virtio-net: switch off offloads on demand if possible on XDP set
[sfrench/cifs-2.6.git] / drivers / net / virtio_net.c
1 /* A network driver using virtio.
2  *
3  * Copyright 2007 Rusty Russell <rusty@rustcorp.com.au> IBM Corporation
4  *
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.
9  *
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.
14  *
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/>.
17  */
18 //#define DEBUG
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>
33
34 static int napi_weight = NAPI_POLL_WEIGHT;
35 module_param(napi_weight, int, 0444);
36
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);
41
42 /* FIXME: MTU in config. */
43 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
44 #define GOOD_COPY_LEN   128
45
46 #define VIRTNET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
47
48 /* Amount of XDP headroom to prepend to packets for use by xdp_adjust_head */
49 #define VIRTIO_XDP_HEADROOM 256
50
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.
55  */
56 DECLARE_EWMA(pkt_len, 0, 64)
57
58 #define VIRTNET_DRIVER_VERSION "1.0.0"
59
60 const unsigned long guest_offloads[] = { VIRTIO_NET_F_GUEST_TSO4,
61                                          VIRTIO_NET_F_GUEST_TSO6,
62                                          VIRTIO_NET_F_GUEST_ECN,
63                                          VIRTIO_NET_F_GUEST_UFO };
64
65 struct virtnet_stats {
66         struct u64_stats_sync tx_syncp;
67         struct u64_stats_sync rx_syncp;
68         u64 tx_bytes;
69         u64 tx_packets;
70
71         u64 rx_bytes;
72         u64 rx_packets;
73 };
74
75 /* Internal representation of a send virtqueue */
76 struct send_queue {
77         /* Virtqueue associated with this send _queue */
78         struct virtqueue *vq;
79
80         /* TX: fragments + linear part + virtio header */
81         struct scatterlist sg[MAX_SKB_FRAGS + 2];
82
83         /* Name of the send queue: output.$index */
84         char name[40];
85
86         struct napi_struct napi;
87 };
88
89 /* Internal representation of a receive virtqueue */
90 struct receive_queue {
91         /* Virtqueue associated with this receive_queue */
92         struct virtqueue *vq;
93
94         struct napi_struct napi;
95
96         struct bpf_prog __rcu *xdp_prog;
97
98         /* Chain pages by the private ptr. */
99         struct page *pages;
100
101         /* Average packet length for mergeable receive buffers. */
102         struct ewma_pkt_len mrg_avg_pkt_len;
103
104         /* Page frag for packet buffer allocation. */
105         struct page_frag alloc_frag;
106
107         /* RX: fragments + linear part + virtio header */
108         struct scatterlist sg[MAX_SKB_FRAGS + 2];
109
110         /* Min single buffer size for mergeable buffers case. */
111         unsigned int min_buf_len;
112
113         /* Name of this receive queue: input.$index */
114         char name[40];
115 };
116
117 struct virtnet_info {
118         struct virtio_device *vdev;
119         struct virtqueue *cvq;
120         struct net_device *dev;
121         struct send_queue *sq;
122         struct receive_queue *rq;
123         unsigned int status;
124
125         /* Max # of queue pairs supported by the device */
126         u16 max_queue_pairs;
127
128         /* # of queue pairs currently used by the driver */
129         u16 curr_queue_pairs;
130
131         /* # of XDP queue pairs currently used by the driver */
132         u16 xdp_queue_pairs;
133
134         /* I like... big packets and I cannot lie! */
135         bool big_packets;
136
137         /* Host will merge rx buffers for big packets (shake it! shake it!) */
138         bool mergeable_rx_bufs;
139
140         /* Has control virtqueue */
141         bool has_cvq;
142
143         /* Host can handle any s/g split between our header and packet data */
144         bool any_header_sg;
145
146         /* Packet virtio header size */
147         u8 hdr_len;
148
149         /* Active statistics */
150         struct virtnet_stats __percpu *stats;
151
152         /* Work struct for refilling if we run low on memory. */
153         struct delayed_work refill;
154
155         /* Work struct for config space updates */
156         struct work_struct config_work;
157
158         /* Does the affinity hint is set for virtqueues? */
159         bool affinity_hint_set;
160
161         /* CPU hotplug instances for online & dead */
162         struct hlist_node node;
163         struct hlist_node node_dead;
164
165         /* Control VQ buffers: protected by the rtnl lock */
166         struct virtio_net_ctrl_hdr ctrl_hdr;
167         virtio_net_ctrl_ack ctrl_status;
168         struct virtio_net_ctrl_mq ctrl_mq;
169         u8 ctrl_promisc;
170         u8 ctrl_allmulti;
171         u16 ctrl_vid;
172         u64 ctrl_offloads;
173
174         /* Ethtool settings */
175         u8 duplex;
176         u32 speed;
177
178         unsigned long guest_offloads;
179 };
180
181 struct padded_vnet_hdr {
182         struct virtio_net_hdr_mrg_rxbuf hdr;
183         /*
184          * hdr is in a separate sg buffer, and data sg buffer shares same page
185          * with this header sg. This padding makes next sg 16 byte aligned
186          * after the header.
187          */
188         char padding[4];
189 };
190
191 /* Converting between virtqueue no. and kernel tx/rx queue no.
192  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
193  */
194 static int vq2txq(struct virtqueue *vq)
195 {
196         return (vq->index - 1) / 2;
197 }
198
199 static int txq2vq(int txq)
200 {
201         return txq * 2 + 1;
202 }
203
204 static int vq2rxq(struct virtqueue *vq)
205 {
206         return vq->index / 2;
207 }
208
209 static int rxq2vq(int rxq)
210 {
211         return rxq * 2;
212 }
213
214 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
215 {
216         return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
217 }
218
219 /*
220  * private is used to chain pages for big packets, put the whole
221  * most recent used list in the beginning for reuse
222  */
223 static void give_pages(struct receive_queue *rq, struct page *page)
224 {
225         struct page *end;
226
227         /* Find end of list, sew whole thing into vi->rq.pages. */
228         for (end = page; end->private; end = (struct page *)end->private);
229         end->private = (unsigned long)rq->pages;
230         rq->pages = page;
231 }
232
233 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
234 {
235         struct page *p = rq->pages;
236
237         if (p) {
238                 rq->pages = (struct page *)p->private;
239                 /* clear private here, it is used to chain pages */
240                 p->private = 0;
241         } else
242                 p = alloc_page(gfp_mask);
243         return p;
244 }
245
246 static void virtqueue_napi_schedule(struct napi_struct *napi,
247                                     struct virtqueue *vq)
248 {
249         if (napi_schedule_prep(napi)) {
250                 virtqueue_disable_cb(vq);
251                 __napi_schedule(napi);
252         }
253 }
254
255 static void virtqueue_napi_complete(struct napi_struct *napi,
256                                     struct virtqueue *vq, int processed)
257 {
258         int opaque;
259
260         opaque = virtqueue_enable_cb_prepare(vq);
261         if (napi_complete_done(napi, processed) &&
262             unlikely(virtqueue_poll(vq, opaque)))
263                 virtqueue_napi_schedule(napi, vq);
264 }
265
266 static void skb_xmit_done(struct virtqueue *vq)
267 {
268         struct virtnet_info *vi = vq->vdev->priv;
269         struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi;
270
271         /* Suppress further interrupts. */
272         virtqueue_disable_cb(vq);
273
274         if (napi->weight)
275                 virtqueue_napi_schedule(napi, vq);
276         else
277                 /* We were probably waiting for more output buffers. */
278                 netif_wake_subqueue(vi->dev, vq2txq(vq));
279 }
280
281 #define MRG_CTX_HEADER_SHIFT 22
282 static void *mergeable_len_to_ctx(unsigned int truesize,
283                                   unsigned int headroom)
284 {
285         return (void *)(unsigned long)((headroom << MRG_CTX_HEADER_SHIFT) | truesize);
286 }
287
288 static unsigned int mergeable_ctx_to_headroom(void *mrg_ctx)
289 {
290         return (unsigned long)mrg_ctx >> MRG_CTX_HEADER_SHIFT;
291 }
292
293 static unsigned int mergeable_ctx_to_truesize(void *mrg_ctx)
294 {
295         return (unsigned long)mrg_ctx & ((1 << MRG_CTX_HEADER_SHIFT) - 1);
296 }
297
298 /* Called from bottom half context */
299 static struct sk_buff *page_to_skb(struct virtnet_info *vi,
300                                    struct receive_queue *rq,
301                                    struct page *page, unsigned int offset,
302                                    unsigned int len, unsigned int truesize)
303 {
304         struct sk_buff *skb;
305         struct virtio_net_hdr_mrg_rxbuf *hdr;
306         unsigned int copy, hdr_len, hdr_padded_len;
307         char *p;
308
309         p = page_address(page) + offset;
310
311         /* copy small packet so we can reuse these pages for small data */
312         skb = napi_alloc_skb(&rq->napi, GOOD_COPY_LEN);
313         if (unlikely(!skb))
314                 return NULL;
315
316         hdr = skb_vnet_hdr(skb);
317
318         hdr_len = vi->hdr_len;
319         if (vi->mergeable_rx_bufs)
320                 hdr_padded_len = sizeof *hdr;
321         else
322                 hdr_padded_len = sizeof(struct padded_vnet_hdr);
323
324         memcpy(hdr, p, hdr_len);
325
326         len -= hdr_len;
327         offset += hdr_padded_len;
328         p += hdr_padded_len;
329
330         copy = len;
331         if (copy > skb_tailroom(skb))
332                 copy = skb_tailroom(skb);
333         skb_put_data(skb, p, copy);
334
335         len -= copy;
336         offset += copy;
337
338         if (vi->mergeable_rx_bufs) {
339                 if (len)
340                         skb_add_rx_frag(skb, 0, page, offset, len, truesize);
341                 else
342                         put_page(page);
343                 return skb;
344         }
345
346         /*
347          * Verify that we can indeed put this data into a skb.
348          * This is here to handle cases when the device erroneously
349          * tries to receive more than is possible. This is usually
350          * the case of a broken device.
351          */
352         if (unlikely(len > MAX_SKB_FRAGS * PAGE_SIZE)) {
353                 net_dbg_ratelimited("%s: too much data\n", skb->dev->name);
354                 dev_kfree_skb(skb);
355                 return NULL;
356         }
357         BUG_ON(offset >= PAGE_SIZE);
358         while (len) {
359                 unsigned int frag_size = min((unsigned)PAGE_SIZE - offset, len);
360                 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags, page, offset,
361                                 frag_size, truesize);
362                 len -= frag_size;
363                 page = (struct page *)page->private;
364                 offset = 0;
365         }
366
367         if (page)
368                 give_pages(rq, page);
369
370         return skb;
371 }
372
373 static bool virtnet_xdp_xmit(struct virtnet_info *vi,
374                              struct receive_queue *rq,
375                              struct xdp_buff *xdp)
376 {
377         struct virtio_net_hdr_mrg_rxbuf *hdr;
378         unsigned int len;
379         struct send_queue *sq;
380         unsigned int qp;
381         void *xdp_sent;
382         int err;
383
384         qp = vi->curr_queue_pairs - vi->xdp_queue_pairs + smp_processor_id();
385         sq = &vi->sq[qp];
386
387         /* Free up any pending old buffers before queueing new ones. */
388         while ((xdp_sent = virtqueue_get_buf(sq->vq, &len)) != NULL) {
389                 struct page *sent_page = virt_to_head_page(xdp_sent);
390
391                 put_page(sent_page);
392         }
393
394         xdp->data -= vi->hdr_len;
395         /* Zero header and leave csum up to XDP layers */
396         hdr = xdp->data;
397         memset(hdr, 0, vi->hdr_len);
398
399         sg_init_one(sq->sg, xdp->data, xdp->data_end - xdp->data);
400
401         err = virtqueue_add_outbuf(sq->vq, sq->sg, 1, xdp->data, GFP_ATOMIC);
402         if (unlikely(err)) {
403                 struct page *page = virt_to_head_page(xdp->data);
404
405                 put_page(page);
406                 return false;
407         }
408
409         virtqueue_kick(sq->vq);
410         return true;
411 }
412
413 static unsigned int virtnet_get_headroom(struct virtnet_info *vi)
414 {
415         return vi->xdp_queue_pairs ? VIRTIO_XDP_HEADROOM : 0;
416 }
417
418 /* We copy the packet for XDP in the following cases:
419  *
420  * 1) Packet is scattered across multiple rx buffers.
421  * 2) Headroom space is insufficient.
422  *
423  * This is inefficient but it's a temporary condition that
424  * we hit right after XDP is enabled and until queue is refilled
425  * with large buffers with sufficient headroom - so it should affect
426  * at most queue size packets.
427  * Afterwards, the conditions to enable
428  * XDP should preclude the underlying device from sending packets
429  * across multiple buffers (num_buf > 1), and we make sure buffers
430  * have enough headroom.
431  */
432 static struct page *xdp_linearize_page(struct receive_queue *rq,
433                                        u16 *num_buf,
434                                        struct page *p,
435                                        int offset,
436                                        int page_off,
437                                        unsigned int *len)
438 {
439         struct page *page = alloc_page(GFP_ATOMIC);
440
441         if (!page)
442                 return NULL;
443
444         memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
445         page_off += *len;
446
447         while (--*num_buf) {
448                 unsigned int buflen;
449                 void *buf;
450                 int off;
451
452                 buf = virtqueue_get_buf(rq->vq, &buflen);
453                 if (unlikely(!buf))
454                         goto err_buf;
455
456                 p = virt_to_head_page(buf);
457                 off = buf - page_address(p);
458
459                 /* guard against a misconfigured or uncooperative backend that
460                  * is sending packet larger than the MTU.
461                  */
462                 if ((page_off + buflen) > PAGE_SIZE) {
463                         put_page(p);
464                         goto err_buf;
465                 }
466
467                 memcpy(page_address(page) + page_off,
468                        page_address(p) + off, buflen);
469                 page_off += buflen;
470                 put_page(p);
471         }
472
473         /* Headroom does not contribute to packet length */
474         *len = page_off - VIRTIO_XDP_HEADROOM;
475         return page;
476 err_buf:
477         __free_pages(page, 0);
478         return NULL;
479 }
480
481 static struct sk_buff *receive_small(struct net_device *dev,
482                                      struct virtnet_info *vi,
483                                      struct receive_queue *rq,
484                                      void *buf, void *ctx,
485                                      unsigned int len)
486 {
487         struct sk_buff *skb;
488         struct bpf_prog *xdp_prog;
489         unsigned int xdp_headroom = (unsigned long)ctx;
490         unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom;
491         unsigned int headroom = vi->hdr_len + header_offset;
492         unsigned int buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
493                               SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
494         struct page *page = virt_to_head_page(buf);
495         unsigned int delta = 0;
496         struct page *xdp_page;
497         len -= vi->hdr_len;
498
499         rcu_read_lock();
500         xdp_prog = rcu_dereference(rq->xdp_prog);
501         if (xdp_prog) {
502                 struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset;
503                 struct xdp_buff xdp;
504                 void *orig_data;
505                 u32 act;
506
507                 if (unlikely(hdr->hdr.gso_type || hdr->hdr.flags))
508                         goto err_xdp;
509
510                 if (unlikely(xdp_headroom < virtnet_get_headroom(vi))) {
511                         int offset = buf - page_address(page) + header_offset;
512                         unsigned int tlen = len + vi->hdr_len;
513                         u16 num_buf = 1;
514
515                         xdp_headroom = virtnet_get_headroom(vi);
516                         header_offset = VIRTNET_RX_PAD + xdp_headroom;
517                         headroom = vi->hdr_len + header_offset;
518                         buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
519                                  SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
520                         xdp_page = xdp_linearize_page(rq, &num_buf, page,
521                                                       offset, header_offset,
522                                                       &tlen);
523                         if (!xdp_page)
524                                 goto err_xdp;
525
526                         buf = page_address(xdp_page);
527                         put_page(page);
528                         page = xdp_page;
529                 }
530
531                 xdp.data_hard_start = buf + VIRTNET_RX_PAD + vi->hdr_len;
532                 xdp.data = xdp.data_hard_start + xdp_headroom;
533                 xdp.data_end = xdp.data + len;
534                 orig_data = xdp.data;
535                 act = bpf_prog_run_xdp(xdp_prog, &xdp);
536
537                 switch (act) {
538                 case XDP_PASS:
539                         /* Recalculate length in case bpf program changed it */
540                         delta = orig_data - xdp.data;
541                         break;
542                 case XDP_TX:
543                         if (unlikely(!virtnet_xdp_xmit(vi, rq, &xdp)))
544                                 trace_xdp_exception(vi->dev, xdp_prog, act);
545                         rcu_read_unlock();
546                         goto xdp_xmit;
547                 default:
548                         bpf_warn_invalid_xdp_action(act);
549                 case XDP_ABORTED:
550                         trace_xdp_exception(vi->dev, xdp_prog, act);
551                 case XDP_DROP:
552                         goto err_xdp;
553                 }
554         }
555         rcu_read_unlock();
556
557         skb = build_skb(buf, buflen);
558         if (!skb) {
559                 put_page(page);
560                 goto err;
561         }
562         skb_reserve(skb, headroom - delta);
563         skb_put(skb, len + delta);
564         if (!delta) {
565                 buf += header_offset;
566                 memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
567         } /* keep zeroed vnet hdr since packet was changed by bpf */
568
569 err:
570         return skb;
571
572 err_xdp:
573         rcu_read_unlock();
574         dev->stats.rx_dropped++;
575         put_page(page);
576 xdp_xmit:
577         return NULL;
578 }
579
580 static struct sk_buff *receive_big(struct net_device *dev,
581                                    struct virtnet_info *vi,
582                                    struct receive_queue *rq,
583                                    void *buf,
584                                    unsigned int len)
585 {
586         struct page *page = buf;
587         struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len, PAGE_SIZE);
588
589         if (unlikely(!skb))
590                 goto err;
591
592         return skb;
593
594 err:
595         dev->stats.rx_dropped++;
596         give_pages(rq, page);
597         return NULL;
598 }
599
600 static struct sk_buff *receive_mergeable(struct net_device *dev,
601                                          struct virtnet_info *vi,
602                                          struct receive_queue *rq,
603                                          void *buf,
604                                          void *ctx,
605                                          unsigned int len)
606 {
607         struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
608         u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
609         struct page *page = virt_to_head_page(buf);
610         int offset = buf - page_address(page);
611         struct sk_buff *head_skb, *curr_skb;
612         struct bpf_prog *xdp_prog;
613         unsigned int truesize;
614         unsigned int headroom = mergeable_ctx_to_headroom(ctx);
615
616         head_skb = NULL;
617
618         rcu_read_lock();
619         xdp_prog = rcu_dereference(rq->xdp_prog);
620         if (xdp_prog) {
621                 struct page *xdp_page;
622                 struct xdp_buff xdp;
623                 void *data;
624                 u32 act;
625
626                 /* This happens when rx buffer size is underestimated */
627                 if (unlikely(num_buf > 1 ||
628                              headroom < virtnet_get_headroom(vi))) {
629                         /* linearize data for XDP */
630                         xdp_page = xdp_linearize_page(rq, &num_buf,
631                                                       page, offset,
632                                                       VIRTIO_XDP_HEADROOM,
633                                                       &len);
634                         if (!xdp_page)
635                                 goto err_xdp;
636                         offset = VIRTIO_XDP_HEADROOM;
637                 } else {
638                         xdp_page = page;
639                 }
640
641                 /* Transient failure which in theory could occur if
642                  * in-flight packets from before XDP was enabled reach
643                  * the receive path after XDP is loaded. In practice I
644                  * was not able to create this condition.
645                  */
646                 if (unlikely(hdr->hdr.gso_type))
647                         goto err_xdp;
648
649                 /* Allow consuming headroom but reserve enough space to push
650                  * the descriptor on if we get an XDP_TX return code.
651                  */
652                 data = page_address(xdp_page) + offset;
653                 xdp.data_hard_start = data - VIRTIO_XDP_HEADROOM + vi->hdr_len;
654                 xdp.data = data + vi->hdr_len;
655                 xdp.data_end = xdp.data + (len - vi->hdr_len);
656                 act = bpf_prog_run_xdp(xdp_prog, &xdp);
657
658                 switch (act) {
659                 case XDP_PASS:
660                         /* recalculate offset to account for any header
661                          * adjustments. Note other cases do not build an
662                          * skb and avoid using offset
663                          */
664                         offset = xdp.data -
665                                         page_address(xdp_page) - vi->hdr_len;
666
667                         /* We can only create skb based on xdp_page. */
668                         if (unlikely(xdp_page != page)) {
669                                 rcu_read_unlock();
670                                 put_page(page);
671                                 head_skb = page_to_skb(vi, rq, xdp_page,
672                                                        offset, len, PAGE_SIZE);
673                                 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, len);
674                                 return head_skb;
675                         }
676                         break;
677                 case XDP_TX:
678                         if (unlikely(!virtnet_xdp_xmit(vi, rq, &xdp)))
679                                 trace_xdp_exception(vi->dev, xdp_prog, act);
680                         ewma_pkt_len_add(&rq->mrg_avg_pkt_len, len);
681                         if (unlikely(xdp_page != page))
682                                 goto err_xdp;
683                         rcu_read_unlock();
684                         goto xdp_xmit;
685                 default:
686                         bpf_warn_invalid_xdp_action(act);
687                 case XDP_ABORTED:
688                         trace_xdp_exception(vi->dev, xdp_prog, act);
689                 case XDP_DROP:
690                         if (unlikely(xdp_page != page))
691                                 __free_pages(xdp_page, 0);
692                         ewma_pkt_len_add(&rq->mrg_avg_pkt_len, len);
693                         goto err_xdp;
694                 }
695         }
696         rcu_read_unlock();
697
698         truesize = mergeable_ctx_to_truesize(ctx);
699         if (unlikely(len > truesize)) {
700                 pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
701                          dev->name, len, (unsigned long)ctx);
702                 dev->stats.rx_length_errors++;
703                 goto err_skb;
704         }
705
706         head_skb = page_to_skb(vi, rq, page, offset, len, truesize);
707         curr_skb = head_skb;
708
709         if (unlikely(!curr_skb))
710                 goto err_skb;
711         while (--num_buf) {
712                 int num_skb_frags;
713
714                 buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx);
715                 if (unlikely(!ctx)) {
716                         pr_debug("%s: rx error: %d buffers out of %d missing\n",
717                                  dev->name, num_buf,
718                                  virtio16_to_cpu(vi->vdev,
719                                                  hdr->num_buffers));
720                         dev->stats.rx_length_errors++;
721                         goto err_buf;
722                 }
723
724                 page = virt_to_head_page(buf);
725
726                 truesize = mergeable_ctx_to_truesize(ctx);
727                 if (unlikely(len > truesize)) {
728                         pr_debug("%s: rx error: len %u exceeds truesize %lu\n",
729                                  dev->name, len, (unsigned long)ctx);
730                         dev->stats.rx_length_errors++;
731                         goto err_skb;
732                 }
733
734                 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
735                 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
736                         struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
737
738                         if (unlikely(!nskb))
739                                 goto err_skb;
740                         if (curr_skb == head_skb)
741                                 skb_shinfo(curr_skb)->frag_list = nskb;
742                         else
743                                 curr_skb->next = nskb;
744                         curr_skb = nskb;
745                         head_skb->truesize += nskb->truesize;
746                         num_skb_frags = 0;
747                 }
748                 if (curr_skb != head_skb) {
749                         head_skb->data_len += len;
750                         head_skb->len += len;
751                         head_skb->truesize += truesize;
752                 }
753                 offset = buf - page_address(page);
754                 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
755                         put_page(page);
756                         skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
757                                              len, truesize);
758                 } else {
759                         skb_add_rx_frag(curr_skb, num_skb_frags, page,
760                                         offset, len, truesize);
761                 }
762         }
763
764         ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
765         return head_skb;
766
767 err_xdp:
768         rcu_read_unlock();
769 err_skb:
770         put_page(page);
771         while (--num_buf) {
772                 buf = virtqueue_get_buf(rq->vq, &len);
773                 if (unlikely(!buf)) {
774                         pr_debug("%s: rx error: %d buffers missing\n",
775                                  dev->name, num_buf);
776                         dev->stats.rx_length_errors++;
777                         break;
778                 }
779                 page = virt_to_head_page(buf);
780                 put_page(page);
781         }
782 err_buf:
783         dev->stats.rx_dropped++;
784         dev_kfree_skb(head_skb);
785 xdp_xmit:
786         return NULL;
787 }
788
789 static int receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
790                        void *buf, unsigned int len, void **ctx)
791 {
792         struct net_device *dev = vi->dev;
793         struct sk_buff *skb;
794         struct virtio_net_hdr_mrg_rxbuf *hdr;
795         int ret;
796
797         if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
798                 pr_debug("%s: short packet %i\n", dev->name, len);
799                 dev->stats.rx_length_errors++;
800                 if (vi->mergeable_rx_bufs) {
801                         put_page(virt_to_head_page(buf));
802                 } else if (vi->big_packets) {
803                         give_pages(rq, buf);
804                 } else {
805                         put_page(virt_to_head_page(buf));
806                 }
807                 return 0;
808         }
809
810         if (vi->mergeable_rx_bufs)
811                 skb = receive_mergeable(dev, vi, rq, buf, ctx, len);
812         else if (vi->big_packets)
813                 skb = receive_big(dev, vi, rq, buf, len);
814         else
815                 skb = receive_small(dev, vi, rq, buf, ctx, len);
816
817         if (unlikely(!skb))
818                 return 0;
819
820         hdr = skb_vnet_hdr(skb);
821
822         ret = skb->len;
823
824         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
825                 skb->ip_summed = CHECKSUM_UNNECESSARY;
826
827         if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
828                                   virtio_is_little_endian(vi->vdev))) {
829                 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
830                                      dev->name, hdr->hdr.gso_type,
831                                      hdr->hdr.gso_size);
832                 goto frame_err;
833         }
834
835         skb->protocol = eth_type_trans(skb, dev);
836         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
837                  ntohs(skb->protocol), skb->len, skb->pkt_type);
838
839         napi_gro_receive(&rq->napi, skb);
840         return ret;
841
842 frame_err:
843         dev->stats.rx_frame_errors++;
844         dev_kfree_skb(skb);
845         return 0;
846 }
847
848 /* Unlike mergeable buffers, all buffers are allocated to the
849  * same size, except for the headroom. For this reason we do
850  * not need to use  mergeable_len_to_ctx here - it is enough
851  * to store the headroom as the context ignoring the truesize.
852  */
853 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
854                              gfp_t gfp)
855 {
856         struct page_frag *alloc_frag = &rq->alloc_frag;
857         char *buf;
858         unsigned int xdp_headroom = virtnet_get_headroom(vi);
859         void *ctx = (void *)(unsigned long)xdp_headroom;
860         int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
861         int err;
862
863         len = SKB_DATA_ALIGN(len) +
864               SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
865         if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
866                 return -ENOMEM;
867
868         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
869         get_page(alloc_frag->page);
870         alloc_frag->offset += len;
871         sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom,
872                     vi->hdr_len + GOOD_PACKET_LEN);
873         err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
874         if (err < 0)
875                 put_page(virt_to_head_page(buf));
876         return err;
877 }
878
879 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
880                            gfp_t gfp)
881 {
882         struct page *first, *list = NULL;
883         char *p;
884         int i, err, offset;
885
886         sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
887
888         /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
889         for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
890                 first = get_a_page(rq, gfp);
891                 if (!first) {
892                         if (list)
893                                 give_pages(rq, list);
894                         return -ENOMEM;
895                 }
896                 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
897
898                 /* chain new page in list head to match sg */
899                 first->private = (unsigned long)list;
900                 list = first;
901         }
902
903         first = get_a_page(rq, gfp);
904         if (!first) {
905                 give_pages(rq, list);
906                 return -ENOMEM;
907         }
908         p = page_address(first);
909
910         /* rq->sg[0], rq->sg[1] share the same page */
911         /* a separated rq->sg[0] for header - required in case !any_header_sg */
912         sg_set_buf(&rq->sg[0], p, vi->hdr_len);
913
914         /* rq->sg[1] for data packet, from offset */
915         offset = sizeof(struct padded_vnet_hdr);
916         sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
917
918         /* chain first in list head */
919         first->private = (unsigned long)list;
920         err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
921                                   first, gfp);
922         if (err < 0)
923                 give_pages(rq, first);
924
925         return err;
926 }
927
928 static unsigned int get_mergeable_buf_len(struct receive_queue *rq,
929                                           struct ewma_pkt_len *avg_pkt_len)
930 {
931         const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
932         unsigned int len;
933
934         len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
935                                 rq->min_buf_len, PAGE_SIZE - hdr_len);
936         return ALIGN(len, L1_CACHE_BYTES);
937 }
938
939 static int add_recvbuf_mergeable(struct virtnet_info *vi,
940                                  struct receive_queue *rq, gfp_t gfp)
941 {
942         struct page_frag *alloc_frag = &rq->alloc_frag;
943         unsigned int headroom = virtnet_get_headroom(vi);
944         char *buf;
945         void *ctx;
946         int err;
947         unsigned int len, hole;
948
949         len = get_mergeable_buf_len(rq, &rq->mrg_avg_pkt_len);
950         if (unlikely(!skb_page_frag_refill(len + headroom, alloc_frag, gfp)))
951                 return -ENOMEM;
952
953         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
954         buf += headroom; /* advance address leaving hole at front of pkt */
955         ctx = mergeable_len_to_ctx(len, headroom);
956         get_page(alloc_frag->page);
957         alloc_frag->offset += len + headroom;
958         hole = alloc_frag->size - alloc_frag->offset;
959         if (hole < len + headroom) {
960                 /* To avoid internal fragmentation, if there is very likely not
961                  * enough space for another buffer, add the remaining space to
962                  * the current buffer. This extra space is not included in
963                  * the truesize stored in ctx.
964                  */
965                 len += hole;
966                 alloc_frag->offset += hole;
967         }
968
969         sg_init_one(rq->sg, buf, len);
970         err = virtqueue_add_inbuf_ctx(rq->vq, rq->sg, 1, buf, ctx, gfp);
971         if (err < 0)
972                 put_page(virt_to_head_page(buf));
973
974         return err;
975 }
976
977 /*
978  * Returns false if we couldn't fill entirely (OOM).
979  *
980  * Normally run in the receive path, but can also be run from ndo_open
981  * before we're receiving packets, or from refill_work which is
982  * careful to disable receiving (using napi_disable).
983  */
984 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
985                           gfp_t gfp)
986 {
987         int err;
988         bool oom;
989
990         gfp |= __GFP_COLD;
991         do {
992                 if (vi->mergeable_rx_bufs)
993                         err = add_recvbuf_mergeable(vi, rq, gfp);
994                 else if (vi->big_packets)
995                         err = add_recvbuf_big(vi, rq, gfp);
996                 else
997                         err = add_recvbuf_small(vi, rq, gfp);
998
999                 oom = err == -ENOMEM;
1000                 if (err)
1001                         break;
1002         } while (rq->vq->num_free);
1003         virtqueue_kick(rq->vq);
1004         return !oom;
1005 }
1006
1007 static void skb_recv_done(struct virtqueue *rvq)
1008 {
1009         struct virtnet_info *vi = rvq->vdev->priv;
1010         struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
1011
1012         virtqueue_napi_schedule(&rq->napi, rvq);
1013 }
1014
1015 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
1016 {
1017         napi_enable(napi);
1018
1019         /* If all buffers were filled by other side before we napi_enabled, we
1020          * won't get another interrupt, so process any outstanding packets now.
1021          * Call local_bh_enable after to trigger softIRQ processing.
1022          */
1023         local_bh_disable();
1024         virtqueue_napi_schedule(napi, vq);
1025         local_bh_enable();
1026 }
1027
1028 static void virtnet_napi_tx_enable(struct virtnet_info *vi,
1029                                    struct virtqueue *vq,
1030                                    struct napi_struct *napi)
1031 {
1032         if (!napi->weight)
1033                 return;
1034
1035         /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
1036          * enable the feature if this is likely affine with the transmit path.
1037          */
1038         if (!vi->affinity_hint_set) {
1039                 napi->weight = 0;
1040                 return;
1041         }
1042
1043         return virtnet_napi_enable(vq, napi);
1044 }
1045
1046 static void virtnet_napi_tx_disable(struct napi_struct *napi)
1047 {
1048         if (napi->weight)
1049                 napi_disable(napi);
1050 }
1051
1052 static void refill_work(struct work_struct *work)
1053 {
1054         struct virtnet_info *vi =
1055                 container_of(work, struct virtnet_info, refill.work);
1056         bool still_empty;
1057         int i;
1058
1059         for (i = 0; i < vi->curr_queue_pairs; i++) {
1060                 struct receive_queue *rq = &vi->rq[i];
1061
1062                 napi_disable(&rq->napi);
1063                 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
1064                 virtnet_napi_enable(rq->vq, &rq->napi);
1065
1066                 /* In theory, this can happen: if we don't get any buffers in
1067                  * we will *never* try to fill again.
1068                  */
1069                 if (still_empty)
1070                         schedule_delayed_work(&vi->refill, HZ/2);
1071         }
1072 }
1073
1074 static int virtnet_receive(struct receive_queue *rq, int budget)
1075 {
1076         struct virtnet_info *vi = rq->vq->vdev->priv;
1077         unsigned int len, received = 0, bytes = 0;
1078         void *buf;
1079         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
1080
1081         if (!vi->big_packets || vi->mergeable_rx_bufs) {
1082                 void *ctx;
1083
1084                 while (received < budget &&
1085                        (buf = virtqueue_get_buf_ctx(rq->vq, &len, &ctx))) {
1086                         bytes += receive_buf(vi, rq, buf, len, ctx);
1087                         received++;
1088                 }
1089         } else {
1090                 while (received < budget &&
1091                        (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
1092                         bytes += receive_buf(vi, rq, buf, len, NULL);
1093                         received++;
1094                 }
1095         }
1096
1097         if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) {
1098                 if (!try_fill_recv(vi, rq, GFP_ATOMIC))
1099                         schedule_delayed_work(&vi->refill, 0);
1100         }
1101
1102         u64_stats_update_begin(&stats->rx_syncp);
1103         stats->rx_bytes += bytes;
1104         stats->rx_packets += received;
1105         u64_stats_update_end(&stats->rx_syncp);
1106
1107         return received;
1108 }
1109
1110 static void free_old_xmit_skbs(struct send_queue *sq)
1111 {
1112         struct sk_buff *skb;
1113         unsigned int len;
1114         struct virtnet_info *vi = sq->vq->vdev->priv;
1115         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
1116         unsigned int packets = 0;
1117         unsigned int bytes = 0;
1118
1119         while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
1120                 pr_debug("Sent skb %p\n", skb);
1121
1122                 bytes += skb->len;
1123                 packets++;
1124
1125                 dev_kfree_skb_any(skb);
1126         }
1127
1128         /* Avoid overhead when no packets have been processed
1129          * happens when called speculatively from start_xmit.
1130          */
1131         if (!packets)
1132                 return;
1133
1134         u64_stats_update_begin(&stats->tx_syncp);
1135         stats->tx_bytes += bytes;
1136         stats->tx_packets += packets;
1137         u64_stats_update_end(&stats->tx_syncp);
1138 }
1139
1140 static void virtnet_poll_cleantx(struct receive_queue *rq)
1141 {
1142         struct virtnet_info *vi = rq->vq->vdev->priv;
1143         unsigned int index = vq2rxq(rq->vq);
1144         struct send_queue *sq = &vi->sq[index];
1145         struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index);
1146
1147         if (!sq->napi.weight)
1148                 return;
1149
1150         if (__netif_tx_trylock(txq)) {
1151                 free_old_xmit_skbs(sq);
1152                 __netif_tx_unlock(txq);
1153         }
1154
1155         if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1156                 netif_tx_wake_queue(txq);
1157 }
1158
1159 static int virtnet_poll(struct napi_struct *napi, int budget)
1160 {
1161         struct receive_queue *rq =
1162                 container_of(napi, struct receive_queue, napi);
1163         unsigned int received;
1164
1165         virtnet_poll_cleantx(rq);
1166
1167         received = virtnet_receive(rq, budget);
1168
1169         /* Out of packets? */
1170         if (received < budget)
1171                 virtqueue_napi_complete(napi, rq->vq, received);
1172
1173         return received;
1174 }
1175
1176 static int virtnet_open(struct net_device *dev)
1177 {
1178         struct virtnet_info *vi = netdev_priv(dev);
1179         int i;
1180
1181         for (i = 0; i < vi->max_queue_pairs; i++) {
1182                 if (i < vi->curr_queue_pairs)
1183                         /* Make sure we have some buffers: if oom use wq. */
1184                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1185                                 schedule_delayed_work(&vi->refill, 0);
1186                 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1187                 virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi);
1188         }
1189
1190         return 0;
1191 }
1192
1193 static int virtnet_poll_tx(struct napi_struct *napi, int budget)
1194 {
1195         struct send_queue *sq = container_of(napi, struct send_queue, napi);
1196         struct virtnet_info *vi = sq->vq->vdev->priv;
1197         struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, vq2txq(sq->vq));
1198
1199         __netif_tx_lock(txq, raw_smp_processor_id());
1200         free_old_xmit_skbs(sq);
1201         __netif_tx_unlock(txq);
1202
1203         virtqueue_napi_complete(napi, sq->vq, 0);
1204
1205         if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1206                 netif_tx_wake_queue(txq);
1207
1208         return 0;
1209 }
1210
1211 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
1212 {
1213         struct virtio_net_hdr_mrg_rxbuf *hdr;
1214         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
1215         struct virtnet_info *vi = sq->vq->vdev->priv;
1216         int num_sg;
1217         unsigned hdr_len = vi->hdr_len;
1218         bool can_push;
1219
1220         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
1221
1222         can_push = vi->any_header_sg &&
1223                 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
1224                 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
1225         /* Even if we can, don't push here yet as this would skew
1226          * csum_start offset below. */
1227         if (can_push)
1228                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1229         else
1230                 hdr = skb_vnet_hdr(skb);
1231
1232         if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1233                                     virtio_is_little_endian(vi->vdev), false))
1234                 BUG();
1235
1236         if (vi->mergeable_rx_bufs)
1237                 hdr->num_buffers = 0;
1238
1239         sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1240         if (can_push) {
1241                 __skb_push(skb, hdr_len);
1242                 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1243                 if (unlikely(num_sg < 0))
1244                         return num_sg;
1245                 /* Pull header back to avoid skew in tx bytes calculations. */
1246                 __skb_pull(skb, hdr_len);
1247         } else {
1248                 sg_set_buf(sq->sg, hdr, hdr_len);
1249                 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len);
1250                 if (unlikely(num_sg < 0))
1251                         return num_sg;
1252                 num_sg++;
1253         }
1254         return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
1255 }
1256
1257 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
1258 {
1259         struct virtnet_info *vi = netdev_priv(dev);
1260         int qnum = skb_get_queue_mapping(skb);
1261         struct send_queue *sq = &vi->sq[qnum];
1262         int err;
1263         struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1264         bool kick = !skb->xmit_more;
1265         bool use_napi = sq->napi.weight;
1266
1267         /* Free up any pending old buffers before queueing new ones. */
1268         free_old_xmit_skbs(sq);
1269
1270         if (use_napi && kick)
1271                 virtqueue_enable_cb_delayed(sq->vq);
1272
1273         /* timestamp packet in software */
1274         skb_tx_timestamp(skb);
1275
1276         /* Try to transmit */
1277         err = xmit_skb(sq, skb);
1278
1279         /* This should not happen! */
1280         if (unlikely(err)) {
1281                 dev->stats.tx_fifo_errors++;
1282                 if (net_ratelimit())
1283                         dev_warn(&dev->dev,
1284                                  "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
1285                 dev->stats.tx_dropped++;
1286                 dev_kfree_skb_any(skb);
1287                 return NETDEV_TX_OK;
1288         }
1289
1290         /* Don't wait up for transmitted skbs to be freed. */
1291         if (!use_napi) {
1292                 skb_orphan(skb);
1293                 nf_reset(skb);
1294         }
1295
1296         /* If running out of space, stop queue to avoid getting packets that we
1297          * are then unable to transmit.
1298          * An alternative would be to force queuing layer to requeue the skb by
1299          * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1300          * returned in a normal path of operation: it means that driver is not
1301          * maintaining the TX queue stop/start state properly, and causes
1302          * the stack to do a non-trivial amount of useless work.
1303          * Since most packets only take 1 or 2 ring slots, stopping the queue
1304          * early means 16 slots are typically wasted.
1305          */
1306         if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
1307                 netif_stop_subqueue(dev, qnum);
1308                 if (!use_napi &&
1309                     unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
1310                         /* More just got used, free them then recheck. */
1311                         free_old_xmit_skbs(sq);
1312                         if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
1313                                 netif_start_subqueue(dev, qnum);
1314                                 virtqueue_disable_cb(sq->vq);
1315                         }
1316                 }
1317         }
1318
1319         if (kick || netif_xmit_stopped(txq))
1320                 virtqueue_kick(sq->vq);
1321
1322         return NETDEV_TX_OK;
1323 }
1324
1325 /*
1326  * Send command via the control virtqueue and check status.  Commands
1327  * supported by the hypervisor, as indicated by feature bits, should
1328  * never fail unless improperly formatted.
1329  */
1330 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
1331                                  struct scatterlist *out)
1332 {
1333         struct scatterlist *sgs[4], hdr, stat;
1334         unsigned out_num = 0, tmp;
1335
1336         /* Caller should know better */
1337         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
1338
1339         vi->ctrl_status = ~0;
1340         vi->ctrl_hdr.class = class;
1341         vi->ctrl_hdr.cmd = cmd;
1342         /* Add header */
1343         sg_init_one(&hdr, &vi->ctrl_hdr, sizeof(vi->ctrl_hdr));
1344         sgs[out_num++] = &hdr;
1345
1346         if (out)
1347                 sgs[out_num++] = out;
1348
1349         /* Add return status. */
1350         sg_init_one(&stat, &vi->ctrl_status, sizeof(vi->ctrl_status));
1351         sgs[out_num] = &stat;
1352
1353         BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1354         virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1355
1356         if (unlikely(!virtqueue_kick(vi->cvq)))
1357                 return vi->ctrl_status == VIRTIO_NET_OK;
1358
1359         /* Spin for a response, the kick causes an ioport write, trapping
1360          * into the hypervisor, so the request should be handled immediately.
1361          */
1362         while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1363                !virtqueue_is_broken(vi->cvq))
1364                 cpu_relax();
1365
1366         return vi->ctrl_status == VIRTIO_NET_OK;
1367 }
1368
1369 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1370 {
1371         struct virtnet_info *vi = netdev_priv(dev);
1372         struct virtio_device *vdev = vi->vdev;
1373         int ret;
1374         struct sockaddr *addr;
1375         struct scatterlist sg;
1376
1377         addr = kmemdup(p, sizeof(*addr), GFP_KERNEL);
1378         if (!addr)
1379                 return -ENOMEM;
1380
1381         ret = eth_prepare_mac_addr_change(dev, addr);
1382         if (ret)
1383                 goto out;
1384
1385         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1386                 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1387                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1388                                           VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1389                         dev_warn(&vdev->dev,
1390                                  "Failed to set mac address by vq command.\n");
1391                         ret = -EINVAL;
1392                         goto out;
1393                 }
1394         } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1395                    !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1396                 unsigned int i;
1397
1398                 /* Naturally, this has an atomicity problem. */
1399                 for (i = 0; i < dev->addr_len; i++)
1400                         virtio_cwrite8(vdev,
1401                                        offsetof(struct virtio_net_config, mac) +
1402                                        i, addr->sa_data[i]);
1403         }
1404
1405         eth_commit_mac_addr_change(dev, p);
1406         ret = 0;
1407
1408 out:
1409         kfree(addr);
1410         return ret;
1411 }
1412
1413 static void virtnet_stats(struct net_device *dev,
1414                           struct rtnl_link_stats64 *tot)
1415 {
1416         struct virtnet_info *vi = netdev_priv(dev);
1417         int cpu;
1418         unsigned int start;
1419
1420         for_each_possible_cpu(cpu) {
1421                 struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
1422                 u64 tpackets, tbytes, rpackets, rbytes;
1423
1424                 do {
1425                         start = u64_stats_fetch_begin_irq(&stats->tx_syncp);
1426                         tpackets = stats->tx_packets;
1427                         tbytes   = stats->tx_bytes;
1428                 } while (u64_stats_fetch_retry_irq(&stats->tx_syncp, start));
1429
1430                 do {
1431                         start = u64_stats_fetch_begin_irq(&stats->rx_syncp);
1432                         rpackets = stats->rx_packets;
1433                         rbytes   = stats->rx_bytes;
1434                 } while (u64_stats_fetch_retry_irq(&stats->rx_syncp, start));
1435
1436                 tot->rx_packets += rpackets;
1437                 tot->tx_packets += tpackets;
1438                 tot->rx_bytes   += rbytes;
1439                 tot->tx_bytes   += tbytes;
1440         }
1441
1442         tot->tx_dropped = dev->stats.tx_dropped;
1443         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1444         tot->rx_dropped = dev->stats.rx_dropped;
1445         tot->rx_length_errors = dev->stats.rx_length_errors;
1446         tot->rx_frame_errors = dev->stats.rx_frame_errors;
1447 }
1448
1449 #ifdef CONFIG_NET_POLL_CONTROLLER
1450 static void virtnet_netpoll(struct net_device *dev)
1451 {
1452         struct virtnet_info *vi = netdev_priv(dev);
1453         int i;
1454
1455         for (i = 0; i < vi->curr_queue_pairs; i++)
1456                 napi_schedule(&vi->rq[i].napi);
1457 }
1458 #endif
1459
1460 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1461 {
1462         rtnl_lock();
1463         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1464                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1465                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1466         rtnl_unlock();
1467 }
1468
1469 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1470 {
1471         struct scatterlist sg;
1472         struct net_device *dev = vi->dev;
1473
1474         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1475                 return 0;
1476
1477         vi->ctrl_mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1478         sg_init_one(&sg, &vi->ctrl_mq, sizeof(vi->ctrl_mq));
1479
1480         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1481                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1482                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1483                          queue_pairs);
1484                 return -EINVAL;
1485         } else {
1486                 vi->curr_queue_pairs = queue_pairs;
1487                 /* virtnet_open() will refill when device is going to up. */
1488                 if (dev->flags & IFF_UP)
1489                         schedule_delayed_work(&vi->refill, 0);
1490         }
1491
1492         return 0;
1493 }
1494
1495 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1496 {
1497         int err;
1498
1499         rtnl_lock();
1500         err = _virtnet_set_queues(vi, queue_pairs);
1501         rtnl_unlock();
1502         return err;
1503 }
1504
1505 static int virtnet_close(struct net_device *dev)
1506 {
1507         struct virtnet_info *vi = netdev_priv(dev);
1508         int i;
1509
1510         /* Make sure refill_work doesn't re-enable napi! */
1511         cancel_delayed_work_sync(&vi->refill);
1512
1513         for (i = 0; i < vi->max_queue_pairs; i++) {
1514                 napi_disable(&vi->rq[i].napi);
1515                 virtnet_napi_tx_disable(&vi->sq[i].napi);
1516         }
1517
1518         return 0;
1519 }
1520
1521 static void virtnet_set_rx_mode(struct net_device *dev)
1522 {
1523         struct virtnet_info *vi = netdev_priv(dev);
1524         struct scatterlist sg[2];
1525         struct virtio_net_ctrl_mac *mac_data;
1526         struct netdev_hw_addr *ha;
1527         int uc_count;
1528         int mc_count;
1529         void *buf;
1530         int i;
1531
1532         /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1533         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1534                 return;
1535
1536         vi->ctrl_promisc = ((dev->flags & IFF_PROMISC) != 0);
1537         vi->ctrl_allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1538
1539         sg_init_one(sg, &vi->ctrl_promisc, sizeof(vi->ctrl_promisc));
1540
1541         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1542                                   VIRTIO_NET_CTRL_RX_PROMISC, sg))
1543                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1544                          vi->ctrl_promisc ? "en" : "dis");
1545
1546         sg_init_one(sg, &vi->ctrl_allmulti, sizeof(vi->ctrl_allmulti));
1547
1548         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1549                                   VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1550                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1551                          vi->ctrl_allmulti ? "en" : "dis");
1552
1553         uc_count = netdev_uc_count(dev);
1554         mc_count = netdev_mc_count(dev);
1555         /* MAC filter - use one buffer for both lists */
1556         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1557                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1558         mac_data = buf;
1559         if (!buf)
1560                 return;
1561
1562         sg_init_table(sg, 2);
1563
1564         /* Store the unicast list and count in the front of the buffer */
1565         mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1566         i = 0;
1567         netdev_for_each_uc_addr(ha, dev)
1568                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1569
1570         sg_set_buf(&sg[0], mac_data,
1571                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1572
1573         /* multicast list and count fill the end */
1574         mac_data = (void *)&mac_data->macs[uc_count][0];
1575
1576         mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1577         i = 0;
1578         netdev_for_each_mc_addr(ha, dev)
1579                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1580
1581         sg_set_buf(&sg[1], mac_data,
1582                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1583
1584         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1585                                   VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1586                 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1587
1588         kfree(buf);
1589 }
1590
1591 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1592                                    __be16 proto, u16 vid)
1593 {
1594         struct virtnet_info *vi = netdev_priv(dev);
1595         struct scatterlist sg;
1596
1597         vi->ctrl_vid = vid;
1598         sg_init_one(&sg, &vi->ctrl_vid, sizeof(vi->ctrl_vid));
1599
1600         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1601                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1602                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1603         return 0;
1604 }
1605
1606 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1607                                     __be16 proto, u16 vid)
1608 {
1609         struct virtnet_info *vi = netdev_priv(dev);
1610         struct scatterlist sg;
1611
1612         vi->ctrl_vid = vid;
1613         sg_init_one(&sg, &vi->ctrl_vid, sizeof(vi->ctrl_vid));
1614
1615         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1616                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1617                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1618         return 0;
1619 }
1620
1621 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1622 {
1623         int i;
1624
1625         if (vi->affinity_hint_set) {
1626                 for (i = 0; i < vi->max_queue_pairs; i++) {
1627                         virtqueue_set_affinity(vi->rq[i].vq, -1);
1628                         virtqueue_set_affinity(vi->sq[i].vq, -1);
1629                 }
1630
1631                 vi->affinity_hint_set = false;
1632         }
1633 }
1634
1635 static void virtnet_set_affinity(struct virtnet_info *vi)
1636 {
1637         int i;
1638         int cpu;
1639
1640         /* In multiqueue mode, when the number of cpu is equal to the number of
1641          * queue pairs, we let the queue pairs to be private to one cpu by
1642          * setting the affinity hint to eliminate the contention.
1643          */
1644         if (vi->curr_queue_pairs == 1 ||
1645             vi->max_queue_pairs != num_online_cpus()) {
1646                 virtnet_clean_affinity(vi, -1);
1647                 return;
1648         }
1649
1650         i = 0;
1651         for_each_online_cpu(cpu) {
1652                 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1653                 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1654                 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
1655                 i++;
1656         }
1657
1658         vi->affinity_hint_set = true;
1659 }
1660
1661 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
1662 {
1663         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1664                                                    node);
1665         virtnet_set_affinity(vi);
1666         return 0;
1667 }
1668
1669 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
1670 {
1671         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1672                                                    node_dead);
1673         virtnet_set_affinity(vi);
1674         return 0;
1675 }
1676
1677 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
1678 {
1679         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1680                                                    node);
1681
1682         virtnet_clean_affinity(vi, cpu);
1683         return 0;
1684 }
1685
1686 static enum cpuhp_state virtionet_online;
1687
1688 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
1689 {
1690         int ret;
1691
1692         ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
1693         if (ret)
1694                 return ret;
1695         ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1696                                                &vi->node_dead);
1697         if (!ret)
1698                 return ret;
1699         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1700         return ret;
1701 }
1702
1703 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
1704 {
1705         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1706         cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1707                                             &vi->node_dead);
1708 }
1709
1710 static void virtnet_get_ringparam(struct net_device *dev,
1711                                 struct ethtool_ringparam *ring)
1712 {
1713         struct virtnet_info *vi = netdev_priv(dev);
1714
1715         ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1716         ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1717         ring->rx_pending = ring->rx_max_pending;
1718         ring->tx_pending = ring->tx_max_pending;
1719 }
1720
1721
1722 static void virtnet_get_drvinfo(struct net_device *dev,
1723                                 struct ethtool_drvinfo *info)
1724 {
1725         struct virtnet_info *vi = netdev_priv(dev);
1726         struct virtio_device *vdev = vi->vdev;
1727
1728         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1729         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1730         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1731
1732 }
1733
1734 /* TODO: Eliminate OOO packets during switching */
1735 static int virtnet_set_channels(struct net_device *dev,
1736                                 struct ethtool_channels *channels)
1737 {
1738         struct virtnet_info *vi = netdev_priv(dev);
1739         u16 queue_pairs = channels->combined_count;
1740         int err;
1741
1742         /* We don't support separate rx/tx channels.
1743          * We don't allow setting 'other' channels.
1744          */
1745         if (channels->rx_count || channels->tx_count || channels->other_count)
1746                 return -EINVAL;
1747
1748         if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
1749                 return -EINVAL;
1750
1751         /* For now we don't support modifying channels while XDP is loaded
1752          * also when XDP is loaded all RX queues have XDP programs so we only
1753          * need to check a single RX queue.
1754          */
1755         if (vi->rq[0].xdp_prog)
1756                 return -EINVAL;
1757
1758         get_online_cpus();
1759         err = _virtnet_set_queues(vi, queue_pairs);
1760         if (!err) {
1761                 netif_set_real_num_tx_queues(dev, queue_pairs);
1762                 netif_set_real_num_rx_queues(dev, queue_pairs);
1763
1764                 virtnet_set_affinity(vi);
1765         }
1766         put_online_cpus();
1767
1768         return err;
1769 }
1770
1771 static void virtnet_get_channels(struct net_device *dev,
1772                                  struct ethtool_channels *channels)
1773 {
1774         struct virtnet_info *vi = netdev_priv(dev);
1775
1776         channels->combined_count = vi->curr_queue_pairs;
1777         channels->max_combined = vi->max_queue_pairs;
1778         channels->max_other = 0;
1779         channels->rx_count = 0;
1780         channels->tx_count = 0;
1781         channels->other_count = 0;
1782 }
1783
1784 /* Check if the user is trying to change anything besides speed/duplex */
1785 static bool
1786 virtnet_validate_ethtool_cmd(const struct ethtool_link_ksettings *cmd)
1787 {
1788         struct ethtool_link_ksettings diff1 = *cmd;
1789         struct ethtool_link_ksettings diff2 = {};
1790
1791         /* cmd is always set so we need to clear it, validate the port type
1792          * and also without autonegotiation we can ignore advertising
1793          */
1794         diff1.base.speed = 0;
1795         diff2.base.port = PORT_OTHER;
1796         ethtool_link_ksettings_zero_link_mode(&diff1, advertising);
1797         diff1.base.duplex = 0;
1798         diff1.base.cmd = 0;
1799         diff1.base.link_mode_masks_nwords = 0;
1800
1801         return !memcmp(&diff1.base, &diff2.base, sizeof(diff1.base)) &&
1802                 bitmap_empty(diff1.link_modes.supported,
1803                              __ETHTOOL_LINK_MODE_MASK_NBITS) &&
1804                 bitmap_empty(diff1.link_modes.advertising,
1805                              __ETHTOOL_LINK_MODE_MASK_NBITS) &&
1806                 bitmap_empty(diff1.link_modes.lp_advertising,
1807                              __ETHTOOL_LINK_MODE_MASK_NBITS);
1808 }
1809
1810 static int virtnet_set_link_ksettings(struct net_device *dev,
1811                                       const struct ethtool_link_ksettings *cmd)
1812 {
1813         struct virtnet_info *vi = netdev_priv(dev);
1814         u32 speed;
1815
1816         speed = cmd->base.speed;
1817         /* don't allow custom speed and duplex */
1818         if (!ethtool_validate_speed(speed) ||
1819             !ethtool_validate_duplex(cmd->base.duplex) ||
1820             !virtnet_validate_ethtool_cmd(cmd))
1821                 return -EINVAL;
1822         vi->speed = speed;
1823         vi->duplex = cmd->base.duplex;
1824
1825         return 0;
1826 }
1827
1828 static int virtnet_get_link_ksettings(struct net_device *dev,
1829                                       struct ethtool_link_ksettings *cmd)
1830 {
1831         struct virtnet_info *vi = netdev_priv(dev);
1832
1833         cmd->base.speed = vi->speed;
1834         cmd->base.duplex = vi->duplex;
1835         cmd->base.port = PORT_OTHER;
1836
1837         return 0;
1838 }
1839
1840 static void virtnet_init_settings(struct net_device *dev)
1841 {
1842         struct virtnet_info *vi = netdev_priv(dev);
1843
1844         vi->speed = SPEED_UNKNOWN;
1845         vi->duplex = DUPLEX_UNKNOWN;
1846 }
1847
1848 static const struct ethtool_ops virtnet_ethtool_ops = {
1849         .get_drvinfo = virtnet_get_drvinfo,
1850         .get_link = ethtool_op_get_link,
1851         .get_ringparam = virtnet_get_ringparam,
1852         .set_channels = virtnet_set_channels,
1853         .get_channels = virtnet_get_channels,
1854         .get_ts_info = ethtool_op_get_ts_info,
1855         .get_link_ksettings = virtnet_get_link_ksettings,
1856         .set_link_ksettings = virtnet_set_link_ksettings,
1857 };
1858
1859 static void virtnet_freeze_down(struct virtio_device *vdev)
1860 {
1861         struct virtnet_info *vi = vdev->priv;
1862         int i;
1863
1864         /* Make sure no work handler is accessing the device */
1865         flush_work(&vi->config_work);
1866
1867         netif_device_detach(vi->dev);
1868         netif_tx_disable(vi->dev);
1869         cancel_delayed_work_sync(&vi->refill);
1870
1871         if (netif_running(vi->dev)) {
1872                 for (i = 0; i < vi->max_queue_pairs; i++) {
1873                         napi_disable(&vi->rq[i].napi);
1874                         virtnet_napi_tx_disable(&vi->sq[i].napi);
1875                 }
1876         }
1877 }
1878
1879 static int init_vqs(struct virtnet_info *vi);
1880
1881 static int virtnet_restore_up(struct virtio_device *vdev)
1882 {
1883         struct virtnet_info *vi = vdev->priv;
1884         int err, i;
1885
1886         err = init_vqs(vi);
1887         if (err)
1888                 return err;
1889
1890         virtio_device_ready(vdev);
1891
1892         if (netif_running(vi->dev)) {
1893                 for (i = 0; i < vi->curr_queue_pairs; i++)
1894                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1895                                 schedule_delayed_work(&vi->refill, 0);
1896
1897                 for (i = 0; i < vi->max_queue_pairs; i++) {
1898                         virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1899                         virtnet_napi_tx_enable(vi, vi->sq[i].vq,
1900                                                &vi->sq[i].napi);
1901                 }
1902         }
1903
1904         netif_device_attach(vi->dev);
1905         return err;
1906 }
1907
1908 static int virtnet_set_guest_offloads(struct virtnet_info *vi, u64 offloads)
1909 {
1910         struct scatterlist sg;
1911         vi->ctrl_offloads = cpu_to_virtio64(vi->vdev, offloads);
1912
1913         sg_init_one(&sg, &vi->ctrl_offloads, sizeof(vi->ctrl_offloads));
1914
1915         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_GUEST_OFFLOADS,
1916                                   VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, &sg)) {
1917                 dev_warn(&vi->dev->dev, "Fail to set guest offload. \n");
1918                 return -EINVAL;
1919         }
1920
1921         return 0;
1922 }
1923
1924 static int virtnet_clear_guest_offloads(struct virtnet_info *vi)
1925 {
1926         u64 offloads = 0;
1927
1928         if (!vi->guest_offloads)
1929                 return 0;
1930
1931         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))
1932                 offloads = 1ULL << VIRTIO_NET_F_GUEST_CSUM;
1933
1934         return virtnet_set_guest_offloads(vi, offloads);
1935 }
1936
1937 static int virtnet_restore_guest_offloads(struct virtnet_info *vi)
1938 {
1939         u64 offloads = vi->guest_offloads;
1940
1941         if (!vi->guest_offloads)
1942                 return 0;
1943         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_CSUM))
1944                 offloads |= 1ULL << VIRTIO_NET_F_GUEST_CSUM;
1945
1946         return virtnet_set_guest_offloads(vi, offloads);
1947 }
1948
1949 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
1950                            struct netlink_ext_ack *extack)
1951 {
1952         unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr);
1953         struct virtnet_info *vi = netdev_priv(dev);
1954         struct bpf_prog *old_prog;
1955         u16 xdp_qp = 0, curr_qp;
1956         int i, err;
1957
1958         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)
1959             && (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1960                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1961                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
1962                 virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO))) {
1963                 NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing LRO, disable LRO first");
1964                 return -EOPNOTSUPP;
1965         }
1966
1967         if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
1968                 NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required");
1969                 return -EINVAL;
1970         }
1971
1972         if (dev->mtu > max_sz) {
1973                 NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP");
1974                 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz);
1975                 return -EINVAL;
1976         }
1977
1978         curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
1979         if (prog)
1980                 xdp_qp = nr_cpu_ids;
1981
1982         /* XDP requires extra queues for XDP_TX */
1983         if (curr_qp + xdp_qp > vi->max_queue_pairs) {
1984                 NL_SET_ERR_MSG_MOD(extack, "Too few free TX rings available");
1985                 netdev_warn(dev, "request %i queues but max is %i\n",
1986                             curr_qp + xdp_qp, vi->max_queue_pairs);
1987                 return -ENOMEM;
1988         }
1989
1990         if (prog) {
1991                 prog = bpf_prog_add(prog, vi->max_queue_pairs - 1);
1992                 if (IS_ERR(prog))
1993                         return PTR_ERR(prog);
1994         }
1995
1996         /* Make sure NAPI is not using any XDP TX queues for RX. */
1997         for (i = 0; i < vi->max_queue_pairs; i++)
1998                 napi_disable(&vi->rq[i].napi);
1999
2000         netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
2001         err = _virtnet_set_queues(vi, curr_qp + xdp_qp);
2002         if (err)
2003                 goto err;
2004         vi->xdp_queue_pairs = xdp_qp;
2005
2006         for (i = 0; i < vi->max_queue_pairs; i++) {
2007                 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2008                 rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
2009                 if (i == 0) {
2010                         if (!old_prog)
2011                                 virtnet_clear_guest_offloads(vi);
2012                         if (!prog)
2013                                 virtnet_restore_guest_offloads(vi);
2014                 }
2015                 if (old_prog)
2016                         bpf_prog_put(old_prog);
2017                 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2018         }
2019
2020         return 0;
2021
2022 err:
2023         for (i = 0; i < vi->max_queue_pairs; i++)
2024                 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
2025         if (prog)
2026                 bpf_prog_sub(prog, vi->max_queue_pairs - 1);
2027         return err;
2028 }
2029
2030 static u32 virtnet_xdp_query(struct net_device *dev)
2031 {
2032         struct virtnet_info *vi = netdev_priv(dev);
2033         const struct bpf_prog *xdp_prog;
2034         int i;
2035
2036         for (i = 0; i < vi->max_queue_pairs; i++) {
2037                 xdp_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2038                 if (xdp_prog)
2039                         return xdp_prog->aux->id;
2040         }
2041         return 0;
2042 }
2043
2044 static int virtnet_xdp(struct net_device *dev, struct netdev_xdp *xdp)
2045 {
2046         switch (xdp->command) {
2047         case XDP_SETUP_PROG:
2048                 return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
2049         case XDP_QUERY_PROG:
2050                 xdp->prog_id = virtnet_xdp_query(dev);
2051                 xdp->prog_attached = !!xdp->prog_id;
2052                 return 0;
2053         default:
2054                 return -EINVAL;
2055         }
2056 }
2057
2058 static const struct net_device_ops virtnet_netdev = {
2059         .ndo_open            = virtnet_open,
2060         .ndo_stop            = virtnet_close,
2061         .ndo_start_xmit      = start_xmit,
2062         .ndo_validate_addr   = eth_validate_addr,
2063         .ndo_set_mac_address = virtnet_set_mac_address,
2064         .ndo_set_rx_mode     = virtnet_set_rx_mode,
2065         .ndo_get_stats64     = virtnet_stats,
2066         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
2067         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
2068 #ifdef CONFIG_NET_POLL_CONTROLLER
2069         .ndo_poll_controller = virtnet_netpoll,
2070 #endif
2071         .ndo_xdp                = virtnet_xdp,
2072         .ndo_features_check     = passthru_features_check,
2073 };
2074
2075 static void virtnet_config_changed_work(struct work_struct *work)
2076 {
2077         struct virtnet_info *vi =
2078                 container_of(work, struct virtnet_info, config_work);
2079         u16 v;
2080
2081         if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
2082                                  struct virtio_net_config, status, &v) < 0)
2083                 return;
2084
2085         if (v & VIRTIO_NET_S_ANNOUNCE) {
2086                 netdev_notify_peers(vi->dev);
2087                 virtnet_ack_link_announce(vi);
2088         }
2089
2090         /* Ignore unknown (future) status bits */
2091         v &= VIRTIO_NET_S_LINK_UP;
2092
2093         if (vi->status == v)
2094                 return;
2095
2096         vi->status = v;
2097
2098         if (vi->status & VIRTIO_NET_S_LINK_UP) {
2099                 netif_carrier_on(vi->dev);
2100                 netif_tx_wake_all_queues(vi->dev);
2101         } else {
2102                 netif_carrier_off(vi->dev);
2103                 netif_tx_stop_all_queues(vi->dev);
2104         }
2105 }
2106
2107 static void virtnet_config_changed(struct virtio_device *vdev)
2108 {
2109         struct virtnet_info *vi = vdev->priv;
2110
2111         schedule_work(&vi->config_work);
2112 }
2113
2114 static void virtnet_free_queues(struct virtnet_info *vi)
2115 {
2116         int i;
2117
2118         for (i = 0; i < vi->max_queue_pairs; i++) {
2119                 napi_hash_del(&vi->rq[i].napi);
2120                 netif_napi_del(&vi->rq[i].napi);
2121                 netif_napi_del(&vi->sq[i].napi);
2122         }
2123
2124         /* We called napi_hash_del() before netif_napi_del(),
2125          * we need to respect an RCU grace period before freeing vi->rq
2126          */
2127         synchronize_net();
2128
2129         kfree(vi->rq);
2130         kfree(vi->sq);
2131 }
2132
2133 static void _free_receive_bufs(struct virtnet_info *vi)
2134 {
2135         struct bpf_prog *old_prog;
2136         int i;
2137
2138         for (i = 0; i < vi->max_queue_pairs; i++) {
2139                 while (vi->rq[i].pages)
2140                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
2141
2142                 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2143                 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
2144                 if (old_prog)
2145                         bpf_prog_put(old_prog);
2146         }
2147 }
2148
2149 static void free_receive_bufs(struct virtnet_info *vi)
2150 {
2151         rtnl_lock();
2152         _free_receive_bufs(vi);
2153         rtnl_unlock();
2154 }
2155
2156 static void free_receive_page_frags(struct virtnet_info *vi)
2157 {
2158         int i;
2159         for (i = 0; i < vi->max_queue_pairs; i++)
2160                 if (vi->rq[i].alloc_frag.page)
2161                         put_page(vi->rq[i].alloc_frag.page);
2162 }
2163
2164 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
2165 {
2166         if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
2167                 return false;
2168         else if (q < vi->curr_queue_pairs)
2169                 return true;
2170         else
2171                 return false;
2172 }
2173
2174 static void free_unused_bufs(struct virtnet_info *vi)
2175 {
2176         void *buf;
2177         int i;
2178
2179         for (i = 0; i < vi->max_queue_pairs; i++) {
2180                 struct virtqueue *vq = vi->sq[i].vq;
2181                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2182                         if (!is_xdp_raw_buffer_queue(vi, i))
2183                                 dev_kfree_skb(buf);
2184                         else
2185                                 put_page(virt_to_head_page(buf));
2186                 }
2187         }
2188
2189         for (i = 0; i < vi->max_queue_pairs; i++) {
2190                 struct virtqueue *vq = vi->rq[i].vq;
2191
2192                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2193                         if (vi->mergeable_rx_bufs) {
2194                                 put_page(virt_to_head_page(buf));
2195                         } else if (vi->big_packets) {
2196                                 give_pages(&vi->rq[i], buf);
2197                         } else {
2198                                 put_page(virt_to_head_page(buf));
2199                         }
2200                 }
2201         }
2202 }
2203
2204 static void virtnet_del_vqs(struct virtnet_info *vi)
2205 {
2206         struct virtio_device *vdev = vi->vdev;
2207
2208         virtnet_clean_affinity(vi, -1);
2209
2210         vdev->config->del_vqs(vdev);
2211
2212         virtnet_free_queues(vi);
2213 }
2214
2215 /* How large should a single buffer be so a queue full of these can fit at
2216  * least one full packet?
2217  * Logic below assumes the mergeable buffer header is used.
2218  */
2219 static unsigned int mergeable_min_buf_len(struct virtnet_info *vi, struct virtqueue *vq)
2220 {
2221         const unsigned int hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2222         unsigned int rq_size = virtqueue_get_vring_size(vq);
2223         unsigned int packet_len = vi->big_packets ? IP_MAX_MTU : vi->dev->max_mtu;
2224         unsigned int buf_len = hdr_len + ETH_HLEN + VLAN_HLEN + packet_len;
2225         unsigned int min_buf_len = DIV_ROUND_UP(buf_len, rq_size);
2226
2227         return max(max(min_buf_len, hdr_len) - hdr_len,
2228                    (unsigned int)GOOD_PACKET_LEN);
2229 }
2230
2231 static int virtnet_find_vqs(struct virtnet_info *vi)
2232 {
2233         vq_callback_t **callbacks;
2234         struct virtqueue **vqs;
2235         int ret = -ENOMEM;
2236         int i, total_vqs;
2237         const char **names;
2238         bool *ctx;
2239
2240         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
2241          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
2242          * possible control vq.
2243          */
2244         total_vqs = vi->max_queue_pairs * 2 +
2245                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
2246
2247         /* Allocate space for find_vqs parameters */
2248         vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
2249         if (!vqs)
2250                 goto err_vq;
2251         callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
2252         if (!callbacks)
2253                 goto err_callback;
2254         names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
2255         if (!names)
2256                 goto err_names;
2257         if (!vi->big_packets || vi->mergeable_rx_bufs) {
2258                 ctx = kzalloc(total_vqs * sizeof(*ctx), GFP_KERNEL);
2259                 if (!ctx)
2260                         goto err_ctx;
2261         } else {
2262                 ctx = NULL;
2263         }
2264
2265         /* Parameters for control virtqueue, if any */
2266         if (vi->has_cvq) {
2267                 callbacks[total_vqs - 1] = NULL;
2268                 names[total_vqs - 1] = "control";
2269         }
2270
2271         /* Allocate/initialize parameters for send/receive virtqueues */
2272         for (i = 0; i < vi->max_queue_pairs; i++) {
2273                 callbacks[rxq2vq(i)] = skb_recv_done;
2274                 callbacks[txq2vq(i)] = skb_xmit_done;
2275                 sprintf(vi->rq[i].name, "input.%d", i);
2276                 sprintf(vi->sq[i].name, "output.%d", i);
2277                 names[rxq2vq(i)] = vi->rq[i].name;
2278                 names[txq2vq(i)] = vi->sq[i].name;
2279                 if (ctx)
2280                         ctx[rxq2vq(i)] = true;
2281         }
2282
2283         ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
2284                                          names, ctx, NULL);
2285         if (ret)
2286                 goto err_find;
2287
2288         if (vi->has_cvq) {
2289                 vi->cvq = vqs[total_vqs - 1];
2290                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
2291                         vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
2292         }
2293
2294         for (i = 0; i < vi->max_queue_pairs; i++) {
2295                 vi->rq[i].vq = vqs[rxq2vq(i)];
2296                 vi->rq[i].min_buf_len = mergeable_min_buf_len(vi, vi->rq[i].vq);
2297                 vi->sq[i].vq = vqs[txq2vq(i)];
2298         }
2299
2300         kfree(names);
2301         kfree(callbacks);
2302         kfree(vqs);
2303         kfree(ctx);
2304
2305         return 0;
2306
2307 err_find:
2308         kfree(ctx);
2309 err_ctx:
2310         kfree(names);
2311 err_names:
2312         kfree(callbacks);
2313 err_callback:
2314         kfree(vqs);
2315 err_vq:
2316         return ret;
2317 }
2318
2319 static int virtnet_alloc_queues(struct virtnet_info *vi)
2320 {
2321         int i;
2322
2323         vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
2324         if (!vi->sq)
2325                 goto err_sq;
2326         vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
2327         if (!vi->rq)
2328                 goto err_rq;
2329
2330         INIT_DELAYED_WORK(&vi->refill, refill_work);
2331         for (i = 0; i < vi->max_queue_pairs; i++) {
2332                 vi->rq[i].pages = NULL;
2333                 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
2334                                napi_weight);
2335                 netif_tx_napi_add(vi->dev, &vi->sq[i].napi, virtnet_poll_tx,
2336                                   napi_tx ? napi_weight : 0);
2337
2338                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
2339                 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
2340                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
2341         }
2342
2343         return 0;
2344
2345 err_rq:
2346         kfree(vi->sq);
2347 err_sq:
2348         return -ENOMEM;
2349 }
2350
2351 static int init_vqs(struct virtnet_info *vi)
2352 {
2353         int ret;
2354
2355         /* Allocate send & receive queues */
2356         ret = virtnet_alloc_queues(vi);
2357         if (ret)
2358                 goto err;
2359
2360         ret = virtnet_find_vqs(vi);
2361         if (ret)
2362                 goto err_free;
2363
2364         get_online_cpus();
2365         virtnet_set_affinity(vi);
2366         put_online_cpus();
2367
2368         return 0;
2369
2370 err_free:
2371         virtnet_free_queues(vi);
2372 err:
2373         return ret;
2374 }
2375
2376 #ifdef CONFIG_SYSFS
2377 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
2378                 struct rx_queue_attribute *attribute, char *buf)
2379 {
2380         struct virtnet_info *vi = netdev_priv(queue->dev);
2381         unsigned int queue_index = get_netdev_rx_queue_index(queue);
2382         struct ewma_pkt_len *avg;
2383
2384         BUG_ON(queue_index >= vi->max_queue_pairs);
2385         avg = &vi->rq[queue_index].mrg_avg_pkt_len;
2386         return sprintf(buf, "%u\n",
2387                        get_mergeable_buf_len(&vi->rq[queue_index], avg));
2388 }
2389
2390 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
2391         __ATTR_RO(mergeable_rx_buffer_size);
2392
2393 static struct attribute *virtio_net_mrg_rx_attrs[] = {
2394         &mergeable_rx_buffer_size_attribute.attr,
2395         NULL
2396 };
2397
2398 static const struct attribute_group virtio_net_mrg_rx_group = {
2399         .name = "virtio_net",
2400         .attrs = virtio_net_mrg_rx_attrs
2401 };
2402 #endif
2403
2404 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
2405                                     unsigned int fbit,
2406                                     const char *fname, const char *dname)
2407 {
2408         if (!virtio_has_feature(vdev, fbit))
2409                 return false;
2410
2411         dev_err(&vdev->dev, "device advertises feature %s but not %s",
2412                 fname, dname);
2413
2414         return true;
2415 }
2416
2417 #define VIRTNET_FAIL_ON(vdev, fbit, dbit)                       \
2418         virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
2419
2420 static bool virtnet_validate_features(struct virtio_device *vdev)
2421 {
2422         if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
2423             (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
2424                              "VIRTIO_NET_F_CTRL_VQ") ||
2425              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
2426                              "VIRTIO_NET_F_CTRL_VQ") ||
2427              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
2428                              "VIRTIO_NET_F_CTRL_VQ") ||
2429              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
2430              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
2431                              "VIRTIO_NET_F_CTRL_VQ"))) {
2432                 return false;
2433         }
2434
2435         return true;
2436 }
2437
2438 #define MIN_MTU ETH_MIN_MTU
2439 #define MAX_MTU ETH_MAX_MTU
2440
2441 static int virtnet_validate(struct virtio_device *vdev)
2442 {
2443         if (!vdev->config->get) {
2444                 dev_err(&vdev->dev, "%s failure: config access disabled\n",
2445                         __func__);
2446                 return -EINVAL;
2447         }
2448
2449         if (!virtnet_validate_features(vdev))
2450                 return -EINVAL;
2451
2452         if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2453                 int mtu = virtio_cread16(vdev,
2454                                          offsetof(struct virtio_net_config,
2455                                                   mtu));
2456                 if (mtu < MIN_MTU)
2457                         __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
2458         }
2459
2460         return 0;
2461 }
2462
2463 static int virtnet_probe(struct virtio_device *vdev)
2464 {
2465         int i, err;
2466         struct net_device *dev;
2467         struct virtnet_info *vi;
2468         u16 max_queue_pairs;
2469         int mtu;
2470
2471         /* Find if host supports multiqueue virtio_net device */
2472         err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
2473                                    struct virtio_net_config,
2474                                    max_virtqueue_pairs, &max_queue_pairs);
2475
2476         /* We need at least 2 queue's */
2477         if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
2478             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
2479             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2480                 max_queue_pairs = 1;
2481
2482         /* Allocate ourselves a network device with room for our info */
2483         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
2484         if (!dev)
2485                 return -ENOMEM;
2486
2487         /* Set up network device as normal. */
2488         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
2489         dev->netdev_ops = &virtnet_netdev;
2490         dev->features = NETIF_F_HIGHDMA;
2491
2492         dev->ethtool_ops = &virtnet_ethtool_ops;
2493         SET_NETDEV_DEV(dev, &vdev->dev);
2494
2495         /* Do we support "hardware" checksums? */
2496         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
2497                 /* This opens up the world of extra features. */
2498                 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2499                 if (csum)
2500                         dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2501
2502                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
2503                         dev->hw_features |= NETIF_F_TSO
2504                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
2505                 }
2506                 /* Individual feature bits: what can host handle? */
2507                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
2508                         dev->hw_features |= NETIF_F_TSO;
2509                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
2510                         dev->hw_features |= NETIF_F_TSO6;
2511                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
2512                         dev->hw_features |= NETIF_F_TSO_ECN;
2513
2514                 dev->features |= NETIF_F_GSO_ROBUST;
2515
2516                 if (gso)
2517                         dev->features |= dev->hw_features & NETIF_F_ALL_TSO;
2518                 /* (!csum && gso) case will be fixed by register_netdev() */
2519         }
2520         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
2521                 dev->features |= NETIF_F_RXCSUM;
2522
2523         dev->vlan_features = dev->features;
2524
2525         /* MTU range: 68 - 65535 */
2526         dev->min_mtu = MIN_MTU;
2527         dev->max_mtu = MAX_MTU;
2528
2529         /* Configuration may specify what MAC to use.  Otherwise random. */
2530         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
2531                 virtio_cread_bytes(vdev,
2532                                    offsetof(struct virtio_net_config, mac),
2533                                    dev->dev_addr, dev->addr_len);
2534         else
2535                 eth_hw_addr_random(dev);
2536
2537         /* Set up our device-specific information */
2538         vi = netdev_priv(dev);
2539         vi->dev = dev;
2540         vi->vdev = vdev;
2541         vdev->priv = vi;
2542         vi->stats = alloc_percpu(struct virtnet_stats);
2543         err = -ENOMEM;
2544         if (vi->stats == NULL)
2545                 goto free;
2546
2547         for_each_possible_cpu(i) {
2548                 struct virtnet_stats *virtnet_stats;
2549                 virtnet_stats = per_cpu_ptr(vi->stats, i);
2550                 u64_stats_init(&virtnet_stats->tx_syncp);
2551                 u64_stats_init(&virtnet_stats->rx_syncp);
2552         }
2553
2554         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
2555
2556         /* If we can receive ANY GSO packets, we must allocate large ones. */
2557         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2558             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2559             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
2560             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
2561                 vi->big_packets = true;
2562
2563         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
2564                 vi->mergeable_rx_bufs = true;
2565
2566         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
2567             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2568                 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2569         else
2570                 vi->hdr_len = sizeof(struct virtio_net_hdr);
2571
2572         if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
2573             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2574                 vi->any_header_sg = true;
2575
2576         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2577                 vi->has_cvq = true;
2578
2579         if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2580                 mtu = virtio_cread16(vdev,
2581                                      offsetof(struct virtio_net_config,
2582                                               mtu));
2583                 if (mtu < dev->min_mtu) {
2584                         /* Should never trigger: MTU was previously validated
2585                          * in virtnet_validate.
2586                          */
2587                         dev_err(&vdev->dev, "device MTU appears to have changed "
2588                                 "it is now %d < %d", mtu, dev->min_mtu);
2589                         goto free_stats;
2590                 }
2591
2592                 dev->mtu = mtu;
2593                 dev->max_mtu = mtu;
2594
2595                 /* TODO: size buffers correctly in this case. */
2596                 if (dev->mtu > ETH_DATA_LEN)
2597                         vi->big_packets = true;
2598         }
2599
2600         if (vi->any_header_sg)
2601                 dev->needed_headroom = vi->hdr_len;
2602
2603         /* Enable multiqueue by default */
2604         if (num_online_cpus() >= max_queue_pairs)
2605                 vi->curr_queue_pairs = max_queue_pairs;
2606         else
2607                 vi->curr_queue_pairs = num_online_cpus();
2608         vi->max_queue_pairs = max_queue_pairs;
2609
2610         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
2611         err = init_vqs(vi);
2612         if (err)
2613                 goto free_stats;
2614
2615 #ifdef CONFIG_SYSFS
2616         if (vi->mergeable_rx_bufs)
2617                 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
2618 #endif
2619         netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
2620         netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
2621
2622         virtnet_init_settings(dev);
2623
2624         err = register_netdev(dev);
2625         if (err) {
2626                 pr_debug("virtio_net: registering device failed\n");
2627                 goto free_vqs;
2628         }
2629
2630         virtio_device_ready(vdev);
2631
2632         err = virtnet_cpu_notif_add(vi);
2633         if (err) {
2634                 pr_debug("virtio_net: registering cpu notifier failed\n");
2635                 goto free_unregister_netdev;
2636         }
2637
2638         virtnet_set_queues(vi, vi->curr_queue_pairs);
2639
2640         /* Assume link up if device can't report link status,
2641            otherwise get link status from config. */
2642         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
2643                 netif_carrier_off(dev);
2644                 schedule_work(&vi->config_work);
2645         } else {
2646                 vi->status = VIRTIO_NET_S_LINK_UP;
2647                 netif_carrier_on(dev);
2648         }
2649
2650         for (i = 0; i < ARRAY_SIZE(guest_offloads); i++)
2651                 if (virtio_has_feature(vi->vdev, guest_offloads[i]))
2652                         set_bit(guest_offloads[i], &vi->guest_offloads);
2653
2654         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
2655                  dev->name, max_queue_pairs);
2656
2657         return 0;
2658
2659 free_unregister_netdev:
2660         vi->vdev->config->reset(vdev);
2661
2662         unregister_netdev(dev);
2663 free_vqs:
2664         cancel_delayed_work_sync(&vi->refill);
2665         free_receive_page_frags(vi);
2666         virtnet_del_vqs(vi);
2667 free_stats:
2668         free_percpu(vi->stats);
2669 free:
2670         free_netdev(dev);
2671         return err;
2672 }
2673
2674 static void remove_vq_common(struct virtnet_info *vi)
2675 {
2676         vi->vdev->config->reset(vi->vdev);
2677
2678         /* Free unused buffers in both send and recv, if any. */
2679         free_unused_bufs(vi);
2680
2681         free_receive_bufs(vi);
2682
2683         free_receive_page_frags(vi);
2684
2685         virtnet_del_vqs(vi);
2686 }
2687
2688 static void virtnet_remove(struct virtio_device *vdev)
2689 {
2690         struct virtnet_info *vi = vdev->priv;
2691
2692         virtnet_cpu_notif_remove(vi);
2693
2694         /* Make sure no work handler is accessing the device. */
2695         flush_work(&vi->config_work);
2696
2697         unregister_netdev(vi->dev);
2698
2699         remove_vq_common(vi);
2700
2701         free_percpu(vi->stats);
2702         free_netdev(vi->dev);
2703 }
2704
2705 #ifdef CONFIG_PM_SLEEP
2706 static int virtnet_freeze(struct virtio_device *vdev)
2707 {
2708         struct virtnet_info *vi = vdev->priv;
2709
2710         virtnet_cpu_notif_remove(vi);
2711         virtnet_freeze_down(vdev);
2712         remove_vq_common(vi);
2713
2714         return 0;
2715 }
2716
2717 static int virtnet_restore(struct virtio_device *vdev)
2718 {
2719         struct virtnet_info *vi = vdev->priv;
2720         int err;
2721
2722         err = virtnet_restore_up(vdev);
2723         if (err)
2724                 return err;
2725         virtnet_set_queues(vi, vi->curr_queue_pairs);
2726
2727         err = virtnet_cpu_notif_add(vi);
2728         if (err)
2729                 return err;
2730
2731         return 0;
2732 }
2733 #endif
2734
2735 static struct virtio_device_id id_table[] = {
2736         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
2737         { 0 },
2738 };
2739
2740 #define VIRTNET_FEATURES \
2741         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
2742         VIRTIO_NET_F_MAC, \
2743         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
2744         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
2745         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
2746         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
2747         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
2748         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
2749         VIRTIO_NET_F_CTRL_MAC_ADDR, \
2750         VIRTIO_NET_F_MTU, VIRTIO_NET_F_CTRL_GUEST_OFFLOADS
2751
2752 static unsigned int features[] = {
2753         VIRTNET_FEATURES,
2754 };
2755
2756 static unsigned int features_legacy[] = {
2757         VIRTNET_FEATURES,
2758         VIRTIO_NET_F_GSO,
2759         VIRTIO_F_ANY_LAYOUT,
2760 };
2761
2762 static struct virtio_driver virtio_net_driver = {
2763         .feature_table = features,
2764         .feature_table_size = ARRAY_SIZE(features),
2765         .feature_table_legacy = features_legacy,
2766         .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
2767         .driver.name =  KBUILD_MODNAME,
2768         .driver.owner = THIS_MODULE,
2769         .id_table =     id_table,
2770         .validate =     virtnet_validate,
2771         .probe =        virtnet_probe,
2772         .remove =       virtnet_remove,
2773         .config_changed = virtnet_config_changed,
2774 #ifdef CONFIG_PM_SLEEP
2775         .freeze =       virtnet_freeze,
2776         .restore =      virtnet_restore,
2777 #endif
2778 };
2779
2780 static __init int virtio_net_driver_init(void)
2781 {
2782         int ret;
2783
2784         ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
2785                                       virtnet_cpu_online,
2786                                       virtnet_cpu_down_prep);
2787         if (ret < 0)
2788                 goto out;
2789         virtionet_online = ret;
2790         ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
2791                                       NULL, virtnet_cpu_dead);
2792         if (ret)
2793                 goto err_dead;
2794
2795         ret = register_virtio_driver(&virtio_net_driver);
2796         if (ret)
2797                 goto err_virtio;
2798         return 0;
2799 err_virtio:
2800         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
2801 err_dead:
2802         cpuhp_remove_multi_state(virtionet_online);
2803 out:
2804         return ret;
2805 }
2806 module_init(virtio_net_driver_init);
2807
2808 static __exit void virtio_net_driver_exit(void)
2809 {
2810         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
2811         cpuhp_remove_multi_state(virtionet_online);
2812         unregister_virtio_driver(&virtio_net_driver);
2813 }
2814 module_exit(virtio_net_driver_exit);
2815
2816 MODULE_DEVICE_TABLE(virtio, id_table);
2817 MODULE_DESCRIPTION("Virtio network driver");
2818 MODULE_LICENSE("GPL");