1 /* A network driver using virtio.
3 * Copyright 2007 Rusty Russell <rusty@rustcorp.com.au> IBM Corporation
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation; either version 2 of the License, or
8 * (at your option) any later version.
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, see <http://www.gnu.org/licenses/>.
19 #include <linux/netdevice.h>
20 #include <linux/etherdevice.h>
21 #include <linux/ethtool.h>
22 #include <linux/module.h>
23 #include <linux/virtio.h>
24 #include <linux/virtio_net.h>
25 #include <linux/bpf.h>
26 #include <linux/bpf_trace.h>
27 #include <linux/scatterlist.h>
28 #include <linux/if_vlan.h>
29 #include <linux/slab.h>
30 #include <linux/cpu.h>
31 #include <linux/average.h>
32 #include <net/route.h>
34 static int napi_weight
= NAPI_POLL_WEIGHT
;
35 module_param(napi_weight
, int, 0444);
37 static bool csum
= true, gso
= true, napi_tx
;
38 module_param(csum
, bool, 0444);
39 module_param(gso
, bool, 0444);
40 module_param(napi_tx
, bool, 0644);
42 /* FIXME: MTU in config. */
43 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
44 #define GOOD_COPY_LEN 128
46 #define VIRTNET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
48 /* Amount of XDP headroom to prepend to packets for use by xdp_adjust_head */
49 #define VIRTIO_XDP_HEADROOM 256
51 /* RX packet size EWMA. The average packet size is used to determine the packet
52 * buffer size when refilling RX rings. As the entire RX ring may be refilled
53 * at once, the weight is chosen so that the EWMA will be insensitive to short-
54 * term, transient changes in packet size.
56 DECLARE_EWMA(pkt_len
, 0, 64)
58 #define VIRTNET_DRIVER_VERSION "1.0.0"
60 struct virtnet_stats
{
61 struct u64_stats_sync tx_syncp
;
62 struct u64_stats_sync rx_syncp
;
70 /* Internal representation of a send virtqueue */
72 /* Virtqueue associated with this send _queue */
75 /* TX: fragments + linear part + virtio header */
76 struct scatterlist sg
[MAX_SKB_FRAGS
+ 2];
78 /* Name of the send queue: output.$index */
81 struct napi_struct napi
;
84 /* Internal representation of a receive virtqueue */
85 struct receive_queue
{
86 /* Virtqueue associated with this receive_queue */
89 struct napi_struct napi
;
91 struct bpf_prog __rcu
*xdp_prog
;
93 /* Chain pages by the private ptr. */
96 /* Average packet length for mergeable receive buffers. */
97 struct ewma_pkt_len mrg_avg_pkt_len
;
99 /* Page frag for packet buffer allocation. */
100 struct page_frag alloc_frag
;
102 /* RX: fragments + linear part + virtio header */
103 struct scatterlist sg
[MAX_SKB_FRAGS
+ 2];
105 /* Min single buffer size for mergeable buffers case. */
106 unsigned int min_buf_len
;
108 /* Name of this receive queue: input.$index */
112 struct virtnet_info
{
113 struct virtio_device
*vdev
;
114 struct virtqueue
*cvq
;
115 struct net_device
*dev
;
116 struct send_queue
*sq
;
117 struct receive_queue
*rq
;
120 /* Max # of queue pairs supported by the device */
123 /* # of queue pairs currently used by the driver */
124 u16 curr_queue_pairs
;
126 /* # of XDP queue pairs currently used by the driver */
129 /* I like... big packets and I cannot lie! */
132 /* Host will merge rx buffers for big packets (shake it! shake it!) */
133 bool mergeable_rx_bufs
;
135 /* Has control virtqueue */
138 /* Host can handle any s/g split between our header and packet data */
141 /* Packet virtio header size */
144 /* Active statistics */
145 struct virtnet_stats __percpu
*stats
;
147 /* Work struct for refilling if we run low on memory. */
148 struct delayed_work refill
;
150 /* Work struct for config space updates */
151 struct work_struct config_work
;
153 /* Does the affinity hint is set for virtqueues? */
154 bool affinity_hint_set
;
156 /* CPU hotplug instances for online & dead */
157 struct hlist_node node
;
158 struct hlist_node node_dead
;
160 /* Control VQ buffers: protected by the rtnl lock */
161 struct virtio_net_ctrl_hdr ctrl_hdr
;
162 virtio_net_ctrl_ack ctrl_status
;
163 struct virtio_net_ctrl_mq ctrl_mq
;
168 /* Ethtool settings */
173 struct padded_vnet_hdr
{
174 struct virtio_net_hdr_mrg_rxbuf hdr
;
176 * hdr is in a separate sg buffer, and data sg buffer shares same page
177 * with this header sg. This padding makes next sg 16 byte aligned
183 /* Converting between virtqueue no. and kernel tx/rx queue no.
184 * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
186 static int vq2txq(struct virtqueue
*vq
)
188 return (vq
->index
- 1) / 2;
191 static int txq2vq(int txq
)
196 static int vq2rxq(struct virtqueue
*vq
)
198 return vq
->index
/ 2;
201 static int rxq2vq(int rxq
)
206 static inline struct virtio_net_hdr_mrg_rxbuf
*skb_vnet_hdr(struct sk_buff
*skb
)
208 return (struct virtio_net_hdr_mrg_rxbuf
*)skb
->cb
;
212 * private is used to chain pages for big packets, put the whole
213 * most recent used list in the beginning for reuse
215 static void give_pages(struct receive_queue
*rq
, struct page
*page
)
219 /* Find end of list, sew whole thing into vi->rq.pages. */
220 for (end
= page
; end
->private; end
= (struct page
*)end
->private);
221 end
->private = (unsigned long)rq
->pages
;
225 static struct page
*get_a_page(struct receive_queue
*rq
, gfp_t gfp_mask
)
227 struct page
*p
= rq
->pages
;
230 rq
->pages
= (struct page
*)p
->private;
231 /* clear private here, it is used to chain pages */
234 p
= alloc_page(gfp_mask
);
238 static void virtqueue_napi_schedule(struct napi_struct
*napi
,
239 struct virtqueue
*vq
)
241 if (napi_schedule_prep(napi
)) {
242 virtqueue_disable_cb(vq
);
243 __napi_schedule(napi
);
247 static void virtqueue_napi_complete(struct napi_struct
*napi
,
248 struct virtqueue
*vq
, int processed
)
252 opaque
= virtqueue_enable_cb_prepare(vq
);
253 if (napi_complete_done(napi
, processed
) &&
254 unlikely(virtqueue_poll(vq
, opaque
)))
255 virtqueue_napi_schedule(napi
, vq
);
258 static void skb_xmit_done(struct virtqueue
*vq
)
260 struct virtnet_info
*vi
= vq
->vdev
->priv
;
261 struct napi_struct
*napi
= &vi
->sq
[vq2txq(vq
)].napi
;
263 /* Suppress further interrupts. */
264 virtqueue_disable_cb(vq
);
267 virtqueue_napi_schedule(napi
, vq
);
269 /* We were probably waiting for more output buffers. */
270 netif_wake_subqueue(vi
->dev
, vq2txq(vq
));
273 #define MRG_CTX_HEADER_SHIFT 22
274 static void *mergeable_len_to_ctx(unsigned int truesize
,
275 unsigned int headroom
)
277 return (void *)(unsigned long)((headroom
<< MRG_CTX_HEADER_SHIFT
) | truesize
);
280 static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx
)
282 return (unsigned long)mrg_ctx
>> MRG_CTX_HEADER_SHIFT
;
285 static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx
)
287 return (unsigned long)mrg_ctx
& ((1 << MRG_CTX_HEADER_SHIFT
) - 1);
290 /* Called from bottom half context */
291 static struct sk_buff
*page_to_skb(struct virtnet_info
*vi
,
292 struct receive_queue
*rq
,
293 struct page
*page
, unsigned int offset
,
294 unsigned int len
, unsigned int truesize
)
297 struct virtio_net_hdr_mrg_rxbuf
*hdr
;
298 unsigned int copy
, hdr_len
, hdr_padded_len
;
301 p
= page_address(page
) + offset
;
303 /* copy small packet so we can reuse these pages for small data */
304 skb
= napi_alloc_skb(&rq
->napi
, GOOD_COPY_LEN
);
308 hdr
= skb_vnet_hdr(skb
);
310 hdr_len
= vi
->hdr_len
;
311 if (vi
->mergeable_rx_bufs
)
312 hdr_padded_len
= sizeof *hdr
;
314 hdr_padded_len
= sizeof(struct padded_vnet_hdr
);
316 memcpy(hdr
, p
, hdr_len
);
319 offset
+= hdr_padded_len
;
323 if (copy
> skb_tailroom(skb
))
324 copy
= skb_tailroom(skb
);
325 skb_put_data(skb
, p
, copy
);
330 if (vi
->mergeable_rx_bufs
) {
332 skb_add_rx_frag(skb
, 0, page
, offset
, len
, truesize
);
339 * Verify that we can indeed put this data into a skb.
340 * This is here to handle cases when the device erroneously
341 * tries to receive more than is possible. This is usually
342 * the case of a broken device.
344 if (unlikely(len
> MAX_SKB_FRAGS
* PAGE_SIZE
)) {
345 net_dbg_ratelimited("%s: too much data\n", skb
->dev
->name
);
349 BUG_ON(offset
>= PAGE_SIZE
);
351 unsigned int frag_size
= min((unsigned)PAGE_SIZE
- offset
, len
);
352 skb_add_rx_frag(skb
, skb_shinfo(skb
)->nr_frags
, page
, offset
,
353 frag_size
, truesize
);
355 page
= (struct page
*)page
->private;
360 give_pages(rq
, page
);
365 static bool virtnet_xdp_xmit(struct virtnet_info
*vi
,
366 struct receive_queue
*rq
,
367 struct xdp_buff
*xdp
)
369 struct virtio_net_hdr_mrg_rxbuf
*hdr
;
371 struct send_queue
*sq
;
376 qp
= vi
->curr_queue_pairs
- vi
->xdp_queue_pairs
+ smp_processor_id();
379 /* Free up any pending old buffers before queueing new ones. */
380 while ((xdp_sent
= virtqueue_get_buf(sq
->vq
, &len
)) != NULL
) {
381 struct page
*sent_page
= virt_to_head_page(xdp_sent
);
386 xdp
->data
-= vi
->hdr_len
;
387 /* Zero header and leave csum up to XDP layers */
389 memset(hdr
, 0, vi
->hdr_len
);
391 sg_init_one(sq
->sg
, xdp
->data
, xdp
->data_end
- xdp
->data
);
393 err
= virtqueue_add_outbuf(sq
->vq
, sq
->sg
, 1, xdp
->data
, GFP_ATOMIC
);
395 struct page
*page
= virt_to_head_page(xdp
->data
);
401 virtqueue_kick(sq
->vq
);
405 static unsigned int virtnet_get_headroom(struct virtnet_info
*vi
)
407 return vi
->xdp_queue_pairs
? VIRTIO_XDP_HEADROOM
: 0;
410 static struct sk_buff
*receive_small(struct net_device
*dev
,
411 struct virtnet_info
*vi
,
412 struct receive_queue
*rq
,
413 void *buf
, unsigned int len
)
416 struct bpf_prog
*xdp_prog
;
417 unsigned int xdp_headroom
= virtnet_get_headroom(vi
);
418 unsigned int header_offset
= VIRTNET_RX_PAD
+ xdp_headroom
;
419 unsigned int headroom
= vi
->hdr_len
+ header_offset
;
420 unsigned int buflen
= SKB_DATA_ALIGN(GOOD_PACKET_LEN
+ headroom
) +
421 SKB_DATA_ALIGN(sizeof(struct skb_shared_info
));
422 unsigned int delta
= 0;
426 xdp_prog
= rcu_dereference(rq
->xdp_prog
);
428 struct virtio_net_hdr_mrg_rxbuf
*hdr
= buf
+ header_offset
;
433 if (unlikely(hdr
->hdr
.gso_type
|| hdr
->hdr
.flags
))
436 xdp
.data_hard_start
= buf
+ VIRTNET_RX_PAD
+ vi
->hdr_len
;
437 xdp
.data
= xdp
.data_hard_start
+ xdp_headroom
;
438 xdp
.data_end
= xdp
.data
+ len
;
439 orig_data
= xdp
.data
;
440 act
= bpf_prog_run_xdp(xdp_prog
, &xdp
);
444 /* Recalculate length in case bpf program changed it */
445 delta
= orig_data
- xdp
.data
;
448 if (unlikely(!virtnet_xdp_xmit(vi
, rq
, &xdp
)))
449 trace_xdp_exception(vi
->dev
, xdp_prog
, act
);
453 bpf_warn_invalid_xdp_action(act
);
455 trace_xdp_exception(vi
->dev
, xdp_prog
, act
);
462 skb
= build_skb(buf
, buflen
);
464 put_page(virt_to_head_page(buf
));
467 skb_reserve(skb
, headroom
- delta
);
468 skb_put(skb
, len
+ delta
);
470 buf
+= header_offset
;
471 memcpy(skb_vnet_hdr(skb
), buf
, vi
->hdr_len
);
472 } /* keep zeroed vnet hdr since packet was changed by bpf */
479 dev
->stats
.rx_dropped
++;
480 put_page(virt_to_head_page(buf
));
485 static struct sk_buff
*receive_big(struct net_device
*dev
,
486 struct virtnet_info
*vi
,
487 struct receive_queue
*rq
,
491 struct page
*page
= buf
;
492 struct sk_buff
*skb
= page_to_skb(vi
, rq
, page
, 0, len
, PAGE_SIZE
);
500 dev
->stats
.rx_dropped
++;
501 give_pages(rq
, page
);
505 /* The conditions to enable XDP should preclude the underlying device from
506 * sending packets across multiple buffers (num_buf > 1). However per spec
507 * it does not appear to be illegal to do so but rather just against convention.
508 * So in order to avoid making a system unresponsive the packets are pushed
509 * into a page and the XDP program is run. This will be extremely slow and we
510 * push a warning to the user to fix this as soon as possible. Fixing this may
511 * require resolving the underlying hardware to determine why multiple buffers
512 * are being received or simply loading the XDP program in the ingress stack
513 * after the skb is built because there is no advantage to running it here
516 static struct page
*xdp_linearize_page(struct receive_queue
*rq
,
522 struct page
*page
= alloc_page(GFP_ATOMIC
);
523 unsigned int page_off
= VIRTIO_XDP_HEADROOM
;
528 memcpy(page_address(page
) + page_off
, page_address(p
) + offset
, *len
);
536 buf
= virtqueue_get_buf(rq
->vq
, &buflen
);
540 p
= virt_to_head_page(buf
);
541 off
= buf
- page_address(p
);
543 /* guard against a misconfigured or uncooperative backend that
544 * is sending packet larger than the MTU.
546 if ((page_off
+ buflen
) > PAGE_SIZE
) {
551 memcpy(page_address(page
) + page_off
,
552 page_address(p
) + off
, buflen
);
557 /* Headroom does not contribute to packet length */
558 *len
= page_off
- VIRTIO_XDP_HEADROOM
;
561 __free_pages(page
, 0);
565 static struct sk_buff
*receive_mergeable(struct net_device
*dev
,
566 struct virtnet_info
*vi
,
567 struct receive_queue
*rq
,
572 struct virtio_net_hdr_mrg_rxbuf
*hdr
= buf
;
573 u16 num_buf
= virtio16_to_cpu(vi
->vdev
, hdr
->num_buffers
);
574 struct page
*page
= virt_to_head_page(buf
);
575 int offset
= buf
- page_address(page
);
576 struct sk_buff
*head_skb
, *curr_skb
;
577 struct bpf_prog
*xdp_prog
;
578 unsigned int truesize
;
583 xdp_prog
= rcu_dereference(rq
->xdp_prog
);
585 struct page
*xdp_page
;
590 /* This happens when rx buffer size is underestimated */
591 if (unlikely(num_buf
> 1)) {
592 /* linearize data for XDP */
593 xdp_page
= xdp_linearize_page(rq
, &num_buf
,
597 offset
= VIRTIO_XDP_HEADROOM
;
602 /* Transient failure which in theory could occur if
603 * in-flight packets from before XDP was enabled reach
604 * the receive path after XDP is loaded. In practice I
605 * was not able to create this condition.
607 if (unlikely(hdr
->hdr
.gso_type
))
610 /* Allow consuming headroom but reserve enough space to push
611 * the descriptor on if we get an XDP_TX return code.
613 data
= page_address(xdp_page
) + offset
;
614 xdp
.data_hard_start
= data
- VIRTIO_XDP_HEADROOM
+ vi
->hdr_len
;
615 xdp
.data
= data
+ vi
->hdr_len
;
616 xdp
.data_end
= xdp
.data
+ (len
- vi
->hdr_len
);
617 act
= bpf_prog_run_xdp(xdp_prog
, &xdp
);
621 /* recalculate offset to account for any header
622 * adjustments. Note other cases do not build an
623 * skb and avoid using offset
626 page_address(xdp_page
) - vi
->hdr_len
;
628 /* We can only create skb based on xdp_page. */
629 if (unlikely(xdp_page
!= page
)) {
632 head_skb
= page_to_skb(vi
, rq
, xdp_page
,
633 offset
, len
, PAGE_SIZE
);
634 ewma_pkt_len_add(&rq
->mrg_avg_pkt_len
, len
);
639 if (unlikely(!virtnet_xdp_xmit(vi
, rq
, &xdp
)))
640 trace_xdp_exception(vi
->dev
, xdp_prog
, act
);
641 ewma_pkt_len_add(&rq
->mrg_avg_pkt_len
, len
);
642 if (unlikely(xdp_page
!= page
))
647 bpf_warn_invalid_xdp_action(act
);
649 trace_xdp_exception(vi
->dev
, xdp_prog
, act
);
651 if (unlikely(xdp_page
!= page
))
652 __free_pages(xdp_page
, 0);
653 ewma_pkt_len_add(&rq
->mrg_avg_pkt_len
, len
);
659 truesize
= mergeable_ctx_to_truesize(ctx
);
660 if (unlikely(len
> truesize
)) {
661 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
662 dev
->name
, len
, (unsigned long)ctx
);
663 dev
->stats
.rx_length_errors
++;
667 head_skb
= page_to_skb(vi
, rq
, page
, offset
, len
, truesize
);
670 if (unlikely(!curr_skb
))
675 buf
= virtqueue_get_buf_ctx(rq
->vq
, &len
, &ctx
);
676 if (unlikely(!ctx
)) {
677 pr_debug("%s: rx error: %d buffers out of %d missing\n",
679 virtio16_to_cpu(vi
->vdev
,
681 dev
->stats
.rx_length_errors
++;
685 page
= virt_to_head_page(buf
);
687 truesize
= mergeable_ctx_to_truesize(ctx
);
688 if (unlikely(len
> truesize
)) {
689 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
690 dev
->name
, len
, (unsigned long)ctx
);
691 dev
->stats
.rx_length_errors
++;
695 num_skb_frags
= skb_shinfo(curr_skb
)->nr_frags
;
696 if (unlikely(num_skb_frags
== MAX_SKB_FRAGS
)) {
697 struct sk_buff
*nskb
= alloc_skb(0, GFP_ATOMIC
);
701 if (curr_skb
== head_skb
)
702 skb_shinfo(curr_skb
)->frag_list
= nskb
;
704 curr_skb
->next
= nskb
;
706 head_skb
->truesize
+= nskb
->truesize
;
709 if (curr_skb
!= head_skb
) {
710 head_skb
->data_len
+= len
;
711 head_skb
->len
+= len
;
712 head_skb
->truesize
+= truesize
;
714 offset
= buf
- page_address(page
);
715 if (skb_can_coalesce(curr_skb
, num_skb_frags
, page
, offset
)) {
717 skb_coalesce_rx_frag(curr_skb
, num_skb_frags
- 1,
720 skb_add_rx_frag(curr_skb
, num_skb_frags
, page
,
721 offset
, len
, truesize
);
725 ewma_pkt_len_add(&rq
->mrg_avg_pkt_len
, head_skb
->len
);
733 buf
= virtqueue_get_buf(rq
->vq
, &len
);
734 if (unlikely(!buf
)) {
735 pr_debug("%s: rx error: %d buffers missing\n",
737 dev
->stats
.rx_length_errors
++;
740 page
= virt_to_head_page(buf
);
744 dev
->stats
.rx_dropped
++;
745 dev_kfree_skb(head_skb
);
750 static int receive_buf(struct virtnet_info
*vi
, struct receive_queue
*rq
,
751 void *buf
, unsigned int len
, void **ctx
)
753 struct net_device
*dev
= vi
->dev
;
755 struct virtio_net_hdr_mrg_rxbuf
*hdr
;
758 if (unlikely(len
< vi
->hdr_len
+ ETH_HLEN
)) {
759 pr_debug("%s: short packet %i\n", dev
->name
, len
);
760 dev
->stats
.rx_length_errors
++;
761 if (vi
->mergeable_rx_bufs
) {
762 put_page(virt_to_head_page(buf
));
763 } else if (vi
->big_packets
) {
766 put_page(virt_to_head_page(buf
));
771 if (vi
->mergeable_rx_bufs
)
772 skb
= receive_mergeable(dev
, vi
, rq
, buf
, ctx
, len
);
773 else if (vi
->big_packets
)
774 skb
= receive_big(dev
, vi
, rq
, buf
, len
);
776 skb
= receive_small(dev
, vi
, rq
, buf
, len
);
781 hdr
= skb_vnet_hdr(skb
);
785 if (hdr
->hdr
.flags
& VIRTIO_NET_HDR_F_DATA_VALID
)
786 skb
->ip_summed
= CHECKSUM_UNNECESSARY
;
788 if (virtio_net_hdr_to_skb(skb
, &hdr
->hdr
,
789 virtio_is_little_endian(vi
->vdev
))) {
790 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
791 dev
->name
, hdr
->hdr
.gso_type
,
796 skb
->protocol
= eth_type_trans(skb
, dev
);
797 pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
798 ntohs(skb
->protocol
), skb
->len
, skb
->pkt_type
);
800 napi_gro_receive(&rq
->napi
, skb
);
804 dev
->stats
.rx_frame_errors
++;
809 static int add_recvbuf_small(struct virtnet_info
*vi
, struct receive_queue
*rq
,
812 struct page_frag
*alloc_frag
= &rq
->alloc_frag
;
814 unsigned int xdp_headroom
= virtnet_get_headroom(vi
);
815 int len
= vi
->hdr_len
+ VIRTNET_RX_PAD
+ GOOD_PACKET_LEN
+ xdp_headroom
;
818 len
= SKB_DATA_ALIGN(len
) +
819 SKB_DATA_ALIGN(sizeof(struct skb_shared_info
));
820 if (unlikely(!skb_page_frag_refill(len
, alloc_frag
, gfp
)))
823 buf
= (char *)page_address(alloc_frag
->page
) + alloc_frag
->offset
;
824 get_page(alloc_frag
->page
);
825 alloc_frag
->offset
+= len
;
826 sg_init_one(rq
->sg
, buf
+ VIRTNET_RX_PAD
+ xdp_headroom
,
827 vi
->hdr_len
+ GOOD_PACKET_LEN
);
828 err
= virtqueue_add_inbuf(rq
->vq
, rq
->sg
, 1, buf
, gfp
);
830 put_page(virt_to_head_page(buf
));
835 static int add_recvbuf_big(struct virtnet_info
*vi
, struct receive_queue
*rq
,
838 struct page
*first
, *list
= NULL
;
842 sg_init_table(rq
->sg
, MAX_SKB_FRAGS
+ 2);
844 /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
845 for (i
= MAX_SKB_FRAGS
+ 1; i
> 1; --i
) {
846 first
= get_a_page(rq
, gfp
);
849 give_pages(rq
, list
);
852 sg_set_buf(&rq
->sg
[i
], page_address(first
), PAGE_SIZE
);
854 /* chain new page in list head to match sg */
855 first
->private = (unsigned long)list
;
859 first
= get_a_page(rq
, gfp
);
861 give_pages(rq
, list
);
864 p
= page_address(first
);
866 /* rq->sg[0], rq->sg[1] share the same page */
867 /* a separated rq->sg[0] for header - required in case !any_header_sg */
868 sg_set_buf(&rq
->sg
[0], p
, vi
->hdr_len
);
870 /* rq->sg[1] for data packet, from offset */
871 offset
= sizeof(struct padded_vnet_hdr
);
872 sg_set_buf(&rq
->sg
[1], p
+ offset
, PAGE_SIZE
- offset
);
874 /* chain first in list head */
875 first
->private = (unsigned long)list
;
876 err
= virtqueue_add_inbuf(rq
->vq
, rq
->sg
, MAX_SKB_FRAGS
+ 2,
879 give_pages(rq
, first
);
884 static unsigned int get_mergeable_buf_len(struct receive_queue
*rq
,
885 struct ewma_pkt_len
*avg_pkt_len
)
887 const size_t hdr_len
= sizeof(struct virtio_net_hdr_mrg_rxbuf
);
890 len
= hdr_len
+ clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len
),
891 rq
->min_buf_len
, PAGE_SIZE
- hdr_len
);
892 return ALIGN(len
, L1_CACHE_BYTES
);
895 static int add_recvbuf_mergeable(struct virtnet_info
*vi
,
896 struct receive_queue
*rq
, gfp_t gfp
)
898 struct page_frag
*alloc_frag
= &rq
->alloc_frag
;
899 unsigned int headroom
= virtnet_get_headroom(vi
);
903 unsigned int len
, hole
;
905 len
= get_mergeable_buf_len(rq
, &rq
->mrg_avg_pkt_len
);
906 if (unlikely(!skb_page_frag_refill(len
+ headroom
, alloc_frag
, gfp
)))
909 buf
= (char *)page_address(alloc_frag
->page
) + alloc_frag
->offset
;
910 buf
+= headroom
; /* advance address leaving hole at front of pkt */
911 ctx
= mergeable_len_to_ctx(len
, headroom
);
912 get_page(alloc_frag
->page
);
913 alloc_frag
->offset
+= len
+ headroom
;
914 hole
= alloc_frag
->size
- alloc_frag
->offset
;
915 if (hole
< len
+ headroom
) {
916 /* To avoid internal fragmentation, if there is very likely not
917 * enough space for another buffer, add the remaining space to
918 * the current buffer. This extra space is not included in
919 * the truesize stored in ctx.
922 alloc_frag
->offset
+= hole
;
925 sg_init_one(rq
->sg
, buf
, len
);
926 err
= virtqueue_add_inbuf_ctx(rq
->vq
, rq
->sg
, 1, buf
, ctx
, gfp
);
928 put_page(virt_to_head_page(buf
));
934 * Returns false if we couldn't fill entirely (OOM).
936 * Normally run in the receive path, but can also be run from ndo_open
937 * before we're receiving packets, or from refill_work which is
938 * careful to disable receiving (using napi_disable).
940 static bool try_fill_recv(struct virtnet_info
*vi
, struct receive_queue
*rq
,
948 if (vi
->mergeable_rx_bufs
)
949 err
= add_recvbuf_mergeable(vi
, rq
, gfp
);
950 else if (vi
->big_packets
)
951 err
= add_recvbuf_big(vi
, rq
, gfp
);
953 err
= add_recvbuf_small(vi
, rq
, gfp
);
955 oom
= err
== -ENOMEM
;
958 } while (rq
->vq
->num_free
);
959 virtqueue_kick(rq
->vq
);
963 static void skb_recv_done(struct virtqueue
*rvq
)
965 struct virtnet_info
*vi
= rvq
->vdev
->priv
;
966 struct receive_queue
*rq
= &vi
->rq
[vq2rxq(rvq
)];
968 virtqueue_napi_schedule(&rq
->napi
, rvq
);
971 static void virtnet_napi_enable(struct virtqueue
*vq
, struct napi_struct
*napi
)
975 /* If all buffers were filled by other side before we napi_enabled, we
976 * won't get another interrupt, so process any outstanding packets now.
977 * Call local_bh_enable after to trigger softIRQ processing.
980 virtqueue_napi_schedule(napi
, vq
);
984 static void virtnet_napi_tx_enable(struct virtnet_info
*vi
,
985 struct virtqueue
*vq
,
986 struct napi_struct
*napi
)
991 /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
992 * enable the feature if this is likely affine with the transmit path.
994 if (!vi
->affinity_hint_set
) {
999 return virtnet_napi_enable(vq
, napi
);
1002 static void virtnet_napi_tx_disable(struct napi_struct
*napi
)
1008 static void refill_work(struct work_struct
*work
)
1010 struct virtnet_info
*vi
=
1011 container_of(work
, struct virtnet_info
, refill
.work
);
1015 for (i
= 0; i
< vi
->curr_queue_pairs
; i
++) {
1016 struct receive_queue
*rq
= &vi
->rq
[i
];
1018 napi_disable(&rq
->napi
);
1019 still_empty
= !try_fill_recv(vi
, rq
, GFP_KERNEL
);
1020 virtnet_napi_enable(rq
->vq
, &rq
->napi
);
1022 /* In theory, this can happen: if we don't get any buffers in
1023 * we will *never* try to fill again.
1026 schedule_delayed_work(&vi
->refill
, HZ
/2);
1030 static int virtnet_receive(struct receive_queue
*rq
, int budget
)
1032 struct virtnet_info
*vi
= rq
->vq
->vdev
->priv
;
1033 unsigned int len
, received
= 0, bytes
= 0;
1035 struct virtnet_stats
*stats
= this_cpu_ptr(vi
->stats
);
1037 if (vi
->mergeable_rx_bufs
) {
1040 while (received
< budget
&&
1041 (buf
= virtqueue_get_buf_ctx(rq
->vq
, &len
, &ctx
))) {
1042 bytes
+= receive_buf(vi
, rq
, buf
, len
, ctx
);
1046 while (received
< budget
&&
1047 (buf
= virtqueue_get_buf(rq
->vq
, &len
)) != NULL
) {
1048 bytes
+= receive_buf(vi
, rq
, buf
, len
, NULL
);
1053 if (rq
->vq
->num_free
> virtqueue_get_vring_size(rq
->vq
) / 2) {
1054 if (!try_fill_recv(vi
, rq
, GFP_ATOMIC
))
1055 schedule_delayed_work(&vi
->refill
, 0);
1058 u64_stats_update_begin(&stats
->rx_syncp
);
1059 stats
->rx_bytes
+= bytes
;
1060 stats
->rx_packets
+= received
;
1061 u64_stats_update_end(&stats
->rx_syncp
);
1066 static void free_old_xmit_skbs(struct send_queue
*sq
)
1068 struct sk_buff
*skb
;
1070 struct virtnet_info
*vi
= sq
->vq
->vdev
->priv
;
1071 struct virtnet_stats
*stats
= this_cpu_ptr(vi
->stats
);
1072 unsigned int packets
= 0;
1073 unsigned int bytes
= 0;
1075 while ((skb
= virtqueue_get_buf(sq
->vq
, &len
)) != NULL
) {
1076 pr_debug("Sent skb %p\n", skb
);
1081 dev_kfree_skb_any(skb
);
1084 /* Avoid overhead when no packets have been processed
1085 * happens when called speculatively from start_xmit.
1090 u64_stats_update_begin(&stats
->tx_syncp
);
1091 stats
->tx_bytes
+= bytes
;
1092 stats
->tx_packets
+= packets
;
1093 u64_stats_update_end(&stats
->tx_syncp
);
1096 static void virtnet_poll_cleantx(struct receive_queue
*rq
)
1098 struct virtnet_info
*vi
= rq
->vq
->vdev
->priv
;
1099 unsigned int index
= vq2rxq(rq
->vq
);
1100 struct send_queue
*sq
= &vi
->sq
[index
];
1101 struct netdev_queue
*txq
= netdev_get_tx_queue(vi
->dev
, index
);
1103 if (!sq
->napi
.weight
)
1106 if (__netif_tx_trylock(txq
)) {
1107 free_old_xmit_skbs(sq
);
1108 __netif_tx_unlock(txq
);
1111 if (sq
->vq
->num_free
>= 2 + MAX_SKB_FRAGS
)
1112 netif_tx_wake_queue(txq
);
1115 static int virtnet_poll(struct napi_struct
*napi
, int budget
)
1117 struct receive_queue
*rq
=
1118 container_of(napi
, struct receive_queue
, napi
);
1119 unsigned int received
;
1121 virtnet_poll_cleantx(rq
);
1123 received
= virtnet_receive(rq
, budget
);
1125 /* Out of packets? */
1126 if (received
< budget
)
1127 virtqueue_napi_complete(napi
, rq
->vq
, received
);
1132 static int virtnet_open(struct net_device
*dev
)
1134 struct virtnet_info
*vi
= netdev_priv(dev
);
1137 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1138 if (i
< vi
->curr_queue_pairs
)
1139 /* Make sure we have some buffers: if oom use wq. */
1140 if (!try_fill_recv(vi
, &vi
->rq
[i
], GFP_KERNEL
))
1141 schedule_delayed_work(&vi
->refill
, 0);
1142 virtnet_napi_enable(vi
->rq
[i
].vq
, &vi
->rq
[i
].napi
);
1143 virtnet_napi_tx_enable(vi
, vi
->sq
[i
].vq
, &vi
->sq
[i
].napi
);
1149 static int virtnet_poll_tx(struct napi_struct
*napi
, int budget
)
1151 struct send_queue
*sq
= container_of(napi
, struct send_queue
, napi
);
1152 struct virtnet_info
*vi
= sq
->vq
->vdev
->priv
;
1153 struct netdev_queue
*txq
= netdev_get_tx_queue(vi
->dev
, vq2txq(sq
->vq
));
1155 __netif_tx_lock(txq
, raw_smp_processor_id());
1156 free_old_xmit_skbs(sq
);
1157 __netif_tx_unlock(txq
);
1159 virtqueue_napi_complete(napi
, sq
->vq
, 0);
1161 if (sq
->vq
->num_free
>= 2 + MAX_SKB_FRAGS
)
1162 netif_tx_wake_queue(txq
);
1167 static int xmit_skb(struct send_queue
*sq
, struct sk_buff
*skb
)
1169 struct virtio_net_hdr_mrg_rxbuf
*hdr
;
1170 const unsigned char *dest
= ((struct ethhdr
*)skb
->data
)->h_dest
;
1171 struct virtnet_info
*vi
= sq
->vq
->vdev
->priv
;
1173 unsigned hdr_len
= vi
->hdr_len
;
1176 pr_debug("%s: xmit %p %pM\n", vi
->dev
->name
, skb
, dest
);
1178 can_push
= vi
->any_header_sg
&&
1179 !((unsigned long)skb
->data
& (__alignof__(*hdr
) - 1)) &&
1180 !skb_header_cloned(skb
) && skb_headroom(skb
) >= hdr_len
;
1181 /* Even if we can, don't push here yet as this would skew
1182 * csum_start offset below. */
1184 hdr
= (struct virtio_net_hdr_mrg_rxbuf
*)(skb
->data
- hdr_len
);
1186 hdr
= skb_vnet_hdr(skb
);
1188 if (virtio_net_hdr_from_skb(skb
, &hdr
->hdr
,
1189 virtio_is_little_endian(vi
->vdev
), false))
1192 if (vi
->mergeable_rx_bufs
)
1193 hdr
->num_buffers
= 0;
1195 sg_init_table(sq
->sg
, skb_shinfo(skb
)->nr_frags
+ (can_push
? 1 : 2));
1197 __skb_push(skb
, hdr_len
);
1198 num_sg
= skb_to_sgvec(skb
, sq
->sg
, 0, skb
->len
);
1199 if (unlikely(num_sg
< 0))
1201 /* Pull header back to avoid skew in tx bytes calculations. */
1202 __skb_pull(skb
, hdr_len
);
1204 sg_set_buf(sq
->sg
, hdr
, hdr_len
);
1205 num_sg
= skb_to_sgvec(skb
, sq
->sg
+ 1, 0, skb
->len
);
1206 if (unlikely(num_sg
< 0))
1210 return virtqueue_add_outbuf(sq
->vq
, sq
->sg
, num_sg
, skb
, GFP_ATOMIC
);
1213 static netdev_tx_t
start_xmit(struct sk_buff
*skb
, struct net_device
*dev
)
1215 struct virtnet_info
*vi
= netdev_priv(dev
);
1216 int qnum
= skb_get_queue_mapping(skb
);
1217 struct send_queue
*sq
= &vi
->sq
[qnum
];
1219 struct netdev_queue
*txq
= netdev_get_tx_queue(dev
, qnum
);
1220 bool kick
= !skb
->xmit_more
;
1221 bool use_napi
= sq
->napi
.weight
;
1223 /* Free up any pending old buffers before queueing new ones. */
1224 free_old_xmit_skbs(sq
);
1226 if (use_napi
&& kick
)
1227 virtqueue_enable_cb_delayed(sq
->vq
);
1229 /* timestamp packet in software */
1230 skb_tx_timestamp(skb
);
1232 /* Try to transmit */
1233 err
= xmit_skb(sq
, skb
);
1235 /* This should not happen! */
1236 if (unlikely(err
)) {
1237 dev
->stats
.tx_fifo_errors
++;
1238 if (net_ratelimit())
1240 "Unexpected TXQ (%d) queue failure: %d\n", qnum
, err
);
1241 dev
->stats
.tx_dropped
++;
1242 dev_kfree_skb_any(skb
);
1243 return NETDEV_TX_OK
;
1246 /* Don't wait up for transmitted skbs to be freed. */
1252 /* If running out of space, stop queue to avoid getting packets that we
1253 * are then unable to transmit.
1254 * An alternative would be to force queuing layer to requeue the skb by
1255 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1256 * returned in a normal path of operation: it means that driver is not
1257 * maintaining the TX queue stop/start state properly, and causes
1258 * the stack to do a non-trivial amount of useless work.
1259 * Since most packets only take 1 or 2 ring slots, stopping the queue
1260 * early means 16 slots are typically wasted.
1262 if (sq
->vq
->num_free
< 2+MAX_SKB_FRAGS
) {
1263 netif_stop_subqueue(dev
, qnum
);
1265 unlikely(!virtqueue_enable_cb_delayed(sq
->vq
))) {
1266 /* More just got used, free them then recheck. */
1267 free_old_xmit_skbs(sq
);
1268 if (sq
->vq
->num_free
>= 2+MAX_SKB_FRAGS
) {
1269 netif_start_subqueue(dev
, qnum
);
1270 virtqueue_disable_cb(sq
->vq
);
1275 if (kick
|| netif_xmit_stopped(txq
))
1276 virtqueue_kick(sq
->vq
);
1278 return NETDEV_TX_OK
;
1282 * Send command via the control virtqueue and check status. Commands
1283 * supported by the hypervisor, as indicated by feature bits, should
1284 * never fail unless improperly formatted.
1286 static bool virtnet_send_command(struct virtnet_info
*vi
, u8
class, u8 cmd
,
1287 struct scatterlist
*out
)
1289 struct scatterlist
*sgs
[4], hdr
, stat
;
1290 unsigned out_num
= 0, tmp
;
1292 /* Caller should know better */
1293 BUG_ON(!virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_CTRL_VQ
));
1295 vi
->ctrl_status
= ~0;
1296 vi
->ctrl_hdr
.class = class;
1297 vi
->ctrl_hdr
.cmd
= cmd
;
1299 sg_init_one(&hdr
, &vi
->ctrl_hdr
, sizeof(vi
->ctrl_hdr
));
1300 sgs
[out_num
++] = &hdr
;
1303 sgs
[out_num
++] = out
;
1305 /* Add return status. */
1306 sg_init_one(&stat
, &vi
->ctrl_status
, sizeof(vi
->ctrl_status
));
1307 sgs
[out_num
] = &stat
;
1309 BUG_ON(out_num
+ 1 > ARRAY_SIZE(sgs
));
1310 virtqueue_add_sgs(vi
->cvq
, sgs
, out_num
, 1, vi
, GFP_ATOMIC
);
1312 if (unlikely(!virtqueue_kick(vi
->cvq
)))
1313 return vi
->ctrl_status
== VIRTIO_NET_OK
;
1315 /* Spin for a response, the kick causes an ioport write, trapping
1316 * into the hypervisor, so the request should be handled immediately.
1318 while (!virtqueue_get_buf(vi
->cvq
, &tmp
) &&
1319 !virtqueue_is_broken(vi
->cvq
))
1322 return vi
->ctrl_status
== VIRTIO_NET_OK
;
1325 static int virtnet_set_mac_address(struct net_device
*dev
, void *p
)
1327 struct virtnet_info
*vi
= netdev_priv(dev
);
1328 struct virtio_device
*vdev
= vi
->vdev
;
1330 struct sockaddr
*addr
;
1331 struct scatterlist sg
;
1333 addr
= kmemdup(p
, sizeof(*addr
), GFP_KERNEL
);
1337 ret
= eth_prepare_mac_addr_change(dev
, addr
);
1341 if (virtio_has_feature(vdev
, VIRTIO_NET_F_CTRL_MAC_ADDR
)) {
1342 sg_init_one(&sg
, addr
->sa_data
, dev
->addr_len
);
1343 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_MAC
,
1344 VIRTIO_NET_CTRL_MAC_ADDR_SET
, &sg
)) {
1345 dev_warn(&vdev
->dev
,
1346 "Failed to set mac address by vq command.\n");
1350 } else if (virtio_has_feature(vdev
, VIRTIO_NET_F_MAC
) &&
1351 !virtio_has_feature(vdev
, VIRTIO_F_VERSION_1
)) {
1354 /* Naturally, this has an atomicity problem. */
1355 for (i
= 0; i
< dev
->addr_len
; i
++)
1356 virtio_cwrite8(vdev
,
1357 offsetof(struct virtio_net_config
, mac
) +
1358 i
, addr
->sa_data
[i
]);
1361 eth_commit_mac_addr_change(dev
, p
);
1369 static void virtnet_stats(struct net_device
*dev
,
1370 struct rtnl_link_stats64
*tot
)
1372 struct virtnet_info
*vi
= netdev_priv(dev
);
1376 for_each_possible_cpu(cpu
) {
1377 struct virtnet_stats
*stats
= per_cpu_ptr(vi
->stats
, cpu
);
1378 u64 tpackets
, tbytes
, rpackets
, rbytes
;
1381 start
= u64_stats_fetch_begin_irq(&stats
->tx_syncp
);
1382 tpackets
= stats
->tx_packets
;
1383 tbytes
= stats
->tx_bytes
;
1384 } while (u64_stats_fetch_retry_irq(&stats
->tx_syncp
, start
));
1387 start
= u64_stats_fetch_begin_irq(&stats
->rx_syncp
);
1388 rpackets
= stats
->rx_packets
;
1389 rbytes
= stats
->rx_bytes
;
1390 } while (u64_stats_fetch_retry_irq(&stats
->rx_syncp
, start
));
1392 tot
->rx_packets
+= rpackets
;
1393 tot
->tx_packets
+= tpackets
;
1394 tot
->rx_bytes
+= rbytes
;
1395 tot
->tx_bytes
+= tbytes
;
1398 tot
->tx_dropped
= dev
->stats
.tx_dropped
;
1399 tot
->tx_fifo_errors
= dev
->stats
.tx_fifo_errors
;
1400 tot
->rx_dropped
= dev
->stats
.rx_dropped
;
1401 tot
->rx_length_errors
= dev
->stats
.rx_length_errors
;
1402 tot
->rx_frame_errors
= dev
->stats
.rx_frame_errors
;
1405 #ifdef CONFIG_NET_POLL_CONTROLLER
1406 static void virtnet_netpoll(struct net_device
*dev
)
1408 struct virtnet_info
*vi
= netdev_priv(dev
);
1411 for (i
= 0; i
< vi
->curr_queue_pairs
; i
++)
1412 napi_schedule(&vi
->rq
[i
].napi
);
1416 static void virtnet_ack_link_announce(struct virtnet_info
*vi
)
1419 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_ANNOUNCE
,
1420 VIRTIO_NET_CTRL_ANNOUNCE_ACK
, NULL
))
1421 dev_warn(&vi
->dev
->dev
, "Failed to ack link announce.\n");
1425 static int _virtnet_set_queues(struct virtnet_info
*vi
, u16 queue_pairs
)
1427 struct scatterlist sg
;
1428 struct net_device
*dev
= vi
->dev
;
1430 if (!vi
->has_cvq
|| !virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_MQ
))
1433 vi
->ctrl_mq
.virtqueue_pairs
= cpu_to_virtio16(vi
->vdev
, queue_pairs
);
1434 sg_init_one(&sg
, &vi
->ctrl_mq
, sizeof(vi
->ctrl_mq
));
1436 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_MQ
,
1437 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET
, &sg
)) {
1438 dev_warn(&dev
->dev
, "Fail to set num of queue pairs to %d\n",
1442 vi
->curr_queue_pairs
= queue_pairs
;
1443 /* virtnet_open() will refill when device is going to up. */
1444 if (dev
->flags
& IFF_UP
)
1445 schedule_delayed_work(&vi
->refill
, 0);
1451 static int virtnet_set_queues(struct virtnet_info
*vi
, u16 queue_pairs
)
1456 err
= _virtnet_set_queues(vi
, queue_pairs
);
1461 static int virtnet_close(struct net_device
*dev
)
1463 struct virtnet_info
*vi
= netdev_priv(dev
);
1466 /* Make sure refill_work doesn't re-enable napi! */
1467 cancel_delayed_work_sync(&vi
->refill
);
1469 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1470 napi_disable(&vi
->rq
[i
].napi
);
1471 virtnet_napi_tx_disable(&vi
->sq
[i
].napi
);
1477 static void virtnet_set_rx_mode(struct net_device
*dev
)
1479 struct virtnet_info
*vi
= netdev_priv(dev
);
1480 struct scatterlist sg
[2];
1481 struct virtio_net_ctrl_mac
*mac_data
;
1482 struct netdev_hw_addr
*ha
;
1488 /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1489 if (!virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_CTRL_RX
))
1492 vi
->ctrl_promisc
= ((dev
->flags
& IFF_PROMISC
) != 0);
1493 vi
->ctrl_allmulti
= ((dev
->flags
& IFF_ALLMULTI
) != 0);
1495 sg_init_one(sg
, &vi
->ctrl_promisc
, sizeof(vi
->ctrl_promisc
));
1497 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_RX
,
1498 VIRTIO_NET_CTRL_RX_PROMISC
, sg
))
1499 dev_warn(&dev
->dev
, "Failed to %sable promisc mode.\n",
1500 vi
->ctrl_promisc
? "en" : "dis");
1502 sg_init_one(sg
, &vi
->ctrl_allmulti
, sizeof(vi
->ctrl_allmulti
));
1504 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_RX
,
1505 VIRTIO_NET_CTRL_RX_ALLMULTI
, sg
))
1506 dev_warn(&dev
->dev
, "Failed to %sable allmulti mode.\n",
1507 vi
->ctrl_allmulti
? "en" : "dis");
1509 uc_count
= netdev_uc_count(dev
);
1510 mc_count
= netdev_mc_count(dev
);
1511 /* MAC filter - use one buffer for both lists */
1512 buf
= kzalloc(((uc_count
+ mc_count
) * ETH_ALEN
) +
1513 (2 * sizeof(mac_data
->entries
)), GFP_ATOMIC
);
1518 sg_init_table(sg
, 2);
1520 /* Store the unicast list and count in the front of the buffer */
1521 mac_data
->entries
= cpu_to_virtio32(vi
->vdev
, uc_count
);
1523 netdev_for_each_uc_addr(ha
, dev
)
1524 memcpy(&mac_data
->macs
[i
++][0], ha
->addr
, ETH_ALEN
);
1526 sg_set_buf(&sg
[0], mac_data
,
1527 sizeof(mac_data
->entries
) + (uc_count
* ETH_ALEN
));
1529 /* multicast list and count fill the end */
1530 mac_data
= (void *)&mac_data
->macs
[uc_count
][0];
1532 mac_data
->entries
= cpu_to_virtio32(vi
->vdev
, mc_count
);
1534 netdev_for_each_mc_addr(ha
, dev
)
1535 memcpy(&mac_data
->macs
[i
++][0], ha
->addr
, ETH_ALEN
);
1537 sg_set_buf(&sg
[1], mac_data
,
1538 sizeof(mac_data
->entries
) + (mc_count
* ETH_ALEN
));
1540 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_MAC
,
1541 VIRTIO_NET_CTRL_MAC_TABLE_SET
, sg
))
1542 dev_warn(&dev
->dev
, "Failed to set MAC filter table.\n");
1547 static int virtnet_vlan_rx_add_vid(struct net_device
*dev
,
1548 __be16 proto
, u16 vid
)
1550 struct virtnet_info
*vi
= netdev_priv(dev
);
1551 struct scatterlist sg
;
1554 sg_init_one(&sg
, &vi
->ctrl_vid
, sizeof(vi
->ctrl_vid
));
1556 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_VLAN
,
1557 VIRTIO_NET_CTRL_VLAN_ADD
, &sg
))
1558 dev_warn(&dev
->dev
, "Failed to add VLAN ID %d.\n", vid
);
1562 static int virtnet_vlan_rx_kill_vid(struct net_device
*dev
,
1563 __be16 proto
, u16 vid
)
1565 struct virtnet_info
*vi
= netdev_priv(dev
);
1566 struct scatterlist sg
;
1569 sg_init_one(&sg
, &vi
->ctrl_vid
, sizeof(vi
->ctrl_vid
));
1571 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_VLAN
,
1572 VIRTIO_NET_CTRL_VLAN_DEL
, &sg
))
1573 dev_warn(&dev
->dev
, "Failed to kill VLAN ID %d.\n", vid
);
1577 static void virtnet_clean_affinity(struct virtnet_info
*vi
, long hcpu
)
1581 if (vi
->affinity_hint_set
) {
1582 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1583 virtqueue_set_affinity(vi
->rq
[i
].vq
, -1);
1584 virtqueue_set_affinity(vi
->sq
[i
].vq
, -1);
1587 vi
->affinity_hint_set
= false;
1591 static void virtnet_set_affinity(struct virtnet_info
*vi
)
1596 /* In multiqueue mode, when the number of cpu is equal to the number of
1597 * queue pairs, we let the queue pairs to be private to one cpu by
1598 * setting the affinity hint to eliminate the contention.
1600 if (vi
->curr_queue_pairs
== 1 ||
1601 vi
->max_queue_pairs
!= num_online_cpus()) {
1602 virtnet_clean_affinity(vi
, -1);
1607 for_each_online_cpu(cpu
) {
1608 virtqueue_set_affinity(vi
->rq
[i
].vq
, cpu
);
1609 virtqueue_set_affinity(vi
->sq
[i
].vq
, cpu
);
1610 netif_set_xps_queue(vi
->dev
, cpumask_of(cpu
), i
);
1614 vi
->affinity_hint_set
= true;
1617 static int virtnet_cpu_online(unsigned int cpu
, struct hlist_node
*node
)
1619 struct virtnet_info
*vi
= hlist_entry_safe(node
, struct virtnet_info
,
1621 virtnet_set_affinity(vi
);
1625 static int virtnet_cpu_dead(unsigned int cpu
, struct hlist_node
*node
)
1627 struct virtnet_info
*vi
= hlist_entry_safe(node
, struct virtnet_info
,
1629 virtnet_set_affinity(vi
);
1633 static int virtnet_cpu_down_prep(unsigned int cpu
, struct hlist_node
*node
)
1635 struct virtnet_info
*vi
= hlist_entry_safe(node
, struct virtnet_info
,
1638 virtnet_clean_affinity(vi
, cpu
);
1642 static enum cpuhp_state virtionet_online
;
1644 static int virtnet_cpu_notif_add(struct virtnet_info
*vi
)
1648 ret
= cpuhp_state_add_instance_nocalls(virtionet_online
, &vi
->node
);
1651 ret
= cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD
,
1655 cpuhp_state_remove_instance_nocalls(virtionet_online
, &vi
->node
);
1659 static void virtnet_cpu_notif_remove(struct virtnet_info
*vi
)
1661 cpuhp_state_remove_instance_nocalls(virtionet_online
, &vi
->node
);
1662 cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD
,
1666 static void virtnet_get_ringparam(struct net_device
*dev
,
1667 struct ethtool_ringparam
*ring
)
1669 struct virtnet_info
*vi
= netdev_priv(dev
);
1671 ring
->rx_max_pending
= virtqueue_get_vring_size(vi
->rq
[0].vq
);
1672 ring
->tx_max_pending
= virtqueue_get_vring_size(vi
->sq
[0].vq
);
1673 ring
->rx_pending
= ring
->rx_max_pending
;
1674 ring
->tx_pending
= ring
->tx_max_pending
;
1678 static void virtnet_get_drvinfo(struct net_device
*dev
,
1679 struct ethtool_drvinfo
*info
)
1681 struct virtnet_info
*vi
= netdev_priv(dev
);
1682 struct virtio_device
*vdev
= vi
->vdev
;
1684 strlcpy(info
->driver
, KBUILD_MODNAME
, sizeof(info
->driver
));
1685 strlcpy(info
->version
, VIRTNET_DRIVER_VERSION
, sizeof(info
->version
));
1686 strlcpy(info
->bus_info
, virtio_bus_name(vdev
), sizeof(info
->bus_info
));
1690 /* TODO: Eliminate OOO packets during switching */
1691 static int virtnet_set_channels(struct net_device
*dev
,
1692 struct ethtool_channels
*channels
)
1694 struct virtnet_info
*vi
= netdev_priv(dev
);
1695 u16 queue_pairs
= channels
->combined_count
;
1698 /* We don't support separate rx/tx channels.
1699 * We don't allow setting 'other' channels.
1701 if (channels
->rx_count
|| channels
->tx_count
|| channels
->other_count
)
1704 if (queue_pairs
> vi
->max_queue_pairs
|| queue_pairs
== 0)
1707 /* For now we don't support modifying channels while XDP is loaded
1708 * also when XDP is loaded all RX queues have XDP programs so we only
1709 * need to check a single RX queue.
1711 if (vi
->rq
[0].xdp_prog
)
1715 err
= _virtnet_set_queues(vi
, queue_pairs
);
1717 netif_set_real_num_tx_queues(dev
, queue_pairs
);
1718 netif_set_real_num_rx_queues(dev
, queue_pairs
);
1720 virtnet_set_affinity(vi
);
1727 static void virtnet_get_channels(struct net_device
*dev
,
1728 struct ethtool_channels
*channels
)
1730 struct virtnet_info
*vi
= netdev_priv(dev
);
1732 channels
->combined_count
= vi
->curr_queue_pairs
;
1733 channels
->max_combined
= vi
->max_queue_pairs
;
1734 channels
->max_other
= 0;
1735 channels
->rx_count
= 0;
1736 channels
->tx_count
= 0;
1737 channels
->other_count
= 0;
1740 /* Check if the user is trying to change anything besides speed/duplex */
1742 virtnet_validate_ethtool_cmd(const struct ethtool_link_ksettings
*cmd
)
1744 struct ethtool_link_ksettings diff1
= *cmd
;
1745 struct ethtool_link_ksettings diff2
= {};
1747 /* cmd is always set so we need to clear it, validate the port type
1748 * and also without autonegotiation we can ignore advertising
1750 diff1
.base
.speed
= 0;
1751 diff2
.base
.port
= PORT_OTHER
;
1752 ethtool_link_ksettings_zero_link_mode(&diff1
, advertising
);
1753 diff1
.base
.duplex
= 0;
1755 diff1
.base
.link_mode_masks_nwords
= 0;
1757 return !memcmp(&diff1
.base
, &diff2
.base
, sizeof(diff1
.base
)) &&
1758 bitmap_empty(diff1
.link_modes
.supported
,
1759 __ETHTOOL_LINK_MODE_MASK_NBITS
) &&
1760 bitmap_empty(diff1
.link_modes
.advertising
,
1761 __ETHTOOL_LINK_MODE_MASK_NBITS
) &&
1762 bitmap_empty(diff1
.link_modes
.lp_advertising
,
1763 __ETHTOOL_LINK_MODE_MASK_NBITS
);
1766 static int virtnet_set_link_ksettings(struct net_device
*dev
,
1767 const struct ethtool_link_ksettings
*cmd
)
1769 struct virtnet_info
*vi
= netdev_priv(dev
);
1772 speed
= cmd
->base
.speed
;
1773 /* don't allow custom speed and duplex */
1774 if (!ethtool_validate_speed(speed
) ||
1775 !ethtool_validate_duplex(cmd
->base
.duplex
) ||
1776 !virtnet_validate_ethtool_cmd(cmd
))
1779 vi
->duplex
= cmd
->base
.duplex
;
1784 static int virtnet_get_link_ksettings(struct net_device
*dev
,
1785 struct ethtool_link_ksettings
*cmd
)
1787 struct virtnet_info
*vi
= netdev_priv(dev
);
1789 cmd
->base
.speed
= vi
->speed
;
1790 cmd
->base
.duplex
= vi
->duplex
;
1791 cmd
->base
.port
= PORT_OTHER
;
1796 static void virtnet_init_settings(struct net_device
*dev
)
1798 struct virtnet_info
*vi
= netdev_priv(dev
);
1800 vi
->speed
= SPEED_UNKNOWN
;
1801 vi
->duplex
= DUPLEX_UNKNOWN
;
1804 static const struct ethtool_ops virtnet_ethtool_ops
= {
1805 .get_drvinfo
= virtnet_get_drvinfo
,
1806 .get_link
= ethtool_op_get_link
,
1807 .get_ringparam
= virtnet_get_ringparam
,
1808 .set_channels
= virtnet_set_channels
,
1809 .get_channels
= virtnet_get_channels
,
1810 .get_ts_info
= ethtool_op_get_ts_info
,
1811 .get_link_ksettings
= virtnet_get_link_ksettings
,
1812 .set_link_ksettings
= virtnet_set_link_ksettings
,
1815 static void virtnet_freeze_down(struct virtio_device
*vdev
)
1817 struct virtnet_info
*vi
= vdev
->priv
;
1820 /* Make sure no work handler is accessing the device */
1821 flush_work(&vi
->config_work
);
1823 netif_device_detach(vi
->dev
);
1824 netif_tx_disable(vi
->dev
);
1825 cancel_delayed_work_sync(&vi
->refill
);
1827 if (netif_running(vi
->dev
)) {
1828 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1829 napi_disable(&vi
->rq
[i
].napi
);
1830 virtnet_napi_tx_disable(&vi
->sq
[i
].napi
);
1835 static int init_vqs(struct virtnet_info
*vi
);
1836 static void _remove_vq_common(struct virtnet_info
*vi
);
1838 static int virtnet_restore_up(struct virtio_device
*vdev
)
1840 struct virtnet_info
*vi
= vdev
->priv
;
1847 virtio_device_ready(vdev
);
1849 if (netif_running(vi
->dev
)) {
1850 for (i
= 0; i
< vi
->curr_queue_pairs
; i
++)
1851 if (!try_fill_recv(vi
, &vi
->rq
[i
], GFP_KERNEL
))
1852 schedule_delayed_work(&vi
->refill
, 0);
1854 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1855 virtnet_napi_enable(vi
->rq
[i
].vq
, &vi
->rq
[i
].napi
);
1856 virtnet_napi_tx_enable(vi
, vi
->sq
[i
].vq
,
1861 netif_device_attach(vi
->dev
);
1865 static int virtnet_reset(struct virtnet_info
*vi
, int curr_qp
, int xdp_qp
)
1867 struct virtio_device
*dev
= vi
->vdev
;
1870 virtio_config_disable(dev
);
1871 dev
->failed
= dev
->config
->get_status(dev
) & VIRTIO_CONFIG_S_FAILED
;
1872 virtnet_freeze_down(dev
);
1873 _remove_vq_common(vi
);
1875 virtio_add_status(dev
, VIRTIO_CONFIG_S_ACKNOWLEDGE
);
1876 virtio_add_status(dev
, VIRTIO_CONFIG_S_DRIVER
);
1878 ret
= virtio_finalize_features(dev
);
1882 vi
->xdp_queue_pairs
= xdp_qp
;
1883 ret
= virtnet_restore_up(dev
);
1886 ret
= _virtnet_set_queues(vi
, curr_qp
);
1890 virtio_add_status(dev
, VIRTIO_CONFIG_S_DRIVER_OK
);
1891 virtio_config_enable(dev
);
1894 virtio_add_status(dev
, VIRTIO_CONFIG_S_FAILED
);
1898 static int virtnet_xdp_set(struct net_device
*dev
, struct bpf_prog
*prog
,
1899 struct netlink_ext_ack
*extack
)
1901 unsigned long int max_sz
= PAGE_SIZE
- sizeof(struct padded_vnet_hdr
);
1902 struct virtnet_info
*vi
= netdev_priv(dev
);
1903 struct bpf_prog
*old_prog
;
1904 u16 xdp_qp
= 0, curr_qp
;
1907 if (virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_TSO4
) ||
1908 virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_TSO6
) ||
1909 virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_ECN
) ||
1910 virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_UFO
)) {
1911 NL_SET_ERR_MSG_MOD(extack
, "Can't set XDP while host is implementing LRO, disable LRO first");
1915 if (vi
->mergeable_rx_bufs
&& !vi
->any_header_sg
) {
1916 NL_SET_ERR_MSG_MOD(extack
, "XDP expects header/data in single page, any_header_sg required");
1920 if (dev
->mtu
> max_sz
) {
1921 NL_SET_ERR_MSG_MOD(extack
, "MTU too large to enable XDP");
1922 netdev_warn(dev
, "XDP requires MTU less than %lu\n", max_sz
);
1926 curr_qp
= vi
->curr_queue_pairs
- vi
->xdp_queue_pairs
;
1928 xdp_qp
= nr_cpu_ids
;
1930 /* XDP requires extra queues for XDP_TX */
1931 if (curr_qp
+ xdp_qp
> vi
->max_queue_pairs
) {
1932 NL_SET_ERR_MSG_MOD(extack
, "Too few free TX rings available");
1933 netdev_warn(dev
, "request %i queues but max is %i\n",
1934 curr_qp
+ xdp_qp
, vi
->max_queue_pairs
);
1939 prog
= bpf_prog_add(prog
, vi
->max_queue_pairs
- 1);
1941 return PTR_ERR(prog
);
1944 /* Changing the headroom in buffers is a disruptive operation because
1945 * existing buffers must be flushed and reallocated. This will happen
1946 * when a xdp program is initially added or xdp is disabled by removing
1947 * the xdp program resulting in number of XDP queues changing.
1949 if (vi
->xdp_queue_pairs
!= xdp_qp
) {
1950 err
= virtnet_reset(vi
, curr_qp
+ xdp_qp
, xdp_qp
);
1952 dev_warn(&dev
->dev
, "XDP reset failure.\n");
1953 goto virtio_reset_err
;
1957 netif_set_real_num_rx_queues(dev
, curr_qp
+ xdp_qp
);
1959 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1960 old_prog
= rtnl_dereference(vi
->rq
[i
].xdp_prog
);
1961 rcu_assign_pointer(vi
->rq
[i
].xdp_prog
, prog
);
1963 bpf_prog_put(old_prog
);
1969 /* On reset error do our best to unwind XDP changes inflight and return
1970 * error up to user space for resolution. The underlying reset hung on
1971 * us so not much we can do here.
1974 bpf_prog_sub(prog
, vi
->max_queue_pairs
- 1);
1978 static u32
virtnet_xdp_query(struct net_device
*dev
)
1980 struct virtnet_info
*vi
= netdev_priv(dev
);
1981 const struct bpf_prog
*xdp_prog
;
1984 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1985 xdp_prog
= rtnl_dereference(vi
->rq
[i
].xdp_prog
);
1987 return xdp_prog
->aux
->id
;
1992 static int virtnet_xdp(struct net_device
*dev
, struct netdev_xdp
*xdp
)
1994 switch (xdp
->command
) {
1995 case XDP_SETUP_PROG
:
1996 return virtnet_xdp_set(dev
, xdp
->prog
, xdp
->extack
);
1997 case XDP_QUERY_PROG
:
1998 xdp
->prog_id
= virtnet_xdp_query(dev
);
1999 xdp
->prog_attached
= !!xdp
->prog_id
;
2006 static const struct net_device_ops virtnet_netdev
= {
2007 .ndo_open
= virtnet_open
,
2008 .ndo_stop
= virtnet_close
,
2009 .ndo_start_xmit
= start_xmit
,
2010 .ndo_validate_addr
= eth_validate_addr
,
2011 .ndo_set_mac_address
= virtnet_set_mac_address
,
2012 .ndo_set_rx_mode
= virtnet_set_rx_mode
,
2013 .ndo_get_stats64
= virtnet_stats
,
2014 .ndo_vlan_rx_add_vid
= virtnet_vlan_rx_add_vid
,
2015 .ndo_vlan_rx_kill_vid
= virtnet_vlan_rx_kill_vid
,
2016 #ifdef CONFIG_NET_POLL_CONTROLLER
2017 .ndo_poll_controller
= virtnet_netpoll
,
2019 .ndo_xdp
= virtnet_xdp
,
2020 .ndo_features_check
= passthru_features_check
,
2023 static void virtnet_config_changed_work(struct work_struct
*work
)
2025 struct virtnet_info
*vi
=
2026 container_of(work
, struct virtnet_info
, config_work
);
2029 if (virtio_cread_feature(vi
->vdev
, VIRTIO_NET_F_STATUS
,
2030 struct virtio_net_config
, status
, &v
) < 0)
2033 if (v
& VIRTIO_NET_S_ANNOUNCE
) {
2034 netdev_notify_peers(vi
->dev
);
2035 virtnet_ack_link_announce(vi
);
2038 /* Ignore unknown (future) status bits */
2039 v
&= VIRTIO_NET_S_LINK_UP
;
2041 if (vi
->status
== v
)
2046 if (vi
->status
& VIRTIO_NET_S_LINK_UP
) {
2047 netif_carrier_on(vi
->dev
);
2048 netif_tx_wake_all_queues(vi
->dev
);
2050 netif_carrier_off(vi
->dev
);
2051 netif_tx_stop_all_queues(vi
->dev
);
2055 static void virtnet_config_changed(struct virtio_device
*vdev
)
2057 struct virtnet_info
*vi
= vdev
->priv
;
2059 schedule_work(&vi
->config_work
);
2062 static void virtnet_free_queues(struct virtnet_info
*vi
)
2066 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2067 napi_hash_del(&vi
->rq
[i
].napi
);
2068 netif_napi_del(&vi
->rq
[i
].napi
);
2069 netif_napi_del(&vi
->sq
[i
].napi
);
2072 /* We called napi_hash_del() before netif_napi_del(),
2073 * we need to respect an RCU grace period before freeing vi->rq
2081 static void _free_receive_bufs(struct virtnet_info
*vi
)
2083 struct bpf_prog
*old_prog
;
2086 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2087 while (vi
->rq
[i
].pages
)
2088 __free_pages(get_a_page(&vi
->rq
[i
], GFP_KERNEL
), 0);
2090 old_prog
= rtnl_dereference(vi
->rq
[i
].xdp_prog
);
2091 RCU_INIT_POINTER(vi
->rq
[i
].xdp_prog
, NULL
);
2093 bpf_prog_put(old_prog
);
2097 static void free_receive_bufs(struct virtnet_info
*vi
)
2100 _free_receive_bufs(vi
);
2104 static void free_receive_page_frags(struct virtnet_info
*vi
)
2107 for (i
= 0; i
< vi
->max_queue_pairs
; i
++)
2108 if (vi
->rq
[i
].alloc_frag
.page
)
2109 put_page(vi
->rq
[i
].alloc_frag
.page
);
2112 static bool is_xdp_raw_buffer_queue(struct virtnet_info
*vi
, int q
)
2114 if (q
< (vi
->curr_queue_pairs
- vi
->xdp_queue_pairs
))
2116 else if (q
< vi
->curr_queue_pairs
)
2122 static void free_unused_bufs(struct virtnet_info
*vi
)
2127 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2128 struct virtqueue
*vq
= vi
->sq
[i
].vq
;
2129 while ((buf
= virtqueue_detach_unused_buf(vq
)) != NULL
) {
2130 if (!is_xdp_raw_buffer_queue(vi
, i
))
2133 put_page(virt_to_head_page(buf
));
2137 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2138 struct virtqueue
*vq
= vi
->rq
[i
].vq
;
2140 while ((buf
= virtqueue_detach_unused_buf(vq
)) != NULL
) {
2141 if (vi
->mergeable_rx_bufs
) {
2142 put_page(virt_to_head_page(buf
));
2143 } else if (vi
->big_packets
) {
2144 give_pages(&vi
->rq
[i
], buf
);
2146 put_page(virt_to_head_page(buf
));
2152 static void virtnet_del_vqs(struct virtnet_info
*vi
)
2154 struct virtio_device
*vdev
= vi
->vdev
;
2156 virtnet_clean_affinity(vi
, -1);
2158 vdev
->config
->del_vqs(vdev
);
2160 virtnet_free_queues(vi
);
2163 /* How large should a single buffer be so a queue full of these can fit at
2164 * least one full packet?
2165 * Logic below assumes the mergeable buffer header is used.
2167 static unsigned int mergeable_min_buf_len(struct virtnet_info
*vi
, struct virtqueue
*vq
)
2169 const unsigned int hdr_len
= sizeof(struct virtio_net_hdr_mrg_rxbuf
);
2170 unsigned int rq_size
= virtqueue_get_vring_size(vq
);
2171 unsigned int packet_len
= vi
->big_packets
? IP_MAX_MTU
: vi
->dev
->max_mtu
;
2172 unsigned int buf_len
= hdr_len
+ ETH_HLEN
+ VLAN_HLEN
+ packet_len
;
2173 unsigned int min_buf_len
= DIV_ROUND_UP(buf_len
, rq_size
);
2175 return max(max(min_buf_len
, hdr_len
) - hdr_len
,
2176 (unsigned int)GOOD_PACKET_LEN
);
2179 static int virtnet_find_vqs(struct virtnet_info
*vi
)
2181 vq_callback_t
**callbacks
;
2182 struct virtqueue
**vqs
;
2188 /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
2189 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
2190 * possible control vq.
2192 total_vqs
= vi
->max_queue_pairs
* 2 +
2193 virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_CTRL_VQ
);
2195 /* Allocate space for find_vqs parameters */
2196 vqs
= kzalloc(total_vqs
* sizeof(*vqs
), GFP_KERNEL
);
2199 callbacks
= kmalloc(total_vqs
* sizeof(*callbacks
), GFP_KERNEL
);
2202 names
= kmalloc(total_vqs
* sizeof(*names
), GFP_KERNEL
);
2205 if (vi
->mergeable_rx_bufs
) {
2206 ctx
= kzalloc(total_vqs
* sizeof(*ctx
), GFP_KERNEL
);
2213 /* Parameters for control virtqueue, if any */
2215 callbacks
[total_vqs
- 1] = NULL
;
2216 names
[total_vqs
- 1] = "control";
2219 /* Allocate/initialize parameters for send/receive virtqueues */
2220 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2221 callbacks
[rxq2vq(i
)] = skb_recv_done
;
2222 callbacks
[txq2vq(i
)] = skb_xmit_done
;
2223 sprintf(vi
->rq
[i
].name
, "input.%d", i
);
2224 sprintf(vi
->sq
[i
].name
, "output.%d", i
);
2225 names
[rxq2vq(i
)] = vi
->rq
[i
].name
;
2226 names
[txq2vq(i
)] = vi
->sq
[i
].name
;
2228 ctx
[rxq2vq(i
)] = true;
2231 ret
= vi
->vdev
->config
->find_vqs(vi
->vdev
, total_vqs
, vqs
, callbacks
,
2237 vi
->cvq
= vqs
[total_vqs
- 1];
2238 if (virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_CTRL_VLAN
))
2239 vi
->dev
->features
|= NETIF_F_HW_VLAN_CTAG_FILTER
;
2242 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2243 vi
->rq
[i
].vq
= vqs
[rxq2vq(i
)];
2244 vi
->rq
[i
].min_buf_len
= mergeable_min_buf_len(vi
, vi
->rq
[i
].vq
);
2245 vi
->sq
[i
].vq
= vqs
[txq2vq(i
)];
2267 static int virtnet_alloc_queues(struct virtnet_info
*vi
)
2271 vi
->sq
= kzalloc(sizeof(*vi
->sq
) * vi
->max_queue_pairs
, GFP_KERNEL
);
2274 vi
->rq
= kzalloc(sizeof(*vi
->rq
) * vi
->max_queue_pairs
, GFP_KERNEL
);
2278 INIT_DELAYED_WORK(&vi
->refill
, refill_work
);
2279 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2280 vi
->rq
[i
].pages
= NULL
;
2281 netif_napi_add(vi
->dev
, &vi
->rq
[i
].napi
, virtnet_poll
,
2283 netif_tx_napi_add(vi
->dev
, &vi
->sq
[i
].napi
, virtnet_poll_tx
,
2284 napi_tx
? napi_weight
: 0);
2286 sg_init_table(vi
->rq
[i
].sg
, ARRAY_SIZE(vi
->rq
[i
].sg
));
2287 ewma_pkt_len_init(&vi
->rq
[i
].mrg_avg_pkt_len
);
2288 sg_init_table(vi
->sq
[i
].sg
, ARRAY_SIZE(vi
->sq
[i
].sg
));
2299 static int init_vqs(struct virtnet_info
*vi
)
2303 /* Allocate send & receive queues */
2304 ret
= virtnet_alloc_queues(vi
);
2308 ret
= virtnet_find_vqs(vi
);
2313 virtnet_set_affinity(vi
);
2319 virtnet_free_queues(vi
);
2325 static ssize_t
mergeable_rx_buffer_size_show(struct netdev_rx_queue
*queue
,
2326 struct rx_queue_attribute
*attribute
, char *buf
)
2328 struct virtnet_info
*vi
= netdev_priv(queue
->dev
);
2329 unsigned int queue_index
= get_netdev_rx_queue_index(queue
);
2330 struct ewma_pkt_len
*avg
;
2332 BUG_ON(queue_index
>= vi
->max_queue_pairs
);
2333 avg
= &vi
->rq
[queue_index
].mrg_avg_pkt_len
;
2334 return sprintf(buf
, "%u\n",
2335 get_mergeable_buf_len(&vi
->rq
[queue_index
], avg
));
2338 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute
=
2339 __ATTR_RO(mergeable_rx_buffer_size
);
2341 static struct attribute
*virtio_net_mrg_rx_attrs
[] = {
2342 &mergeable_rx_buffer_size_attribute
.attr
,
2346 static const struct attribute_group virtio_net_mrg_rx_group
= {
2347 .name
= "virtio_net",
2348 .attrs
= virtio_net_mrg_rx_attrs
2352 static bool virtnet_fail_on_feature(struct virtio_device
*vdev
,
2354 const char *fname
, const char *dname
)
2356 if (!virtio_has_feature(vdev
, fbit
))
2359 dev_err(&vdev
->dev
, "device advertises feature %s but not %s",
2365 #define VIRTNET_FAIL_ON(vdev, fbit, dbit) \
2366 virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
2368 static bool virtnet_validate_features(struct virtio_device
*vdev
)
2370 if (!virtio_has_feature(vdev
, VIRTIO_NET_F_CTRL_VQ
) &&
2371 (VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_CTRL_RX
,
2372 "VIRTIO_NET_F_CTRL_VQ") ||
2373 VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_CTRL_VLAN
,
2374 "VIRTIO_NET_F_CTRL_VQ") ||
2375 VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_GUEST_ANNOUNCE
,
2376 "VIRTIO_NET_F_CTRL_VQ") ||
2377 VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_MQ
, "VIRTIO_NET_F_CTRL_VQ") ||
2378 VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_CTRL_MAC_ADDR
,
2379 "VIRTIO_NET_F_CTRL_VQ"))) {
2386 #define MIN_MTU ETH_MIN_MTU
2387 #define MAX_MTU ETH_MAX_MTU
2389 static int virtnet_validate(struct virtio_device
*vdev
)
2391 if (!vdev
->config
->get
) {
2392 dev_err(&vdev
->dev
, "%s failure: config access disabled\n",
2397 if (!virtnet_validate_features(vdev
))
2400 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MTU
)) {
2401 int mtu
= virtio_cread16(vdev
,
2402 offsetof(struct virtio_net_config
,
2405 __virtio_clear_bit(vdev
, VIRTIO_NET_F_MTU
);
2411 static int virtnet_probe(struct virtio_device
*vdev
)
2414 struct net_device
*dev
;
2415 struct virtnet_info
*vi
;
2416 u16 max_queue_pairs
;
2419 /* Find if host supports multiqueue virtio_net device */
2420 err
= virtio_cread_feature(vdev
, VIRTIO_NET_F_MQ
,
2421 struct virtio_net_config
,
2422 max_virtqueue_pairs
, &max_queue_pairs
);
2424 /* We need at least 2 queue's */
2425 if (err
|| max_queue_pairs
< VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN
||
2426 max_queue_pairs
> VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX
||
2427 !virtio_has_feature(vdev
, VIRTIO_NET_F_CTRL_VQ
))
2428 max_queue_pairs
= 1;
2430 /* Allocate ourselves a network device with room for our info */
2431 dev
= alloc_etherdev_mq(sizeof(struct virtnet_info
), max_queue_pairs
);
2435 /* Set up network device as normal. */
2436 dev
->priv_flags
|= IFF_UNICAST_FLT
| IFF_LIVE_ADDR_CHANGE
;
2437 dev
->netdev_ops
= &virtnet_netdev
;
2438 dev
->features
= NETIF_F_HIGHDMA
;
2440 dev
->ethtool_ops
= &virtnet_ethtool_ops
;
2441 SET_NETDEV_DEV(dev
, &vdev
->dev
);
2443 /* Do we support "hardware" checksums? */
2444 if (virtio_has_feature(vdev
, VIRTIO_NET_F_CSUM
)) {
2445 /* This opens up the world of extra features. */
2446 dev
->hw_features
|= NETIF_F_HW_CSUM
| NETIF_F_SG
;
2448 dev
->features
|= NETIF_F_HW_CSUM
| NETIF_F_SG
;
2450 if (virtio_has_feature(vdev
, VIRTIO_NET_F_GSO
)) {
2451 dev
->hw_features
|= NETIF_F_TSO
2452 | NETIF_F_TSO_ECN
| NETIF_F_TSO6
;
2454 /* Individual feature bits: what can host handle? */
2455 if (virtio_has_feature(vdev
, VIRTIO_NET_F_HOST_TSO4
))
2456 dev
->hw_features
|= NETIF_F_TSO
;
2457 if (virtio_has_feature(vdev
, VIRTIO_NET_F_HOST_TSO6
))
2458 dev
->hw_features
|= NETIF_F_TSO6
;
2459 if (virtio_has_feature(vdev
, VIRTIO_NET_F_HOST_ECN
))
2460 dev
->hw_features
|= NETIF_F_TSO_ECN
;
2462 dev
->features
|= NETIF_F_GSO_ROBUST
;
2465 dev
->features
|= dev
->hw_features
& NETIF_F_ALL_TSO
;
2466 /* (!csum && gso) case will be fixed by register_netdev() */
2468 if (virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_CSUM
))
2469 dev
->features
|= NETIF_F_RXCSUM
;
2471 dev
->vlan_features
= dev
->features
;
2473 /* MTU range: 68 - 65535 */
2474 dev
->min_mtu
= MIN_MTU
;
2475 dev
->max_mtu
= MAX_MTU
;
2477 /* Configuration may specify what MAC to use. Otherwise random. */
2478 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MAC
))
2479 virtio_cread_bytes(vdev
,
2480 offsetof(struct virtio_net_config
, mac
),
2481 dev
->dev_addr
, dev
->addr_len
);
2483 eth_hw_addr_random(dev
);
2485 /* Set up our device-specific information */
2486 vi
= netdev_priv(dev
);
2490 vi
->stats
= alloc_percpu(struct virtnet_stats
);
2492 if (vi
->stats
== NULL
)
2495 for_each_possible_cpu(i
) {
2496 struct virtnet_stats
*virtnet_stats
;
2497 virtnet_stats
= per_cpu_ptr(vi
->stats
, i
);
2498 u64_stats_init(&virtnet_stats
->tx_syncp
);
2499 u64_stats_init(&virtnet_stats
->rx_syncp
);
2502 INIT_WORK(&vi
->config_work
, virtnet_config_changed_work
);
2504 /* If we can receive ANY GSO packets, we must allocate large ones. */
2505 if (virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_TSO4
) ||
2506 virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_TSO6
) ||
2507 virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_ECN
) ||
2508 virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_UFO
))
2509 vi
->big_packets
= true;
2511 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MRG_RXBUF
))
2512 vi
->mergeable_rx_bufs
= true;
2514 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MRG_RXBUF
) ||
2515 virtio_has_feature(vdev
, VIRTIO_F_VERSION_1
))
2516 vi
->hdr_len
= sizeof(struct virtio_net_hdr_mrg_rxbuf
);
2518 vi
->hdr_len
= sizeof(struct virtio_net_hdr
);
2520 if (virtio_has_feature(vdev
, VIRTIO_F_ANY_LAYOUT
) ||
2521 virtio_has_feature(vdev
, VIRTIO_F_VERSION_1
))
2522 vi
->any_header_sg
= true;
2524 if (virtio_has_feature(vdev
, VIRTIO_NET_F_CTRL_VQ
))
2527 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MTU
)) {
2528 mtu
= virtio_cread16(vdev
,
2529 offsetof(struct virtio_net_config
,
2531 if (mtu
< dev
->min_mtu
) {
2532 /* Should never trigger: MTU was previously validated
2533 * in virtnet_validate.
2535 dev_err(&vdev
->dev
, "device MTU appears to have changed "
2536 "it is now %d < %d", mtu
, dev
->min_mtu
);
2543 /* TODO: size buffers correctly in this case. */
2544 if (dev
->mtu
> ETH_DATA_LEN
)
2545 vi
->big_packets
= true;
2548 if (vi
->any_header_sg
)
2549 dev
->needed_headroom
= vi
->hdr_len
;
2551 /* Enable multiqueue by default */
2552 if (num_online_cpus() >= max_queue_pairs
)
2553 vi
->curr_queue_pairs
= max_queue_pairs
;
2555 vi
->curr_queue_pairs
= num_online_cpus();
2556 vi
->max_queue_pairs
= max_queue_pairs
;
2558 /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
2564 if (vi
->mergeable_rx_bufs
)
2565 dev
->sysfs_rx_queue_group
= &virtio_net_mrg_rx_group
;
2567 netif_set_real_num_tx_queues(dev
, vi
->curr_queue_pairs
);
2568 netif_set_real_num_rx_queues(dev
, vi
->curr_queue_pairs
);
2570 virtnet_init_settings(dev
);
2572 err
= register_netdev(dev
);
2574 pr_debug("virtio_net: registering device failed\n");
2578 virtio_device_ready(vdev
);
2580 err
= virtnet_cpu_notif_add(vi
);
2582 pr_debug("virtio_net: registering cpu notifier failed\n");
2583 goto free_unregister_netdev
;
2586 virtnet_set_queues(vi
, vi
->curr_queue_pairs
);
2588 /* Assume link up if device can't report link status,
2589 otherwise get link status from config. */
2590 if (virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_STATUS
)) {
2591 netif_carrier_off(dev
);
2592 schedule_work(&vi
->config_work
);
2594 vi
->status
= VIRTIO_NET_S_LINK_UP
;
2595 netif_carrier_on(dev
);
2598 pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
2599 dev
->name
, max_queue_pairs
);
2603 free_unregister_netdev
:
2604 vi
->vdev
->config
->reset(vdev
);
2606 unregister_netdev(dev
);
2608 cancel_delayed_work_sync(&vi
->refill
);
2609 free_receive_page_frags(vi
);
2610 virtnet_del_vqs(vi
);
2612 free_percpu(vi
->stats
);
2618 static void _remove_vq_common(struct virtnet_info
*vi
)
2620 vi
->vdev
->config
->reset(vi
->vdev
);
2621 free_unused_bufs(vi
);
2622 _free_receive_bufs(vi
);
2623 free_receive_page_frags(vi
);
2624 virtnet_del_vqs(vi
);
2627 static void remove_vq_common(struct virtnet_info
*vi
)
2629 vi
->vdev
->config
->reset(vi
->vdev
);
2631 /* Free unused buffers in both send and recv, if any. */
2632 free_unused_bufs(vi
);
2634 free_receive_bufs(vi
);
2636 free_receive_page_frags(vi
);
2638 virtnet_del_vqs(vi
);
2641 static void virtnet_remove(struct virtio_device
*vdev
)
2643 struct virtnet_info
*vi
= vdev
->priv
;
2645 virtnet_cpu_notif_remove(vi
);
2647 /* Make sure no work handler is accessing the device. */
2648 flush_work(&vi
->config_work
);
2650 unregister_netdev(vi
->dev
);
2652 remove_vq_common(vi
);
2654 free_percpu(vi
->stats
);
2655 free_netdev(vi
->dev
);
2658 #ifdef CONFIG_PM_SLEEP
2659 static int virtnet_freeze(struct virtio_device
*vdev
)
2661 struct virtnet_info
*vi
= vdev
->priv
;
2663 virtnet_cpu_notif_remove(vi
);
2664 virtnet_freeze_down(vdev
);
2665 remove_vq_common(vi
);
2670 static int virtnet_restore(struct virtio_device
*vdev
)
2672 struct virtnet_info
*vi
= vdev
->priv
;
2675 err
= virtnet_restore_up(vdev
);
2678 virtnet_set_queues(vi
, vi
->curr_queue_pairs
);
2680 err
= virtnet_cpu_notif_add(vi
);
2688 static struct virtio_device_id id_table
[] = {
2689 { VIRTIO_ID_NET
, VIRTIO_DEV_ANY_ID
},
2693 #define VIRTNET_FEATURES \
2694 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
2696 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
2697 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
2698 VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
2699 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
2700 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
2701 VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
2702 VIRTIO_NET_F_CTRL_MAC_ADDR, \
2705 static unsigned int features
[] = {
2709 static unsigned int features_legacy
[] = {
2712 VIRTIO_F_ANY_LAYOUT
,
2715 static struct virtio_driver virtio_net_driver
= {
2716 .feature_table
= features
,
2717 .feature_table_size
= ARRAY_SIZE(features
),
2718 .feature_table_legacy
= features_legacy
,
2719 .feature_table_size_legacy
= ARRAY_SIZE(features_legacy
),
2720 .driver
.name
= KBUILD_MODNAME
,
2721 .driver
.owner
= THIS_MODULE
,
2722 .id_table
= id_table
,
2723 .validate
= virtnet_validate
,
2724 .probe
= virtnet_probe
,
2725 .remove
= virtnet_remove
,
2726 .config_changed
= virtnet_config_changed
,
2727 #ifdef CONFIG_PM_SLEEP
2728 .freeze
= virtnet_freeze
,
2729 .restore
= virtnet_restore
,
2733 static __init
int virtio_net_driver_init(void)
2737 ret
= cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN
, "virtio/net:online",
2739 virtnet_cpu_down_prep
);
2742 virtionet_online
= ret
;
2743 ret
= cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD
, "virtio/net:dead",
2744 NULL
, virtnet_cpu_dead
);
2748 ret
= register_virtio_driver(&virtio_net_driver
);
2753 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD
);
2755 cpuhp_remove_multi_state(virtionet_online
);
2759 module_init(virtio_net_driver_init
);
2761 static __exit
void virtio_net_driver_exit(void)
2763 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD
);
2764 cpuhp_remove_multi_state(virtionet_online
);
2765 unregister_virtio_driver(&virtio_net_driver
);
2767 module_exit(virtio_net_driver_exit
);
2769 MODULE_DEVICE_TABLE(virtio
, id_table
);
2770 MODULE_DESCRIPTION("Virtio network driver");
2771 MODULE_LICENSE("GPL");