));
811 if (vi
->mergeable_rx_bufs
)
812 skb
= receive_mergeable(dev
, vi
, rq
, buf
, ctx
, len
);
813 else if (vi
->big_packets
)
814 skb
= receive_big(dev
, vi
, rq
, buf
, len
);
816 skb
= receive_small(dev
, vi
, rq
, buf
, ctx
, len
);
821 hdr
= skb_vnet_hdr(skb
);
825 if (hdr
->hdr
.flags
& VIRTIO_NET_HDR_F_DATA_VALID
)
826 skb
->ip_summed
= CHECKSUM_UNNECESSARY
;
828 if (virtio_net_hdr_to_skb(skb
, &hdr
->hdr
,
829 virtio_is_little_endian(vi
->vdev
))) {
830 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
831 dev
->name
, hdr
->hdr
.gso_type
,
836 skb
->protocol
= eth_type_trans(skb
, dev
);
837 pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
838 ntohs(skb
->protocol
), skb
->len
, skb
->pkt_type
);
840 napi_gro_receive(&rq
->napi
, skb
);
844 dev
->stats
.rx_frame_errors
++;
849 /* Unlike mergeable buffers, all buffers are allocated to the
850 * same size, except for the headroom. For this reason we do
851 * not need to use mergeable_len_to_ctx here - it is enough
852 * to store the headroom as the context ignoring the truesize.
854 static int add_recvbuf_small(struct virtnet_info
*vi
, struct receive_queue
*rq
,
857 struct page_frag
*alloc_frag
= &rq
->alloc_frag
;
859 unsigned int xdp_headroom
= virtnet_get_headroom(vi
);
860 void *ctx
= (void *)(unsigned long)xdp_headroom
;
861 int len
= vi
->hdr_len
+ VIRTNET_RX_PAD
+ GOOD_PACKET_LEN
+ xdp_headroom
;
864 len
= SKB_DATA_ALIGN(len
) +
865 SKB_DATA_ALIGN(sizeof(struct skb_shared_info
));
866 if (unlikely(!skb_page_frag_refill(len
, alloc_frag
, gfp
)))
869 buf
= (char *)page_address(alloc_frag
->page
) + alloc_frag
->offset
;
870 get_page(alloc_frag
->page
);
871 alloc_frag
->offset
+= len
;
872 sg_init_one(rq
->sg
, buf
+ VIRTNET_RX_PAD
+ xdp_headroom
,
873 vi
->hdr_len
+ GOOD_PACKET_LEN
);
874 err
= virtqueue_add_inbuf_ctx(rq
->vq
, rq
->sg
, 1, buf
, ctx
, gfp
);
876 put_page(virt_to_head_page(buf
));
880 static int add_recvbuf_big(struct virtnet_info
*vi
, struct receive_queue
*rq
,
883 struct page
*first
, *list
= NULL
;
887 sg_init_table(rq
->sg
, MAX_SKB_FRAGS
+ 2);
889 /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
890 for (i
= MAX_SKB_FRAGS
+ 1; i
> 1; --i
) {
891 first
= get_a_page(rq
, gfp
);
894 give_pages(rq
, list
);
897 sg_set_buf(&rq
->sg
[i
], page_address(first
), PAGE_SIZE
);
899 /* chain new page in list head to match sg */
900 first
->private = (unsigned long)list
;
904 first
= get_a_page(rq
, gfp
);
906 give_pages(rq
, list
);
909 p
= page_address(first
);
911 /* rq->sg[0], rq->sg[1] share the same page */
912 /* a separated rq->sg[0] for header - required in case !any_header_sg */
913 sg_set_buf(&rq
->sg
[0], p
, vi
->hdr_len
);
915 /* rq->sg[1] for data packet, from offset */
916 offset
= sizeof(struct padded_vnet_hdr
);
917 sg_set_buf(&rq
->sg
[1], p
+ offset
, PAGE_SIZE
- offset
);
919 /* chain first in list head */
920 first
->private = (unsigned long)list
;
921 err
= virtqueue_add_inbuf(rq
->vq
, rq
->sg
, MAX_SKB_FRAGS
+ 2,
924 give_pages(rq
, first
);
929 static unsigned int get_mergeable_buf_len(struct receive_queue
*rq
,
930 struct ewma_pkt_len
*avg_pkt_len
)
932 const size_t hdr_len
= sizeof(struct virtio_net_hdr_mrg_rxbuf
);
935 len
= hdr_len
+ clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len
),
936 rq
->min_buf_len
, PAGE_SIZE
- hdr_len
);
937 return ALIGN(len
, L1_CACHE_BYTES
);
940 static int add_recvbuf_mergeable(struct virtnet_info
*vi
,
941 struct receive_queue
*rq
, gfp_t gfp
)
943 struct page_frag
*alloc_frag
= &rq
->alloc_frag
;
944 unsigned int headroom
= virtnet_get_headroom(vi
);
948 unsigned int len
, hole
;
950 len
= get_mergeable_buf_len(rq
, &rq
->mrg_avg_pkt_len
);
951 if (unlikely(!skb_page_frag_refill(len
+ headroom
, alloc_frag
, gfp
)))
954 buf
= (char *)page_address(alloc_frag
->page
) + alloc_frag
->offset
;
955 buf
+= headroom
; /* advance address leaving hole at front of pkt */
956 get_page(alloc_frag
->page
);
957 alloc_frag
->offset
+= len
+ headroom
;
958 hole
= alloc_frag
->size
- alloc_frag
->offset
;
959 if (hole
< len
+ headroom
) {
960 /* To avoid internal fragmentation, if there is very likely not
961 * enough space for another buffer, add the remaining space to
962 * the current buffer.
965 alloc_frag
->offset
+= hole
;
968 sg_init_one(rq
->sg
, buf
, len
);
969 ctx
= mergeable_len_to_ctx(len
, headroom
);
970 err
= virtqueue_add_inbuf_ctx(rq
->vq
, rq
->sg
, 1, buf
, ctx
, gfp
);
972 put_page(virt_to_head_page(buf
));
978 * Returns false if we couldn't fill entirely (OOM).
980 * Normally run in the receive path, but can also be run from ndo_open
981 * before we're receiving packets, or from refill_work which is
982 * careful to disable receiving (using napi_disable).
984 static bool try_fill_recv(struct virtnet_info
*vi
, struct receive_queue
*rq
,
992 if (vi
->mergeable_rx_bufs
)
993 err
= add_recvbuf_mergeable(vi
, rq
, gfp
);
994 else if (vi
->big_packets
)
995 err
= add_recvbuf_big(vi
, rq
, gfp
);
997 err
= add_recvbuf_small(vi
, rq
, gfp
);
999 oom
= err
== -ENOMEM
;
1002 } while (rq
->vq
->num_free
);
1003 virtqueue_kick(rq
->vq
);
1007 static void skb_recv_done(struct virtqueue
*rvq
)
1009 struct virtnet_info
*vi
= rvq
->vdev
->priv
;
1010 struct receive_queue
*rq
= &vi
->rq
[vq2rxq(rvq
)];
1012 virtqueue_napi_schedule(&rq
->napi
, rvq
);
1015 static void virtnet_napi_enable(struct virtqueue
*vq
, struct napi_struct
*napi
)
1019 /* If all buffers were filled by other side before we napi_enabled, we
1020 * won't get another interrupt, so process any outstanding packets now.
1021 * Call local_bh_enable after to trigger softIRQ processing.
1024 virtqueue_napi_schedule(napi
, vq
);
1028 static void virtnet_napi_tx_enable(struct virtnet_info
*vi
,
1029 struct virtqueue
*vq
,
1030 struct napi_struct
*napi
)
1035 /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
1036 * enable the feature if this is likely affine with the transmit path.
1038 if (!vi
->affinity_hint_set
) {
1043 return virtnet_napi_enable(vq
, napi
);
1046 static void virtnet_napi_tx_disable(struct napi_struct
*napi
)
1052 static void refill_work(struct work_struct
*work
)
1054 struct virtnet_info
*vi
=
1055 container_of(work
, struct virtnet_info
, refill
.work
);
1059 for (i
= 0; i
< vi
->curr_queue_pairs
; i
++) {
1060 struct receive_queue
*rq
= &vi
->rq
[i
];
1062 napi_disable(&rq
->napi
);
1063 still_empty
= !try_fill_recv(vi
, rq
, GFP_KERNEL
);
1064 virtnet_napi_enable(rq
->vq
, &rq
->napi
);
1066 /* In theory, this can happen: if we don't get any buffers in
1067 * we will *never* try to fill again.
1070 schedule_delayed_work(&vi
->refill
, HZ
/2);
1074 static int virtnet_receive(struct receive_queue
*rq
, int budget
)
1076 struct virtnet_info
*vi
= rq
->vq
->vdev
->priv
;
1077 unsigned int len
, received
= 0, bytes
= 0;
1079 struct virtnet_stats
*stats
= this_cpu_ptr(vi
->stats
);
1081 if (!vi
->big_packets
|| vi
->mergeable_rx_bufs
) {
1084 while (received
< budget
&&
1085 (buf
= virtqueue_get_buf_ctx(rq
->vq
, &len
, &ctx
))) {
1086 bytes
+= receive_buf(vi
, rq
, buf
, len
, ctx
);
1090 while (received
< budget
&&
1091 (buf
= virtqueue_get_buf(rq
->vq
, &len
)) != NULL
) {
1092 bytes
+= receive_buf(vi
, rq
, buf
, len
, NULL
);
1097 if (rq
->vq
->num_free
> virtqueue_get_vring_size(rq
->vq
) / 2) {
1098 if (!try_fill_recv(vi
, rq
, GFP_ATOMIC
))
1099 schedule_delayed_work(&vi
->refill
, 0);
1102 u64_stats_update_begin(&stats
->rx_syncp
);
1103 stats
->rx_bytes
+= bytes
;
1104 stats
->rx_packets
+= received
;
1105 u64_stats_update_end(&stats
->rx_syncp
);
1110 static void free_old_xmit_skbs(struct send_queue
*sq
)
1112 struct sk_buff
*skb
;
1114 struct virtnet_info
*vi
= sq
->vq
->vdev
->priv
;
1115 struct virtnet_stats
*stats
= this_cpu_ptr(vi
->stats
);
1116 unsigned int packets
= 0;
1117 unsigned int bytes
= 0;
1119 while ((skb
= virtqueue_get_buf(sq
->vq
, &len
)) != NULL
) {
1120 pr_debug("Sent skb %p\n", skb
);
1125 dev_consume_skb_any(skb
);
1128 /* Avoid overhead when no packets have been processed
1129 * happens when called speculatively from start_xmit.
1134 u64_stats_update_begin(&stats
->tx_syncp
);
1135 stats
->tx_bytes
+= bytes
;
1136 stats
->tx_packets
+= packets
;
1137 u64_stats_update_end(&stats
->tx_syncp
);
1140 static void virtnet_poll_cleantx(struct receive_queue
*rq
)
1142 struct virtnet_info
*vi
= rq
->vq
->vdev
->priv
;
1143 unsigned int index
= vq2rxq(rq
->vq
);
1144 struct send_queue
*sq
= &vi
->sq
[index
];
1145 struct netdev_queue
*txq
= netdev_get_tx_queue(vi
->dev
, index
);
1147 if (!sq
->napi
.weight
)
1150 if (__netif_tx_trylock(txq
)) {
1151 free_old_xmit_skbs(sq
);
1152 __netif_tx_unlock(txq
);
1155 if (sq
->vq
->num_free
>= 2 + MAX_SKB_FRAGS
)
1156 netif_tx_wake_queue(txq
);
1159 static int virtnet_poll(struct napi_struct
*napi
, int budget
)
1161 struct receive_queue
*rq
=
1162 container_of(napi
, struct receive_queue
, napi
);
1163 unsigned int received
;
1165 virtnet_poll_cleantx(rq
);
1167 received
= virtnet_receive(rq
, budget
);
1169 /* Out of packets? */
1170 if (received
< budget
)
1171 virtqueue_napi_complete(napi
, rq
->vq
, received
);
1176 static int virtnet_open(struct net_device
*dev
)
1178 struct virtnet_info
*vi
= netdev_priv(dev
);
1181 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1182 if (i
< vi
->curr_queue_pairs
)
1183 /* Make sure we have some buffers: if oom use wq. */
1184 if (!try_fill_recv(vi
, &vi
->rq
[i
], GFP_KERNEL
))
1185 schedule_delayed_work(&vi
->refill
, 0);
1186 virtnet_napi_enable(vi
->rq
[i
].vq
, &vi
->rq
[i
].napi
);
1187 virtnet_napi_tx_enable(vi
, vi
->sq
[i
].vq
, &vi
->sq
[i
].napi
);
1193 static int virtnet_poll_tx(struct napi_struct
*napi
, int budget
)
1195 struct send_queue
*sq
= container_of(napi
, struct send_queue
, napi
);
1196 struct virtnet_info
*vi
= sq
->vq
->vdev
->priv
;
1197 struct netdev_queue
*txq
= netdev_get_tx_queue(vi
->dev
, vq2txq(sq
->vq
));
1199 __netif_tx_lock(txq
, raw_smp_processor_id());
1200 free_old_xmit_skbs(sq
);
1201 __netif_tx_unlock(txq
);
1203 virtqueue_napi_complete(napi
, sq
->vq
, 0);
1205 if (sq
->vq
->num_free
>= 2 + MAX_SKB_FRAGS
)
1206 netif_tx_wake_queue(txq
);
1211 static int xmit_skb(struct send_queue
*sq
, struct sk_buff
*skb
)
1213 struct virtio_net_hdr_mrg_rxbuf
*hdr
;
1214 const unsigned char *dest
= ((struct ethhdr
*)skb
->data
)->h_dest
;
1215 struct virtnet_info
*vi
= sq
->vq
->vdev
->priv
;
1217 unsigned hdr_len
= vi
->hdr_len
;
1220 pr_debug("%s: xmit %p %pM\n", vi
->dev
->name
, skb
, dest
);
1222 can_push
= vi
->any_header_sg
&&
1223 !((unsigned long)skb
->data
& (__alignof__(*hdr
) - 1)) &&
1224 !skb_header_cloned(skb
) && skb_headroom(skb
) >= hdr_len
;
1225 /* Even if we can, don't push here yet as this would skew
1226 * csum_start offset below. */
1228 hdr
= (struct virtio_net_hdr_mrg_rxbuf
*)(skb
->data
- hdr_len
);
1230 hdr
= skb_vnet_hdr(skb
);
1232 if (virtio_net_hdr_from_skb(skb
, &hdr
->hdr
,
1233 virtio_is_little_endian(vi
->vdev
), false))
1236 if (vi
->mergeable_rx_bufs
)
1237 hdr
->num_buffers
= 0;
1239 sg_init_table(sq
->sg
, skb_shinfo(skb
)->nr_frags
+ (can_push
? 1 : 2));
1241 __skb_push(skb
, hdr_len
);
1242 num_sg
= skb_to_sgvec(skb
, sq
->sg
, 0, skb
->len
);
1243 if (unlikely(num_sg
< 0))
1245 /* Pull header back to avoid skew in tx bytes calculations. */
1246 __skb_pull(skb
, hdr_len
);
1248 sg_set_buf(sq
->sg
, hdr
, hdr_len
);
1249 num_sg
= skb_to_sgvec(skb
, sq
->sg
+ 1, 0, skb
->len
);
1250 if (unlikely(num_sg
< 0))
1254 return virtqueue_add_outbuf(sq
->vq
, sq
->sg
, num_sg
, skb
, GFP_ATOMIC
);
1257 static netdev_tx_t
start_xmit(struct sk_buff
*skb
, struct net_device
*dev
)
1259 struct virtnet_info
*vi
= netdev_priv(dev
);
1260 int qnum
= skb_get_queue_mapping(skb
);
1261 struct send_queue
*sq
= &vi
->sq
[qnum
];
1263 struct netdev_queue
*txq
= netdev_get_tx_queue(dev
, qnum
);
1264 bool kick
= !skb
->xmit_more
;
1265 bool use_napi
= sq
->napi
.weight
;
1267 /* Free up any pending old buffers before queueing new ones. */
1268 free_old_xmit_skbs(sq
);
1270 if (use_napi
&& kick
)
1271 virtqueue_enable_cb_delayed(sq
->vq
);
1273 /* timestamp packet in software */
1274 skb_tx_timestamp(skb
);
1276 /* Try to transmit */
1277 err
= xmit_skb(sq
, skb
);
1279 /* This should not happen! */
1280 if (unlikely(err
)) {
1281 dev
->stats
.tx_fifo_errors
++;
1282 if (net_ratelimit())
1284 "Unexpected TXQ (%d) queue failure: %d\n", qnum
, err
);
1285 dev
->stats
.tx_dropped
++;
1286 dev_kfree_skb_any(skb
);
1287 return NETDEV_TX_OK
;
1290 /* Don't wait up for transmitted skbs to be freed. */
1296 /* If running out of space, stop queue to avoid getting packets that we
1297 * are then unable to transmit.
1298 * An alternative would be to force queuing layer to requeue the skb by
1299 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1300 * returned in a normal path of operation: it means that driver is not
1301 * maintaining the TX queue stop/start state properly, and causes
1302 * the stack to do a non-trivial amount of useless work.
1303 * Since most packets only take 1 or 2 ring slots, stopping the queue
1304 * early means 16 slots are typically wasted.
1306 if (sq
->vq
->num_free
< 2+MAX_SKB_FRAGS
) {
1307 netif_stop_subqueue(dev
, qnum
);
1309 unlikely(!virtqueue_enable_cb_delayed(sq
->vq
))) {
1310 /* More just got used, free them then recheck. */
1311 free_old_xmit_skbs(sq
);
1312 if (sq
->vq
->num_free
>= 2+MAX_SKB_FRAGS
) {
1313 netif_start_subqueue(dev
, qnum
);
1314 virtqueue_disable_cb(sq
->vq
);
1319 if (kick
|| netif_xmit_stopped(txq
))
1320 virtqueue_kick(sq
->vq
);
1322 return NETDEV_TX_OK
;
1326 * Send command via the control virtqueue and check status. Commands
1327 * supported by the hypervisor, as indicated by feature bits, should
1328 * never fail unless improperly formatted.
1330 static bool virtnet_send_command(struct virtnet_info
*vi
, u8
class, u8 cmd
,
1331 struct scatterlist
*out
)
1333 struct scatterlist
*sgs
[4], hdr
, stat
;
1334 unsigned out_num
= 0, tmp
;
1336 /* Caller should know better */
1337 BUG_ON(!virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_CTRL_VQ
));
1339 vi
->ctrl_status
= ~0;
1340 vi
->ctrl_hdr
.class = class;
1341 vi
->ctrl_hdr
.cmd
= cmd
;
1343 sg_init_one(&hdr
, &vi
->ctrl_hdr
, sizeof(vi
->ctrl_hdr
));
1344 sgs
[out_num
++] = &hdr
;
1347 sgs
[out_num
++] = out
;
1349 /* Add return status. */
1350 sg_init_one(&stat
, &vi
->ctrl_status
, sizeof(vi
->ctrl_status
));
1351 sgs
[out_num
] = &stat
;
1353 BUG_ON(out_num
+ 1 > ARRAY_SIZE(sgs
));
1354 virtqueue_add_sgs(vi
->cvq
, sgs
, out_num
, 1, vi
, GFP_ATOMIC
);
1356 if (unlikely(!virtqueue_kick(vi
->cvq
)))
1357 return vi
->ctrl_status
== VIRTIO_NET_OK
;
1359 /* Spin for a response, the kick causes an ioport write, trapping
1360 * into the hypervisor, so the request should be handled immediately.
1362 while (!virtqueue_get_buf(vi
->cvq
, &tmp
) &&
1363 !virtqueue_is_broken(vi
->cvq
))
1366 return vi
->ctrl_status
== VIRTIO_NET_OK
;
1369 static int virtnet_set_mac_address(struct net_device
*dev
, void *p
)
1371 struct virtnet_info
*vi
= netdev_priv(dev
);
1372 struct virtio_device
*vdev
= vi
->vdev
;
1374 struct sockaddr
*addr
;
1375 struct scatterlist sg
;
1377 addr
= kmemdup(p
, sizeof(*addr
), GFP_KERNEL
);
1381 ret
= eth_prepare_mac_addr_change(dev
, addr
);
1385 if (virtio_has_feature(vdev
, VIRTIO_NET_F_CTRL_MAC_ADDR
)) {
1386 sg_init_one(&sg
, addr
->sa_data
, dev
->addr_len
);
1387 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_MAC
,
1388 VIRTIO_NET_CTRL_MAC_ADDR_SET
, &sg
)) {
1389 dev_warn(&vdev
->dev
,
1390 "Failed to set mac address by vq command.\n");
1394 } else if (virtio_has_feature(vdev
, VIRTIO_NET_F_MAC
) &&
1395 !virtio_has_feature(vdev
, VIRTIO_F_VERSION_1
)) {
1398 /* Naturally, this has an atomicity problem. */
1399 for (i
= 0; i
< dev
->addr_len
; i
++)
1400 virtio_cwrite8(vdev
,
1401 offsetof(struct virtio_net_config
, mac
) +
1402 i
, addr
->sa_data
[i
]);
1405 eth_commit_mac_addr_change(dev
, p
);
1413 static void virtnet_stats(struct net_device
*dev
,
1414 struct rtnl_link_stats64
*tot
)
1416 struct virtnet_info
*vi
= netdev_priv(dev
);
1420 for_each_possible_cpu(cpu
) {
1421 struct virtnet_stats
*stats
= per_cpu_ptr(vi
->stats
, cpu
);
1422 u64 tpackets
, tbytes
, rpackets
, rbytes
;
1425 start
= u64_stats_fetch_begin_irq(&stats
->tx_syncp
);
1426 tpackets
= stats
->tx_packets
;
1427 tbytes
= stats
->tx_bytes
;
1428 } while (u64_stats_fetch_retry_irq(&stats
->tx_syncp
, start
));
1431 start
= u64_stats_fetch_begin_irq(&stats
->rx_syncp
);
1432 rpackets
= stats
->rx_packets
;
1433 rbytes
= stats
->rx_bytes
;
1434 } while (u64_stats_fetch_retry_irq(&stats
->rx_syncp
, start
));
1436 tot
->rx_packets
+= rpackets
;
1437 tot
->tx_packets
+= tpackets
;
1438 tot
->rx_bytes
+= rbytes
;
1439 tot
->tx_bytes
+= tbytes
;
1442 tot
->tx_dropped
= dev
->stats
.tx_dropped
;
1443 tot
->tx_fifo_errors
= dev
->stats
.tx_fifo_errors
;
1444 tot
->rx_dropped
= dev
->stats
.rx_dropped
;
1445 tot
->rx_length_errors
= dev
->stats
.rx_length_errors
;
1446 tot
->rx_frame_errors
= dev
->stats
.rx_frame_errors
;
1449 #ifdef CONFIG_NET_POLL_CONTROLLER
1450 static void virtnet_netpoll(struct net_device
*dev
)
1452 struct virtnet_info
*vi
= netdev_priv(dev
);
1455 for (i
= 0; i
< vi
->curr_queue_pairs
; i
++)
1456 napi_schedule(&vi
->rq
[i
].napi
);
1460 static void virtnet_ack_link_announce(struct virtnet_info
*vi
)
1463 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_ANNOUNCE
,
1464 VIRTIO_NET_CTRL_ANNOUNCE_ACK
, NULL
))
1465 dev_warn(&vi
->dev
->dev
, "Failed to ack link announce.\n");
1469 static int _virtnet_set_queues(struct virtnet_info
*vi
, u16 queue_pairs
)
1471 struct scatterlist sg
;
1472 struct net_device
*dev
= vi
->dev
;
1474 if (!vi
->has_cvq
|| !virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_MQ
))
1477 vi
->ctrl_mq
.virtqueue_pairs
= cpu_to_virtio16(vi
->vdev
, queue_pairs
);
1478 sg_init_one(&sg
, &vi
->ctrl_mq
, sizeof(vi
->ctrl_mq
));
1480 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_MQ
,
1481 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET
, &sg
)) {
1482 dev_warn(&dev
->dev
, "Fail to set num of queue pairs to %d\n",
1486 vi
->curr_queue_pairs
= queue_pairs
;
1487 /* virtnet_open() will refill when device is going to up. */
1488 if (dev
->flags
& IFF_UP
)
1489 schedule_delayed_work(&vi
->refill
, 0);
1495 static int virtnet_set_queues(struct virtnet_info
*vi
, u16 queue_pairs
)
1500 err
= _virtnet_set_queues(vi
, queue_pairs
);
1505 static int virtnet_close(struct net_device
*dev
)
1507 struct virtnet_info
*vi
= netdev_priv(dev
);
1510 /* Make sure refill_work doesn't re-enable napi! */
1511 cancel_delayed_work_sync(&vi
->refill
);
1513 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1514 napi_disable(&vi
->rq
[i
].napi
);
1515 virtnet_napi_tx_disable(&vi
->sq
[i
].napi
);
1521 static void virtnet_set_rx_mode(struct net_device
*dev
)
1523 struct virtnet_info
*vi
= netdev_priv(dev
);
1524 struct scatterlist sg
[2];
1525 struct virtio_net_ctrl_mac
*mac_data
;
1526 struct netdev_hw_addr
*ha
;
1532 /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1533 if (!virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_CTRL_RX
))
1536 vi
->ctrl_promisc
= ((dev
->flags
& IFF_PROMISC
) != 0);
1537 vi
->ctrl_allmulti
= ((dev
->flags
& IFF_ALLMULTI
) != 0);
1539 sg_init_one(sg
, &vi
->ctrl_promisc
, sizeof(vi
->ctrl_promisc
));
1541 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_RX
,
1542 VIRTIO_NET_CTRL_RX_PROMISC
, sg
))
1543 dev_warn(&dev
->dev
, "Failed to %sable promisc mode.\n",
1544 vi
->ctrl_promisc
? "en" : "dis");
1546 sg_init_one(sg
, &vi
->ctrl_allmulti
, sizeof(vi
->ctrl_allmulti
));
1548 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_RX
,
1549 VIRTIO_NET_CTRL_RX_ALLMULTI
, sg
))
1550 dev_warn(&dev
->dev
, "Failed to %sable allmulti mode.\n",
1551 vi
->ctrl_allmulti
? "en" : "dis");
1553 uc_count
= netdev_uc_count(dev
);
1554 mc_count
= netdev_mc_count(dev
);
1555 /* MAC filter - use one buffer for both lists */
1556 buf
= kzalloc(((uc_count
+ mc_count
) * ETH_ALEN
) +
1557 (2 * sizeof(mac_data
->entries
)), GFP_ATOMIC
);
1562 sg_init_table(sg
, 2);
1564 /* Store the unicast list and count in the front of the buffer */
1565 mac_data
->entries
= cpu_to_virtio32(vi
->vdev
, uc_count
);
1567 netdev_for_each_uc_addr(ha
, dev
)
1568 memcpy(&mac_data
->macs
[i
++][0], ha
->addr
, ETH_ALEN
);
1570 sg_set_buf(&sg
[0], mac_data
,
1571 sizeof(mac_data
->entries
) + (uc_count
* ETH_ALEN
));
1573 /* multicast list and count fill the end */
1574 mac_data
= (void *)&mac_data
->macs
[uc_count
][0];
1576 mac_data
->entries
= cpu_to_virtio32(vi
->vdev
, mc_count
);
1578 netdev_for_each_mc_addr(ha
, dev
)
1579 memcpy(&mac_data
->macs
[i
++][0], ha
->addr
, ETH_ALEN
);
1581 sg_set_buf(&sg
[1], mac_data
,
1582 sizeof(mac_data
->entries
) + (mc_count
* ETH_ALEN
));
1584 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_MAC
,
1585 VIRTIO_NET_CTRL_MAC_TABLE_SET
, sg
))
1586 dev_warn(&dev
->dev
, "Failed to set MAC filter table.\n");
1591 static int virtnet_vlan_rx_add_vid(struct net_device
*dev
,
1592 __be16 proto
, u16 vid
)
1594 struct virtnet_info
*vi
= netdev_priv(dev
);
1595 struct scatterlist sg
;
1598 sg_init_one(&sg
, &vi
->ctrl_vid
, sizeof(vi
->ctrl_vid
));
1600 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_VLAN
,
1601 VIRTIO_NET_CTRL_VLAN_ADD
, &sg
))
1602 dev_warn(&dev
->dev
, "Failed to add VLAN ID %d.\n", vid
);
1606 static int virtnet_vlan_rx_kill_vid(struct net_device
*dev
,
1607 __be16 proto
, u16 vid
)
1609 struct virtnet_info
*vi
= netdev_priv(dev
);
1610 struct scatterlist sg
;
1613 sg_init_one(&sg
, &vi
->ctrl_vid
, sizeof(vi
->ctrl_vid
));
1615 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_VLAN
,
1616 VIRTIO_NET_CTRL_VLAN_DEL
, &sg
))
1617 dev_warn(&dev
->dev
, "Failed to kill VLAN ID %d.\n", vid
);
1621 static void virtnet_clean_affinity(struct virtnet_info
*vi
, long hcpu
)
1625 if (vi
->affinity_hint_set
) {
1626 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1627 virtqueue_set_affinity(vi
->rq
[i
].vq
, -1);
1628 virtqueue_set_affinity(vi
->sq
[i
].vq
, -1);
1631 vi
->affinity_hint_set
= false;
1635 static void virtnet_set_affinity(struct virtnet_info
*vi
)
1640 /* In multiqueue mode, when the number of cpu is equal to the number of
1641 * queue pairs, we let the queue pairs to be private to one cpu by
1642 * setting the affinity hint to eliminate the contention.
1644 if (vi
->curr_queue_pairs
== 1 ||
1645 vi
->max_queue_pairs
!= num_online_cpus()) {
1646 virtnet_clean_affinity(vi
, -1);
1651 for_each_online_cpu(cpu
) {
1652 virtqueue_set_affinity(vi
->rq
[i
].vq
, cpu
);
1653 virtqueue_set_affinity(vi
->sq
[i
].vq
, cpu
);
1654 netif_set_xps_queue(vi
->dev
, cpumask_of(cpu
), i
);
1658 vi
->affinity_hint_set
= true;
1661 static int virtnet_cpu_online(unsigned int cpu
, struct hlist_node
*node
)
1663 struct virtnet_info
*vi
= hlist_entry_safe(node
, struct virtnet_info
,
1665 virtnet_set_affinity(vi
);
1669 static int virtnet_cpu_dead(unsigned int cpu
, struct hlist_node
*node
)
1671 struct virtnet_info
*vi
= hlist_entry_safe(node
, struct virtnet_info
,
1673 virtnet_set_affinity(vi
);
1677 static int virtnet_cpu_down_prep(unsigned int cpu
, struct hlist_node
*node
)
1679 struct virtnet_info
*vi
= hlist_entry_safe(node
, struct virtnet_info
,
1682 virtnet_clean_affinity(vi
, cpu
);
1686 static enum cpuhp_state virtionet_online
;
1688 static int virtnet_cpu_notif_add(struct virtnet_info
*vi
)
1692 ret
= cpuhp_state_add_instance_nocalls(virtionet_online
, &vi
->node
);
1695 ret
= cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD
,
1699 cpuhp_state_remove_instance_nocalls(virtionet_online
, &vi
->node
);
1703 static void virtnet_cpu_notif_remove(struct virtnet_info
*vi
)
1705 cpuhp_state_remove_instance_nocalls(virtionet_online
, &vi
->node
);
1706 cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD
,
1710 static void virtnet_get_ringparam(struct net_device
*dev
,
1711 struct ethtool_ringparam
*ring
)
1713 struct virtnet_info
*vi
= netdev_priv(dev
);
1715 ring
->rx_max_pending
= virtqueue_get_vring_size(vi
->rq
[0].vq
);
1716 ring
->tx_max_pending
= virtqueue_get_vring_size(vi
->sq
[0].vq
);
1717 ring
->rx_pending
= ring
->rx_max_pending
;
1718 ring
->tx_pending
= ring
->tx_max_pending
;
1722 static void virtnet_get_drvinfo(struct net_device
*dev
,
1723 struct ethtool_drvinfo
*info
)
1725 struct virtnet_info
*vi
= netdev_priv(dev
);
1726 struct virtio_device
*vdev
= vi
->vdev
;
1728 strlcpy(info
->driver
, KBUILD_MODNAME
, sizeof(info
->driver
));
1729 strlcpy(info
->version
, VIRTNET_DRIVER_VERSION
, sizeof(info
->version
));
1730 strlcpy(info
->bus_info
, virtio_bus_name(vdev
), sizeof(info
->bus_info
));
1734 /* TODO: Eliminate OOO packets during switching */
1735 static int virtnet_set_channels(struct net_device
*dev
,
1736 struct ethtool_channels
*channels
)
1738 struct virtnet_info
*vi
= netdev_priv(dev
);
1739 u16 queue_pairs
= channels
->combined_count
;
1742 /* We don't support separate rx/tx channels.
1743 * We don't allow setting 'other' channels.
1745 if (channels
->rx_count
|| channels
->tx_count
|| channels
->other_count
)
1748 if (queue_pairs
> vi
->max_queue_pairs
|| queue_pairs
== 0)
1751 /* For now we don't support modifying channels while XDP is loaded
1752 * also when XDP is loaded all RX queues have XDP programs so we only
1753 * need to check a single RX queue.
1755 if (vi
->rq
[0].xdp_prog
)
1759 err
= _virtnet_set_queues(vi
, queue_pairs
);
1761 netif_set_real_num_tx_queues(dev
, queue_pairs
);
1762 netif_set_real_num_rx_queues(dev
, queue_pairs
);
1764 virtnet_set_affinity(vi
);
1771 static void virtnet_get_channels(struct net_device
*dev
,
1772 struct ethtool_channels
*channels
)
1774 struct virtnet_info
*vi
= netdev_priv(dev
);
1776 channels
->combined_count
= vi
->curr_queue_pairs
;
1777 channels
->max_combined
= vi
->max_queue_pairs
;
1778 channels
->max_other
= 0;
1779 channels
->rx_count
= 0;
1780 channels
->tx_count
= 0;
1781 channels
->other_count
= 0;
1784 /* Check if the user is trying to change anything besides speed/duplex */
1786 virtnet_validate_ethtool_cmd(const struct ethtool_link_ksettings
*cmd
)
1788 struct ethtool_link_ksettings diff1
= *cmd
;
1789 struct ethtool_link_ksettings diff2
= {};
1791 /* cmd is always set so we need to clear it, validate the port type
1792 * and also without autonegotiation we can ignore advertising
1794 diff1
.base
.speed
= 0;
1795 diff2
.base
.port
= PORT_OTHER
;
1796 ethtool_link_ksettings_zero_link_mode(&diff1
, advertising
);
1797 diff1
.base
.duplex
= 0;
1799 diff1
.base
.link_mode_masks_nwords
= 0;
1801 return !memcmp(&diff1
.base
, &diff2
.base
, sizeof(diff1
.base
)) &&
1802 bitmap_empty(diff1
.link_modes
.supported
,
1803 __ETHTOOL_LINK_MODE_MASK_NBITS
) &&
1804 bitmap_empty(diff1
.link_modes
.advertising
,
1805 __ETHTOOL_LINK_MODE_MASK_NBITS
) &&
1806 bitmap_empty(diff1
.link_modes
.lp_advertising
,
1807 __ETHTOOL_LINK_MODE_MASK_NBITS
);
1810 static int virtnet_set_link_ksettings(struct net_device
*dev
,
1811 const struct ethtool_link_ksettings
*cmd
)
1813 struct virtnet_info
*vi
= netdev_priv(dev
);
1816 speed
= cmd
->base
.speed
;
1817 /* don't allow custom speed and duplex */
1818 if (!ethtool_validate_speed(speed
) ||
1819 !ethtool_validate_duplex(cmd
->base
.duplex
) ||
1820 !virtnet_validate_ethtool_cmd(cmd
))
1823 vi
->duplex
= cmd
->base
.duplex
;
1828 static int virtnet_get_link_ksettings(struct net_device
*dev
,
1829 struct ethtool_link_ksettings
*cmd
)
1831 struct virtnet_info
*vi
= netdev_priv(dev
);
1833 cmd
->base
.speed
= vi
->speed
;
1834 cmd
->base
.duplex
= vi
->duplex
;
1835 cmd
->base
.port
= PORT_OTHER
;
1840 static void virtnet_init_settings(struct net_device
*dev
)
1842 struct virtnet_info
*vi
= netdev_priv(dev
);
1844 vi
->speed
= SPEED_UNKNOWN
;
1845 vi
->duplex
= DUPLEX_UNKNOWN
;
1848 static const struct ethtool_ops virtnet_ethtool_ops
= {
1849 .get_drvinfo
= virtnet_get_drvinfo
,
1850 .get_link
= ethtool_op_get_link
,
1851 .get_ringparam
= virtnet_get_ringparam
,
1852 .set_channels
= virtnet_set_channels
,
1853 .get_channels
= virtnet_get_channels
,
1854 .get_ts_info
= ethtool_op_get_ts_info
,
1855 .get_link_ksettings
= virtnet_get_link_ksettings
,
1856 .set_link_ksettings
= virtnet_set_link_ksettings
,
1859 static void virtnet_freeze_down(struct virtio_device
*vdev
)
1861 struct virtnet_info
*vi
= vdev
->priv
;
1864 /* Make sure no work handler is accessing the device */
1865 flush_work(&vi
->config_work
);
1867 netif_device_detach(vi
->dev
);
1868 netif_tx_disable(vi
->dev
);
1869 cancel_delayed_work_sync(&vi
->refill
);
1871 if (netif_running(vi
->dev
)) {
1872 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1873 napi_disable(&vi
->rq
[i
].napi
);
1874 virtnet_napi_tx_disable(&vi
->sq
[i
].napi
);
1879 static int init_vqs(struct virtnet_info
*vi
);
1881 static int virtnet_restore_up(struct virtio_device
*vdev
)
1883 struct virtnet_info
*vi
= vdev
->priv
;
1890 virtio_device_ready(vdev
);
1892 if (netif_running(vi
->dev
)) {
1893 for (i
= 0; i
< vi
->curr_queue_pairs
; i
++)
1894 if (!try_fill_recv(vi
, &vi
->rq
[i
], GFP_KERNEL
))
1895 schedule_delayed_work(&vi
->refill
, 0);
1897 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1898 virtnet_napi_enable(vi
->rq
[i
].vq
, &vi
->rq
[i
].napi
);
1899 virtnet_napi_tx_enable(vi
, vi
->sq
[i
].vq
,
1904 netif_device_attach(vi
->dev
);
1908 static int virtnet_set_guest_offloads(struct virtnet_info
*vi
, u64 offloads
)
1910 struct scatterlist sg
;
1911 vi
->ctrl_offloads
= cpu_to_virtio64(vi
->vdev
, offloads
);
1913 sg_init_one(&sg
, &vi
->ctrl_offloads
, sizeof(vi
->ctrl_offloads
));
1915 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_GUEST_OFFLOADS
,
1916 VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET
, &sg
)) {
1917 dev_warn(&vi
->dev
->dev
, "Fail to set guest offload. \n");
1924 static int virtnet_clear_guest_offloads(struct virtnet_info
*vi
)
1928 if (!vi
->guest_offloads
)
1931 if (virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_CSUM
))
1932 offloads
= 1ULL << VIRTIO_NET_F_GUEST_CSUM
;
1934 return virtnet_set_guest_offloads(vi
, offloads
);
1937 static int virtnet_restore_guest_offloads(struct virtnet_info
*vi
)
1939 u64 offloads
= vi
->guest_offloads
;
1941 if (!vi
->guest_offloads
)
1943 if (virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_CSUM
))
1944 offloads
|= 1ULL << VIRTIO_NET_F_GUEST_CSUM
;
1946 return virtnet_set_guest_offloads(vi
, offloads
);
1949 static int virtnet_xdp_set(struct net_device
*dev
, struct bpf_prog
*prog
,
1950 struct netlink_ext_ack
*extack
)
1952 unsigned long int max_sz
= PAGE_SIZE
- sizeof(struct padded_vnet_hdr
);
1953 struct virtnet_info
*vi
= netdev_priv(dev
);
1954 struct bpf_prog
*old_prog
;
1955 u16 xdp_qp
= 0, curr_qp
;
1958 if (!virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS
)
1959 && (virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_TSO4
) ||
1960 virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_TSO6
) ||
1961 virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_ECN
) ||
1962 virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_UFO
))) {
1963 NL_SET_ERR_MSG_MOD(extack
, "Can't set XDP while host is implementing LRO, disable LRO first");
1967 if (vi
->mergeable_rx_bufs
&& !vi
->any_header_sg
) {
1968 NL_SET_ERR_MSG_MOD(extack
, "XDP expects header/data in single page, any_header_sg required");
1972 if (dev
->mtu
> max_sz
) {
1973 NL_SET_ERR_MSG_MOD(extack
, "MTU too large to enable XDP");
1974 netdev_warn(dev
, "XDP requires MTU less than %lu\n", max_sz
);
1978 curr_qp
= vi
->curr_queue_pairs
- vi
->xdp_queue_pairs
;
1980 xdp_qp
= nr_cpu_ids
;
1982 /* XDP requires extra queues for XDP_TX */
1983 if (curr_qp
+ xdp_qp
> vi
->max_queue_pairs
) {
1984 NL_SET_ERR_MSG_MOD(extack
, "Too few free TX rings available");
1985 netdev_warn(dev
, "request %i queues but max is %i\n",
1986 curr_qp
+ xdp_qp
, vi
->max_queue_pairs
);
1991 prog
= bpf_prog_add(prog
, vi
->max_queue_pairs
- 1);
1993 return PTR_ERR(prog
);
1996 /* Make sure NAPI is not using any XDP TX queues for RX. */
1997 for (i
= 0; i
< vi
->max_queue_pairs
; i
++)
1998 napi_disable(&vi
->rq
[i
].napi
);
2000 netif_set_real_num_rx_queues(dev
, curr_qp
+ xdp_qp
);
2001 err
= _virtnet_set_queues(vi
, curr_qp
+ xdp_qp
);
2004 vi
->xdp_queue_pairs
= xdp_qp
;
2006 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2007 old_prog
= rtnl_dereference(vi
->rq
[i
].xdp_prog
);
2008 rcu_assign_pointer(vi
->rq
[i
].xdp_prog
, prog
);
2011 virtnet_clear_guest_offloads(vi
);
2013 virtnet_restore_guest_offloads(vi
);
2016 bpf_prog_put(old_prog
);
2017 virtnet_napi_enable(vi
->rq
[i
].vq
, &vi
->rq
[i
].napi
);
2023 for (i
= 0; i
< vi
->max_queue_pairs
; i
++)
2024 virtnet_napi_enable(vi
->rq
[i
].vq
, &vi
->rq
[i
].napi
);
2026 bpf_prog_sub(prog
, vi
->max_queue_pairs
- 1);
2030 static u32
virtnet_xdp_query(struct net_device
*dev
)
2032 struct virtnet_info
*vi
= netdev_priv(dev
);
2033 const struct bpf_prog
*xdp_prog
;
2036 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2037 xdp_prog
= rtnl_dereference(vi
->rq
[i
].xdp_prog
);
2039 return xdp_prog
->aux
->id
;
2044 static int virtnet_xdp(struct net_device
*dev
, struct netdev_xdp
*xdp
)
2046 switch (xdp
->command
) {
2047 case XDP_SETUP_PROG
:
2048 return virtnet_xdp_set(dev
, xdp
->prog
, xdp
->extack
);
2049 case XDP_QUERY_PROG
:
2050 xdp
->prog_id
= virtnet_xdp_query(dev
);
2051 xdp
->prog_attached
= !!xdp
->prog_id
;
2058 static const struct net_device_ops virtnet_netdev
= {
2059 .ndo_open
= virtnet_open
,
2060 .ndo_stop
= virtnet_close
,
2061 .ndo_start_xmit
= start_xmit
,
2062 .ndo_validate_addr
= eth_validate_addr
,
2063 .ndo_set_mac_address
= virtnet_set_mac_address
,
2064 .ndo_set_rx_mode
= virtnet_set_rx_mode
,
2065 .ndo_get_stats64
= virtnet_stats
,
2066 .ndo_vlan_rx_add_vid
= virtnet_vlan_rx_add_vid
,
2067 .ndo_vlan_rx_kill_vid
= virtnet_vlan_rx_kill_vid
,
2068 #ifdef CONFIG_NET_POLL_CONTROLLER
2069 .ndo_poll_controller
= virtnet_netpoll
,
2071 .ndo_xdp
= virtnet_xdp
,
2072 .ndo_features_check
= passthru_features_check
,
2075 static void virtnet_config_changed_work(struct work_struct
*work
)
2077 struct virtnet_info
*vi
=
2078 container_of(work
, struct virtnet_info
, config_work
);
2081 if (virtio_cread_feature(vi
->vdev
, VIRTIO_NET_F_STATUS
,
2082 struct virtio_net_config
, status
, &v
) < 0)
2085 if (v
& VIRTIO_NET_S_ANNOUNCE
) {
2086 netdev_notify_peers(vi
->dev
);
2087 virtnet_ack_link_announce(vi
);
2090 /* Ignore unknown (future) status bits */
2091 v
&= VIRTIO_NET_S_LINK_UP
;
2093 if (vi
->status
== v
)
2098 if (vi
->status
& VIRTIO_NET_S_LINK_UP
) {
2099 netif_carrier_on(vi
->dev
);
2100 netif_tx_wake_all_queues(vi
->dev
);
2102 netif_carrier_off(vi
->dev
);
2103 netif_tx_stop_all_queues(vi
->dev
);
2107 static void virtnet_config_changed(struct virtio_device
*vdev
)
2109 struct virtnet_info
*vi
= vdev
->priv
;
2111 schedule_work(&vi
->config_work
);
2114 static void virtnet_free_queues(struct virtnet_info
*vi
)
2118 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2119 napi_hash_del(&vi
->rq
[i
].napi
);
2120 netif_napi_del(&vi
->rq
[i
].napi
);
2121 netif_napi_del(&vi
->sq
[i
].napi
);
2124 /* We called napi_hash_del() before netif_napi_del(),
2125 * we need to respect an RCU grace period before freeing vi->rq
2133 static void _free_receive_bufs(struct virtnet_info
*vi
)
2135 struct bpf_prog
*old_prog
;
2138 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2139 while (vi
->rq
[i
].pages
)
2140 __free_pages(get_a_page(&vi
->rq
[i
], GFP_KERNEL
), 0);
2142 old_prog
= rtnl_dereference(vi
->rq
[i
].xdp_prog
);
2143 RCU_INIT_POINTER(vi
->rq
[i
].xdp_prog
, NULL
);
2145 bpf_prog_put(old_prog
);
2149 static void free_receive_bufs(struct virtnet_info
*vi
)
2152 _free_receive_bufs(vi
);
2156 static void free_receive_page_frags(struct virtnet_info
*vi
)
2159 for (i
= 0; i
< vi
->max_queue_pairs
; i
++)
2160 if (vi
->rq
[i
].alloc_frag
.page
)
2161 put_page(vi
->rq
[i
].alloc_frag
.page
);
2164 static bool is_xdp_raw_buffer_queue(struct virtnet_info
*vi
, int q
)
2166 if (q
< (vi
->curr_queue_pairs
- vi
->xdp_queue_pairs
))
2168 else if (q
< vi
->curr_queue_pairs
)
2174 static void free_unused_bufs(struct virtnet_info
*vi
)
2179 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2180 struct virtqueue
*vq
= vi
->sq
[i
].vq
;
2181 while ((buf
= virtqueue_detach_unused_buf(vq
)) != NULL
) {
2182 if (!is_xdp_raw_buffer_queue(vi
, i
))
2185 put_page(virt_to_head_page(buf
));
2189 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2190 struct virtqueue
*vq
= vi
->rq
[i
].vq
;
2192 while ((buf
= virtqueue_detach_unused_buf(vq
)) != NULL
) {
2193 if (vi
->mergeable_rx_bufs
) {
2194 put_page(virt_to_head_page(buf
));
2195 } else if (vi
->big_packets
) {
2196 give_pages(&vi
->rq
[i
], buf
);
2198 put_page(virt_to_head_page(buf
));
2204 static void virtnet_del_vqs(struct virtnet_info
*vi
)
2206 struct virtio_device
*vdev
= vi
->vdev
;
2208 virtnet_clean_affinity(vi
, -1);
2210 vdev
->config
->del_vqs(vdev
);
2212 virtnet_free_queues(vi
);
2215 /* How large should a single buffer be so a queue full of these can fit at
2216 * least one full packet?
2217 * Logic below assumes the mergeable buffer header is used.
2219 static unsigned int mergeable_min_buf_len(struct virtnet_info
*vi
, struct virtqueue
*vq
)
2221 const unsigned int hdr_len
= sizeof(struct virtio_net_hdr_mrg_rxbuf
);
2222 unsigned int rq_size
= virtqueue_get_vring_size(vq
);
2223 unsigned int packet_len
= vi
->big_packets
? IP_MAX_MTU
: vi
->dev
->max_mtu
;
2224 unsigned int buf_len
= hdr_len
+ ETH_HLEN
+ VLAN_HLEN
+ packet_len
;
2225 unsigned int min_buf_len
= DIV_ROUND_UP(buf_len
, rq_size
);
2227 return max(max(min_buf_len
, hdr_len
) - hdr_len
,
2228 (unsigned int)GOOD_PACKET_LEN
);
2231 static int virtnet_find_vqs(struct virtnet_info
*vi
)
2233 vq_callback_t
**callbacks
;
2234 struct virtqueue
**vqs
;
2240 /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
2241 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
2242 * possible control vq.
2244 total_vqs
= vi
->max_queue_pairs
* 2 +
2245 virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_CTRL_VQ
);
2247 /* Allocate space for find_vqs parameters */
2248 vqs
= kzalloc(total_vqs
* sizeof(*vqs
), GFP_KERNEL
);
2251 callbacks
= kmalloc(total_vqs
* sizeof(*callbacks
), GFP_KERNEL
);
2254 names
= kmalloc(total_vqs
* sizeof(*names
), GFP_KERNEL
);
2257 if (!vi
->big_packets
|| vi
->mergeable_rx_bufs
) {
2258 ctx
= kzalloc(total_vqs
* sizeof(*ctx
), GFP_KERNEL
);
2265 /* Parameters for control virtqueue, if any */
2267 callbacks
[total_vqs
- 1] = NULL
;
2268 names
[total_vqs
- 1] = "control";
2271 /* Allocate/initialize parameters for send/receive virtqueues */
2272 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2273 callbacks
[rxq2vq(i
)] = skb_recv_done
;
2274 callbacks
[txq2vq(i
)] = skb_xmit_done
;
2275 sprintf(vi
->rq
[i
].name
, "input.%d", i
);
2276 sprintf(vi
->sq
[i
].name
, "output.%d", i
);
2277 names
[rxq2vq(i
)] = vi
->rq
[i
].name
;
2278 names
[txq2vq(i
)] = vi
->sq
[i
].name
;
2280 ctx
[rxq2vq(i
)] = true;
2283 ret
= vi
->vdev
->config
->find_vqs(vi
->vdev
, total_vqs
, vqs
, callbacks
,
2289 vi
->cvq
= vqs
[total_vqs
- 1];
2290 if (virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_CTRL_VLAN
))
2291 vi
->dev
->features
|= NETIF_F_HW_VLAN_CTAG_FILTER
;
2294 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2295 vi
->rq
[i
].vq
= vqs
[rxq2vq(i
)];
2296 vi
->rq
[i
].min_buf_len
= mergeable_min_buf_len(vi
, vi
->rq
[i
].vq
);
2297 vi
->sq
[i
].vq
= vqs
[txq2vq(i
)];
2319 static int virtnet_alloc_queues(struct virtnet_info
*vi
)
2323 vi
->sq
= kzalloc(sizeof(*vi
->sq
) * vi
->max_queue_pairs
, GFP_KERNEL
);
2326 vi
->rq
= kzalloc(sizeof(*vi
->rq
) * vi
->max_queue_pairs
, GFP_KERNEL
);
2330 INIT_DELAYED_WORK(&vi
->refill
, refill_work
);
2331 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2332 vi
->rq
[i
].pages
= NULL
;
2333 netif_napi_add(vi
->dev
, &vi
->rq
[i
].napi
, virtnet_poll
,
2335 netif_tx_napi_add(vi
->dev
, &vi
->sq
[i
].napi
, virtnet_poll_tx
,
2336 napi_tx
? napi_weight
: 0);
2338 sg_init_table(vi
->rq
[i
].sg
, ARRAY_SIZE(vi
->rq
[i
].sg
));
2339 ewma_pkt_len_init(&vi
->rq
[i
].mrg_avg_pkt_len
);
2340 sg_init_table(vi
->sq
[i
].sg
, ARRAY_SIZE(vi
->sq
[i
].sg
));
2351 static int init_vqs(struct virtnet_info
*vi
)
2355 /* Allocate send & receive queues */
2356 ret
= virtnet_alloc_queues(vi
);
2360 ret
= virtnet_find_vqs(vi
);
2365 virtnet_set_affinity(vi
);
2371 virtnet_free_queues(vi
);
2377 static ssize_t
mergeable_rx_buffer_size_show(struct netdev_rx_queue
*queue
,
2380 struct virtnet_info
*vi
= netdev_priv(queue
->dev
);
2381 unsigned int queue_index
= get_netdev_rx_queue_index(queue
);
2382 struct ewma_pkt_len
*avg
;
2384 BUG_ON(queue_index
>= vi
->max_queue_pairs
);
2385 avg
= &vi
->rq
[queue_index
].mrg_avg_pkt_len
;
2386 return sprintf(buf
, "%u\n",
2387 get_mergeable_buf_len(&vi
->rq
[queue_index
], avg
));
2390 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute
=
2391 __ATTR_RO(mergeable_rx_buffer_size
);
2393 static struct attribute
*virtio_net_mrg_rx_attrs
[] = {
2394 &mergeable_rx_buffer_size_attribute
.attr
,
2398 static const struct attribute_group virtio_net_mrg_rx_group
= {
2399 .name
= "virtio_net",
2400 .attrs
= virtio_net_mrg_rx_attrs
2404 static bool virtnet_fail_on_feature(struct virtio_device
*vdev
,
2406 const char *fname
, const char *dname
)
2408 if (!virtio_has_feature(vdev
, fbit
))
2411 dev_err(&vdev
->dev
, "device advertises feature %s but not %s",
2417 #define VIRTNET_FAIL_ON(vdev, fbit, dbit) \
2418 virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
2420 static bool virtnet_validate_features(struct virtio_device
*vdev
)
2422 if (!virtio_has_feature(vdev
, VIRTIO_NET_F_CTRL_VQ
) &&
2423 (VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_CTRL_RX
,
2424 "VIRTIO_NET_F_CTRL_VQ") ||
2425 VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_CTRL_VLAN
,
2426 "VIRTIO_NET_F_CTRL_VQ") ||
2427 VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_GUEST_ANNOUNCE
,
2428 "VIRTIO_NET_F_CTRL_VQ") ||
2429 VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_MQ
, "VIRTIO_NET_F_CTRL_VQ") ||
2430 VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_CTRL_MAC_ADDR
,
2431 "VIRTIO_NET_F_CTRL_VQ"))) {
2438 #define MIN_MTU ETH_MIN_MTU
2439 #define MAX_MTU ETH_MAX_MTU
2441 static int virtnet_validate(struct virtio_device
*vdev
)
2443 if (!vdev
->config
->get
) {
2444 dev_err(&vdev
->dev
, "%s failure: config access disabled\n",
2449 if (!virtnet_validate_features(vdev
))
2452 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MTU
)) {
2453 int mtu
= virtio_cread16(vdev
,
2454 offsetof(struct virtio_net_config
,
2457 __virtio_clear_bit(vdev
, VIRTIO_NET_F_MTU
);
2463 static int virtnet_probe(struct virtio_device
*vdev
)
2466 struct net_device
*dev
;
2467 struct virtnet_info
*vi
;
2468 u16 max_queue_pairs
;
2471 /* Find if host supports multiqueue virtio_net device */
2472 err
= virtio_cread_feature(vdev
, VIRTIO_NET_F_MQ
,
2473 struct virtio_net_config
,
2474 max_virtqueue_pairs
, &max_queue_pairs
);
2476 /* We need at least 2 queue's */
2477 if (err
|| max_queue_pairs
< VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN
||
2478 max_queue_pairs
> VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX
||
2479 !virtio_has_feature(vdev
, VIRTIO_NET_F_CTRL_VQ
))
2480 max_queue_pairs
= 1;
2482 /* Allocate ourselves a network device with room for our info */
2483 dev
= alloc_etherdev_mq(sizeof(struct virtnet_info
), max_queue_pairs
);
2487 /* Set up network device as normal. */
2488 dev
->priv_flags
|= IFF_UNICAST_FLT
| IFF_LIVE_ADDR_CHANGE
;
2489 dev
->netdev_ops
= &virtnet_netdev
;
2490 dev
->features
= NETIF_F_HIGHDMA
;
2492 dev
->ethtool_ops
= &virtnet_ethtool_ops
;
2493 SET_NETDEV_DEV(dev
, &vdev
->dev
);
2495 /* Do we support "hardware" checksums? */
2496 if (virtio_has_feature(vdev
, VIRTIO_NET_F_CSUM
)) {
2497 /* This opens up the world of extra features. */
2498 dev
->hw_features
|= NETIF_F_HW_CSUM
| NETIF_F_SG
;
2500 dev
->features
|= NETIF_F_HW_CSUM
| NETIF_F_SG
;
2502 if (virtio_has_feature(vdev
, VIRTIO_NET_F_GSO
)) {
2503 dev
->hw_features
|= NETIF_F_TSO
2504 | NETIF_F_TSO_ECN
| NETIF_F_TSO6
;
2506 /* Individual feature bits: what can host handle? */
2507 if (virtio_has_feature(vdev
, VIRTIO_NET_F_HOST_TSO4
))
2508 dev
->hw_features
|= NETIF_F_TSO
;
2509 if (virtio_has_feature(vdev
, VIRTIO_NET_F_HOST_TSO6
))
2510 dev
->hw_features
|= NETIF_F_TSO6
;
2511 if (virtio_has_feature(vdev
, VIRTIO_NET_F_HOST_ECN
))
2512 dev
->hw_features
|= NETIF_F_TSO_ECN
;
2514 dev
->features
|= NETIF_F_GSO_ROBUST
;
2517 dev
->features
|= dev
->hw_features
& NETIF_F_ALL_TSO
;
2518 /* (!csum && gso) case will be fixed by register_netdev() */
2520 if (virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_CSUM
))
2521 dev
->features
|= NETIF_F_RXCSUM
;
2523 dev
->vlan_features
= dev
->features
;
2525 /* MTU range: 68 - 65535 */
2526 dev
->min_mtu
= MIN_MTU
;
2527 dev
->max_mtu
= MAX_MTU
;
2529 /* Configuration may specify what MAC to use. Otherwise random. */
2530 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MAC
))
2531 virtio_cread_bytes(vdev
,
2532 offsetof(struct virtio_net_config
, mac
),
2533 dev
->dev_addr
, dev
->addr_len
);
2535 eth_hw_addr_random(dev
);
2537 /* Set up our device-specific information */
2538 vi
= netdev_priv(dev
);
2542 vi
->stats
= alloc_percpu(struct virtnet_stats
);
2544 if (vi
->stats
== NULL
)
2547 for_each_possible_cpu(i
) {
2548 struct virtnet_stats
*virtnet_stats
;
2549 virtnet_stats
= per_cpu_ptr(vi
->stats
, i
);
2550 u64_stats_init(&virtnet_stats
->tx_syncp
);
2551 u64_stats_init(&virtnet_stats
->rx_syncp
);
2554 INIT_WORK(&vi
->config_work
, virtnet_config_changed_work
);
2556 /* If we can receive ANY GSO packets, we must allocate large ones. */
2557 if (virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_TSO4
) ||
2558 virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_TSO6
) ||
2559 virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_ECN
) ||
2560 virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_UFO
))
2561 vi
->big_packets
= true;
2563 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MRG_RXBUF
))
2564 vi
->mergeable_rx_bufs
= true;
2566 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MRG_RXBUF
) ||
2567 virtio_has_feature(vdev
, VIRTIO_F_VERSION_1
))
2568 vi
->hdr_len
= sizeof(struct virtio_net_hdr_mrg_rxbuf
);
2570 vi
->hdr_len
= sizeof(struct virtio_net_hdr
);
2572 if (virtio_has_feature(vdev
, VIRTIO_F_ANY_LAYOUT
) ||
2573 virtio_has_feature(vdev
, VIRTIO_F_VERSION_1
))
2574 vi
->any_header_sg
= true;
2576 if (virtio_has_feature(vdev
, VIRTIO_NET_F_CTRL_VQ
))
2579 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MTU
)) {
2580 mtu
= virtio_cread16(vdev
,
2581 offsetof(struct virtio_net_config
,
2583 if (mtu
< dev
->min_mtu
) {
2584 /* Should never trigger: MTU was previously validated
2585 * in virtnet_validate.
2587 dev_err(&vdev
->dev
, "device MTU appears to have changed "
2588 "it is now %d < %d", mtu
, dev
->min_mtu
);
2595 /* TODO: size buffers correctly in this case. */
2596 if (dev
->mtu
> ETH_DATA_LEN
)
2597 vi
->big_packets
= true;
2600 if (vi
->any_header_sg
)
2601 dev
->needed_headroom
= vi
->hdr_len
;
2603 /* Enable multiqueue by default */
2604 if (num_online_cpus() >= max_queue_pairs
)
2605 vi
->curr_queue_pairs
= max_queue_pairs
;
2607 vi
->curr_queue_pairs
= num_online_cpus();
2608 vi
->max_queue_pairs
= max_queue_pairs
;
2610 /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
2616 if (vi
->mergeable_rx_bufs
)
2617 dev
->sysfs_rx_queue_group
= &virtio_net_mrg_rx_group
;
2619 netif_set_real_num_tx_queues(dev
, vi
->curr_queue_pairs
);
2620 netif_set_real_num_rx_queues(dev
, vi
->curr_queue_pairs
);
2622 virtnet_init_settings(dev
);
2624 err
= register_netdev(dev
);
2626 pr_debug("virtio_net: registering device failed\n");
2630 virtio_device_ready(vdev
);
2632 err
= virtnet_cpu_notif_add(vi
);
2634 pr_debug("virtio_net: registering cpu notifier failed\n");
2635 goto free_unregister_netdev
;
2638 virtnet_set_queues(vi
, vi
->curr_queue_pairs
);
2640 /* Assume link up if device can't report link status,
2641 otherwise get link status from config. */
2642 if (virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_STATUS
)) {
2643 netif_carrier_off(dev
);
2644 schedule_work(&vi
->config_work
);
2646 vi
->status
= VIRTIO_NET_S_LINK_UP
;
2647 netif_carrier_on(dev
);
2650 for (i
= 0; i
< ARRAY_SIZE(guest_offloads
); i
++)
2651 if (virtio_has_feature(vi
->vdev
, guest_offloads
[i
]))
2652 set_bit(guest_offloads
[i
], &vi
->guest_offloads
);
2654 pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
2655 dev
->name
, max_queue_pairs
);
2659 free_unregister_netdev
:
2660 vi
->vdev
->config
->reset(vdev
);
2662 unregister_netdev(dev
);
2664 cancel_delayed_work_sync(&vi
->refill
);
2665 free_receive_page_frags(vi
);
2666 virtnet_del_vqs(vi
);
2668 free_percpu(vi
->stats
);
2674 static void remove_vq_common(struct virtnet_info
*vi
)
2676 vi
->vdev
->config
->reset(vi
->vdev
);
2678 /* Free unused buffers in both send and recv, if any. */
2679 free_unused_bufs(vi
);
2681 free_receive_bufs(vi
);
2683 free_receive_page_frags(vi
);
2685 virtnet_del_vqs(vi
);
2688 static void virtnet_remove(struct virtio_device
*vdev
)
2690 struct virtnet_info
*vi
= vdev
->priv
;
2692 virtnet_cpu_notif_remove(vi
);
2694 /* Make sure no work handler is accessing the device. */
2695 flush_work(&vi
->config_work
);
2697 unregister_netdev(vi
->dev
);
2699 remove_vq_common(vi
);
2701 free_percpu(vi
->stats
);
2702 free_netdev(vi
->dev
);
2705 static __maybe_unused
int virtnet_freeze(struct virtio_device
*vdev
)
2707 struct virtnet_info
*vi
= vdev
->priv
;
2709 virtnet_cpu_notif_remove(vi
);
2710 virtnet_freeze_down(vdev
);
2711 remove_vq_common(vi
);
2716 static __maybe_unused
int virtnet_restore(struct virtio_device
*vdev
)
2718 struct virtnet_info
*vi
= vdev
->priv
;
2721 err
= virtnet_restore_up(vdev
);
2724 virtnet_set_queues(vi
, vi
->curr_queue_pairs
);
2726 err
= virtnet_cpu_notif_add(vi
);
2733 static struct virtio_device_id id_table
[] = {
2734 { VIRTIO_ID_NET
, VIRTIO_DEV_ANY_ID
},
2738 #define VIRTNET_FEATURES \
2739 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
2741 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
2742 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
2743 VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
2744 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
2745 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
2746 VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
2747 VIRTIO_NET_F_CTRL_MAC_ADDR, \
2748 VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS
2750 static unsigned int features
[] = {
2754 static unsigned int features_legacy
[] = {
2757 VIRTIO_F_ANY_LAYOUT
,
2760 static struct virtio_driver virtio_net_driver
= {
2761 .feature_table
= features
,
2762 .feature_table_size
= ARRAY_SIZE(features
),
2763 .feature_table_legacy
= features_legacy
,
2764 .feature_table_size_legacy
= ARRAY_SIZE(features_legacy
),
2765 .driver
.name
= KBUILD_MODNAME
,
2766 .driver
.owner
= THIS_MODULE
,
2767 .id_table
= id_table
,
2768 .validate
= virtnet_validate
,
2769 .probe
= virtnet_probe
,
2770 .remove
= virtnet_remove
,
2771 .config_changed
= virtnet_config_changed
,
2772 #ifdef CONFIG_PM_SLEEP
2773 .freeze
= virtnet_freeze
,
2774 .restore
= virtnet_restore
,
2778 static __init
int virtio_net_driver_init(void)
2782 ret
= cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN
, "virtio/net:online",
2784 virtnet_cpu_down_prep
);
2787 virtionet_online
= ret
;
2788 ret
= cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD
, "virtio/net:dead",
2789 NULL
, virtnet_cpu_dead
);
2793 ret
= register_virtio_driver(&virtio_net_driver
);
2798 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD
);
2800 cpuhp_remove_multi_state(virtionet_online
);
2804 module_init(virtio_net_driver_init
);
2806 static __exit
void virtio_net_driver_exit(void)
2808 unregister_virtio_driver(&virtio_net_driver
);
2809 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD
);
2810 cpuhp_remove_multi_state(virtionet_online
);
2812 module_exit(virtio_net_driver_exit
);
2814 MODULE_DEVICE_TABLE(virtio
, id_table
);
2815 MODULE_DESCRIPTION("Virtio network driver");
2816 MODULE_LICENSE("GPL");