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 <linux/filter.h>
33 #include <linux/kernel.h>
34 #include <linux/pci.h>
35 #include <net/route.h>
37 #include <net/net_failover.h>
39 static int napi_weight
= NAPI_POLL_WEIGHT
;
40 module_param(napi_weight
, int, 0444);
42 static bool csum
= true, gso
= true, napi_tx
;
43 module_param(csum
, bool, 0444);
44 module_param(gso
, bool, 0444);
45 module_param(napi_tx
, bool, 0644);
47 /* FIXME: MTU in config. */
48 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
49 #define GOOD_COPY_LEN 128
51 #define VIRTNET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
53 /* Amount of XDP headroom to prepend to packets for use by xdp_adjust_head */
54 #define VIRTIO_XDP_HEADROOM 256
56 /* Separating two types of XDP xmit */
57 #define VIRTIO_XDP_TX BIT(0)
58 #define VIRTIO_XDP_REDIR BIT(1)
60 /* RX packet size EWMA. The average packet size is used to determine the packet
61 * buffer size when refilling RX rings. As the entire RX ring may be refilled
62 * at once, the weight is chosen so that the EWMA will be insensitive to short-
63 * term, transient changes in packet size.
65 DECLARE_EWMA(pkt_len
, 0, 64)
67 #define VIRTNET_DRIVER_VERSION "1.0.0"
69 static const unsigned long guest_offloads
[] = {
70 VIRTIO_NET_F_GUEST_TSO4
,
71 VIRTIO_NET_F_GUEST_TSO6
,
72 VIRTIO_NET_F_GUEST_ECN
,
73 VIRTIO_NET_F_GUEST_UFO
76 struct virtnet_stat_desc
{
77 char desc
[ETH_GSTRING_LEN
];
81 struct virtnet_sq_stats
{
82 struct u64_stats_sync syncp
;
90 struct virtnet_rq_stats
{
91 struct u64_stats_sync syncp
;
102 #define VIRTNET_SQ_STAT(m) offsetof(struct virtnet_sq_stats, m)
103 #define VIRTNET_RQ_STAT(m) offsetof(struct virtnet_rq_stats, m)
105 static const struct virtnet_stat_desc virtnet_sq_stats_desc
[] = {
106 { "packets", VIRTNET_SQ_STAT(packets
) },
107 { "bytes", VIRTNET_SQ_STAT(bytes
) },
108 { "xdp_tx", VIRTNET_SQ_STAT(xdp_tx
) },
109 { "xdp_tx_drops", VIRTNET_SQ_STAT(xdp_tx_drops
) },
110 { "kicks", VIRTNET_SQ_STAT(kicks
) },
113 static const struct virtnet_stat_desc virtnet_rq_stats_desc
[] = {
114 { "packets", VIRTNET_RQ_STAT(packets
) },
115 { "bytes", VIRTNET_RQ_STAT(bytes
) },
116 { "drops", VIRTNET_RQ_STAT(drops
) },
117 { "xdp_packets", VIRTNET_RQ_STAT(xdp_packets
) },
118 { "xdp_tx", VIRTNET_RQ_STAT(xdp_tx
) },
119 { "xdp_redirects", VIRTNET_RQ_STAT(xdp_redirects
) },
120 { "xdp_drops", VIRTNET_RQ_STAT(xdp_drops
) },
121 { "kicks", VIRTNET_RQ_STAT(kicks
) },
124 #define VIRTNET_SQ_STATS_LEN ARRAY_SIZE(virtnet_sq_stats_desc)
125 #define VIRTNET_RQ_STATS_LEN ARRAY_SIZE(virtnet_rq_stats_desc)
127 /* Internal representation of a send virtqueue */
129 /* Virtqueue associated with this send _queue */
130 struct virtqueue
*vq
;
132 /* TX: fragments + linear part + virtio header */
133 struct scatterlist sg
[MAX_SKB_FRAGS
+ 2];
135 /* Name of the send queue: output.$index */
138 struct virtnet_sq_stats stats
;
140 struct napi_struct napi
;
143 /* Internal representation of a receive virtqueue */
144 struct receive_queue
{
145 /* Virtqueue associated with this receive_queue */
146 struct virtqueue
*vq
;
148 struct napi_struct napi
;
150 struct bpf_prog __rcu
*xdp_prog
;
152 struct virtnet_rq_stats stats
;
154 /* Chain pages by the private ptr. */
157 /* Average packet length for mergeable receive buffers. */
158 struct ewma_pkt_len mrg_avg_pkt_len
;
160 /* Page frag for packet buffer allocation. */
161 struct page_frag alloc_frag
;
163 /* RX: fragments + linear part + virtio header */
164 struct scatterlist sg
[MAX_SKB_FRAGS
+ 2];
166 /* Min single buffer size for mergeable buffers case. */
167 unsigned int min_buf_len
;
169 /* Name of this receive queue: input.$index */
172 struct xdp_rxq_info xdp_rxq
;
175 /* Control VQ buffers: protected by the rtnl lock */
177 struct virtio_net_ctrl_hdr hdr
;
178 virtio_net_ctrl_ack status
;
179 struct virtio_net_ctrl_mq mq
;
186 struct virtnet_info
{
187 struct virtio_device
*vdev
;
188 struct virtqueue
*cvq
;
189 struct net_device
*dev
;
190 struct send_queue
*sq
;
191 struct receive_queue
*rq
;
194 /* Max # of queue pairs supported by the device */
197 /* # of queue pairs currently used by the driver */
198 u16 curr_queue_pairs
;
200 /* # of XDP queue pairs currently used by the driver */
203 /* I like... big packets and I cannot lie! */
206 /* Host will merge rx buffers for big packets (shake it! shake it!) */
207 bool mergeable_rx_bufs
;
209 /* Has control virtqueue */
212 /* Host can handle any s/g split between our header and packet data */
215 /* Packet virtio header size */
218 /* Work struct for refilling if we run low on memory. */
219 struct delayed_work refill
;
221 /* Work struct for config space updates */
222 struct work_struct config_work
;
224 /* Does the affinity hint is set for virtqueues? */
225 bool affinity_hint_set
;
227 /* CPU hotplug instances for online & dead */
228 struct hlist_node node
;
229 struct hlist_node node_dead
;
231 struct control_buf
*ctrl
;
233 /* Ethtool settings */
237 unsigned long guest_offloads
;
239 /* failover when STANDBY feature enabled */
240 struct failover
*failover
;
243 struct padded_vnet_hdr
{
244 struct virtio_net_hdr_mrg_rxbuf hdr
;
246 * hdr is in a separate sg buffer, and data sg buffer shares same page
247 * with this header sg. This padding makes next sg 16 byte aligned
253 /* Converting between virtqueue no. and kernel tx/rx queue no.
254 * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
256 static int vq2txq(struct virtqueue
*vq
)
258 return (vq
->index
- 1) / 2;
261 static int txq2vq(int txq
)
266 static int vq2rxq(struct virtqueue
*vq
)
268 return vq
->index
/ 2;
271 static int rxq2vq(int rxq
)
276 static inline struct virtio_net_hdr_mrg_rxbuf
*skb_vnet_hdr(struct sk_buff
*skb
)
278 return (struct virtio_net_hdr_mrg_rxbuf
*)skb
->cb
;
282 * private is used to chain pages for big packets, put the whole
283 * most recent used list in the beginning for reuse
285 static void give_pages(struct receive_queue
*rq
, struct page
*page
)
289 /* Find end of list, sew whole thing into vi->rq.pages. */
290 for (end
= page
; end
->private; end
= (struct page
*)end
->private);
291 end
->private = (unsigned long)rq
->pages
;
295 static struct page
*get_a_page(struct receive_queue
*rq
, gfp_t gfp_mask
)
297 struct page
*p
= rq
->pages
;
300 rq
->pages
= (struct page
*)p
->private;
301 /* clear private here, it is used to chain pages */
304 p
= alloc_page(gfp_mask
);
308 static void virtqueue_napi_schedule(struct napi_struct
*napi
,
309 struct virtqueue
*vq
)
311 if (napi_schedule_prep(napi
)) {
312 virtqueue_disable_cb(vq
);
313 __napi_schedule(napi
);
317 static void virtqueue_napi_complete(struct napi_struct
*napi
,
318 struct virtqueue
*vq
, int processed
)
322 opaque
= virtqueue_enable_cb_prepare(vq
);
323 if (napi_complete_done(napi
, processed
)) {
324 if (unlikely(virtqueue_poll(vq
, opaque
)))
325 virtqueue_napi_schedule(napi
, vq
);
327 virtqueue_disable_cb(vq
);
331 static void skb_xmit_done(struct virtqueue
*vq
)
333 struct virtnet_info
*vi
= vq
->vdev
->priv
;
334 struct napi_struct
*napi
= &vi
->sq
[vq2txq(vq
)].napi
;
336 /* Suppress further interrupts. */
337 virtqueue_disable_cb(vq
);
340 virtqueue_napi_schedule(napi
, vq
);
342 /* We were probably waiting for more output buffers. */
343 netif_wake_subqueue(vi
->dev
, vq2txq(vq
));
346 #define MRG_CTX_HEADER_SHIFT 22
347 static void *mergeable_len_to_ctx(unsigned int truesize
,
348 unsigned int headroom
)
350 return (void *)(unsigned long)((headroom
<< MRG_CTX_HEADER_SHIFT
) | truesize
);
353 static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx
)
355 return (unsigned long)mrg_ctx
>> MRG_CTX_HEADER_SHIFT
;
358 static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx
)
360 return (unsigned long)mrg_ctx
& ((1 << MRG_CTX_HEADER_SHIFT
) - 1);
363 /* Called from bottom half context */
364 static struct sk_buff
*page_to_skb(struct virtnet_info
*vi
,
365 struct receive_queue
*rq
,
366 struct page
*page
, unsigned int offset
,
367 unsigned int len
, unsigned int truesize
)
370 struct virtio_net_hdr_mrg_rxbuf
*hdr
;
371 unsigned int copy
, hdr_len
, hdr_padded_len
;
374 p
= page_address(page
) + offset
;
376 /* copy small packet so we can reuse these pages for small data */
377 skb
= napi_alloc_skb(&rq
->napi
, GOOD_COPY_LEN
);
381 hdr
= skb_vnet_hdr(skb
);
383 hdr_len
= vi
->hdr_len
;
384 if (vi
->mergeable_rx_bufs
)
385 hdr_padded_len
= sizeof(*hdr
);
387 hdr_padded_len
= sizeof(struct padded_vnet_hdr
);
389 memcpy(hdr
, p
, hdr_len
);
392 offset
+= hdr_padded_len
;
396 if (copy
> skb_tailroom(skb
))
397 copy
= skb_tailroom(skb
);
398 skb_put_data(skb
, p
, copy
);
403 if (vi
->mergeable_rx_bufs
) {
405 skb_add_rx_frag(skb
, 0, page
, offset
, len
, truesize
);
412 * Verify that we can indeed put this data into a skb.
413 * This is here to handle cases when the device erroneously
414 * tries to receive more than is possible. This is usually
415 * the case of a broken device.
417 if (unlikely(len
> MAX_SKB_FRAGS
* PAGE_SIZE
)) {
418 net_dbg_ratelimited("%s: too much data\n", skb
->dev
->name
);
422 BUG_ON(offset
>= PAGE_SIZE
);
424 unsigned int frag_size
= min((unsigned)PAGE_SIZE
- offset
, len
);
425 skb_add_rx_frag(skb
, skb_shinfo(skb
)->nr_frags
, page
, offset
,
426 frag_size
, truesize
);
428 page
= (struct page
*)page
->private;
433 give_pages(rq
, page
);
438 static int __virtnet_xdp_xmit_one(struct virtnet_info
*vi
,
439 struct send_queue
*sq
,
440 struct xdp_frame
*xdpf
)
442 struct virtio_net_hdr_mrg_rxbuf
*hdr
;
445 /* virtqueue want to use data area in-front of packet */
446 if (unlikely(xdpf
->metasize
> 0))
449 if (unlikely(xdpf
->headroom
< vi
->hdr_len
))
452 /* Make room for virtqueue hdr (also change xdpf->headroom?) */
453 xdpf
->data
-= vi
->hdr_len
;
454 /* Zero header and leave csum up to XDP layers */
456 memset(hdr
, 0, vi
->hdr_len
);
457 xdpf
->len
+= vi
->hdr_len
;
459 sg_init_one(sq
->sg
, xdpf
->data
, xdpf
->len
);
461 err
= virtqueue_add_outbuf(sq
->vq
, sq
->sg
, 1, xdpf
, GFP_ATOMIC
);
463 return -ENOSPC
; /* Caller handle free/refcnt */
468 static struct send_queue
*virtnet_xdp_sq(struct virtnet_info
*vi
)
472 qp
= vi
->curr_queue_pairs
- vi
->xdp_queue_pairs
+ smp_processor_id();
476 static int virtnet_xdp_xmit(struct net_device
*dev
,
477 int n
, struct xdp_frame
**frames
, u32 flags
)
479 struct virtnet_info
*vi
= netdev_priv(dev
);
480 struct receive_queue
*rq
= vi
->rq
;
481 struct xdp_frame
*xdpf_sent
;
482 struct bpf_prog
*xdp_prog
;
483 struct send_queue
*sq
;
490 sq
= virtnet_xdp_sq(vi
);
492 if (unlikely(flags
& ~XDP_XMIT_FLAGS_MASK
)) {
498 /* Only allow ndo_xdp_xmit if XDP is loaded on dev, as this
499 * indicate XDP resources have been successfully allocated.
501 xdp_prog
= rcu_dereference(rq
->xdp_prog
);
508 /* Free up any pending old buffers before queueing new ones. */
509 while ((xdpf_sent
= virtqueue_get_buf(sq
->vq
, &len
)) != NULL
)
510 xdp_return_frame(xdpf_sent
);
512 for (i
= 0; i
< n
; i
++) {
513 struct xdp_frame
*xdpf
= frames
[i
];
515 err
= __virtnet_xdp_xmit_one(vi
, sq
, xdpf
);
517 xdp_return_frame_rx_napi(xdpf
);
523 if (flags
& XDP_XMIT_FLUSH
) {
524 if (virtqueue_kick_prepare(sq
->vq
) && virtqueue_notify(sq
->vq
))
528 u64_stats_update_begin(&sq
->stats
.syncp
);
529 sq
->stats
.xdp_tx
+= n
;
530 sq
->stats
.xdp_tx_drops
+= drops
;
531 sq
->stats
.kicks
+= kicks
;
532 u64_stats_update_end(&sq
->stats
.syncp
);
537 static unsigned int virtnet_get_headroom(struct virtnet_info
*vi
)
539 return vi
->xdp_queue_pairs
? VIRTIO_XDP_HEADROOM
: 0;
542 /* We copy the packet for XDP in the following cases:
544 * 1) Packet is scattered across multiple rx buffers.
545 * 2) Headroom space is insufficient.
547 * This is inefficient but it's a temporary condition that
548 * we hit right after XDP is enabled and until queue is refilled
549 * with large buffers with sufficient headroom - so it should affect
550 * at most queue size packets.
551 * Afterwards, the conditions to enable
552 * XDP should preclude the underlying device from sending packets
553 * across multiple buffers (num_buf > 1), and we make sure buffers
554 * have enough headroom.
556 static struct page
*xdp_linearize_page(struct receive_queue
*rq
,
563 struct page
*page
= alloc_page(GFP_ATOMIC
);
568 memcpy(page_address(page
) + page_off
, page_address(p
) + offset
, *len
);
572 int tailroom
= SKB_DATA_ALIGN(sizeof(struct skb_shared_info
));
577 buf
= virtqueue_get_buf(rq
->vq
, &buflen
);
581 p
= virt_to_head_page(buf
);
582 off
= buf
- page_address(p
);
584 /* guard against a misconfigured or uncooperative backend that
585 * is sending packet larger than the MTU.
587 if ((page_off
+ buflen
+ tailroom
) > PAGE_SIZE
) {
592 memcpy(page_address(page
) + page_off
,
593 page_address(p
) + off
, buflen
);
598 /* Headroom does not contribute to packet length */
599 *len
= page_off
- VIRTIO_XDP_HEADROOM
;
602 __free_pages(page
, 0);
606 static struct sk_buff
*receive_small(struct net_device
*dev
,
607 struct virtnet_info
*vi
,
608 struct receive_queue
*rq
,
609 void *buf
, void *ctx
,
611 unsigned int *xdp_xmit
,
612 struct virtnet_rq_stats
*stats
)
615 struct bpf_prog
*xdp_prog
;
616 unsigned int xdp_headroom
= (unsigned long)ctx
;
617 unsigned int header_offset
= VIRTNET_RX_PAD
+ xdp_headroom
;
618 unsigned int headroom
= vi
->hdr_len
+ header_offset
;
619 unsigned int buflen
= SKB_DATA_ALIGN(GOOD_PACKET_LEN
+ headroom
) +
620 SKB_DATA_ALIGN(sizeof(struct skb_shared_info
));
621 struct page
*page
= virt_to_head_page(buf
);
622 unsigned int delta
= 0;
623 struct page
*xdp_page
;
630 xdp_prog
= rcu_dereference(rq
->xdp_prog
);
632 struct virtio_net_hdr_mrg_rxbuf
*hdr
= buf
+ header_offset
;
633 struct xdp_frame
*xdpf
;
638 if (unlikely(hdr
->hdr
.gso_type
))
641 if (unlikely(xdp_headroom
< virtnet_get_headroom(vi
))) {
642 int offset
= buf
- page_address(page
) + header_offset
;
643 unsigned int tlen
= len
+ vi
->hdr_len
;
646 xdp_headroom
= virtnet_get_headroom(vi
);
647 header_offset
= VIRTNET_RX_PAD
+ xdp_headroom
;
648 headroom
= vi
->hdr_len
+ header_offset
;
649 buflen
= SKB_DATA_ALIGN(GOOD_PACKET_LEN
+ headroom
) +
650 SKB_DATA_ALIGN(sizeof(struct skb_shared_info
));
651 xdp_page
= xdp_linearize_page(rq
, &num_buf
, page
,
652 offset
, header_offset
,
657 buf
= page_address(xdp_page
);
662 xdp
.data_hard_start
= buf
+ VIRTNET_RX_PAD
+ vi
->hdr_len
;
663 xdp
.data
= xdp
.data_hard_start
+ xdp_headroom
;
664 xdp_set_data_meta_invalid(&xdp
);
665 xdp
.data_end
= xdp
.data
+ len
;
666 xdp
.rxq
= &rq
->xdp_rxq
;
667 orig_data
= xdp
.data
;
668 act
= bpf_prog_run_xdp(xdp_prog
, &xdp
);
669 stats
->xdp_packets
++;
673 /* Recalculate length in case bpf program changed it */
674 delta
= orig_data
- xdp
.data
;
675 len
= xdp
.data_end
- xdp
.data
;
679 xdpf
= convert_to_xdp_frame(&xdp
);
682 err
= virtnet_xdp_xmit(dev
, 1, &xdpf
, 0);
683 if (unlikely(err
< 0)) {
684 trace_xdp_exception(vi
->dev
, xdp_prog
, act
);
687 *xdp_xmit
|= VIRTIO_XDP_TX
;
691 stats
->xdp_redirects
++;
692 err
= xdp_do_redirect(dev
, &xdp
, xdp_prog
);
695 *xdp_xmit
|= VIRTIO_XDP_REDIR
;
699 bpf_warn_invalid_xdp_action(act
);
702 trace_xdp_exception(vi
->dev
, xdp_prog
, act
);
709 skb
= build_skb(buf
, buflen
);
714 skb_reserve(skb
, headroom
- delta
);
717 buf
+= header_offset
;
718 memcpy(skb_vnet_hdr(skb
), buf
, vi
->hdr_len
);
719 } /* keep zeroed vnet hdr since packet was changed by bpf */
733 static struct sk_buff
*receive_big(struct net_device
*dev
,
734 struct virtnet_info
*vi
,
735 struct receive_queue
*rq
,
738 struct virtnet_rq_stats
*stats
)
740 struct page
*page
= buf
;
741 struct sk_buff
*skb
= page_to_skb(vi
, rq
, page
, 0, len
, PAGE_SIZE
);
743 stats
->bytes
+= len
- vi
->hdr_len
;
751 give_pages(rq
, page
);
755 static struct sk_buff
*receive_mergeable(struct net_device
*dev
,
756 struct virtnet_info
*vi
,
757 struct receive_queue
*rq
,
761 unsigned int *xdp_xmit
,
762 struct virtnet_rq_stats
*stats
)
764 struct virtio_net_hdr_mrg_rxbuf
*hdr
= buf
;
765 u16 num_buf
= virtio16_to_cpu(vi
->vdev
, hdr
->num_buffers
);
766 struct page
*page
= virt_to_head_page(buf
);
767 int offset
= buf
- page_address(page
);
768 struct sk_buff
*head_skb
, *curr_skb
;
769 struct bpf_prog
*xdp_prog
;
770 unsigned int truesize
;
771 unsigned int headroom
= mergeable_ctx_to_headroom(ctx
);
775 stats
->bytes
+= len
- vi
->hdr_len
;
778 xdp_prog
= rcu_dereference(rq
->xdp_prog
);
780 struct xdp_frame
*xdpf
;
781 struct page
*xdp_page
;
786 /* Transient failure which in theory could occur if
787 * in-flight packets from before XDP was enabled reach
788 * the receive path after XDP is loaded.
790 if (unlikely(hdr
->hdr
.gso_type
))
793 /* This happens when rx buffer size is underestimated
794 * or headroom is not enough because of the buffer
795 * was refilled before XDP is set. This should only
796 * happen for the first several packets, so we don't
797 * care much about its performance.
799 if (unlikely(num_buf
> 1 ||
800 headroom
< virtnet_get_headroom(vi
))) {
801 /* linearize data for XDP */
802 xdp_page
= xdp_linearize_page(rq
, &num_buf
,
808 offset
= VIRTIO_XDP_HEADROOM
;
813 /* Allow consuming headroom but reserve enough space to push
814 * the descriptor on if we get an XDP_TX return code.
816 data
= page_address(xdp_page
) + offset
;
817 xdp
.data_hard_start
= data
- VIRTIO_XDP_HEADROOM
+ vi
->hdr_len
;
818 xdp
.data
= data
+ vi
->hdr_len
;
819 xdp_set_data_meta_invalid(&xdp
);
820 xdp
.data_end
= xdp
.data
+ (len
- vi
->hdr_len
);
821 xdp
.rxq
= &rq
->xdp_rxq
;
823 act
= bpf_prog_run_xdp(xdp_prog
, &xdp
);
824 stats
->xdp_packets
++;
828 /* recalculate offset to account for any header
829 * adjustments. Note other cases do not build an
830 * skb and avoid using offset
833 page_address(xdp_page
) - vi
->hdr_len
;
835 /* recalculate len if xdp.data or xdp.data_end were
838 len
= xdp
.data_end
- xdp
.data
+ vi
->hdr_len
;
839 /* We can only create skb based on xdp_page. */
840 if (unlikely(xdp_page
!= page
)) {
843 head_skb
= page_to_skb(vi
, rq
, xdp_page
,
844 offset
, len
, PAGE_SIZE
);
850 xdpf
= convert_to_xdp_frame(&xdp
);
853 err
= virtnet_xdp_xmit(dev
, 1, &xdpf
, 0);
854 if (unlikely(err
< 0)) {
855 trace_xdp_exception(vi
->dev
, xdp_prog
, act
);
856 if (unlikely(xdp_page
!= page
))
860 *xdp_xmit
|= VIRTIO_XDP_TX
;
861 if (unlikely(xdp_page
!= page
))
866 stats
->xdp_redirects
++;
867 err
= xdp_do_redirect(dev
, &xdp
, xdp_prog
);
869 if (unlikely(xdp_page
!= page
))
873 *xdp_xmit
|= VIRTIO_XDP_REDIR
;
874 if (unlikely(xdp_page
!= page
))
879 bpf_warn_invalid_xdp_action(act
);
882 trace_xdp_exception(vi
->dev
, xdp_prog
, act
);
885 if (unlikely(xdp_page
!= page
))
886 __free_pages(xdp_page
, 0);
892 truesize
= mergeable_ctx_to_truesize(ctx
);
893 if (unlikely(len
> truesize
)) {
894 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
895 dev
->name
, len
, (unsigned long)ctx
);
896 dev
->stats
.rx_length_errors
++;
900 head_skb
= page_to_skb(vi
, rq
, page
, offset
, len
, truesize
);
903 if (unlikely(!curr_skb
))
908 buf
= virtqueue_get_buf_ctx(rq
->vq
, &len
, &ctx
);
909 if (unlikely(!buf
)) {
910 pr_debug("%s: rx error: %d buffers out of %d missing\n",
912 virtio16_to_cpu(vi
->vdev
,
914 dev
->stats
.rx_length_errors
++;
919 page
= virt_to_head_page(buf
);
921 truesize
= mergeable_ctx_to_truesize(ctx
);
922 if (unlikely(len
> truesize
)) {
923 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
924 dev
->name
, len
, (unsigned long)ctx
);
925 dev
->stats
.rx_length_errors
++;
929 num_skb_frags
= skb_shinfo(curr_skb
)->nr_frags
;
930 if (unlikely(num_skb_frags
== MAX_SKB_FRAGS
)) {
931 struct sk_buff
*nskb
= alloc_skb(0, GFP_ATOMIC
);
935 if (curr_skb
== head_skb
)
936 skb_shinfo(curr_skb
)->frag_list
= nskb
;
938 curr_skb
->next
= nskb
;
940 head_skb
->truesize
+= nskb
->truesize
;
943 if (curr_skb
!= head_skb
) {
944 head_skb
->data_len
+= len
;
945 head_skb
->len
+= len
;
946 head_skb
->truesize
+= truesize
;
948 offset
= buf
- page_address(page
);
949 if (skb_can_coalesce(curr_skb
, num_skb_frags
, page
, offset
)) {
951 skb_coalesce_rx_frag(curr_skb
, num_skb_frags
- 1,
954 skb_add_rx_frag(curr_skb
, num_skb_frags
, page
,
955 offset
, len
, truesize
);
959 ewma_pkt_len_add(&rq
->mrg_avg_pkt_len
, head_skb
->len
);
967 while (num_buf
-- > 1) {
968 buf
= virtqueue_get_buf(rq
->vq
, &len
);
969 if (unlikely(!buf
)) {
970 pr_debug("%s: rx error: %d buffers missing\n",
972 dev
->stats
.rx_length_errors
++;
976 page
= virt_to_head_page(buf
);
981 dev_kfree_skb(head_skb
);
986 static void receive_buf(struct virtnet_info
*vi
, struct receive_queue
*rq
,
987 void *buf
, unsigned int len
, void **ctx
,
988 unsigned int *xdp_xmit
,
989 struct virtnet_rq_stats
*stats
)
991 struct net_device
*dev
= vi
->dev
;
993 struct virtio_net_hdr_mrg_rxbuf
*hdr
;
995 if (unlikely(len
< vi
->hdr_len
+ ETH_HLEN
)) {
996 pr_debug("%s: short packet %i\n", dev
->name
, len
);
997 dev
->stats
.rx_length_errors
++;
998 if (vi
->mergeable_rx_bufs
) {
999 put_page(virt_to_head_page(buf
));
1000 } else if (vi
->big_packets
) {
1001 give_pages(rq
, buf
);
1003 put_page(virt_to_head_page(buf
));
1008 if (vi
->mergeable_rx_bufs
)
1009 skb
= receive_mergeable(dev
, vi
, rq
, buf
, ctx
, len
, xdp_xmit
,
1011 else if (vi
->big_packets
)
1012 skb
= receive_big(dev
, vi
, rq
, buf
, len
, stats
);
1014 skb
= receive_small(dev
, vi
, rq
, buf
, ctx
, len
, xdp_xmit
, stats
);
1019 hdr
= skb_vnet_hdr(skb
);
1021 if (hdr
->hdr
.flags
& VIRTIO_NET_HDR_F_DATA_VALID
)
1022 skb
->ip_summed
= CHECKSUM_UNNECESSARY
;
1024 if (virtio_net_hdr_to_skb(skb
, &hdr
->hdr
,
1025 virtio_is_little_endian(vi
->vdev
))) {
1026 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
1027 dev
->name
, hdr
->hdr
.gso_type
,
1032 skb
->protocol
= eth_type_trans(skb
, dev
);
1033 pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
1034 ntohs(skb
->protocol
), skb
->len
, skb
->pkt_type
);
1036 napi_gro_receive(&rq
->napi
, skb
);
1040 dev
->stats
.rx_frame_errors
++;
1044 /* Unlike mergeable buffers, all buffers are allocated to the
1045 * same size, except for the headroom. For this reason we do
1046 * not need to use mergeable_len_to_ctx here - it is enough
1047 * to store the headroom as the context ignoring the truesize.
1049 static int add_recvbuf_small(struct virtnet_info
*vi
, struct receive_queue
*rq
,
1052 struct page_frag
*alloc_frag
= &rq
->alloc_frag
;
1054 unsigned int xdp_headroom
= virtnet_get_headroom(vi
);
1055 void *ctx
= (void *)(unsigned long)xdp_headroom
;
1056 int len
= vi
->hdr_len
+ VIRTNET_RX_PAD
+ GOOD_PACKET_LEN
+ xdp_headroom
;
1059 len
= SKB_DATA_ALIGN(len
) +
1060 SKB_DATA_ALIGN(sizeof(struct skb_shared_info
));
1061 if (unlikely(!skb_page_frag_refill(len
, alloc_frag
, gfp
)))
1064 buf
= (char *)page_address(alloc_frag
->page
) + alloc_frag
->offset
;
1065 get_page(alloc_frag
->page
);
1066 alloc_frag
->offset
+= len
;
1067 sg_init_one(rq
->sg
, buf
+ VIRTNET_RX_PAD
+ xdp_headroom
,
1068 vi
->hdr_len
+ GOOD_PACKET_LEN
);
1069 err
= virtqueue_add_inbuf_ctx(rq
->vq
, rq
->sg
, 1, buf
, ctx
, gfp
);
1071 put_page(virt_to_head_page(buf
));
1075 static int add_recvbuf_big(struct virtnet_info
*vi
, struct receive_queue
*rq
,
1078 struct page
*first
, *list
= NULL
;
1082 sg_init_table(rq
->sg
, MAX_SKB_FRAGS
+ 2);
1084 /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
1085 for (i
= MAX_SKB_FRAGS
+ 1; i
> 1; --i
) {
1086 first
= get_a_page(rq
, gfp
);
1089 give_pages(rq
, list
);
1092 sg_set_buf(&rq
->sg
[i
], page_address(first
), PAGE_SIZE
);
1094 /* chain new page in list head to match sg */
1095 first
->private = (unsigned long)list
;
1099 first
= get_a_page(rq
, gfp
);
1101 give_pages(rq
, list
);
1104 p
= page_address(first
);
1106 /* rq->sg[0], rq->sg[1] share the same page */
1107 /* a separated rq->sg[0] for header - required in case !any_header_sg */
1108 sg_set_buf(&rq
->sg
[0], p
, vi
->hdr_len
);
1110 /* rq->sg[1] for data packet, from offset */
1111 offset
= sizeof(struct padded_vnet_hdr
);
1112 sg_set_buf(&rq
->sg
[1], p
+ offset
, PAGE_SIZE
- offset
);
1114 /* chain first in list head */
1115 first
->private = (unsigned long)list
;
1116 err
= virtqueue_add_inbuf(rq
->vq
, rq
->sg
, MAX_SKB_FRAGS
+ 2,
1119 give_pages(rq
, first
);
1124 static unsigned int get_mergeable_buf_len(struct receive_queue
*rq
,
1125 struct ewma_pkt_len
*avg_pkt_len
,
1128 const size_t hdr_len
= sizeof(struct virtio_net_hdr_mrg_rxbuf
);
1132 return PAGE_SIZE
- room
;
1134 len
= hdr_len
+ clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len
),
1135 rq
->min_buf_len
, PAGE_SIZE
- hdr_len
);
1137 return ALIGN(len
, L1_CACHE_BYTES
);
1140 static int add_recvbuf_mergeable(struct virtnet_info
*vi
,
1141 struct receive_queue
*rq
, gfp_t gfp
)
1143 struct page_frag
*alloc_frag
= &rq
->alloc_frag
;
1144 unsigned int headroom
= virtnet_get_headroom(vi
);
1145 unsigned int tailroom
= headroom
? sizeof(struct skb_shared_info
) : 0;
1146 unsigned int room
= SKB_DATA_ALIGN(headroom
+ tailroom
);
1150 unsigned int len
, hole
;
1152 /* Extra tailroom is needed to satisfy XDP's assumption. This
1153 * means rx frags coalescing won't work, but consider we've
1154 * disabled GSO for XDP, it won't be a big issue.
1156 len
= get_mergeable_buf_len(rq
, &rq
->mrg_avg_pkt_len
, room
);
1157 if (unlikely(!skb_page_frag_refill(len
+ room
, alloc_frag
, gfp
)))
1160 buf
= (char *)page_address(alloc_frag
->page
) + alloc_frag
->offset
;
1161 buf
+= headroom
; /* advance address leaving hole at front of pkt */
1162 get_page(alloc_frag
->page
);
1163 alloc_frag
->offset
+= len
+ room
;
1164 hole
= alloc_frag
->size
- alloc_frag
->offset
;
1165 if (hole
< len
+ room
) {
1166 /* To avoid internal fragmentation, if there is very likely not
1167 * enough space for another buffer, add the remaining space to
1168 * the current buffer.
1171 alloc_frag
->offset
+= hole
;
1174 sg_init_one(rq
->sg
, buf
, len
);
1175 ctx
= mergeable_len_to_ctx(len
, headroom
);
1176 err
= virtqueue_add_inbuf_ctx(rq
->vq
, rq
->sg
, 1, buf
, ctx
, gfp
);
1178 put_page(virt_to_head_page(buf
));
1184 * Returns false if we couldn't fill entirely (OOM).
1186 * Normally run in the receive path, but can also be run from ndo_open
1187 * before we're receiving packets, or from refill_work which is
1188 * careful to disable receiving (using napi_disable).
1190 static bool try_fill_recv(struct virtnet_info
*vi
, struct receive_queue
*rq
,
1197 if (vi
->mergeable_rx_bufs
)
1198 err
= add_recvbuf_mergeable(vi
, rq
, gfp
);
1199 else if (vi
->big_packets
)
1200 err
= add_recvbuf_big(vi
, rq
, gfp
);
1202 err
= add_recvbuf_small(vi
, rq
, gfp
);
1204 oom
= err
== -ENOMEM
;
1207 } while (rq
->vq
->num_free
);
1208 if (virtqueue_kick_prepare(rq
->vq
) && virtqueue_notify(rq
->vq
)) {
1209 u64_stats_update_begin(&rq
->stats
.syncp
);
1211 u64_stats_update_end(&rq
->stats
.syncp
);
1217 static void skb_recv_done(struct virtqueue
*rvq
)
1219 struct virtnet_info
*vi
= rvq
->vdev
->priv
;
1220 struct receive_queue
*rq
= &vi
->rq
[vq2rxq(rvq
)];
1222 virtqueue_napi_schedule(&rq
->napi
, rvq
);
1225 static void virtnet_napi_enable(struct virtqueue
*vq
, struct napi_struct
*napi
)
1229 /* If all buffers were filled by other side before we napi_enabled, we
1230 * won't get another interrupt, so process any outstanding packets now.
1231 * Call local_bh_enable after to trigger softIRQ processing.
1234 virtqueue_napi_schedule(napi
, vq
);
1238 static void virtnet_napi_tx_enable(struct virtnet_info
*vi
,
1239 struct virtqueue
*vq
,
1240 struct napi_struct
*napi
)
1245 /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
1246 * enable the feature if this is likely affine with the transmit path.
1248 if (!vi
->affinity_hint_set
) {
1253 return virtnet_napi_enable(vq
, napi
);
1256 static void virtnet_napi_tx_disable(struct napi_struct
*napi
)
1262 static void refill_work(struct work_struct
*work
)
1264 struct virtnet_info
*vi
=
1265 container_of(work
, struct virtnet_info
, refill
.work
);
1269 for (i
= 0; i
< vi
->curr_queue_pairs
; i
++) {
1270 struct receive_queue
*rq
= &vi
->rq
[i
];
1272 napi_disable(&rq
->napi
);
1273 still_empty
= !try_fill_recv(vi
, rq
, GFP_KERNEL
);
1274 virtnet_napi_enable(rq
->vq
, &rq
->napi
);
1276 /* In theory, this can happen: if we don't get any buffers in
1277 * we will *never* try to fill again.
1280 schedule_delayed_work(&vi
->refill
, HZ
/2);
1284 static int virtnet_receive(struct receive_queue
*rq
, int budget
,
1285 unsigned int *xdp_xmit
)
1287 struct virtnet_info
*vi
= rq
->vq
->vdev
->priv
;
1288 struct virtnet_rq_stats stats
= {};
1293 if (!vi
->big_packets
|| vi
->mergeable_rx_bufs
) {
1296 while (stats
.packets
< budget
&&
1297 (buf
= virtqueue_get_buf_ctx(rq
->vq
, &len
, &ctx
))) {
1298 receive_buf(vi
, rq
, buf
, len
, ctx
, xdp_xmit
, &stats
);
1302 while (stats
.packets
< budget
&&
1303 (buf
= virtqueue_get_buf(rq
->vq
, &len
)) != NULL
) {
1304 receive_buf(vi
, rq
, buf
, len
, NULL
, xdp_xmit
, &stats
);
1309 if (rq
->vq
->num_free
> virtqueue_get_vring_size(rq
->vq
) / 2) {
1310 if (!try_fill_recv(vi
, rq
, GFP_ATOMIC
))
1311 schedule_delayed_work(&vi
->refill
, 0);
1314 u64_stats_update_begin(&rq
->stats
.syncp
);
1315 for (i
= 0; i
< VIRTNET_RQ_STATS_LEN
; i
++) {
1316 size_t offset
= virtnet_rq_stats_desc
[i
].offset
;
1319 item
= (u64
*)((u8
*)&rq
->stats
+ offset
);
1320 *item
+= *(u64
*)((u8
*)&stats
+ offset
);
1322 u64_stats_update_end(&rq
->stats
.syncp
);
1324 return stats
.packets
;
1327 static void free_old_xmit_skbs(struct send_queue
*sq
)
1329 struct sk_buff
*skb
;
1331 unsigned int packets
= 0;
1332 unsigned int bytes
= 0;
1334 while ((skb
= virtqueue_get_buf(sq
->vq
, &len
)) != NULL
) {
1335 pr_debug("Sent skb %p\n", skb
);
1340 dev_consume_skb_any(skb
);
1343 /* Avoid overhead when no packets have been processed
1344 * happens when called speculatively from start_xmit.
1349 u64_stats_update_begin(&sq
->stats
.syncp
);
1350 sq
->stats
.bytes
+= bytes
;
1351 sq
->stats
.packets
+= packets
;
1352 u64_stats_update_end(&sq
->stats
.syncp
);
1355 static void virtnet_poll_cleantx(struct receive_queue
*rq
)
1357 struct virtnet_info
*vi
= rq
->vq
->vdev
->priv
;
1358 unsigned int index
= vq2rxq(rq
->vq
);
1359 struct send_queue
*sq
= &vi
->sq
[index
];
1360 struct netdev_queue
*txq
= netdev_get_tx_queue(vi
->dev
, index
);
1362 if (!sq
->napi
.weight
)
1365 if (__netif_tx_trylock(txq
)) {
1366 free_old_xmit_skbs(sq
);
1367 __netif_tx_unlock(txq
);
1370 if (sq
->vq
->num_free
>= 2 + MAX_SKB_FRAGS
)
1371 netif_tx_wake_queue(txq
);
1374 static int virtnet_poll(struct napi_struct
*napi
, int budget
)
1376 struct receive_queue
*rq
=
1377 container_of(napi
, struct receive_queue
, napi
);
1378 struct virtnet_info
*vi
= rq
->vq
->vdev
->priv
;
1379 struct send_queue
*sq
;
1380 unsigned int received
;
1381 unsigned int xdp_xmit
= 0;
1383 virtnet_poll_cleantx(rq
);
1385 received
= virtnet_receive(rq
, budget
, &xdp_xmit
);
1387 /* Out of packets? */
1388 if (received
< budget
)
1389 virtqueue_napi_complete(napi
, rq
->vq
, received
);
1391 if (xdp_xmit
& VIRTIO_XDP_REDIR
)
1394 if (xdp_xmit
& VIRTIO_XDP_TX
) {
1395 sq
= virtnet_xdp_sq(vi
);
1396 if (virtqueue_kick_prepare(sq
->vq
) && virtqueue_notify(sq
->vq
)) {
1397 u64_stats_update_begin(&sq
->stats
.syncp
);
1399 u64_stats_update_end(&sq
->stats
.syncp
);
1406 static int virtnet_open(struct net_device
*dev
)
1408 struct virtnet_info
*vi
= netdev_priv(dev
);
1411 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1412 if (i
< vi
->curr_queue_pairs
)
1413 /* Make sure we have some buffers: if oom use wq. */
1414 if (!try_fill_recv(vi
, &vi
->rq
[i
], GFP_KERNEL
))
1415 schedule_delayed_work(&vi
->refill
, 0);
1417 err
= xdp_rxq_info_reg(&vi
->rq
[i
].xdp_rxq
, dev
, i
);
1421 err
= xdp_rxq_info_reg_mem_model(&vi
->rq
[i
].xdp_rxq
,
1422 MEM_TYPE_PAGE_SHARED
, NULL
);
1424 xdp_rxq_info_unreg(&vi
->rq
[i
].xdp_rxq
);
1428 virtnet_napi_enable(vi
->rq
[i
].vq
, &vi
->rq
[i
].napi
);
1429 virtnet_napi_tx_enable(vi
, vi
->sq
[i
].vq
, &vi
->sq
[i
].napi
);
1435 static int virtnet_poll_tx(struct napi_struct
*napi
, int budget
)
1437 struct send_queue
*sq
= container_of(napi
, struct send_queue
, napi
);
1438 struct virtnet_info
*vi
= sq
->vq
->vdev
->priv
;
1439 struct netdev_queue
*txq
= netdev_get_tx_queue(vi
->dev
, vq2txq(sq
->vq
));
1441 __netif_tx_lock(txq
, raw_smp_processor_id());
1442 free_old_xmit_skbs(sq
);
1443 __netif_tx_unlock(txq
);
1445 virtqueue_napi_complete(napi
, sq
->vq
, 0);
1447 if (sq
->vq
->num_free
>= 2 + MAX_SKB_FRAGS
)
1448 netif_tx_wake_queue(txq
);
1453 static int xmit_skb(struct send_queue
*sq
, struct sk_buff
*skb
)
1455 struct virtio_net_hdr_mrg_rxbuf
*hdr
;
1456 const unsigned char *dest
= ((struct ethhdr
*)skb
->data
)->h_dest
;
1457 struct virtnet_info
*vi
= sq
->vq
->vdev
->priv
;
1459 unsigned hdr_len
= vi
->hdr_len
;
1462 pr_debug("%s: xmit %p %pM\n", vi
->dev
->name
, skb
, dest
);
1464 can_push
= vi
->any_header_sg
&&
1465 !((unsigned long)skb
->data
& (__alignof__(*hdr
) - 1)) &&
1466 !skb_header_cloned(skb
) && skb_headroom(skb
) >= hdr_len
;
1467 /* Even if we can, don't push here yet as this would skew
1468 * csum_start offset below. */
1470 hdr
= (struct virtio_net_hdr_mrg_rxbuf
*)(skb
->data
- hdr_len
);
1472 hdr
= skb_vnet_hdr(skb
);
1474 if (virtio_net_hdr_from_skb(skb
, &hdr
->hdr
,
1475 virtio_is_little_endian(vi
->vdev
), false,
1479 if (vi
->mergeable_rx_bufs
)
1480 hdr
->num_buffers
= 0;
1482 sg_init_table(sq
->sg
, skb_shinfo(skb
)->nr_frags
+ (can_push
? 1 : 2));
1484 __skb_push(skb
, hdr_len
);
1485 num_sg
= skb_to_sgvec(skb
, sq
->sg
, 0, skb
->len
);
1486 if (unlikely(num_sg
< 0))
1488 /* Pull header back to avoid skew in tx bytes calculations. */
1489 __skb_pull(skb
, hdr_len
);
1491 sg_set_buf(sq
->sg
, hdr
, hdr_len
);
1492 num_sg
= skb_to_sgvec(skb
, sq
->sg
+ 1, 0, skb
->len
);
1493 if (unlikely(num_sg
< 0))
1497 return virtqueue_add_outbuf(sq
->vq
, sq
->sg
, num_sg
, skb
, GFP_ATOMIC
);
1500 static netdev_tx_t
start_xmit(struct sk_buff
*skb
, struct net_device
*dev
)
1502 struct virtnet_info
*vi
= netdev_priv(dev
);
1503 int qnum
= skb_get_queue_mapping(skb
);
1504 struct send_queue
*sq
= &vi
->sq
[qnum
];
1506 struct netdev_queue
*txq
= netdev_get_tx_queue(dev
, qnum
);
1507 bool kick
= !skb
->xmit_more
;
1508 bool use_napi
= sq
->napi
.weight
;
1510 /* Free up any pending old buffers before queueing new ones. */
1511 free_old_xmit_skbs(sq
);
1513 if (use_napi
&& kick
)
1514 virtqueue_enable_cb_delayed(sq
->vq
);
1516 /* timestamp packet in software */
1517 skb_tx_timestamp(skb
);
1519 /* Try to transmit */
1520 err
= xmit_skb(sq
, skb
);
1522 /* This should not happen! */
1523 if (unlikely(err
)) {
1524 dev
->stats
.tx_fifo_errors
++;
1525 if (net_ratelimit())
1527 "Unexpected TXQ (%d) queue failure: %d\n", qnum
, err
);
1528 dev
->stats
.tx_dropped
++;
1529 dev_kfree_skb_any(skb
);
1530 return NETDEV_TX_OK
;
1533 /* Don't wait up for transmitted skbs to be freed. */
1539 /* If running out of space, stop queue to avoid getting packets that we
1540 * are then unable to transmit.
1541 * An alternative would be to force queuing layer to requeue the skb by
1542 * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1543 * returned in a normal path of operation: it means that driver is not
1544 * maintaining the TX queue stop/start state properly, and causes
1545 * the stack to do a non-trivial amount of useless work.
1546 * Since most packets only take 1 or 2 ring slots, stopping the queue
1547 * early means 16 slots are typically wasted.
1549 if (sq
->vq
->num_free
< 2+MAX_SKB_FRAGS
) {
1550 netif_stop_subqueue(dev
, qnum
);
1552 unlikely(!virtqueue_enable_cb_delayed(sq
->vq
))) {
1553 /* More just got used, free them then recheck. */
1554 free_old_xmit_skbs(sq
);
1555 if (sq
->vq
->num_free
>= 2+MAX_SKB_FRAGS
) {
1556 netif_start_subqueue(dev
, qnum
);
1557 virtqueue_disable_cb(sq
->vq
);
1562 if (kick
|| netif_xmit_stopped(txq
)) {
1563 if (virtqueue_kick_prepare(sq
->vq
) && virtqueue_notify(sq
->vq
)) {
1564 u64_stats_update_begin(&sq
->stats
.syncp
);
1566 u64_stats_update_end(&sq
->stats
.syncp
);
1570 return NETDEV_TX_OK
;
1574 * Send command via the control virtqueue and check status. Commands
1575 * supported by the hypervisor, as indicated by feature bits, should
1576 * never fail unless improperly formatted.
1578 static bool virtnet_send_command(struct virtnet_info
*vi
, u8
class, u8 cmd
,
1579 struct scatterlist
*out
)
1581 struct scatterlist
*sgs
[4], hdr
, stat
;
1582 unsigned out_num
= 0, tmp
;
1584 /* Caller should know better */
1585 BUG_ON(!virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_CTRL_VQ
));
1587 vi
->ctrl
->status
= ~0;
1588 vi
->ctrl
->hdr
.class = class;
1589 vi
->ctrl
->hdr
.cmd
= cmd
;
1591 sg_init_one(&hdr
, &vi
->ctrl
->hdr
, sizeof(vi
->ctrl
->hdr
));
1592 sgs
[out_num
++] = &hdr
;
1595 sgs
[out_num
++] = out
;
1597 /* Add return status. */
1598 sg_init_one(&stat
, &vi
->ctrl
->status
, sizeof(vi
->ctrl
->status
));
1599 sgs
[out_num
] = &stat
;
1601 BUG_ON(out_num
+ 1 > ARRAY_SIZE(sgs
));
1602 virtqueue_add_sgs(vi
->cvq
, sgs
, out_num
, 1, vi
, GFP_ATOMIC
);
1604 if (unlikely(!virtqueue_kick(vi
->cvq
)))
1605 return vi
->ctrl
->status
== VIRTIO_NET_OK
;
1607 /* Spin for a response, the kick causes an ioport write, trapping
1608 * into the hypervisor, so the request should be handled immediately.
1610 while (!virtqueue_get_buf(vi
->cvq
, &tmp
) &&
1611 !virtqueue_is_broken(vi
->cvq
))
1614 return vi
->ctrl
->status
== VIRTIO_NET_OK
;
1617 static int virtnet_set_mac_address(struct net_device
*dev
, void *p
)
1619 struct virtnet_info
*vi
= netdev_priv(dev
);
1620 struct virtio_device
*vdev
= vi
->vdev
;
1622 struct sockaddr
*addr
;
1623 struct scatterlist sg
;
1625 if (virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_STANDBY
))
1628 addr
= kmemdup(p
, sizeof(*addr
), GFP_KERNEL
);
1632 ret
= eth_prepare_mac_addr_change(dev
, addr
);
1636 if (virtio_has_feature(vdev
, VIRTIO_NET_F_CTRL_MAC_ADDR
)) {
1637 sg_init_one(&sg
, addr
->sa_data
, dev
->addr_len
);
1638 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_MAC
,
1639 VIRTIO_NET_CTRL_MAC_ADDR_SET
, &sg
)) {
1640 dev_warn(&vdev
->dev
,
1641 "Failed to set mac address by vq command.\n");
1645 } else if (virtio_has_feature(vdev
, VIRTIO_NET_F_MAC
) &&
1646 !virtio_has_feature(vdev
, VIRTIO_F_VERSION_1
)) {
1649 /* Naturally, this has an atomicity problem. */
1650 for (i
= 0; i
< dev
->addr_len
; i
++)
1651 virtio_cwrite8(vdev
,
1652 offsetof(struct virtio_net_config
, mac
) +
1653 i
, addr
->sa_data
[i
]);
1656 eth_commit_mac_addr_change(dev
, p
);
1664 static void virtnet_stats(struct net_device
*dev
,
1665 struct rtnl_link_stats64
*tot
)
1667 struct virtnet_info
*vi
= netdev_priv(dev
);
1671 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1672 u64 tpackets
, tbytes
, rpackets
, rbytes
, rdrops
;
1673 struct receive_queue
*rq
= &vi
->rq
[i
];
1674 struct send_queue
*sq
= &vi
->sq
[i
];
1677 start
= u64_stats_fetch_begin_irq(&sq
->stats
.syncp
);
1678 tpackets
= sq
->stats
.packets
;
1679 tbytes
= sq
->stats
.bytes
;
1680 } while (u64_stats_fetch_retry_irq(&sq
->stats
.syncp
, start
));
1683 start
= u64_stats_fetch_begin_irq(&rq
->stats
.syncp
);
1684 rpackets
= rq
->stats
.packets
;
1685 rbytes
= rq
->stats
.bytes
;
1686 rdrops
= rq
->stats
.drops
;
1687 } while (u64_stats_fetch_retry_irq(&rq
->stats
.syncp
, start
));
1689 tot
->rx_packets
+= rpackets
;
1690 tot
->tx_packets
+= tpackets
;
1691 tot
->rx_bytes
+= rbytes
;
1692 tot
->tx_bytes
+= tbytes
;
1693 tot
->rx_dropped
+= rdrops
;
1696 tot
->tx_dropped
= dev
->stats
.tx_dropped
;
1697 tot
->tx_fifo_errors
= dev
->stats
.tx_fifo_errors
;
1698 tot
->rx_length_errors
= dev
->stats
.rx_length_errors
;
1699 tot
->rx_frame_errors
= dev
->stats
.rx_frame_errors
;
1702 static void virtnet_ack_link_announce(struct virtnet_info
*vi
)
1705 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_ANNOUNCE
,
1706 VIRTIO_NET_CTRL_ANNOUNCE_ACK
, NULL
))
1707 dev_warn(&vi
->dev
->dev
, "Failed to ack link announce.\n");
1711 static int _virtnet_set_queues(struct virtnet_info
*vi
, u16 queue_pairs
)
1713 struct scatterlist sg
;
1714 struct net_device
*dev
= vi
->dev
;
1716 if (!vi
->has_cvq
|| !virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_MQ
))
1719 vi
->ctrl
->mq
.virtqueue_pairs
= cpu_to_virtio16(vi
->vdev
, queue_pairs
);
1720 sg_init_one(&sg
, &vi
->ctrl
->mq
, sizeof(vi
->ctrl
->mq
));
1722 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_MQ
,
1723 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET
, &sg
)) {
1724 dev_warn(&dev
->dev
, "Fail to set num of queue pairs to %d\n",
1728 vi
->curr_queue_pairs
= queue_pairs
;
1729 /* virtnet_open() will refill when device is going to up. */
1730 if (dev
->flags
& IFF_UP
)
1731 schedule_delayed_work(&vi
->refill
, 0);
1737 static int virtnet_set_queues(struct virtnet_info
*vi
, u16 queue_pairs
)
1742 err
= _virtnet_set_queues(vi
, queue_pairs
);
1747 static int virtnet_close(struct net_device
*dev
)
1749 struct virtnet_info
*vi
= netdev_priv(dev
);
1752 /* Make sure refill_work doesn't re-enable napi! */
1753 cancel_delayed_work_sync(&vi
->refill
);
1755 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1756 xdp_rxq_info_unreg(&vi
->rq
[i
].xdp_rxq
);
1757 napi_disable(&vi
->rq
[i
].napi
);
1758 virtnet_napi_tx_disable(&vi
->sq
[i
].napi
);
1764 static void virtnet_set_rx_mode(struct net_device
*dev
)
1766 struct virtnet_info
*vi
= netdev_priv(dev
);
1767 struct scatterlist sg
[2];
1768 struct virtio_net_ctrl_mac
*mac_data
;
1769 struct netdev_hw_addr
*ha
;
1775 /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1776 if (!virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_CTRL_RX
))
1779 vi
->ctrl
->promisc
= ((dev
->flags
& IFF_PROMISC
) != 0);
1780 vi
->ctrl
->allmulti
= ((dev
->flags
& IFF_ALLMULTI
) != 0);
1782 sg_init_one(sg
, &vi
->ctrl
->promisc
, sizeof(vi
->ctrl
->promisc
));
1784 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_RX
,
1785 VIRTIO_NET_CTRL_RX_PROMISC
, sg
))
1786 dev_warn(&dev
->dev
, "Failed to %sable promisc mode.\n",
1787 vi
->ctrl
->promisc
? "en" : "dis");
1789 sg_init_one(sg
, &vi
->ctrl
->allmulti
, sizeof(vi
->ctrl
->allmulti
));
1791 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_RX
,
1792 VIRTIO_NET_CTRL_RX_ALLMULTI
, sg
))
1793 dev_warn(&dev
->dev
, "Failed to %sable allmulti mode.\n",
1794 vi
->ctrl
->allmulti
? "en" : "dis");
1796 uc_count
= netdev_uc_count(dev
);
1797 mc_count
= netdev_mc_count(dev
);
1798 /* MAC filter - use one buffer for both lists */
1799 buf
= kzalloc(((uc_count
+ mc_count
) * ETH_ALEN
) +
1800 (2 * sizeof(mac_data
->entries
)), GFP_ATOMIC
);
1805 sg_init_table(sg
, 2);
1807 /* Store the unicast list and count in the front of the buffer */
1808 mac_data
->entries
= cpu_to_virtio32(vi
->vdev
, uc_count
);
1810 netdev_for_each_uc_addr(ha
, dev
)
1811 memcpy(&mac_data
->macs
[i
++][0], ha
->addr
, ETH_ALEN
);
1813 sg_set_buf(&sg
[0], mac_data
,
1814 sizeof(mac_data
->entries
) + (uc_count
* ETH_ALEN
));
1816 /* multicast list and count fill the end */
1817 mac_data
= (void *)&mac_data
->macs
[uc_count
][0];
1819 mac_data
->entries
= cpu_to_virtio32(vi
->vdev
, mc_count
);
1821 netdev_for_each_mc_addr(ha
, dev
)
1822 memcpy(&mac_data
->macs
[i
++][0], ha
->addr
, ETH_ALEN
);
1824 sg_set_buf(&sg
[1], mac_data
,
1825 sizeof(mac_data
->entries
) + (mc_count
* ETH_ALEN
));
1827 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_MAC
,
1828 VIRTIO_NET_CTRL_MAC_TABLE_SET
, sg
))
1829 dev_warn(&dev
->dev
, "Failed to set MAC filter table.\n");
1834 static int virtnet_vlan_rx_add_vid(struct net_device
*dev
,
1835 __be16 proto
, u16 vid
)
1837 struct virtnet_info
*vi
= netdev_priv(dev
);
1838 struct scatterlist sg
;
1840 vi
->ctrl
->vid
= cpu_to_virtio16(vi
->vdev
, vid
);
1841 sg_init_one(&sg
, &vi
->ctrl
->vid
, sizeof(vi
->ctrl
->vid
));
1843 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_VLAN
,
1844 VIRTIO_NET_CTRL_VLAN_ADD
, &sg
))
1845 dev_warn(&dev
->dev
, "Failed to add VLAN ID %d.\n", vid
);
1849 static int virtnet_vlan_rx_kill_vid(struct net_device
*dev
,
1850 __be16 proto
, u16 vid
)
1852 struct virtnet_info
*vi
= netdev_priv(dev
);
1853 struct scatterlist sg
;
1855 vi
->ctrl
->vid
= cpu_to_virtio16(vi
->vdev
, vid
);
1856 sg_init_one(&sg
, &vi
->ctrl
->vid
, sizeof(vi
->ctrl
->vid
));
1858 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_VLAN
,
1859 VIRTIO_NET_CTRL_VLAN_DEL
, &sg
))
1860 dev_warn(&dev
->dev
, "Failed to kill VLAN ID %d.\n", vid
);
1864 static void virtnet_clean_affinity(struct virtnet_info
*vi
, long hcpu
)
1868 if (vi
->affinity_hint_set
) {
1869 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
1870 virtqueue_set_affinity(vi
->rq
[i
].vq
, NULL
);
1871 virtqueue_set_affinity(vi
->sq
[i
].vq
, NULL
);
1874 vi
->affinity_hint_set
= false;
1878 static void virtnet_set_affinity(struct virtnet_info
*vi
)
1887 if (!zalloc_cpumask_var(&mask
, GFP_KERNEL
)) {
1888 virtnet_clean_affinity(vi
, -1);
1892 num_cpu
= num_online_cpus();
1893 stride
= max_t(int, num_cpu
/ vi
->curr_queue_pairs
, 1);
1894 stragglers
= num_cpu
>= vi
->curr_queue_pairs
?
1895 num_cpu
% vi
->curr_queue_pairs
:
1897 cpu
= cpumask_next(-1, cpu_online_mask
);
1899 for (i
= 0; i
< vi
->curr_queue_pairs
; i
++) {
1900 group_size
= stride
+ (i
< stragglers
? 1 : 0);
1902 for (j
= 0; j
< group_size
; j
++) {
1903 cpumask_set_cpu(cpu
, mask
);
1904 cpu
= cpumask_next_wrap(cpu
, cpu_online_mask
,
1907 virtqueue_set_affinity(vi
->rq
[i
].vq
, mask
);
1908 virtqueue_set_affinity(vi
->sq
[i
].vq
, mask
);
1909 __netif_set_xps_queue(vi
->dev
, cpumask_bits(mask
), i
, false);
1910 cpumask_clear(mask
);
1913 vi
->affinity_hint_set
= true;
1914 free_cpumask_var(mask
);
1917 static int virtnet_cpu_online(unsigned int cpu
, struct hlist_node
*node
)
1919 struct virtnet_info
*vi
= hlist_entry_safe(node
, struct virtnet_info
,
1921 virtnet_set_affinity(vi
);
1925 static int virtnet_cpu_dead(unsigned int cpu
, struct hlist_node
*node
)
1927 struct virtnet_info
*vi
= hlist_entry_safe(node
, struct virtnet_info
,
1929 virtnet_set_affinity(vi
);
1933 static int virtnet_cpu_down_prep(unsigned int cpu
, struct hlist_node
*node
)
1935 struct virtnet_info
*vi
= hlist_entry_safe(node
, struct virtnet_info
,
1938 virtnet_clean_affinity(vi
, cpu
);
1942 static enum cpuhp_state virtionet_online
;
1944 static int virtnet_cpu_notif_add(struct virtnet_info
*vi
)
1948 ret
= cpuhp_state_add_instance_nocalls(virtionet_online
, &vi
->node
);
1951 ret
= cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD
,
1955 cpuhp_state_remove_instance_nocalls(virtionet_online
, &vi
->node
);
1959 static void virtnet_cpu_notif_remove(struct virtnet_info
*vi
)
1961 cpuhp_state_remove_instance_nocalls(virtionet_online
, &vi
->node
);
1962 cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD
,
1966 static void virtnet_get_ringparam(struct net_device
*dev
,
1967 struct ethtool_ringparam
*ring
)
1969 struct virtnet_info
*vi
= netdev_priv(dev
);
1971 ring
->rx_max_pending
= virtqueue_get_vring_size(vi
->rq
[0].vq
);
1972 ring
->tx_max_pending
= virtqueue_get_vring_size(vi
->sq
[0].vq
);
1973 ring
->rx_pending
= ring
->rx_max_pending
;
1974 ring
->tx_pending
= ring
->tx_max_pending
;
1978 static void virtnet_get_drvinfo(struct net_device
*dev
,
1979 struct ethtool_drvinfo
*info
)
1981 struct virtnet_info
*vi
= netdev_priv(dev
);
1982 struct virtio_device
*vdev
= vi
->vdev
;
1984 strlcpy(info
->driver
, KBUILD_MODNAME
, sizeof(info
->driver
));
1985 strlcpy(info
->version
, VIRTNET_DRIVER_VERSION
, sizeof(info
->version
));
1986 strlcpy(info
->bus_info
, virtio_bus_name(vdev
), sizeof(info
->bus_info
));
1990 /* TODO: Eliminate OOO packets during switching */
1991 static int virtnet_set_channels(struct net_device
*dev
,
1992 struct ethtool_channels
*channels
)
1994 struct virtnet_info
*vi
= netdev_priv(dev
);
1995 u16 queue_pairs
= channels
->combined_count
;
1998 /* We don't support separate rx/tx channels.
1999 * We don't allow setting 'other' channels.
2001 if (channels
->rx_count
|| channels
->tx_count
|| channels
->other_count
)
2004 if (queue_pairs
> vi
->max_queue_pairs
|| queue_pairs
== 0)
2007 /* For now we don't support modifying channels while XDP is loaded
2008 * also when XDP is loaded all RX queues have XDP programs so we only
2009 * need to check a single RX queue.
2011 if (vi
->rq
[0].xdp_prog
)
2015 err
= _virtnet_set_queues(vi
, queue_pairs
);
2017 netif_set_real_num_tx_queues(dev
, queue_pairs
);
2018 netif_set_real_num_rx_queues(dev
, queue_pairs
);
2020 virtnet_set_affinity(vi
);
2027 static void virtnet_get_strings(struct net_device
*dev
, u32 stringset
, u8
*data
)
2029 struct virtnet_info
*vi
= netdev_priv(dev
);
2030 char *p
= (char *)data
;
2033 switch (stringset
) {
2035 for (i
= 0; i
< vi
->curr_queue_pairs
; i
++) {
2036 for (j
= 0; j
< VIRTNET_RQ_STATS_LEN
; j
++) {
2037 snprintf(p
, ETH_GSTRING_LEN
, "rx_queue_%u_%s",
2038 i
, virtnet_rq_stats_desc
[j
].desc
);
2039 p
+= ETH_GSTRING_LEN
;
2043 for (i
= 0; i
< vi
->curr_queue_pairs
; i
++) {
2044 for (j
= 0; j
< VIRTNET_SQ_STATS_LEN
; j
++) {
2045 snprintf(p
, ETH_GSTRING_LEN
, "tx_queue_%u_%s",
2046 i
, virtnet_sq_stats_desc
[j
].desc
);
2047 p
+= ETH_GSTRING_LEN
;
2054 static int virtnet_get_sset_count(struct net_device
*dev
, int sset
)
2056 struct virtnet_info
*vi
= netdev_priv(dev
);
2060 return vi
->curr_queue_pairs
* (VIRTNET_RQ_STATS_LEN
+
2061 VIRTNET_SQ_STATS_LEN
);
2067 static void virtnet_get_ethtool_stats(struct net_device
*dev
,
2068 struct ethtool_stats
*stats
, u64
*data
)
2070 struct virtnet_info
*vi
= netdev_priv(dev
);
2071 unsigned int idx
= 0, start
, i
, j
;
2072 const u8
*stats_base
;
2075 for (i
= 0; i
< vi
->curr_queue_pairs
; i
++) {
2076 struct receive_queue
*rq
= &vi
->rq
[i
];
2078 stats_base
= (u8
*)&rq
->stats
;
2080 start
= u64_stats_fetch_begin_irq(&rq
->stats
.syncp
);
2081 for (j
= 0; j
< VIRTNET_RQ_STATS_LEN
; j
++) {
2082 offset
= virtnet_rq_stats_desc
[j
].offset
;
2083 data
[idx
+ j
] = *(u64
*)(stats_base
+ offset
);
2085 } while (u64_stats_fetch_retry_irq(&rq
->stats
.syncp
, start
));
2086 idx
+= VIRTNET_RQ_STATS_LEN
;
2089 for (i
= 0; i
< vi
->curr_queue_pairs
; i
++) {
2090 struct send_queue
*sq
= &vi
->sq
[i
];
2092 stats_base
= (u8
*)&sq
->stats
;
2094 start
= u64_stats_fetch_begin_irq(&sq
->stats
.syncp
);
2095 for (j
= 0; j
< VIRTNET_SQ_STATS_LEN
; j
++) {
2096 offset
= virtnet_sq_stats_desc
[j
].offset
;
2097 data
[idx
+ j
] = *(u64
*)(stats_base
+ offset
);
2099 } while (u64_stats_fetch_retry_irq(&sq
->stats
.syncp
, start
));
2100 idx
+= VIRTNET_SQ_STATS_LEN
;
2104 static void virtnet_get_channels(struct net_device
*dev
,
2105 struct ethtool_channels
*channels
)
2107 struct virtnet_info
*vi
= netdev_priv(dev
);
2109 channels
->combined_count
= vi
->curr_queue_pairs
;
2110 channels
->max_combined
= vi
->max_queue_pairs
;
2111 channels
->max_other
= 0;
2112 channels
->rx_count
= 0;
2113 channels
->tx_count
= 0;
2114 channels
->other_count
= 0;
2117 /* Check if the user is trying to change anything besides speed/duplex */
2119 virtnet_validate_ethtool_cmd(const struct ethtool_link_ksettings
*cmd
)
2121 struct ethtool_link_ksettings diff1
= *cmd
;
2122 struct ethtool_link_ksettings diff2
= {};
2124 /* cmd is always set so we need to clear it, validate the port type
2125 * and also without autonegotiation we can ignore advertising
2127 diff1
.base
.speed
= 0;
2128 diff2
.base
.port
= PORT_OTHER
;
2129 ethtool_link_ksettings_zero_link_mode(&diff1
, advertising
);
2130 diff1
.base
.duplex
= 0;
2132 diff1
.base
.link_mode_masks_nwords
= 0;
2134 return !memcmp(&diff1
.base
, &diff2
.base
, sizeof(diff1
.base
)) &&
2135 bitmap_empty(diff1
.link_modes
.supported
,
2136 __ETHTOOL_LINK_MODE_MASK_NBITS
) &&
2137 bitmap_empty(diff1
.link_modes
.advertising
,
2138 __ETHTOOL_LINK_MODE_MASK_NBITS
) &&
2139 bitmap_empty(diff1
.link_modes
.lp_advertising
,
2140 __ETHTOOL_LINK_MODE_MASK_NBITS
);
2143 static int virtnet_set_link_ksettings(struct net_device
*dev
,
2144 const struct ethtool_link_ksettings
*cmd
)
2146 struct virtnet_info
*vi
= netdev_priv(dev
);
2149 speed
= cmd
->base
.speed
;
2150 /* don't allow custom speed and duplex */
2151 if (!ethtool_validate_speed(speed
) ||
2152 !ethtool_validate_duplex(cmd
->base
.duplex
) ||
2153 !virtnet_validate_ethtool_cmd(cmd
))
2156 vi
->duplex
= cmd
->base
.duplex
;
2161 static int virtnet_get_link_ksettings(struct net_device
*dev
,
2162 struct ethtool_link_ksettings
*cmd
)
2164 struct virtnet_info
*vi
= netdev_priv(dev
);
2166 cmd
->base
.speed
= vi
->speed
;
2167 cmd
->base
.duplex
= vi
->duplex
;
2168 cmd
->base
.port
= PORT_OTHER
;
2173 static void virtnet_init_settings(struct net_device
*dev
)
2175 struct virtnet_info
*vi
= netdev_priv(dev
);
2177 vi
->speed
= SPEED_UNKNOWN
;
2178 vi
->duplex
= DUPLEX_UNKNOWN
;
2181 static void virtnet_update_settings(struct virtnet_info
*vi
)
2186 if (!virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_SPEED_DUPLEX
))
2189 speed
= virtio_cread32(vi
->vdev
, offsetof(struct virtio_net_config
,
2191 if (ethtool_validate_speed(speed
))
2193 duplex
= virtio_cread8(vi
->vdev
, offsetof(struct virtio_net_config
,
2195 if (ethtool_validate_duplex(duplex
))
2196 vi
->duplex
= duplex
;
2199 static const struct ethtool_ops virtnet_ethtool_ops
= {
2200 .get_drvinfo
= virtnet_get_drvinfo
,
2201 .get_link
= ethtool_op_get_link
,
2202 .get_ringparam
= virtnet_get_ringparam
,
2203 .get_strings
= virtnet_get_strings
,
2204 .get_sset_count
= virtnet_get_sset_count
,
2205 .get_ethtool_stats
= virtnet_get_ethtool_stats
,
2206 .set_channels
= virtnet_set_channels
,
2207 .get_channels
= virtnet_get_channels
,
2208 .get_ts_info
= ethtool_op_get_ts_info
,
2209 .get_link_ksettings
= virtnet_get_link_ksettings
,
2210 .set_link_ksettings
= virtnet_set_link_ksettings
,
2213 static void virtnet_freeze_down(struct virtio_device
*vdev
)
2215 struct virtnet_info
*vi
= vdev
->priv
;
2218 /* Make sure no work handler is accessing the device */
2219 flush_work(&vi
->config_work
);
2221 netif_device_detach(vi
->dev
);
2222 netif_tx_disable(vi
->dev
);
2223 cancel_delayed_work_sync(&vi
->refill
);
2225 if (netif_running(vi
->dev
)) {
2226 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2227 napi_disable(&vi
->rq
[i
].napi
);
2228 virtnet_napi_tx_disable(&vi
->sq
[i
].napi
);
2233 static int init_vqs(struct virtnet_info
*vi
);
2235 static int virtnet_restore_up(struct virtio_device
*vdev
)
2237 struct virtnet_info
*vi
= vdev
->priv
;
2244 virtio_device_ready(vdev
);
2246 if (netif_running(vi
->dev
)) {
2247 for (i
= 0; i
< vi
->curr_queue_pairs
; i
++)
2248 if (!try_fill_recv(vi
, &vi
->rq
[i
], GFP_KERNEL
))
2249 schedule_delayed_work(&vi
->refill
, 0);
2251 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2252 virtnet_napi_enable(vi
->rq
[i
].vq
, &vi
->rq
[i
].napi
);
2253 virtnet_napi_tx_enable(vi
, vi
->sq
[i
].vq
,
2258 netif_device_attach(vi
->dev
);
2262 static int virtnet_set_guest_offloads(struct virtnet_info
*vi
, u64 offloads
)
2264 struct scatterlist sg
;
2265 vi
->ctrl
->offloads
= cpu_to_virtio64(vi
->vdev
, offloads
);
2267 sg_init_one(&sg
, &vi
->ctrl
->offloads
, sizeof(vi
->ctrl
->offloads
));
2269 if (!virtnet_send_command(vi
, VIRTIO_NET_CTRL_GUEST_OFFLOADS
,
2270 VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET
, &sg
)) {
2271 dev_warn(&vi
->dev
->dev
, "Fail to set guest offload. \n");
2278 static int virtnet_clear_guest_offloads(struct virtnet_info
*vi
)
2282 if (!vi
->guest_offloads
)
2285 if (virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_CSUM
))
2286 offloads
= 1ULL << VIRTIO_NET_F_GUEST_CSUM
;
2288 return virtnet_set_guest_offloads(vi
, offloads
);
2291 static int virtnet_restore_guest_offloads(struct virtnet_info
*vi
)
2293 u64 offloads
= vi
->guest_offloads
;
2295 if (!vi
->guest_offloads
)
2297 if (virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_CSUM
))
2298 offloads
|= 1ULL << VIRTIO_NET_F_GUEST_CSUM
;
2300 return virtnet_set_guest_offloads(vi
, offloads
);
2303 static int virtnet_xdp_set(struct net_device
*dev
, struct bpf_prog
*prog
,
2304 struct netlink_ext_ack
*extack
)
2306 unsigned long int max_sz
= PAGE_SIZE
- sizeof(struct padded_vnet_hdr
);
2307 struct virtnet_info
*vi
= netdev_priv(dev
);
2308 struct bpf_prog
*old_prog
;
2309 u16 xdp_qp
= 0, curr_qp
;
2312 if (!virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS
)
2313 && (virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_TSO4
) ||
2314 virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_TSO6
) ||
2315 virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_ECN
) ||
2316 virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_GUEST_UFO
))) {
2317 NL_SET_ERR_MSG_MOD(extack
, "Can't set XDP while host is implementing LRO, disable LRO first");
2321 if (vi
->mergeable_rx_bufs
&& !vi
->any_header_sg
) {
2322 NL_SET_ERR_MSG_MOD(extack
, "XDP expects header/data in single page, any_header_sg required");
2326 if (dev
->mtu
> max_sz
) {
2327 NL_SET_ERR_MSG_MOD(extack
, "MTU too large to enable XDP");
2328 netdev_warn(dev
, "XDP requires MTU less than %lu\n", max_sz
);
2332 curr_qp
= vi
->curr_queue_pairs
- vi
->xdp_queue_pairs
;
2334 xdp_qp
= nr_cpu_ids
;
2336 /* XDP requires extra queues for XDP_TX */
2337 if (curr_qp
+ xdp_qp
> vi
->max_queue_pairs
) {
2338 NL_SET_ERR_MSG_MOD(extack
, "Too few free TX rings available");
2339 netdev_warn(dev
, "request %i queues but max is %i\n",
2340 curr_qp
+ xdp_qp
, vi
->max_queue_pairs
);
2345 prog
= bpf_prog_add(prog
, vi
->max_queue_pairs
- 1);
2347 return PTR_ERR(prog
);
2350 /* Make sure NAPI is not using any XDP TX queues for RX. */
2351 if (netif_running(dev
))
2352 for (i
= 0; i
< vi
->max_queue_pairs
; i
++)
2353 napi_disable(&vi
->rq
[i
].napi
);
2355 netif_set_real_num_rx_queues(dev
, curr_qp
+ xdp_qp
);
2356 err
= _virtnet_set_queues(vi
, curr_qp
+ xdp_qp
);
2359 vi
->xdp_queue_pairs
= xdp_qp
;
2361 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2362 old_prog
= rtnl_dereference(vi
->rq
[i
].xdp_prog
);
2363 rcu_assign_pointer(vi
->rq
[i
].xdp_prog
, prog
);
2366 virtnet_clear_guest_offloads(vi
);
2368 virtnet_restore_guest_offloads(vi
);
2371 bpf_prog_put(old_prog
);
2372 if (netif_running(dev
))
2373 virtnet_napi_enable(vi
->rq
[i
].vq
, &vi
->rq
[i
].napi
);
2379 for (i
= 0; i
< vi
->max_queue_pairs
; i
++)
2380 virtnet_napi_enable(vi
->rq
[i
].vq
, &vi
->rq
[i
].napi
);
2382 bpf_prog_sub(prog
, vi
->max_queue_pairs
- 1);
2386 static u32
virtnet_xdp_query(struct net_device
*dev
)
2388 struct virtnet_info
*vi
= netdev_priv(dev
);
2389 const struct bpf_prog
*xdp_prog
;
2392 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2393 xdp_prog
= rtnl_dereference(vi
->rq
[i
].xdp_prog
);
2395 return xdp_prog
->aux
->id
;
2400 static int virtnet_xdp(struct net_device
*dev
, struct netdev_bpf
*xdp
)
2402 switch (xdp
->command
) {
2403 case XDP_SETUP_PROG
:
2404 return virtnet_xdp_set(dev
, xdp
->prog
, xdp
->extack
);
2405 case XDP_QUERY_PROG
:
2406 xdp
->prog_id
= virtnet_xdp_query(dev
);
2413 static int virtnet_get_phys_port_name(struct net_device
*dev
, char *buf
,
2416 struct virtnet_info
*vi
= netdev_priv(dev
);
2419 if (!virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_STANDBY
))
2422 ret
= snprintf(buf
, len
, "sby");
2429 static const struct net_device_ops virtnet_netdev
= {
2430 .ndo_open
= virtnet_open
,
2431 .ndo_stop
= virtnet_close
,
2432 .ndo_start_xmit
= start_xmit
,
2433 .ndo_validate_addr
= eth_validate_addr
,
2434 .ndo_set_mac_address
= virtnet_set_mac_address
,
2435 .ndo_set_rx_mode
= virtnet_set_rx_mode
,
2436 .ndo_get_stats64
= virtnet_stats
,
2437 .ndo_vlan_rx_add_vid
= virtnet_vlan_rx_add_vid
,
2438 .ndo_vlan_rx_kill_vid
= virtnet_vlan_rx_kill_vid
,
2439 .ndo_bpf
= virtnet_xdp
,
2440 .ndo_xdp_xmit
= virtnet_xdp_xmit
,
2441 .ndo_features_check
= passthru_features_check
,
2442 .ndo_get_phys_port_name
= virtnet_get_phys_port_name
,
2445 static void virtnet_config_changed_work(struct work_struct
*work
)
2447 struct virtnet_info
*vi
=
2448 container_of(work
, struct virtnet_info
, config_work
);
2451 if (virtio_cread_feature(vi
->vdev
, VIRTIO_NET_F_STATUS
,
2452 struct virtio_net_config
, status
, &v
) < 0)
2455 if (v
& VIRTIO_NET_S_ANNOUNCE
) {
2456 netdev_notify_peers(vi
->dev
);
2457 virtnet_ack_link_announce(vi
);
2460 /* Ignore unknown (future) status bits */
2461 v
&= VIRTIO_NET_S_LINK_UP
;
2463 if (vi
->status
== v
)
2468 if (vi
->status
& VIRTIO_NET_S_LINK_UP
) {
2469 virtnet_update_settings(vi
);
2470 netif_carrier_on(vi
->dev
);
2471 netif_tx_wake_all_queues(vi
->dev
);
2473 netif_carrier_off(vi
->dev
);
2474 netif_tx_stop_all_queues(vi
->dev
);
2478 static void virtnet_config_changed(struct virtio_device
*vdev
)
2480 struct virtnet_info
*vi
= vdev
->priv
;
2482 schedule_work(&vi
->config_work
);
2485 static void virtnet_free_queues(struct virtnet_info
*vi
)
2489 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2490 napi_hash_del(&vi
->rq
[i
].napi
);
2491 netif_napi_del(&vi
->rq
[i
].napi
);
2492 netif_napi_del(&vi
->sq
[i
].napi
);
2495 /* We called napi_hash_del() before netif_napi_del(),
2496 * we need to respect an RCU grace period before freeing vi->rq
2505 static void _free_receive_bufs(struct virtnet_info
*vi
)
2507 struct bpf_prog
*old_prog
;
2510 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2511 while (vi
->rq
[i
].pages
)
2512 __free_pages(get_a_page(&vi
->rq
[i
], GFP_KERNEL
), 0);
2514 old_prog
= rtnl_dereference(vi
->rq
[i
].xdp_prog
);
2515 RCU_INIT_POINTER(vi
->rq
[i
].xdp_prog
, NULL
);
2517 bpf_prog_put(old_prog
);
2521 static void free_receive_bufs(struct virtnet_info
*vi
)
2524 _free_receive_bufs(vi
);
2528 static void free_receive_page_frags(struct virtnet_info
*vi
)
2531 for (i
= 0; i
< vi
->max_queue_pairs
; i
++)
2532 if (vi
->rq
[i
].alloc_frag
.page
)
2533 put_page(vi
->rq
[i
].alloc_frag
.page
);
2536 static bool is_xdp_raw_buffer_queue(struct virtnet_info
*vi
, int q
)
2538 if (q
< (vi
->curr_queue_pairs
- vi
->xdp_queue_pairs
))
2540 else if (q
< vi
->curr_queue_pairs
)
2546 static void free_unused_bufs(struct virtnet_info
*vi
)
2551 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2552 struct virtqueue
*vq
= vi
->sq
[i
].vq
;
2553 while ((buf
= virtqueue_detach_unused_buf(vq
)) != NULL
) {
2554 if (!is_xdp_raw_buffer_queue(vi
, i
))
2557 put_page(virt_to_head_page(buf
));
2561 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2562 struct virtqueue
*vq
= vi
->rq
[i
].vq
;
2564 while ((buf
= virtqueue_detach_unused_buf(vq
)) != NULL
) {
2565 if (vi
->mergeable_rx_bufs
) {
2566 put_page(virt_to_head_page(buf
));
2567 } else if (vi
->big_packets
) {
2568 give_pages(&vi
->rq
[i
], buf
);
2570 put_page(virt_to_head_page(buf
));
2576 static void virtnet_del_vqs(struct virtnet_info
*vi
)
2578 struct virtio_device
*vdev
= vi
->vdev
;
2580 virtnet_clean_affinity(vi
, -1);
2582 vdev
->config
->del_vqs(vdev
);
2584 virtnet_free_queues(vi
);
2587 /* How large should a single buffer be so a queue full of these can fit at
2588 * least one full packet?
2589 * Logic below assumes the mergeable buffer header is used.
2591 static unsigned int mergeable_min_buf_len(struct virtnet_info
*vi
, struct virtqueue
*vq
)
2593 const unsigned int hdr_len
= sizeof(struct virtio_net_hdr_mrg_rxbuf
);
2594 unsigned int rq_size
= virtqueue_get_vring_size(vq
);
2595 unsigned int packet_len
= vi
->big_packets
? IP_MAX_MTU
: vi
->dev
->max_mtu
;
2596 unsigned int buf_len
= hdr_len
+ ETH_HLEN
+ VLAN_HLEN
+ packet_len
;
2597 unsigned int min_buf_len
= DIV_ROUND_UP(buf_len
, rq_size
);
2599 return max(max(min_buf_len
, hdr_len
) - hdr_len
,
2600 (unsigned int)GOOD_PACKET_LEN
);
2603 static int virtnet_find_vqs(struct virtnet_info
*vi
)
2605 vq_callback_t
**callbacks
;
2606 struct virtqueue
**vqs
;
2612 /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
2613 * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
2614 * possible control vq.
2616 total_vqs
= vi
->max_queue_pairs
* 2 +
2617 virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_CTRL_VQ
);
2619 /* Allocate space for find_vqs parameters */
2620 vqs
= kcalloc(total_vqs
, sizeof(*vqs
), GFP_KERNEL
);
2623 callbacks
= kmalloc_array(total_vqs
, sizeof(*callbacks
), GFP_KERNEL
);
2626 names
= kmalloc_array(total_vqs
, sizeof(*names
), GFP_KERNEL
);
2629 if (!vi
->big_packets
|| vi
->mergeable_rx_bufs
) {
2630 ctx
= kcalloc(total_vqs
, sizeof(*ctx
), GFP_KERNEL
);
2637 /* Parameters for control virtqueue, if any */
2639 callbacks
[total_vqs
- 1] = NULL
;
2640 names
[total_vqs
- 1] = "control";
2643 /* Allocate/initialize parameters for send/receive virtqueues */
2644 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2645 callbacks
[rxq2vq(i
)] = skb_recv_done
;
2646 callbacks
[txq2vq(i
)] = skb_xmit_done
;
2647 sprintf(vi
->rq
[i
].name
, "input.%d", i
);
2648 sprintf(vi
->sq
[i
].name
, "output.%d", i
);
2649 names
[rxq2vq(i
)] = vi
->rq
[i
].name
;
2650 names
[txq2vq(i
)] = vi
->sq
[i
].name
;
2652 ctx
[rxq2vq(i
)] = true;
2655 ret
= vi
->vdev
->config
->find_vqs(vi
->vdev
, total_vqs
, vqs
, callbacks
,
2661 vi
->cvq
= vqs
[total_vqs
- 1];
2662 if (virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_CTRL_VLAN
))
2663 vi
->dev
->features
|= NETIF_F_HW_VLAN_CTAG_FILTER
;
2666 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2667 vi
->rq
[i
].vq
= vqs
[rxq2vq(i
)];
2668 vi
->rq
[i
].min_buf_len
= mergeable_min_buf_len(vi
, vi
->rq
[i
].vq
);
2669 vi
->sq
[i
].vq
= vqs
[txq2vq(i
)];
2672 /* run here: ret == 0. */
2687 static int virtnet_alloc_queues(struct virtnet_info
*vi
)
2691 vi
->ctrl
= kzalloc(sizeof(*vi
->ctrl
), GFP_KERNEL
);
2694 vi
->sq
= kcalloc(vi
->max_queue_pairs
, sizeof(*vi
->sq
), GFP_KERNEL
);
2697 vi
->rq
= kcalloc(vi
->max_queue_pairs
, sizeof(*vi
->rq
), GFP_KERNEL
);
2701 INIT_DELAYED_WORK(&vi
->refill
, refill_work
);
2702 for (i
= 0; i
< vi
->max_queue_pairs
; i
++) {
2703 vi
->rq
[i
].pages
= NULL
;
2704 netif_napi_add(vi
->dev
, &vi
->rq
[i
].napi
, virtnet_poll
,
2706 netif_tx_napi_add(vi
->dev
, &vi
->sq
[i
].napi
, virtnet_poll_tx
,
2707 napi_tx
? napi_weight
: 0);
2709 sg_init_table(vi
->rq
[i
].sg
, ARRAY_SIZE(vi
->rq
[i
].sg
));
2710 ewma_pkt_len_init(&vi
->rq
[i
].mrg_avg_pkt_len
);
2711 sg_init_table(vi
->sq
[i
].sg
, ARRAY_SIZE(vi
->sq
[i
].sg
));
2713 u64_stats_init(&vi
->rq
[i
].stats
.syncp
);
2714 u64_stats_init(&vi
->sq
[i
].stats
.syncp
);
2727 static int init_vqs(struct virtnet_info
*vi
)
2731 /* Allocate send & receive queues */
2732 ret
= virtnet_alloc_queues(vi
);
2736 ret
= virtnet_find_vqs(vi
);
2741 virtnet_set_affinity(vi
);
2747 virtnet_free_queues(vi
);
2753 static ssize_t
mergeable_rx_buffer_size_show(struct netdev_rx_queue
*queue
,
2756 struct virtnet_info
*vi
= netdev_priv(queue
->dev
);
2757 unsigned int queue_index
= get_netdev_rx_queue_index(queue
);
2758 unsigned int headroom
= virtnet_get_headroom(vi
);
2759 unsigned int tailroom
= headroom
? sizeof(struct skb_shared_info
) : 0;
2760 struct ewma_pkt_len
*avg
;
2762 BUG_ON(queue_index
>= vi
->max_queue_pairs
);
2763 avg
= &vi
->rq
[queue_index
].mrg_avg_pkt_len
;
2764 return sprintf(buf
, "%u\n",
2765 get_mergeable_buf_len(&vi
->rq
[queue_index
], avg
,
2766 SKB_DATA_ALIGN(headroom
+ tailroom
)));
2769 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute
=
2770 __ATTR_RO(mergeable_rx_buffer_size
);
2772 static struct attribute
*virtio_net_mrg_rx_attrs
[] = {
2773 &mergeable_rx_buffer_size_attribute
.attr
,
2777 static const struct attribute_group virtio_net_mrg_rx_group
= {
2778 .name
= "virtio_net",
2779 .attrs
= virtio_net_mrg_rx_attrs
2783 static bool virtnet_fail_on_feature(struct virtio_device
*vdev
,
2785 const char *fname
, const char *dname
)
2787 if (!virtio_has_feature(vdev
, fbit
))
2790 dev_err(&vdev
->dev
, "device advertises feature %s but not %s",
2796 #define VIRTNET_FAIL_ON(vdev, fbit, dbit) \
2797 virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
2799 static bool virtnet_validate_features(struct virtio_device
*vdev
)
2801 if (!virtio_has_feature(vdev
, VIRTIO_NET_F_CTRL_VQ
) &&
2802 (VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_CTRL_RX
,
2803 "VIRTIO_NET_F_CTRL_VQ") ||
2804 VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_CTRL_VLAN
,
2805 "VIRTIO_NET_F_CTRL_VQ") ||
2806 VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_GUEST_ANNOUNCE
,
2807 "VIRTIO_NET_F_CTRL_VQ") ||
2808 VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_MQ
, "VIRTIO_NET_F_CTRL_VQ") ||
2809 VIRTNET_FAIL_ON(vdev
, VIRTIO_NET_F_CTRL_MAC_ADDR
,
2810 "VIRTIO_NET_F_CTRL_VQ"))) {
2817 #define MIN_MTU ETH_MIN_MTU
2818 #define MAX_MTU ETH_MAX_MTU
2820 static int virtnet_validate(struct virtio_device
*vdev
)
2822 if (!vdev
->config
->get
) {
2823 dev_err(&vdev
->dev
, "%s failure: config access disabled\n",
2828 if (!virtnet_validate_features(vdev
))
2831 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MTU
)) {
2832 int mtu
= virtio_cread16(vdev
,
2833 offsetof(struct virtio_net_config
,
2836 __virtio_clear_bit(vdev
, VIRTIO_NET_F_MTU
);
2842 static int virtnet_probe(struct virtio_device
*vdev
)
2844 int i
, err
= -ENOMEM
;
2845 struct net_device
*dev
;
2846 struct virtnet_info
*vi
;
2847 u16 max_queue_pairs
;
2850 /* Find if host supports multiqueue virtio_net device */
2851 err
= virtio_cread_feature(vdev
, VIRTIO_NET_F_MQ
,
2852 struct virtio_net_config
,
2853 max_virtqueue_pairs
, &max_queue_pairs
);
2855 /* We need at least 2 queue's */
2856 if (err
|| max_queue_pairs
< VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN
||
2857 max_queue_pairs
> VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX
||
2858 !virtio_has_feature(vdev
, VIRTIO_NET_F_CTRL_VQ
))
2859 max_queue_pairs
= 1;
2861 /* Allocate ourselves a network device with room for our info */
2862 dev
= alloc_etherdev_mq(sizeof(struct virtnet_info
), max_queue_pairs
);
2866 /* Set up network device as normal. */
2867 dev
->priv_flags
|= IFF_UNICAST_FLT
| IFF_LIVE_ADDR_CHANGE
;
2868 dev
->netdev_ops
= &virtnet_netdev
;
2869 dev
->features
= NETIF_F_HIGHDMA
;
2871 dev
->ethtool_ops
= &virtnet_ethtool_ops
;
2872 SET_NETDEV_DEV(dev
, &vdev
->dev
);
2874 /* Do we support "hardware" checksums? */
2875 if (virtio_has_feature(vdev
, VIRTIO_NET_F_CSUM
)) {
2876 /* This opens up the world of extra features. */
2877 dev
->hw_features
|= NETIF_F_HW_CSUM
| NETIF_F_SG
;
2879 dev
->features
|= NETIF_F_HW_CSUM
| NETIF_F_SG
;
2881 if (virtio_has_feature(vdev
, VIRTIO_NET_F_GSO
)) {
2882 dev
->hw_features
|= NETIF_F_TSO
2883 | NETIF_F_TSO_ECN
| NETIF_F_TSO6
;
2885 /* Individual feature bits: what can host handle? */
2886 if (virtio_has_feature(vdev
, VIRTIO_NET_F_HOST_TSO4
))
2887 dev
->hw_features
|= NETIF_F_TSO
;
2888 if (virtio_has_feature(vdev
, VIRTIO_NET_F_HOST_TSO6
))
2889 dev
->hw_features
|= NETIF_F_TSO6
;
2890 if (virtio_has_feature(vdev
, VIRTIO_NET_F_HOST_ECN
))
2891 dev
->hw_features
|= NETIF_F_TSO_ECN
;
2893 dev
->features
|= NETIF_F_GSO_ROBUST
;
2896 dev
->features
|= dev
->hw_features
& NETIF_F_ALL_TSO
;
2897 /* (!csum && gso) case will be fixed by register_netdev() */
2899 if (virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_CSUM
))
2900 dev
->features
|= NETIF_F_RXCSUM
;
2902 dev
->vlan_features
= dev
->features
;
2904 /* MTU range: 68 - 65535 */
2905 dev
->min_mtu
= MIN_MTU
;
2906 dev
->max_mtu
= MAX_MTU
;
2908 /* Configuration may specify what MAC to use. Otherwise random. */
2909 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MAC
))
2910 virtio_cread_bytes(vdev
,
2911 offsetof(struct virtio_net_config
, mac
),
2912 dev
->dev_addr
, dev
->addr_len
);
2914 eth_hw_addr_random(dev
);
2916 /* Set up our device-specific information */
2917 vi
= netdev_priv(dev
);
2922 INIT_WORK(&vi
->config_work
, virtnet_config_changed_work
);
2924 /* If we can receive ANY GSO packets, we must allocate large ones. */
2925 if (virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_TSO4
) ||
2926 virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_TSO6
) ||
2927 virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_ECN
) ||
2928 virtio_has_feature(vdev
, VIRTIO_NET_F_GUEST_UFO
))
2929 vi
->big_packets
= true;
2931 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MRG_RXBUF
))
2932 vi
->mergeable_rx_bufs
= true;
2934 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MRG_RXBUF
) ||
2935 virtio_has_feature(vdev
, VIRTIO_F_VERSION_1
))
2936 vi
->hdr_len
= sizeof(struct virtio_net_hdr_mrg_rxbuf
);
2938 vi
->hdr_len
= sizeof(struct virtio_net_hdr
);
2940 if (virtio_has_feature(vdev
, VIRTIO_F_ANY_LAYOUT
) ||
2941 virtio_has_feature(vdev
, VIRTIO_F_VERSION_1
))
2942 vi
->any_header_sg
= true;
2944 if (virtio_has_feature(vdev
, VIRTIO_NET_F_CTRL_VQ
))
2947 if (virtio_has_feature(vdev
, VIRTIO_NET_F_MTU
)) {
2948 mtu
= virtio_cread16(vdev
,
2949 offsetof(struct virtio_net_config
,
2951 if (mtu
< dev
->min_mtu
) {
2952 /* Should never trigger: MTU was previously validated
2953 * in virtnet_validate.
2955 dev_err(&vdev
->dev
, "device MTU appears to have changed "
2956 "it is now %d < %d", mtu
, dev
->min_mtu
);
2963 /* TODO: size buffers correctly in this case. */
2964 if (dev
->mtu
> ETH_DATA_LEN
)
2965 vi
->big_packets
= true;
2968 if (vi
->any_header_sg
)
2969 dev
->needed_headroom
= vi
->hdr_len
;
2971 /* Enable multiqueue by default */
2972 if (num_online_cpus() >= max_queue_pairs
)
2973 vi
->curr_queue_pairs
= max_queue_pairs
;
2975 vi
->curr_queue_pairs
= num_online_cpus();
2976 vi
->max_queue_pairs
= max_queue_pairs
;
2978 /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
2984 if (vi
->mergeable_rx_bufs
)
2985 dev
->sysfs_rx_queue_group
= &virtio_net_mrg_rx_group
;
2987 netif_set_real_num_tx_queues(dev
, vi
->curr_queue_pairs
);
2988 netif_set_real_num_rx_queues(dev
, vi
->curr_queue_pairs
);
2990 virtnet_init_settings(dev
);
2992 if (virtio_has_feature(vdev
, VIRTIO_NET_F_STANDBY
)) {
2993 vi
->failover
= net_failover_create(vi
->dev
);
2994 if (IS_ERR(vi
->failover
)) {
2995 err
= PTR_ERR(vi
->failover
);
3000 err
= register_netdev(dev
);
3002 pr_debug("virtio_net: registering device failed\n");
3006 virtio_device_ready(vdev
);
3008 err
= virtnet_cpu_notif_add(vi
);
3010 pr_debug("virtio_net: registering cpu notifier failed\n");
3011 goto free_unregister_netdev
;
3014 virtnet_set_queues(vi
, vi
->curr_queue_pairs
);
3016 /* Assume link up if device can't report link status,
3017 otherwise get link status from config. */
3018 netif_carrier_off(dev
);
3019 if (virtio_has_feature(vi
->vdev
, VIRTIO_NET_F_STATUS
)) {
3020 schedule_work(&vi
->config_work
);
3022 vi
->status
= VIRTIO_NET_S_LINK_UP
;
3023 virtnet_update_settings(vi
);
3024 netif_carrier_on(dev
);
3027 for (i
= 0; i
< ARRAY_SIZE(guest_offloads
); i
++)
3028 if (virtio_has_feature(vi
->vdev
, guest_offloads
[i
]))
3029 set_bit(guest_offloads
[i
], &vi
->guest_offloads
);
3031 pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
3032 dev
->name
, max_queue_pairs
);
3036 free_unregister_netdev
:
3037 vi
->vdev
->config
->reset(vdev
);
3039 unregister_netdev(dev
);
3041 net_failover_destroy(vi
->failover
);
3043 cancel_delayed_work_sync(&vi
->refill
);
3044 free_receive_page_frags(vi
);
3045 virtnet_del_vqs(vi
);
3051 static void remove_vq_common(struct virtnet_info
*vi
)
3053 vi
->vdev
->config
->reset(vi
->vdev
);
3055 /* Free unused buffers in both send and recv, if any. */
3056 free_unused_bufs(vi
);
3058 free_receive_bufs(vi
);
3060 free_receive_page_frags(vi
);
3062 virtnet_del_vqs(vi
);
3065 static void virtnet_remove(struct virtio_device
*vdev
)
3067 struct virtnet_info
*vi
= vdev
->priv
;
3069 virtnet_cpu_notif_remove(vi
);
3071 /* Make sure no work handler is accessing the device. */
3072 flush_work(&vi
->config_work
);
3074 unregister_netdev(vi
->dev
);
3076 net_failover_destroy(vi
->failover
);
3078 remove_vq_common(vi
);
3080 free_netdev(vi
->dev
);
3083 static __maybe_unused
int virtnet_freeze(struct virtio_device
*vdev
)
3085 struct virtnet_info
*vi
= vdev
->priv
;
3087 virtnet_cpu_notif_remove(vi
);
3088 virtnet_freeze_down(vdev
);
3089 remove_vq_common(vi
);
3094 static __maybe_unused
int virtnet_restore(struct virtio_device
*vdev
)
3096 struct virtnet_info
*vi
= vdev
->priv
;
3099 err
= virtnet_restore_up(vdev
);
3102 virtnet_set_queues(vi
, vi
->curr_queue_pairs
);
3104 err
= virtnet_cpu_notif_add(vi
);
3111 static struct virtio_device_id id_table
[] = {
3112 { VIRTIO_ID_NET
, VIRTIO_DEV_ANY_ID
},
3116 #define VIRTNET_FEATURES \
3117 VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
3119 VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
3120 VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
3121 VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
3122 VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
3123 VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
3124 VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
3125 VIRTIO_NET_F_CTRL_MAC_ADDR, \
3126 VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, \
3127 VIRTIO_NET_F_SPEED_DUPLEX, VIRTIO_NET_F_STANDBY
3129 static unsigned int features
[] = {
3133 static unsigned int features_legacy
[] = {
3136 VIRTIO_F_ANY_LAYOUT
,
3139 static struct virtio_driver virtio_net_driver
= {
3140 .feature_table
= features
,
3141 .feature_table_size
= ARRAY_SIZE(features
),
3142 .feature_table_legacy
= features_legacy
,
3143 .feature_table_size_legacy
= ARRAY_SIZE(features_legacy
),
3144 .driver
.name
= KBUILD_MODNAME
,
3145 .driver
.owner
= THIS_MODULE
,
3146 .id_table
= id_table
,
3147 .validate
= virtnet_validate
,
3148 .probe
= virtnet_probe
,
3149 .remove
= virtnet_remove
,
3150 .config_changed
= virtnet_config_changed
,
3151 #ifdef CONFIG_PM_SLEEP
3152 .freeze
= virtnet_freeze
,
3153 .restore
= virtnet_restore
,
3157 static __init
int virtio_net_driver_init(void)
3161 ret
= cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN
, "virtio/net:online",
3163 virtnet_cpu_down_prep
);
3166 virtionet_online
= ret
;
3167 ret
= cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD
, "virtio/net:dead",
3168 NULL
, virtnet_cpu_dead
);
3172 ret
= register_virtio_driver(&virtio_net_driver
);
3177 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD
);
3179 cpuhp_remove_multi_state(virtionet_online
);
3183 module_init(virtio_net_driver_init
);
3185 static __exit
void virtio_net_driver_exit(void)
3187 unregister_virtio_driver(&virtio_net_driver
);
3188 cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD
);
3189 cpuhp_remove_multi_state(virtionet_online
);
3191 module_exit(virtio_net_driver_exit
);
3193 MODULE_DEVICE_TABLE(virtio
, id_table
);
3194 MODULE_DESCRIPTION("Virtio network driver");
3195 MODULE_LICENSE("GPL");