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 /* We copy the packet for XDP in the following cases:
412 * 1) Packet is scattered across multiple rx buffers.
413 * 2) Headroom space is insufficient.
415 * This is inefficient but it's a temporary condition that
416 * we hit right after XDP is enabled and until queue is refilled
417 * with large buffers with sufficient headroom - so it should affect
418 * at most queue size packets.
419 * Afterwards, the conditions to enable
420 * XDP should preclude the underlying device from sending packets
421 * across multiple buffers (num_buf > 1), and we make sure buffers
422 * have enough headroom.
424 static struct page
*xdp_linearize_page(struct receive_queue
*rq
,
431 struct page
*page
= alloc_page(GFP_ATOMIC
);
436 memcpy(page_address(page
) + page_off
, page_address(p
) + offset
, *len
);
444 buf
= virtqueue_get_buf(rq
->vq
, &buflen
);
448 p
= virt_to_head_page(buf
);
449 off
= buf
- page_address(p
);
451 /* guard against a misconfigured or uncooperative backend that
452 * is sending packet larger than the MTU.
454 if ((page_off
+ buflen
) > PAGE_SIZE
) {
459 memcpy(page_address(page
) + page_off
,
460 page_address(p
) + off
, buflen
);
465 /* Headroom does not contribute to packet length */
466 *len
= page_off
- VIRTIO_XDP_HEADROOM
;
469 __free_pages(page
, 0);
473 static struct sk_buff
*receive_small(struct net_device
*dev
,
474 struct virtnet_info
*vi
,
475 struct receive_queue
*rq
,
476 void *buf
, void *ctx
,
480 struct bpf_prog
*xdp_prog
;
481 unsigned int xdp_headroom
= (unsigned long)ctx
;
482 unsigned int header_offset
= VIRTNET_RX_PAD
+ xdp_headroom
;
483 unsigned int headroom
= vi
->hdr_len
+ header_offset
;
484 unsigned int buflen
= SKB_DATA_ALIGN(GOOD_PACKET_LEN
+ headroom
) +
485 SKB_DATA_ALIGN(sizeof(struct skb_shared_info
));
486 struct page
*page
= virt_to_head_page(buf
);
487 unsigned int delta
= 0;
488 struct page
*xdp_page
;
492 xdp_prog
= rcu_dereference(rq
->xdp_prog
);
494 struct virtio_net_hdr_mrg_rxbuf
*hdr
= buf
+ header_offset
;
499 if (unlikely(hdr
->hdr
.gso_type
|| hdr
->hdr
.flags
))
502 if (unlikely(xdp_headroom
< virtnet_get_headroom(vi
))) {
503 int offset
= buf
- page_address(page
) + header_offset
;
504 unsigned int tlen
= len
+ vi
->hdr_len
;
507 xdp_headroom
= virtnet_get_headroom(vi
);
508 header_offset
= VIRTNET_RX_PAD
+ xdp_headroom
;
509 headroom
= vi
->hdr_len
+ header_offset
;
510 buflen
= SKB_DATA_ALIGN(GOOD_PACKET_LEN
+ headroom
) +
511 SKB_DATA_ALIGN(sizeof(struct skb_shared_info
));
512 xdp_page
= xdp_linearize_page(rq
, &num_buf
, page
,
513 offset
, header_offset
,
518 buf
= page_address(xdp_page
);
523 xdp
.data_hard_start
= buf
+ VIRTNET_RX_PAD
+ vi
->hdr_len
;
524 xdp
.data
= xdp
.data_hard_start
+ xdp_headroom
;
525 xdp
.data_end
= xdp
.data
+ len
;
526 orig_data
= xdp
.data
;
527 act
= bpf_prog_run_xdp(xdp_prog
, &xdp
);
531 /* Recalculate length in case bpf program changed it */
532 delta
= orig_data
- xdp
.data
;
535 if (unlikely(!virtnet_xdp_xmit(vi
, rq
, &xdp
)))
536 trace_xdp_exception(vi
->dev
, xdp_prog
, act
);
540 bpf_warn_invalid_xdp_action(act
);
542 trace_xdp_exception(vi
->dev
, xdp_prog
, act
);
549 skb
= build_skb(buf
, buflen
);
554 skb_reserve(skb
, headroom
- delta
);
555 skb_put(skb
, len
+ delta
);
557 buf
+= header_offset
;
558 memcpy(skb_vnet_hdr(skb
), buf
, vi
->hdr_len
);
559 } /* keep zeroed vnet hdr since packet was changed by bpf */
566 dev
->stats
.rx_dropped
++;
572 static struct sk_buff
*receive_big(struct net_device
*dev
,
573 struct virtnet_info
*vi
,
574 struct receive_queue
*rq
,
578 struct page
*page
= buf
;
579 struct sk_buff
*skb
= page_to_skb(vi
, rq
, page
, 0, len
, PAGE_SIZE
);
587 dev
->stats
.rx_dropped
++;
588 give_pages(rq
, page
);
592 static struct sk_buff
*receive_mergeable(struct net_device
*dev
,
593 struct virtnet_info
*vi
,
594 struct receive_queue
*rq
,
599 struct virtio_net_hdr_mrg_rxbuf
*hdr
= buf
;
600 u16 num_buf
= virtio16_to_cpu(vi
->vdev
, hdr
->num_buffers
);
601 struct page
*page
= virt_to_head_page(buf
);
602 int offset
= buf
- page_address(page
);
603 struct sk_buff
*head_skb
, *curr_skb
;
604 struct bpf_prog
*xdp_prog
;
605 unsigned int truesize
;
606 unsigned int headroom
= mergeable_ctx_to_headroom(ctx
);
611 xdp_prog
= rcu_dereference(rq
->xdp_prog
);
613 struct page
*xdp_page
;
618 /* This happens when rx buffer size is underestimated */
619 if (unlikely(num_buf
> 1 ||
620 headroom
< virtnet_get_headroom(vi
))) {
621 /* linearize data for XDP */
622 xdp_page
= xdp_linearize_page(rq
, &num_buf
,
628 offset
= VIRTIO_XDP_HEADROOM
;
633 /* Transient failure which in theory could occur if
634 * in-flight packets from before XDP was enabled reach
635 * the receive path after XDP is loaded. In practice I
636 * was not able to create this condition.
638 if (unlikely(hdr
->hdr
.gso_type
))
641 /* Allow consuming headroom but reserve enough space to push
642 * the descriptor on if we get an XDP_TX return code.
644 data
= page_address(xdp_page
) + offset
;
645 xdp
.data_hard_start
= data
- VIRTIO_XDP_HEADROOM
+ vi
->hdr_len
;
646 xdp
.data
= data
+ vi
->hdr_len
;
647 xdp
.data_end
= xdp
.data
+ (len
- vi
->hdr_len
);
648 act
= bpf_prog_run_xdp(xdp_prog
, &xdp
);
652 /* recalculate offset to account for any header
653 * adjustments. Note other cases do not build an
654 * skb and avoid using offset
657 page_address(xdp_page
) - vi
->hdr_len
;
659 /* We can only create skb based on xdp_page. */
660 if (unlikely(xdp_page
!= page
)) {
663 head_skb
= page_to_skb(vi
, rq
, xdp_page
,
664 offset
, len
, PAGE_SIZE
);
665 ewma_pkt_len_add(&rq
->mrg_avg_pkt_len
, len
);
670 if (unlikely(!virtnet_xdp_xmit(vi
, rq
, &xdp
)))
671 trace_xdp_exception(vi
->dev
, xdp_prog
, act
);
672 ewma_pkt_len_add(&rq
->mrg_avg_pkt_len
, len
);
673 if (unlikely(xdp_page
!= page
))
678 bpf_warn_invalid_xdp_action(act
);
680 trace_xdp_exception(vi
->dev
, xdp_prog
, act
);
682 if (unlikely(xdp_page
!= page
))
683 __free_pages(xdp_page
, 0);
684 ewma_pkt_len_add(&rq
->mrg_avg_pkt_len
, len
);
690 truesize
= mergeable_ctx_to_truesize(ctx
);
691 if (unlikely(len
> truesize
)) {
692 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
693 dev
->name
, len
, (unsigned long)ctx
);
694 dev
->stats
.rx_length_errors
++;
698 head_skb
= page_to_skb(vi
, rq
, page
, offset
, len
, truesize
);
701 if (unlikely(!curr_skb
))
706 buf
= virtqueue_get_buf_ctx(rq
->vq
, &len
, &ctx
);
707 if (unlikely(!ctx
)) {
708 pr_debug("%s: rx error: %d buffers out of %d missing\n",
710 virtio16_to_cpu(vi
->vdev
,
712 dev
->stats
.rx_length_errors
++;
716 page
= virt_to_head_page(buf
);
718 truesize
= mergeable_ctx_to_truesize(ctx
);
719 if (unlikely(len
> truesize
)) {
720 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
721 dev
->name
, len
, (unsigned long)ctx
);
722 dev
->stats
.rx_length_errors
++;
726 num_skb_frags
= skb_shinfo(curr_skb
)->nr_frags
;
727 if (unlikely(num_skb_frags
== MAX_SKB_FRAGS
)) {
728 struct sk_buff
*nskb
= alloc_skb(0, GFP_ATOMIC
);
732 if (curr_skb
== head_skb
)
733 skb_shinfo(curr_skb
)->frag_list
= nskb
;
735 curr_skb
->next
= nskb
;
737 head_skb
->truesize
+= nskb
->truesize
;
740 if (curr_skb
!= head_skb
) {
741 head_skb
->data_len
+= len
;
742 head_skb
->len
+= len
;
743 head_skb
->truesize
+= truesize
;
745 offset
= buf
- page_address(page
);
746 if (skb_can_coalesce(curr_skb
, num_skb_frags
, page
, offset
)) {
748 skb_coalesce_rx_frag(curr_skb
, num_skb_frags
- 1,
751 skb_add_rx_frag(curr_skb
, num_skb_frags
, page
,
752 offset
, len
, truesize
);
756 ewma_pkt_len_add(&rq
->mrg_avg_pkt_len
, head_skb
->len
);
764 buf
= virtqueue_get_buf(rq
->vq
, &len
);
765 if (unlikely(!buf
)) {
766 pr_debug("%s: rx error: %d buffers missing\n",
768 dev
->stats
.rx_length_errors
++;
771 page
= virt_to_head_page(buf
);
775 dev
->stats
.rx_dropped
++;
776 dev_kfree_skb(head_skb
);
781 static int receive_buf(struct virtnet_info
*vi
, struct receive_queue
*rq
,
782 void *buf
, unsigned int len
, void **ctx
)
784 struct net_device
*dev
= vi
->dev
;
786 struct virtio_net_hdr_mrg_rxbuf
*hdr
;
789 if (unlikely(len
< vi
->hdr_len
+ ETH_HLEN
)) {
790 pr_debug("%s: short packet %i\n", dev
->name
, len
);
791 dev
->stats
.rx_length_errors
++;
792 if (vi
->mergeable_rx_bufs
) {
793 put_page(virt_to_head_page(buf
));
794 } else if (vi
->big_packets
) {
797 put_page(virt_to_head_page(buf
));
802 if (vi
->mergeable_rx_bufs
)
803 skb
= receive_mergeable(dev
, vi
, rq
, buf
, ctx
, len
);
804 else if (vi
->big_packets
)
805 skb
= receive_big(dev
, vi
, rq
, buf
, len
);
807 skb
= receive_small(dev
, vi
, rq
, buf
, ctx
, len
);
812 hdr
= skb_vnet_hdr(skb
);
816 if (hdr
->hdr
.flags
& VIRTIO_NET_HDR_F_DATA_VALID
)
817 skb
->ip_summed
= CHECKSUM_UNNECESSARY
;
819 if (virtio_net_hdr_to_skb(skb
, &hdr
->hdr
,
820 virtio_is_little_endian(vi
->vdev
))) {
821 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
822 dev
->name
, hdr
->hdr
.gso_type
,
827 skb
->protocol
= eth_type_trans(skb
, dev
);
828 pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
829 ntohs(skb
->protocol
), skb
->len
, skb
->pkt_type
);
831 napi_gro_receive(&rq
->napi
, skb
);
835 dev
->stats
.rx_frame_errors
++;
840 /* Unlike mergeable buffers, all buffers are allocated to the
841 * same size, except for the headroom. For this reason we do
842 * not need to use mergeable_len_to_ctx here - it is enough
843 * to store the headroom as the context ignoring the truesize.
845 static int add_recvbuf_small(struct virtnet_info
*vi
, struct receive_queue
*rq
,
848 struct page_frag
*alloc_frag
= &rq
->alloc_frag
;
850 unsigned int xdp_headroom
= virtnet_get_headroom(vi
);
851 void *ctx
= (void *)(unsigned long)xdp_headroom
;
852 int len
= vi
->hdr_len
+ VIRTNET_RX_PAD
+ GOOD_PACKET_LEN
+ xdp_headroom
;
855 len
= SKB_DATA_ALIGN(len
) +
856 SKB_DATA_ALIGN(sizeof(struct skb_shared_info
));
857 if (unlikely(!skb_page_frag_refill(len
, alloc_frag
, gfp
)))
860 buf
= (char *)page_address(alloc_frag
->page
) + alloc_frag
->offset
;
861 get_page(alloc_frag
->page
);
862 alloc_frag
->offset
+= len
;
863 sg_init_one(rq
->sg
, buf
+ VIRTNET_RX_PAD
+ xdp_headroom
,
864 vi
->hdr_len
+ GOOD_PACKET_LEN
);
865 err
= virtqueue_add_inbuf_ctx(rq
->vq
, rq
->sg
, 1, buf
, ctx
, gfp
);
867 put_page(virt_to_head_page(buf
));
871 static int add_recvbuf_big(struct virtnet_info
*vi
, struct receive_queue
*rq
,
874 struct page
*first
, *list
= NULL
;
878 sg_init_table(rq
->sg
, MAX_SKB_FRAGS
+ 2);
880 /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
881 for (i
= MAX_SKB_FRAGS
+ 1; i
> 1; --i
) {
882 first
= get_a_page(rq
, gfp
);
885 give_pages(rq
, list
);
888 sg_set_buf(&rq
->sg
[i
], page_address(first
), PAGE_SIZE
);
890 /* chain new page in list head to match sg */
891 first
->private = (unsigned long)list
;
895 first
= get_a_page(rq
, gfp
);
897 give_pages(rq
, list
);
900 p
= page_address(first
);
902 /* rq->sg[0], rq->sg[1] share the same page */
903 /* a separated rq->sg[0] for header - required in case !any_header_sg */
904 sg_set_buf(&rq
->sg
[0], p
, vi
->hdr_len
);
906 /* rq->sg[1] for data packet, from offset */
907 offset
= sizeof(struct padded_vnet_hdr
);
908 sg_set_buf(&rq
->sg
[1], p
+ offset
, PAGE_SIZE
- offset
);
910 /* chain first in list head */
911 first
->private = (unsigned long)list
;
912 err
= virtqueue_add_inbuf(rq
->vq
, rq
->sg
, MAX_SKB_FRAGS
+ 2,
915 give_pages(rq
, first
);
920 static unsigned int get_mergeable_buf_len(struct receive_queue
*rq
,
921 struct ewma_pkt_len
*avg_pkt_len
)
923 const size_t hdr_len
= sizeof(struct virtio_net_hdr_mrg_rxbuf
);
926 len
= hdr_len
+ clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len
),
927 rq
->min_buf_len
, PAGE_SIZE
- hdr_len
);
928 return ALIGN(len
, L1_CACHE_BYTES
);
931 static int add_recvbuf_mergeable(struct virtnet_info
*vi
,
932 struct receive_queue
*rq
, gfp_t gfp
)
934 struct page_frag
*alloc_frag
= &rq
->alloc_frag
;
935 unsigned int headroom
= virtnet_get_headroom(vi
);
939 unsigned int len
, hole
;
941 len
= get_mergeable_buf_len(rq
, &rq
->mrg_avg_pkt_len
);
942 if (unlikely(!skb_page_frag_refill(len
+ headroom
, alloc_frag
, gfp
)))
945 buf
= (char *)page_address(alloc_frag
->page
) + alloc_frag
->offset
;
946 buf
+= headroom
; /* advance address leaving hole at front of pkt */
947 ctx
= mergeable_len_to_ctx(len
, headroom
);
948 get_page(alloc_frag
->page
);
949 alloc_frag
->offset
+= len
+ headroom
;
950 hole
= alloc_frag
->size
- alloc_frag
->offset
;
951 if (hole
< len
+ headroom
) {
952 /* To avoid internal fragmentation, if there is very likely not
953 * enough space for another buffer, add the remaining space to
954 * the current buffer. This extra space is not included in
955 * the truesize stored in ctx.
958 alloc_frag
->offset
+= hole
;
961 sg_init_one(rq
->sg
, buf
, len
);
962 err
= virtqueue_add_inbuf_ctx(rq
->vq
, rq
->sg
, 1, buf
, ctx
, gfp
);
964 put_page(virt_to_head_page(buf
));
970 * Returns false if we couldn't fill entirely (OOM).
972 * Normally run in the receive path, but can also be run from ndo_open
973 * before we're receiving packets, or from refill_work which is
974 * careful to disable receiving (using napi_disable).
976 static bool try_fill_recv(struct virtnet_info
*vi
, struct receive_queue
*rq
,
984 if (vi
->mergeable_rx_bufs
)
985 err
= add_recvbuf_mergeable(vi
, rq
, gfp
);
986 else if (vi
->big_packets
)
987 err
= add_recvbuf_big(vi
, rq
, gfp
);
989 err
= add_recvbuf_small(vi
, rq
, gfp
);
991 oom
= err
== -ENOMEM
;
994 } while (rq
->vq
->num_free
);
995 virtqueue_kick(rq
->vq
);
999 static void skb_recv_done(struct virtqueue
*rvq
)
1001 struct virtnet_info
*vi
= rvq
->vdev
->priv
;
1002 struct receive_queue
*rq
= &vi
->rq
[vq2rxq(rvq
)];
1004 virtqueue_napi_schedule(&rq
->napi
, rvq
);
1007 static void virtnet_napi_enable(struct virtqueue
*vq
, struct napi_struct
*napi
)
1011 /* If all buffers were filled by other side before we napi_enabled, we
1012 * won't get another interrupt, so process any outstanding packets now.
1013 * Call local_bh_enable after to trigger softIRQ processing.
1016 virtqueue_napi_schedule(napi
, vq
);
1020 static void virtnet_napi_tx_enable(struct virtnet_info
*vi
,
1021 struct virtqueue
*vq
,
1022 struct napi_struct
*napi
)
1027 /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
1028 * enable the feature if this is likely affine with the transmit path.
1030 if (!vi
->affinity_hint_set
) {
1035 return virtnet_napi_enable(vq
, napi
);
1038 static void virtnet_napi_tx_disable(struct napi_struct
*napi
)
1044 static void refill_work(struct work_struct
*work
)
1046 struct virtnet_info
*vi
=
1047 container_of(work
, struct virtnet_info
, refill
.work
);
1051 for (i
= 0; i
< vi
->curr_queue_pairs
; i
++) {
1052 struct receive_queue
*rq
= &vi
->rq
[i
];
1054 napi_disable(&rq
->napi
);
1055 still_empty
= !try_fill_recv(vi
, rq
, GFP_KERNEL
);
1056 virtnet_napi_enable(rq
->vq
, &rq
->napi
);
1058 /* In theory, this can happen: if we don't get any buffers in
1059 * we will *never* try to fill again.
1062 schedule_delayed_work(&vi
->refill
, HZ
/2);
1066 static int virtnet_receive(struct receive_queue
*rq
, int budget
)
1068 struct virtnet_info
*vi
= rq
->vq
->vdev
->priv
;
1069 unsigned int len
, received
= 0, bytes
= 0;
1071 struct virtnet_stats
*stats
= this_cpu_ptr(vi
->stats
);
1073 if (!vi
->big_packets
|| vi
->mergeable_rx_bufs
) {
1076 while (received
< budget
&&
1077 (buf
= virtqueue_get_buf_ctx(rq
->vq
, &len
, &ctx
))) {
1078 bytes
+= receive_buf(vi
, rq
, buf
, len
, ctx
);
1082 while (received
< budget
&&
1083 (buf
= virtqueue_get_buf(rq
->vq
, &len
)) != NULL
) {
1084 bytes
+= receive_buf(vi
, rq
, buf
, len
, NULL
);
1089 if (rq
->vq
->num_free
> virtqueue_get_vring_size(rq
->vq
) / 2) {
1090 if (!try_fill_recv(vi
, rq
, GFP_ATOMIC
))
1091 schedule_delayed_work(&vi
->refill
, 0);
1094 u64_stats_update_begin(&stats
->rx_syncp
);
1095 stats
->rx_bytes
+= bytes
;
1096 stats
->rx_packets
+= received
;
1097 u64_stats_update_end(&stats
->rx_syncp
);
1102 static void free_old_xmit_skbs(struct send_queue
*sq
)
1104 struct sk_buff
*skb
;
1106 struct virtnet_info
*vi
= sq
->vq
->vdev
->priv
;
1107 struct virtnet_stats
*stats
= this_cpu_ptr(vi
->stats
);
1108 unsigned int packets
= 0;
1109 unsigned int bytes
= 0;
1111 while ((skb
= virtqueue_get_buf(sq
->vq
, &len
)) != NULL
) {
1112 pr_debug("Sent skb %p\n", skb
);
1117 dev_kfree_skb_any(skb
);
1120 /* Avoid overhead when no packets have been processed
1121 * happens when called speculatively from start_xmit.
1126 u64_stats_update_begin(&stats
->tx_syncp
);
1127 stats
->tx_bytes
+= bytes
;
1128 stats
->tx_packets
+= packets
;
1129 u64_stats_update_end(&stats
->tx_syncp
);
1132 static void virtnet_poll_cleantx(struct receive_queue
*rq
)
1134 struct virtnet_info
*vi
= rq
->vq
->vdev
->priv
;
1135 unsigned int index
= vq2rxq(rq
->vq
);
1136 struct send_queue
*sq
= &vi
->sq
[index
];
1137 struct netdev_queue
*txq
= netdev_get_tx_queue(vi
->dev
, index
);
1139 if (!sq
->napi
.weight
)
1142 if (__netif_tx_trylock(txq
)) {
1143 free_old_xmit_skbs(sq
);
1144 __netif_tx_unlock(txq
);
1147 if (sq
->vq
->num_free
>= 2 + MAX_SKB_FRAGS
)
1148 netif_tx_wake_queue(txq
);
1151 static int virtnet_poll(struct napi_struct
*napi
, int budget
)
1153 struct receive_queue
*rq
=
1154 container_of(napi
, struct receive_queue
, napi
);
1155 unsigned int received
;
1157 virtnet_poll_cleantx(rq
);
1159 received
= virtnet_receive(rq
, budget
);
1161 /* Out of packets? */
1162 if (received
< budget
)
1163 virtqueue_napi_complete(napi
, rq
->vq
, received
);
1168 static int virtnet_open(struct net_device
*dev
)
1170 struct virtnet_info
*vi
= netdev_priv(dev
);
1173 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1174 if (i
< vi
->curr_queue_pairs
)
1175 /* Make sure we have some buffers: if oom use wq. */
1176 if (!try_fill_recv(vi
, &vi
->rq
[i
], GFP_KERNEL
))
1177 schedule_delayed_work(&vi
->refill
, 0);
1178 virtnet_napi_enable(vi
->rq
[i
].vq
, &vi
->rq
[i
].napi
);
1179 virtnet_napi_tx_enable(vi
, vi
->sq
[i
].vq
, &vi
->sq
[i
].napi
);
1185 static int virtnet_poll_tx(struct napi_struct
*napi
, int budget
)
1187 struct send_queue
*sq
= container_of(napi
, struct send_queue
, napi
);
1188 struct virtnet_info
*vi
= sq
->vq
->vdev
->priv
;
1189 struct netdev_queue
*txq
= netdev_get_tx_queue(vi
->dev
, vq2txq(sq
->vq
));
1191 __netif_tx_lock(txq
, raw_smp_processor_id());
1192 free_old_xmit_skbs(sq
);
1193 __netif_tx_unlock(txq
);
1195 virtqueue_napi_complete(napi
, sq
->vq
, 0);
1197 if (sq
->vq
->num_free
>= 2 + MAX_SKB_FRAGS
)
1198 netif_tx_wake_queue(txq
);
1203 static int xmit_skb(struct send_queue
*sq
, struct sk_buff
*skb
)
1205 struct virtio_net_hdr_mrg_rxbuf
*hdr
;
1206 const unsigned char *dest
= ((struct ethhdr
*)skb
->data
)->h_dest
;
1207 struct virtnet_info
*vi
= sq
->vq
->vdev
->priv
;
1209 unsigned hdr_len
= vi
->hdr_len
;
1212 pr_debug("%s: xmit %p %pM\n", vi
->dev
->name
, skb
, dest
);
1214 can_push
= vi
->any_header_sg
&&
1215 !((unsigned long)skb
->data
& (__alignof__(*hdr
) - 1)) &&
1216 !skb_header_cloned(skb
) && skb_headroom(skb
) >= hdr_len
;
1217 /* Even if we can, don't push here yet as this would skew
1218 * csum_start offset below. */
1220 hdr
= (struct virtio_net_hdr_mrg_rxbuf
*)(skb
->data
- hdr_len
);
1222 hdr
= skb_vnet_hdr(skb
);
1224 if (virtio_net_hdr_from_skb(skb
, &hdr
->hdr
,
1225 virtio_is_little_endian(vi
->vdev
), false))
1228 if (vi
->mergeable_rx_bufs
)
1229 hdr
->num_buffers
= 0;
1231 sg_init_table(sq
->sg
, skb_shinfo(skb
)->nr_frags
+ (can_push
? 1 : 2));
1233 __skb_push(skb
, hdr_len
);
1234 num_sg
= skb_to_sgvec(skb
, sq
->sg
, 0, skb
->len
);
1235 if (unlikely(num_sg
< 0))
1237 /* Pull header back to avoid skew in tx bytes calculations. */
1238 __skb_pull(skb
, hdr_len
);
1240 sg_set_buf(sq
->sg
, hdr
, hdr_len
);
1241 num_sg
= skb_to_sgvec(skb
, sq
->sg
+ 1, 0, skb
->len
);
1242 if (unlikely(num_sg
< 0))
1246 return virtqueue_add_outbuf(sq
->vq
, sq
->sg
, num_sg
, skb
, GFP_ATOMIC
);
1249 static netdev_tx_t
start_xmit(struct sk_buff
*skb
, struct net_device
*dev
)
1251 struct virtnet_info
*vi
= netdev_priv(dev
);
1252 int qnum
= skb_get_queue_mapping(skb
);
1253 struct send_queue
*sq
= &vi
->sq
[qnum
];
1255 struct netdev_queue
*txq
= netdev_get_tx_queue(dev
, qnum
);
1256 bool kick
= !skb
->xmit_more
;
1257 bool use_napi
= sq
->napi
.weight
;
1259 /* Free up any pending old buffers before queueing new ones. */
1260 free_old_xmit_skbs(sq
);
1262 if (use_napi
&& kick
)
1263 virtqueue_enable_cb_delayed(sq
->vq
);
1265 /* timestamp packet in software */
1266 skb_tx_timestamp(skb
);
1268 /* Try to transmit */
1269 err
= xmit_skb(sq
, skb
);
1271 /* This should not happen! */
1272 if (unlikely(err
)) {
1273 dev
->stats
.tx_fifo_errors
++;
1274 if (net_ratelimit())
1276 "Unexpected TXQ (%d) queue failure: %d\n", qnum
, err
);
1277 dev
->stats
.tx_dropped
++;
1278 dev_kfree_skb_any(skb
);
1279 return NETDEV_TX_OK
;
1282 /* Don't wait up for transmitted skbs to be freed. */
1288 /* If running out of space, stop queue to avoid getting packets that we
1289 * are then unable to transmit.
1290 * An alternative would be to force queuing layer to requeue the skb by
1291 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1292 * returned in a normal path of operation: it means that driver is not
1293 * maintaining the TX queue stop/start state properly, and causes
1294 * the stack to do a non-trivial amount of useless work.
1295 * Since most packets only take 1 or 2 ring slots, stopping the queue
1296 * early means 16 slots are typically wasted.
1298 if (sq
->vq
->num_free
< 2+MAX_SKB_FRAGS
) {
1299 netif_stop_subqueue(dev
, qnum
);
1301 unlikely(!virtqueue_enable_cb_delayed(sq
->vq
))) {
1302 /* More just got used, free them then recheck. */
1303 free_old_xmit_skbs(sq
);
1304 if (sq
->vq
->num_free
>= 2+MAX_SKB_FRAGS
) {
1305 netif_start_subqueue(dev
, qnum
);
1306 virtqueue_disable_cb(sq
->vq
);
1311 if (kick
|| netif_xmit_stopped(txq
))
1312 virtqueue_kick(sq
->vq
);
1314 return NETDEV_TX_OK
;
1318 * Send command via the control virtqueue and check status. Commands
1319 * supported by the hypervisor, as indicated by feature bits, should
1320 * never fail unless improperly formatted.
1322 static bool virtnet_send_command(struct virtnet_info
*vi
, u8
class, u8 cmd
,
1323 struct scatterlist
*out
)
1325 struct scatterlist
*sgs
[4], hdr
, stat
;
1326 unsigned out_num
= 0, tmp
;
1328 /* Caller should know better */
1329 BUG_ON(!virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_CTRL_VQ
));
1331 vi
->ctrl_status
= ~0;
1332 vi
->ctrl_hdr
.class = class;
1333 vi
->ctrl_hdr
.cmd
= cmd
;
1335 sg_init_one(&hdr
, &vi
->ctrl_hdr
, sizeof(vi
->ctrl_hdr
));
1336 sgs
[out_num
++] = &hdr
;
1339 sgs
[out_num
++] = out
;
1341 /* Add return status. */
1342 sg_init_one(&stat
, &vi
->ctrl_status
, sizeof(vi
->ctrl_status
));
1343 sgs
[out_num
] = &stat
;
1345 BUG_ON(out_num
+ 1 > ARRAY_SIZE(sgs
));
1346 virtqueue_add_sgs(vi
->cvq
, sgs
, out_num
, 1, vi
, GFP_ATOMIC
);
1348 if (unlikely(!virtqueue_kick(vi
->cvq
)))
1349 return vi
->ctrl_status
== VIRTIO_NET_OK
;
1351 /* Spin for a response, the kick causes an ioport write, trapping
1352 * into the hypervisor, so the request should be handled immediately.
1354 while (!virtqueue_get_buf(vi
->cvq
, &tmp
) &&
1355 !virtqueue_is_broken(vi
->cvq
))
1358 return vi
->ctrl_status
== VIRTIO_NET_OK
;
1361 static int virtnet_set_mac_address(struct net_device
*dev
, void *p
)
1363 struct virtnet_info
*vi
= netdev_priv(dev
);
1364 struct virtio_device
*vdev
= vi
->vdev
;
1366 struct sockaddr
*addr
;
1367 struct scatterlist sg
;
1369 addr
= kmemdup(p
, sizeof(*addr
), GFP_KERNEL
);
1373 ret
= eth_prepare_mac_addr_change(dev
, addr
);
1377 if (virtio_has_feature(vdev
, VIRTIO_NET_F_CTRL_MAC_ADDR
)) {
1378 sg_init_one(&sg
, addr
->sa_data
, dev
->addr_len
);
1379 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_MAC
,
1380 VIRTIO_NET_CTRL_MAC_ADDR_SET
, &sg
)) {
1381 dev_warn(&vdev
->dev
,
1382 "Failed to set mac address by vq command.\n");
1386 } else if (virtio_has_feature(vdev
, VIRTIO_NET_F_MAC
) &&
1387 !virtio_has_feature(vdev
, VIRTIO_F_VERSION_1
)) {
1390 /* Naturally, this has an atomicity problem. */
1391 for (i
= 0; i
< dev
->addr_len
; i
++)
1392 virtio_cwrite8(vdev
,
1393 offsetof(struct virtio_net_config
, mac
) +
1394 i
, addr
->sa_data
[i
]);
1397 eth_commit_mac_addr_change(dev
, p
);
1405 static void virtnet_stats(struct net_device
*dev
,
1406 struct rtnl_link_stats64
*tot
)
1408 struct virtnet_info
*vi
= netdev_priv(dev
);
1412 for_each_possible_cpu(cpu
) {
1413 struct virtnet_stats
*stats
= per_cpu_ptr(vi
->stats
, cpu
);
1414 u64 tpackets
, tbytes
, rpackets
, rbytes
;
1417 start
= u64_stats_fetch_begin_irq(&stats
->tx_syncp
);
1418 tpackets
= stats
->tx_packets
;
1419 tbytes
= stats
->tx_bytes
;
1420 } while (u64_stats_fetch_retry_irq(&stats
->tx_syncp
, start
));
1423 start
= u64_stats_fetch_begin_irq(&stats
->rx_syncp
);
1424 rpackets
= stats
->rx_packets
;
1425 rbytes
= stats
->rx_bytes
;
1426 } while (u64_stats_fetch_retry_irq(&stats
->rx_syncp
, start
));
1428 tot
->rx_packets
+= rpackets
;
1429 tot
->tx_packets
+= tpackets
;
1430 tot
->rx_bytes
+= rbytes
;
1431 tot
->tx_bytes
+= tbytes
;
1434 tot
->tx_dropped
= dev
->stats
.tx_dropped
;
1435 tot
->tx_fifo_errors
= dev
->stats
.tx_fifo_errors
;
1436 tot
->rx_dropped
= dev
->stats
.rx_dropped
;
1437 tot
->rx_length_errors
= dev
->stats
.rx_length_errors
;
1438 tot
->rx_frame_errors
= dev
->stats
.rx_frame_errors
;
1441 #ifdef CONFIG_NET_POLL_CONTROLLER
1442 static void virtnet_netpoll(struct net_device
*dev
)
1444 struct virtnet_info
*vi
= netdev_priv(dev
);
1447 for (i
= 0; i
< vi
->curr_queue_pairs
; i
++)
1448 napi_schedule(&vi
->rq
[i
].napi
);
1452 static void virtnet_ack_link_announce(struct virtnet_info
*vi
)
1455 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_ANNOUNCE
,
1456 VIRTIO_NET_CTRL_ANNOUNCE_ACK
, NULL
))
1457 dev_warn(&vi
->dev
->dev
, "Failed to ack link announce.\n");
1461 static int _virtnet_set_queues(struct virtnet_info
*vi
, u16 queue_pairs
)
1463 struct scatterlist sg
;
1464 struct net_device
*dev
= vi
->dev
;
1466 if (!vi
->has_cvq
|| !virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_MQ
))
1469 vi
->ctrl_mq
.virtqueue_pairs
= cpu_to_virtio16(vi
->vdev
, queue_pairs
);
1470 sg_init_one(&sg
, &vi
->ctrl_mq
, sizeof(vi
->ctrl_mq
));
1472 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_MQ
,
1473 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET
, &sg
)) {
1474 dev_warn(&dev
->dev
, "Fail to set num of queue pairs to %d\n",
1478 vi
->curr_queue_pairs
= queue_pairs
;
1479 /* virtnet_open() will refill when device is going to up. */
1480 if (dev
->flags
& IFF_UP
)
1481 schedule_delayed_work(&vi
->refill
, 0);
1487 static int virtnet_set_queues(struct virtnet_info
*vi
, u16 queue_pairs
)
1492 err
= _virtnet_set_queues(vi
, queue_pairs
);
1497 static int virtnet_close(struct net_device
*dev
)
1499 struct virtnet_info
*vi
= netdev_priv(dev
);
1502 /* Make sure refill_work doesn't re-enable napi! */
1503 cancel_delayed_work_sync(&vi
->refill
);
1505 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1506 napi_disable(&vi
->rq
[i
].napi
);
1507 virtnet_napi_tx_disable(&vi
->sq
[i
].napi
);
1513 static void virtnet_set_rx_mode(struct net_device
*dev
)
1515 struct virtnet_info
*vi
= netdev_priv(dev
);
1516 struct scatterlist sg
[2];
1517 struct virtio_net_ctrl_mac
*mac_data
;
1518 struct netdev_hw_addr
*ha
;
1524 /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1525 if (!virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_CTRL_RX
))
1528 vi
->ctrl_promisc
= ((dev
->flags
& IFF_PROMISC
) != 0);
1529 vi
->ctrl_allmulti
= ((dev
->flags
& IFF_ALLMULTI
) != 0);
1531 sg_init_one(sg
, &vi
->ctrl_promisc
, sizeof(vi
->ctrl_promisc
));
1533 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_RX
,
1534 VIRTIO_NET_CTRL_RX_PROMISC
, sg
))
1535 dev_warn(&dev
->dev
, "Failed to %sable promisc mode.\n",
1536 vi
->ctrl_promisc
? "en" : "dis");
1538 sg_init_one(sg
, &vi
->ctrl_allmulti
, sizeof(vi
->ctrl_allmulti
));
1540 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_RX
,
1541 VIRTIO_NET_CTRL_RX_ALLMULTI
, sg
))
1542 dev_warn(&dev
->dev
, "Failed to %sable allmulti mode.\n",
1543 vi
->ctrl_allmulti
? "en" : "dis");
1545 uc_count
= netdev_uc_count(dev
);
1546 mc_count
= netdev_mc_count(dev
);
1547 /* MAC filter - use one buffer for both lists */
1548 buf
= kzalloc(((uc_count
+ mc_count
) * ETH_ALEN
) +
1549 (2 * sizeof(mac_data
->entries
)), GFP_ATOMIC
);
1554 sg_init_table(sg
, 2);
1556 /* Store the unicast list and count in the front of the buffer */
1557 mac_data
->entries
= cpu_to_virtio32(vi
->vdev
, uc_count
);
1559 netdev_for_each_uc_addr(ha
, dev
)
1560 memcpy(&mac_data
->macs
[i
++][0], ha
->addr
, ETH_ALEN
);
1562 sg_set_buf(&sg
[0], mac_data
,
1563 sizeof(mac_data
->entries
) + (uc_count
* ETH_ALEN
));
1565 /* multicast list and count fill the end */
1566 mac_data
= (void *)&mac_data
->macs
[uc_count
][0];
1568 mac_data
->entries
= cpu_to_virtio32(vi
->vdev
, mc_count
);
1570 netdev_for_each_mc_addr(ha
, dev
)
1571 memcpy(&mac_data
->macs
[i
++][0], ha
->addr
, ETH_ALEN
);
1573 sg_set_buf(&sg
[1], mac_data
,
1574 sizeof(mac_data
->entries
) + (mc_count
* ETH_ALEN
));
1576 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_MAC
,
1577 VIRTIO_NET_CTRL_MAC_TABLE_SET
, sg
))
1578 dev_warn(&dev
->dev
, "Failed to set MAC filter table.\n");
1583 static int virtnet_vlan_rx_add_vid(struct net_device
*dev
,
1584 __be16 proto
, u16 vid
)
1586 struct virtnet_info
*vi
= netdev_priv(dev
);
1587 struct scatterlist sg
;
1590 sg_init_one(&sg
, &vi
->ctrl_vid
, sizeof(vi
->ctrl_vid
));
1592 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_VLAN
,
1593 VIRTIO_NET_CTRL_VLAN_ADD
, &sg
))
1594 dev_warn(&dev
->dev
, "Failed to add VLAN ID %d.\n", vid
);
1598 static int virtnet_vlan_rx_kill_vid(struct net_device
*dev
,
1599 __be16 proto
, u16 vid
)
1601 struct virtnet_info
*vi
= netdev_priv(dev
);
1602 struct scatterlist sg
;
1605 sg_init_one(&sg
, &vi
->ctrl_vid
, sizeof(vi
->ctrl_vid
));
1607 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_VLAN
,
1608 VIRTIO_NET_CTRL_VLAN_DEL
, &sg
))
1609 dev_warn(&dev
->dev
, "Failed to kill VLAN ID %d.\n", vid
);
1613 static void virtnet_clean_affinity(struct virtnet_info
*vi
, long hcpu
)
1617 if (vi
->affinity_hint_set
) {
1618 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1619 virtqueue_set_affinity(vi
->rq
[i
].vq
, -1);
1620 virtqueue_set_affinity(vi
->sq
[i
].vq
, -1);
1623 vi
->affinity_hint_set
= false;
1627 static void virtnet_set_affinity(struct virtnet_info
*vi
)
1632 /* In multiqueue mode, when the number of cpu is equal to the number of
1633 * queue pairs, we let the queue pairs to be private to one cpu by
1634 * setting the affinity hint to eliminate the contention.
1636 if (vi
->curr_queue_pairs
== 1 ||
1637 vi
->max_queue_pairs
!= num_online_cpus()) {
1638 virtnet_clean_affinity(vi
, -1);
1643 for_each_online_cpu(cpu
) {
1644 virtqueue_set_affinity(vi
->rq
[i
].vq
, cpu
);
1645 virtqueue_set_affinity(vi
->sq
[i
].vq
, cpu
);
1646 netif_set_xps_queue(vi
->dev
, cpumask_of(cpu
), i
);
1650 vi
->affinity_hint_set
= true;
1653 static int virtnet_cpu_online(unsigned int cpu
, struct hlist_node
*node
)
1655 struct virtnet_info
*vi
= hlist_entry_safe(node
, struct virtnet_info
,
1657 virtnet_set_affinity(vi
);
1661 static int virtnet_cpu_dead(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_down_prep(unsigned int cpu
, struct hlist_node
*node
)
1671 struct virtnet_info
*vi
= hlist_entry_safe(node
, struct virtnet_info
,
1674 virtnet_clean_affinity(vi
, cpu
);
1678 static enum cpuhp_state virtionet_online
;
1680 static int virtnet_cpu_notif_add(struct virtnet_info
*vi
)
1684 ret
= cpuhp_state_add_instance_nocalls(virtionet_online
, &vi
->node
);
1687 ret
= cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD
,
1691 cpuhp_state_remove_instance_nocalls(virtionet_online
, &vi
->node
);
1695 static void virtnet_cpu_notif_remove(struct virtnet_info
*vi
)
1697 cpuhp_state_remove_instance_nocalls(virtionet_online
, &vi
->node
);
1698 cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD
,
1702 static void virtnet_get_ringparam(struct net_device
*dev
,
1703 struct ethtool_ringparam
*ring
)
1705 struct virtnet_info
*vi
= netdev_priv(dev
);
1707 ring
->rx_max_pending
= virtqueue_get_vring_size(vi
->rq
[0].vq
);
1708 ring
->tx_max_pending
= virtqueue_get_vring_size(vi
->sq
[0].vq
);
1709 ring
->rx_pending
= ring
->rx_max_pending
;
1710 ring
->tx_pending
= ring
->tx_max_pending
;
1714 static void virtnet_get_drvinfo(struct net_device
*dev
,
1715 struct ethtool_drvinfo
*info
)
1717 struct virtnet_info
*vi
= netdev_priv(dev
);
1718 struct virtio_device
*vdev
= vi
->vdev
;
1720 strlcpy(info
->driver
, KBUILD_MODNAME
, sizeof(info
->driver
));
1721 strlcpy(info
->version
, VIRTNET_DRIVER_VERSION
, sizeof(info
->version
));
1722 strlcpy(info
->bus_info
, virtio_bus_name(vdev
), sizeof(info
->bus_info
));
1726 /* TODO: Eliminate OOO packets during switching */
1727 static int virtnet_set_channels(struct net_device
*dev
,
1728 struct ethtool_channels
*channels
)
1730 struct virtnet_info
*vi
= netdev_priv(dev
);
1731 u16 queue_pairs
= channels
->combined_count
;
1734 /* We don't support separate rx/tx channels.
1735 * We don't allow setting 'other' channels.
1737 if (channels
->rx_count
|| channels
->tx_count
|| channels
->other_count
)
1740 if (queue_pairs
> vi
->max_queue_pairs
|| queue_pairs
== 0)
1743 /* For now we don't support modifying channels while XDP is loaded
1744 * also when XDP is loaded all RX queues have XDP programs so we only
1745 * need to check a single RX queue.
1747 if (vi
->rq
[0].xdp_prog
)
1751 err
= _virtnet_set_queues(vi
, queue_pairs
);
1753 netif_set_real_num_tx_queues(dev
, queue_pairs
);
1754 netif_set_real_num_rx_queues(dev
, queue_pairs
);
1756 virtnet_set_affinity(vi
);
1763 static void virtnet_get_channels(struct net_device
*dev
,
1764 struct ethtool_channels
*channels
)
1766 struct virtnet_info
*vi
= netdev_priv(dev
);
1768 channels
->combined_count
= vi
->curr_queue_pairs
;
1769 channels
->max_combined
= vi
->max_queue_pairs
;
1770 channels
->max_other
= 0;
1771 channels
->rx_count
= 0;
1772 channels
->tx_count
= 0;
1773 channels
->other_count
= 0;
1776 /* Check if the user is trying to change anything besides speed/duplex */
1778 virtnet_validate_ethtool_cmd(const struct ethtool_link_ksettings
*cmd
)
1780 struct ethtool_link_ksettings diff1
= *cmd
;
1781 struct ethtool_link_ksettings diff2
= {};
1783 /* cmd is always set so we need to clear it, validate the port type
1784 * and also without autonegotiation we can ignore advertising
1786 diff1
.base
.speed
= 0;
1787 diff2
.base
.port
= PORT_OTHER
;
1788 ethtool_link_ksettings_zero_link_mode(&diff1
, advertising
);
1789 diff1
.base
.duplex
= 0;
1791 diff1
.base
.link_mode_masks_nwords
= 0;
1793 return !memcmp(&diff1
.base
, &diff2
.base
, sizeof(diff1
.base
)) &&
1794 bitmap_empty(diff1
.link_modes
.supported
,
1795 __ETHTOOL_LINK_MODE_MASK_NBITS
) &&
1796 bitmap_empty(diff1
.link_modes
.advertising
,
1797 __ETHTOOL_LINK_MODE_MASK_NBITS
) &&
1798 bitmap_empty(diff1
.link_modes
.lp_advertising
,
1799 __ETHTOOL_LINK_MODE_MASK_NBITS
);
1802 static int virtnet_set_link_ksettings(struct net_device
*dev
,
1803 const struct ethtool_link_ksettings
*cmd
)
1805 struct virtnet_info
*vi
= netdev_priv(dev
);
1808 speed
= cmd
->base
.speed
;
1809 /* don't allow custom speed and duplex */
1810 if (!ethtool_validate_speed(speed
) ||
1811 !ethtool_validate_duplex(cmd
->base
.duplex
) ||
1812 !virtnet_validate_ethtool_cmd(cmd
))
1815 vi
->duplex
= cmd
->base
.duplex
;
1820 static int virtnet_get_link_ksettings(struct net_device
*dev
,
1821 struct ethtool_link_ksettings
*cmd
)
1823 struct virtnet_info
*vi
= netdev_priv(dev
);
1825 cmd
->base
.speed
= vi
->speed
;
1826 cmd
->base
.duplex
= vi
->duplex
;
1827 cmd
->base
.port
= PORT_OTHER
;
1832 static void virtnet_init_settings(struct net_device
*dev
)
1834 struct virtnet_info
*vi
= netdev_priv(dev
);
1836 vi
->speed
= SPEED_UNKNOWN
;
1837 vi
->duplex
= DUPLEX_UNKNOWN
;
1840 static const struct ethtool_ops virtnet_ethtool_ops
= {
1841 .get_drvinfo
= virtnet_get_drvinfo
,
1842 .get_link
= ethtool_op_get_link
,
1843 .get_ringparam
= virtnet_get_ringparam
,
1844 .set_channels
= virtnet_set_channels
,
1845 .get_channels
= virtnet_get_channels
,
1846 .get_ts_info
= ethtool_op_get_ts_info
,
1847 .get_link_ksettings
= virtnet_get_link_ksettings
,
1848 .set_link_ksettings
= virtnet_set_link_ksettings
,
1851 static void virtnet_freeze_down(struct virtio_device
*vdev
)
1853 struct virtnet_info
*vi
= vdev
->priv
;
1856 /* Make sure no work handler is accessing the device */
1857 flush_work(&vi
->config_work
);
1859 netif_device_detach(vi
->dev
);
1860 netif_tx_disable(vi
->dev
);
1861 cancel_delayed_work_sync(&vi
->refill
);
1863 if (netif_running(vi
->dev
)) {
1864 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1865 napi_disable(&vi
->rq
[i
].napi
);
1866 virtnet_napi_tx_disable(&vi
->sq
[i
].napi
);
1871 static int init_vqs(struct virtnet_info
*vi
);
1873 static int virtnet_restore_up(struct virtio_device
*vdev
)
1875 struct virtnet_info
*vi
= vdev
->priv
;
1882 virtio_device_ready(vdev
);
1884 if (netif_running(vi
->dev
)) {
1885 for (i
= 0; i
< vi
->curr_queue_pairs
; i
++)
1886 if (!try_fill_recv(vi
, &vi
->rq
[i
], GFP_KERNEL
))
1887 schedule_delayed_work(&vi
->refill
, 0);
1889 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1890 virtnet_napi_enable(vi
->rq
[i
].vq
, &vi
->rq
[i
].napi
);
1891 virtnet_napi_tx_enable(vi
, vi
->sq
[i
].vq
,
1896 netif_device_attach(vi
->dev
);
1900 static int virtnet_xdp_set(struct net_device
*dev
, struct bpf_prog
*prog
,
1901 struct netlink_ext_ack
*extack
)
1903 unsigned long int max_sz
= PAGE_SIZE
- sizeof(struct padded_vnet_hdr
);
1904 struct virtnet_info
*vi
= netdev_priv(dev
);
1905 struct bpf_prog
*old_prog
;
1906 u16 xdp_qp
= 0, curr_qp
;
1909 if (virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_TSO4
) ||
1910 virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_TSO6
) ||
1911 virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_ECN
) ||
1912 virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_UFO
)) {
1913 NL_SET_ERR_MSG_MOD(extack
, "Can't set XDP while host is implementing LRO, disable LRO first");
1917 if (vi
->mergeable_rx_bufs
&& !vi
->any_header_sg
) {
1918 NL_SET_ERR_MSG_MOD(extack
, "XDP expects header/data in single page, any_header_sg required");
1922 if (dev
->mtu
> max_sz
) {
1923 NL_SET_ERR_MSG_MOD(extack
, "MTU too large to enable XDP");
1924 netdev_warn(dev
, "XDP requires MTU less than %lu\n", max_sz
);
1928 curr_qp
= vi
->curr_queue_pairs
- vi
->xdp_queue_pairs
;
1930 xdp_qp
= nr_cpu_ids
;
1932 /* XDP requires extra queues for XDP_TX */
1933 if (curr_qp
+ xdp_qp
> vi
->max_queue_pairs
) {
1934 NL_SET_ERR_MSG_MOD(extack
, "Too few free TX rings available");
1935 netdev_warn(dev
, "request %i queues but max is %i\n",
1936 curr_qp
+ xdp_qp
, vi
->max_queue_pairs
);
1941 prog
= bpf_prog_add(prog
, vi
->max_queue_pairs
- 1);
1943 return PTR_ERR(prog
);
1946 /* Make sure NAPI is not using any XDP TX queues for RX. */
1947 for (i
= 0; i
< vi
->max_queue_pairs
; i
++)
1948 napi_disable(&vi
->rq
[i
].napi
);
1950 netif_set_real_num_rx_queues(dev
, curr_qp
+ xdp_qp
);
1951 err
= _virtnet_set_queues(vi
, curr_qp
+ xdp_qp
);
1954 vi
->xdp_queue_pairs
= xdp_qp
;
1956 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1957 old_prog
= rtnl_dereference(vi
->rq
[i
].xdp_prog
);
1958 rcu_assign_pointer(vi
->rq
[i
].xdp_prog
, prog
);
1960 bpf_prog_put(old_prog
);
1961 virtnet_napi_enable(vi
->rq
[i
].vq
, &vi
->rq
[i
].napi
);
1967 for (i
= 0; i
< vi
->max_queue_pairs
; i
++)
1968 virtnet_napi_enable(vi
->rq
[i
].vq
, &vi
->rq
[i
].napi
);
1970 bpf_prog_sub(prog
, vi
->max_queue_pairs
- 1);
1974 static u32
virtnet_xdp_query(struct net_device
*dev
)
1976 struct virtnet_info
*vi
= netdev_priv(dev
);
1977 const struct bpf_prog
*xdp_prog
;
1980 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1981 xdp_prog
= rtnl_dereference(vi
->rq
[i
].xdp_prog
);
1983 return xdp_prog
->aux
->id
;
1988 static int virtnet_xdp(struct net_device
*dev
, struct netdev_xdp
*xdp
)
1990 switch (xdp
->command
) {
1991 case XDP_SETUP_PROG
:
1992 return virtnet_xdp_set(dev
, xdp
->prog
, xdp
->extack
);
1993 case XDP_QUERY_PROG
:
1994 xdp
->prog_id
= virtnet_xdp_query(dev
);
1995 xdp
->prog_attached
= !!xdp
->prog_id
;
2002 static const struct net_device_ops virtnet_netdev
= {
2003 .ndo_open
= virtnet_open
,
2004 .ndo_stop
= virtnet_close
,
2005 .ndo_start_xmit
= start_xmit
,
2006 .ndo_validate_addr
= eth_validate_addr
,
2007 .ndo_set_mac_address
= virtnet_set_mac_address
,
2008 .ndo_set_rx_mode
= virtnet_set_rx_mode
,
2009 .ndo_get_stats64
= virtnet_stats
,
2010 .ndo_vlan_rx_add_vid
= virtnet_vlan_rx_add_vid
,
2011 .ndo_vlan_rx_kill_vid
= virtnet_vlan_rx_kill_vid
,
2012 #ifdef CONFIG_NET_POLL_CONTROLLER
2013 .ndo_poll_controller
= virtnet_netpoll
,
2015 .ndo_xdp
= virtnet_xdp
,
2016 .ndo_features_check
= passthru_features_check
,
2019 static void virtnet_config_changed_work(struct work_struct
*work
)
2021 struct virtnet_info
*vi
=
2022 container_of(work
, struct virtnet_info
, config_work
);
2025 if (virtio_cread_feature(vi
->vdev
, VIRTIO_NET_F_STATUS
,
2026 struct virtio_net_config
, status
, &v
) < 0)
2029 if (v
& VIRTIO_NET_S_ANNOUNCE
) {
2030 netdev_notify_peers(vi
->dev
);
2031 virtnet_ack_link_announce(vi
);
2034 /* Ignore unknown (future) status bits */
2035 v
&= VIRTIO_NET_S_LINK_UP
;
2037 if (vi
->status
== v
)
2042 if (vi
->status
& VIRTIO_NET_S_LINK_UP
) {
2043 netif_carrier_on(vi
->dev
);
2044 netif_tx_wake_all_queues(vi
->dev
);
2046 netif_carrier_off(vi
->dev
);
2047 netif_tx_stop_all_queues(vi
->dev
);
2051 static void virtnet_config_changed(struct virtio_device
*vdev
)
2053 struct virtnet_info
*vi
= vdev
->priv
;
2055 schedule_work(&vi
->config_work
);
2058 static void virtnet_free_queues(struct virtnet_info
*vi
)
2062 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2063 napi_hash_del(&vi
->rq
[i
].napi
);
2064 netif_napi_del(&vi
->rq
[i
].napi
);
2065 netif_napi_del(&vi
->sq
[i
].napi
);
2068 /* We called napi_hash_del() before netif_napi_del(),
2069 * we need to respect an RCU grace period before freeing vi->rq
2077 static void _free_receive_bufs(struct virtnet_info
*vi
)
2079 struct bpf_prog
*old_prog
;
2082 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2083 while (vi
->rq
[i
].pages
)
2084 __free_pages(get_a_page(&vi
->rq
[i
], GFP_KERNEL
), 0);
2086 old_prog
= rtnl_dereference(vi
->rq
[i
].xdp_prog
);
2087 RCU_INIT_POINTER(vi
->rq
[i
].xdp_prog
, NULL
);
2089 bpf_prog_put(old_prog
);
2093 static void free_receive_bufs(struct virtnet_info
*vi
)
2096 _free_receive_bufs(vi
);
2100 static void free_receive_page_frags(struct virtnet_info
*vi
)
2103 for (i
= 0; i
< vi
->max_queue_pairs
; i
++)
2104 if (vi
->rq
[i
].alloc_frag
.page
)
2105 put_page(vi
->rq
[i
].alloc_frag
.page
);
2108 static bool is_xdp_raw_buffer_queue(struct virtnet_info
*vi
, int q
)
2110 if (q
< (vi
->curr_queue_pairs
- vi
->xdp_queue_pairs
))
2112 else if (q
< vi
->curr_queue_pairs
)
2118 static void free_unused_bufs(struct virtnet_info
*vi
)
2123 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2124 struct virtqueue
*vq
= vi
->sq
[i
].vq
;
2125 while ((buf
= virtqueue_detach_unused_buf(vq
)) != NULL
) {
2126 if (!is_xdp_raw_buffer_queue(vi
, i
))
2129 put_page(virt_to_head_page(buf
));
2133 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2134 struct virtqueue
*vq
= vi
->rq
[i
].vq
;
2136 while ((buf
= virtqueue_detach_unused_buf(vq
)) != NULL
) {
2137 if (vi
->mergeable_rx_bufs
) {
2138 put_page(virt_to_head_page(buf
));
2139 } else if (vi
->big_packets
) {
2140 give_pages(&vi
->rq
[i
], buf
);
2142 put_page(virt_to_head_page(buf
));
2148 static void virtnet_del_vqs(struct virtnet_info
*vi
)
2150 struct virtio_device
*vdev
= vi
->vdev
;
2152 virtnet_clean_affinity(vi
, -1);
2154 vdev
->config
->del_vqs(vdev
);
2156 virtnet_free_queues(vi
);
2159 /* How large should a single buffer be so a queue full of these can fit at
2160 * least one full packet?
2161 * Logic below assumes the mergeable buffer header is used.
2163 static unsigned int mergeable_min_buf_len(struct virtnet_info
*vi
, struct virtqueue
*vq
)
2165 const unsigned int hdr_len
= sizeof(struct virtio_net_hdr_mrg_rxbuf
);
2166 unsigned int rq_size
= virtqueue_get_vring_size(vq
);
2167 unsigned int packet_len
= vi
->big_packets
? IP_MAX_MTU
: vi
->dev
->max_mtu
;
2168 unsigned int buf_len
= hdr_len
+ ETH_HLEN
+ VLAN_HLEN
+ packet_len
;
2169 unsigned int min_buf_len
= DIV_ROUND_UP(buf_len
, rq_size
);
2171 return max(max(min_buf_len
, hdr_len
) - hdr_len
,
2172 (unsigned int)GOOD_PACKET_LEN
);
2175 static int virtnet_find_vqs(struct virtnet_info
*vi
)
2177 vq_callback_t
**callbacks
;
2178 struct virtqueue
**vqs
;
2184 /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
2185 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
2186 * possible control vq.
2188 total_vqs
= vi
->max_queue_pairs
* 2 +
2189 virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_CTRL_VQ
);
2191 /* Allocate space for find_vqs parameters */
2192 vqs
= kzalloc(total_vqs
* sizeof(*vqs
), GFP_KERNEL
);
2195 callbacks
= kmalloc(total_vqs
* sizeof(*callbacks
), GFP_KERNEL
);
2198 names
= kmalloc(total_vqs
* sizeof(*names
), GFP_KERNEL
);
2201 if (!vi
->big_packets
|| vi
->mergeable_rx_bufs
) {
2202 ctx
= kzalloc(total_vqs
* sizeof(*ctx
), GFP_KERNEL
);
2209 /* Parameters for control virtqueue, if any */
2211 callbacks
[total_vqs
- 1] = NULL
;
2212 names
[total_vqs
- 1] = "control";
2215 /* Allocate/initialize parameters for send/receive virtqueues */
2216 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2217 callbacks
[rxq2vq(i
)] = skb_recv_done
;
2218 callbacks
[txq2vq(i
)] = skb_xmit_done
;
2219 sprintf(vi
->rq
[i
].name
, "input.%d", i
);
2220 sprintf(vi
->sq
[i
].name
, "output.%d", i
);
2221 names
[rxq2vq(i
)] = vi
->rq
[i
].name
;
2222 names
[txq2vq(i
)] = vi
->sq
[i
].name
;
2224 ctx
[rxq2vq(i
)] = true;
2227 ret
= vi
->vdev
->config
->find_vqs(vi
->vdev
, total_vqs
, vqs
, callbacks
,
2233 vi
->cvq
= vqs
[total_vqs
- 1];
2234 if (virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_CTRL_VLAN
))
2235 vi
->dev
->features
|= NETIF_F_HW_VLAN_CTAG_FILTER
;
2238 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2239 vi
->rq
[i
].vq
= vqs
[rxq2vq(i
)];
2240 vi
->rq
[i
].min_buf_len
= mergeable_min_buf_len(vi
, vi
->rq
[i
].vq
);
2241 vi
->sq
[i
].vq
= vqs
[txq2vq(i
)];
2263 static int virtnet_alloc_queues(struct virtnet_info
*vi
)
2267 vi
->sq
= kzalloc(sizeof(*vi
->sq
) * vi
->max_queue_pairs
, GFP_KERNEL
);
2270 vi
->rq
= kzalloc(sizeof(*vi
->rq
) * vi
->max_queue_pairs
, GFP_KERNEL
);
2274 INIT_DELAYED_WORK(&vi
->refill
, refill_work
);
2275 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2276 vi
->rq
[i
].pages
= NULL
;
2277 netif_napi_add(vi
->dev
, &vi
->rq
[i
].napi
, virtnet_poll
,
2279 netif_tx_napi_add(vi
->dev
, &vi
->sq
[i
].napi
, virtnet_poll_tx
,
2280 napi_tx
? napi_weight
: 0);
2282 sg_init_table(vi
->rq
[i
].sg
, ARRAY_SIZE(vi
->rq
[i
].sg
));
2283 ewma_pkt_len_init(&vi
->rq
[i
].mrg_avg_pkt_len
);
2284 sg_init_table(vi
->sq
[i
].sg
, ARRAY_SIZE(vi
->sq
[i
].sg
));
2295 static int init_vqs(struct virtnet_info
*vi
)
2299 /* Allocate send & receive queues */
2300 ret
= virtnet_alloc_queues(vi
);
2304 ret
= virtnet_find_vqs(vi
);
2309 virtnet_set_affinity(vi
);
2315 virtnet_free_queues(vi
);
2321 static ssize_t
mergeable_rx_buffer_size_show(struct netdev_rx_queue
*queue
,
2322 struct rx_queue_attribute
*attribute
, char *buf
)
2324 struct virtnet_info
*vi
= netdev_priv(queue
->dev
);
2325 unsigned int queue_index
= get_netdev_rx_queue_index(queue
);
2326 struct ewma_pkt_len
*avg
;
2328 BUG_ON(queue_index
>= vi
->max_queue_pairs
);
2329 avg
= &vi
->rq
[queue_index
].mrg_avg_pkt_len
;
2330 return sprintf(buf
, "%u\n",
2331 get_mergeable_buf_len(&vi
->rq
[queue_index
], avg
));
2334 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute
=
2335 __ATTR_RO(mergeable_rx_buffer_size
);
2337 static struct attribute
*virtio_net_mrg_rx_attrs
[] = {
2338 &mergeable_rx_buffer_size_attribute
.attr
,
2342 static const struct attribute_group virtio_net_mrg_rx_group
= {
2343 .name
= "virtio_net",
2344 .attrs
= virtio_net_mrg_rx_attrs
2348 static bool virtnet_fail_on_feature(struct virtio_device
*vdev
,
2350 const char *fname
, const char *dname
)
2352 if (!virtio_has_feature(vdev
, fbit
))
2355 dev_err(&vdev
->dev
, "device advertises feature %s but not %s",
2361 #define VIRTNET_FAIL_ON(vdev, fbit, dbit) \
2362 virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
2364 static bool virtnet_validate_features(struct virtio_device
*vdev
)
2366 if (!virtio_has_feature(vdev
, VIRTIO_NET_F_CTRL_VQ
) &&
2367 (VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_CTRL_RX
,
2368 "VIRTIO_NET_F_CTRL_VQ") ||
2369 VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_CTRL_VLAN
,
2370 "VIRTIO_NET_F_CTRL_VQ") ||
2371 VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_GUEST_ANNOUNCE
,
2372 "VIRTIO_NET_F_CTRL_VQ") ||
2373 VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_MQ
, "VIRTIO_NET_F_CTRL_VQ") ||
2374 VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_CTRL_MAC_ADDR
,
2375 "VIRTIO_NET_F_CTRL_VQ"))) {
2382 #define MIN_MTU ETH_MIN_MTU
2383 #define MAX_MTU ETH_MAX_MTU
2385 static int virtnet_validate(struct virtio_device
*vdev
)
2387 if (!vdev
->config
->get
) {
2388 dev_err(&vdev
->dev
, "%s failure: config access disabled\n",
2393 if (!virtnet_validate_features(vdev
))
2396 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MTU
)) {
2397 int mtu
= virtio_cread16(vdev
,
2398 offsetof(struct virtio_net_config
,
2401 __virtio_clear_bit(vdev
, VIRTIO_NET_F_MTU
);
2407 static int virtnet_probe(struct virtio_device
*vdev
)
2410 struct net_device
*dev
;
2411 struct virtnet_info
*vi
;
2412 u16 max_queue_pairs
;
2415 /* Find if host supports multiqueue virtio_net device */
2416 err
= virtio_cread_feature(vdev
, VIRTIO_NET_F_MQ
,
2417 struct virtio_net_config
,
2418 max_virtqueue_pairs
, &max_queue_pairs
);
2420 /* We need at least 2 queue's */
2421 if (err
|| max_queue_pairs
< VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN
||
2422 max_queue_pairs
> VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX
||
2423 !virtio_has_feature(vdev
, VIRTIO_NET_F_CTRL_VQ
))
2424 max_queue_pairs
= 1;
2426 /* Allocate ourselves a network device with room for our info */
2427 dev
= alloc_etherdev_mq(sizeof(struct virtnet_info
), max_queue_pairs
);
2431 /* Set up network device as normal. */
2432 dev
->priv_flags
|= IFF_UNICAST_FLT
| IFF_LIVE_ADDR_CHANGE
;
2433 dev
->netdev_ops
= &virtnet_netdev
;
2434 dev
->features
= NETIF_F_HIGHDMA
;
2436 dev
->ethtool_ops
= &virtnet_ethtool_ops
;
2437 SET_NETDEV_DEV(dev
, &vdev
->dev
);
2439 /* Do we support "hardware" checksums? */
2440 if (virtio_has_feature(vdev
, VIRTIO_NET_F_CSUM
)) {
2441 /* This opens up the world of extra features. */
2442 dev
->hw_features
|= NETIF_F_HW_CSUM
| NETIF_F_SG
;
2444 dev
->features
|= NETIF_F_HW_CSUM
| NETIF_F_SG
;
2446 if (virtio_has_feature(vdev
, VIRTIO_NET_F_GSO
)) {
2447 dev
->hw_features
|= NETIF_F_TSO
2448 | NETIF_F_TSO_ECN
| NETIF_F_TSO6
;
2450 /* Individual feature bits: what can host handle? */
2451 if (virtio_has_feature(vdev
, VIRTIO_NET_F_HOST_TSO4
))
2452 dev
->hw_features
|= NETIF_F_TSO
;
2453 if (virtio_has_feature(vdev
, VIRTIO_NET_F_HOST_TSO6
))
2454 dev
->hw_features
|= NETIF_F_TSO6
;
2455 if (virtio_has_feature(vdev
, VIRTIO_NET_F_HOST_ECN
))
2456 dev
->hw_features
|= NETIF_F_TSO_ECN
;
2458 dev
->features
|= NETIF_F_GSO_ROBUST
;
2461 dev
->features
|= dev
->hw_features
& NETIF_F_ALL_TSO
;
2462 /* (!csum && gso) case will be fixed by register_netdev() */
2464 if (virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_CSUM
))
2465 dev
->features
|= NETIF_F_RXCSUM
;
2467 dev
->vlan_features
= dev
->features
;
2469 /* MTU range: 68 - 65535 */
2470 dev
->min_mtu
= MIN_MTU
;
2471 dev
->max_mtu
= MAX_MTU
;
2473 /* Configuration may specify what MAC to use. Otherwise random. */
2474 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MAC
))
2475 virtio_cread_bytes(vdev
,
2476 offsetof(struct virtio_net_config
, mac
),
2477 dev
->dev_addr
, dev
->addr_len
);
2479 eth_hw_addr_random(dev
);
2481 /* Set up our device-specific information */
2482 vi
= netdev_priv(dev
);
2486 vi
->stats
= alloc_percpu(struct virtnet_stats
);
2488 if (vi
->stats
== NULL
)
2491 for_each_possible_cpu(i
) {
2492 struct virtnet_stats
*virtnet_stats
;
2493 virtnet_stats
= per_cpu_ptr(vi
->stats
, i
);
2494 u64_stats_init(&virtnet_stats
->tx_syncp
);
2495 u64_stats_init(&virtnet_stats
->rx_syncp
);
2498 INIT_WORK(&vi
->config_work
, virtnet_config_changed_work
);
2500 /* If we can receive ANY GSO packets, we must allocate large ones. */
2501 if (virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_TSO4
) ||
2502 virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_TSO6
) ||
2503 virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_ECN
) ||
2504 virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_UFO
))
2505 vi
->big_packets
= true;
2507 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MRG_RXBUF
))
2508 vi
->mergeable_rx_bufs
= true;
2510 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MRG_RXBUF
) ||
2511 virtio_has_feature(vdev
, VIRTIO_F_VERSION_1
))
2512 vi
->hdr_len
= sizeof(struct virtio_net_hdr_mrg_rxbuf
);
2514 vi
->hdr_len
= sizeof(struct virtio_net_hdr
);
2516 if (virtio_has_feature(vdev
, VIRTIO_F_ANY_LAYOUT
) ||
2517 virtio_has_feature(vdev
, VIRTIO_F_VERSION_1
))
2518 vi
->any_header_sg
= true;
2520 if (virtio_has_feature(vdev
, VIRTIO_NET_F_CTRL_VQ
))
2523 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MTU
)) {
2524 mtu
= virtio_cread16(vdev
,
2525 offsetof(struct virtio_net_config
,
2527 if (mtu
< dev
->min_mtu
) {
2528 /* Should never trigger: MTU was previously validated
2529 * in virtnet_validate.
2531 dev_err(&vdev
->dev
, "device MTU appears to have changed "
2532 "it is now %d < %d", mtu
, dev
->min_mtu
);
2539 /* TODO: size buffers correctly in this case. */
2540 if (dev
->mtu
> ETH_DATA_LEN
)
2541 vi
->big_packets
= true;
2544 if (vi
->any_header_sg
)
2545 dev
->needed_headroom
= vi
->hdr_len
;
2547 /* Enable multiqueue by default */
2548 if (num_online_cpus() >= max_queue_pairs
)
2549 vi
->curr_queue_pairs
= max_queue_pairs
;
2551 vi
->curr_queue_pairs
= num_online_cpus();
2552 vi
->max_queue_pairs
= max_queue_pairs
;
2554 /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
2560 if (vi
->mergeable_rx_bufs
)
2561 dev
->sysfs_rx_queue_group
= &virtio_net_mrg_rx_group
;
2563 netif_set_real_num_tx_queues(dev
, vi
->curr_queue_pairs
);
2564 netif_set_real_num_rx_queues(dev
, vi
->curr_queue_pairs
);
2566 virtnet_init_settings(dev
);
2568 err
= register_netdev(dev
);
2570 pr_debug("virtio_net: registering device failed\n");
2574 virtio_device_ready(vdev
);
2576 err
= virtnet_cpu_notif_add(vi
);
2578 pr_debug("virtio_net: registering cpu notifier failed\n");
2579 goto free_unregister_netdev
;
2582 virtnet_set_queues(vi
, vi
->curr_queue_pairs
);
2584 /* Assume link up if device can't report link status,
2585 otherwise get link status from config. */
2586 if (virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_STATUS
)) {
2587 netif_carrier_off(dev
);
2588 schedule_work(&vi
->config_work
);
2590 vi
->status
= VIRTIO_NET_S_LINK_UP
;
2591 netif_carrier_on(dev
);
2594 pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
2595 dev
->name
, max_queue_pairs
);
2599 free_unregister_netdev
:
2600 vi
->vdev
->config
->reset(vdev
);
2602 unregister_netdev(dev
);
2604 cancel_delayed_work_sync(&vi
->refill
);
2605 free_receive_page_frags(vi
);
2606 virtnet_del_vqs(vi
);
2608 free_percpu(vi
->stats
);
2614 static void remove_vq_common(struct virtnet_info
*vi
)
2616 vi
->vdev
->config
->reset(vi
->vdev
);
2618 /* Free unused buffers in both send and recv, if any. */
2619 free_unused_bufs(vi
);
2621 free_receive_bufs(vi
);
2623 free_receive_page_frags(vi
);
2625 virtnet_del_vqs(vi
);
2628 static void virtnet_remove(struct virtio_device
*vdev
)
2630 struct virtnet_info
*vi
= vdev
->priv
;
2632 virtnet_cpu_notif_remove(vi
);
2634 /* Make sure no work handler is accessing the device. */
2635 flush_work(&vi
->config_work
);
2637 unregister_netdev(vi
->dev
);
2639 remove_vq_common(vi
);
2641 free_percpu(vi
->stats
);
2642 free_netdev(vi
->dev
);
2645 #ifdef CONFIG_PM_SLEEP
2646 static int virtnet_freeze(struct virtio_device
*vdev
)
2648 struct virtnet_info
*vi
= vdev
->priv
;
2650 virtnet_cpu_notif_remove(vi
);
2651 virtnet_freeze_down(vdev
);
2652 remove_vq_common(vi
);
2657 static int virtnet_restore(struct virtio_device
*vdev
)
2659 struct virtnet_info
*vi
= vdev
->priv
;
2662 err
= virtnet_restore_up(vdev
);
2665 virtnet_set_queues(vi
, vi
->curr_queue_pairs
);
2667 err
= virtnet_cpu_notif_add(vi
);
2675 static struct virtio_device_id id_table
[] = {
2676 { VIRTIO_ID_NET
, VIRTIO_DEV_ANY_ID
},
2680 #define VIRTNET_FEATURES \
2681 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
2683 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
2684 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
2685 VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
2686 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
2687 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
2688 VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
2689 VIRTIO_NET_F_CTRL_MAC_ADDR, \
2692 static unsigned int features
[] = {
2696 static unsigned int features_legacy
[] = {
2699 VIRTIO_F_ANY_LAYOUT
,
2702 static struct virtio_driver virtio_net_driver
= {
2703 .feature_table
= features
,
2704 .feature_table_size
= ARRAY_SIZE(features
),
2705 .feature_table_legacy
= features_legacy
,
2706 .feature_table_size_legacy
= ARRAY_SIZE(features_legacy
),
2707 .driver
.name
= KBUILD_MODNAME
,
2708 .driver
.owner
= THIS_MODULE
,
2709 .id_table
= id_table
,
2710 .validate
= virtnet_validate
,
2711 .probe
= virtnet_probe
,
2712 .remove
= virtnet_remove
,
2713 .config_changed
= virtnet_config_changed
,
2714 #ifdef CONFIG_PM_SLEEP
2715 .freeze
= virtnet_freeze
,
2716 .restore
= virtnet_restore
,
2720 static __init
int virtio_net_driver_init(void)
2724 ret
= cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN
, "virtio/net:online",
2726 virtnet_cpu_down_prep
);
2729 virtionet_online
= ret
;
2730 ret
= cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD
, "virtio/net:dead",
2731 NULL
, virtnet_cpu_dead
);
2735 ret
= register_virtio_driver(&virtio_net_driver
);
2740 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD
);
2742 cpuhp_remove_multi_state(virtionet_online
);
2746 module_init(virtio_net_driver_init
);
2748 static __exit
void virtio_net_driver_exit(void)
2750 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD
);
2751 cpuhp_remove_multi_state(virtionet_online
);
2752 unregister_virtio_driver(&virtio_net_driver
);
2754 module_exit(virtio_net_driver_exit
);
2756 MODULE_DEVICE_TABLE(virtio
, id_table
);
2757 MODULE_DESCRIPTION("Virtio network driver");
2758 MODULE_LICENSE("GPL");