Merge tag 'for-linus' of git://git.kernel.org/pub/scm/virt/kvm/kvm
[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
33 static int napi_weight = NAPI_POLL_WEIGHT;
34 module_param(napi_weight, int, 0444);
35
36 static bool csum = true, gso = true, napi_tx;
37 module_param(csum, bool, 0444);
38 module_param(gso, bool, 0444);
39 module_param(napi_tx, bool, 0644);
40
41 /* FIXME: MTU in config. */
42 #define GOOD_PACKET_LEN (ETH_HLEN + VLAN_HLEN + ETH_DATA_LEN)
43 #define GOOD_COPY_LEN   128
44
45 #define VIRTNET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
46
47 /* Amount of XDP headroom to prepend to packets for use by xdp_adjust_head */
48 #define VIRTIO_XDP_HEADROOM 256
49
50 /* RX packet size EWMA. The average packet size is used to determine the packet
51  * buffer size when refilling RX rings. As the entire RX ring may be refilled
52  * at once, the weight is chosen so that the EWMA will be insensitive to short-
53  * term, transient changes in packet size.
54  */
55 DECLARE_EWMA(pkt_len, 0, 64)
56
57 /* With mergeable buffers we align buffer address and use the low bits to
58  * encode its true size. Buffer size is up to 1 page so we need to align to
59  * square root of page size to ensure we reserve enough bits to encode the true
60  * size.
61  */
62 #define MERGEABLE_BUFFER_MIN_ALIGN_SHIFT ((PAGE_SHIFT + 1) / 2)
63
64 /* Minimum alignment for mergeable packet buffers. */
65 #define MERGEABLE_BUFFER_ALIGN max(L1_CACHE_BYTES, \
66                                    1 << MERGEABLE_BUFFER_MIN_ALIGN_SHIFT)
67
68 #define VIRTNET_DRIVER_VERSION "1.0.0"
69
70 struct virtnet_stats {
71         struct u64_stats_sync tx_syncp;
72         struct u64_stats_sync rx_syncp;
73         u64 tx_bytes;
74         u64 tx_packets;
75
76         u64 rx_bytes;
77         u64 rx_packets;
78 };
79
80 /* Internal representation of a send virtqueue */
81 struct send_queue {
82         /* Virtqueue associated with this send _queue */
83         struct virtqueue *vq;
84
85         /* TX: fragments + linear part + virtio header */
86         struct scatterlist sg[MAX_SKB_FRAGS + 2];
87
88         /* Name of the send queue: output.$index */
89         char name[40];
90
91         struct napi_struct napi;
92 };
93
94 /* Internal representation of a receive virtqueue */
95 struct receive_queue {
96         /* Virtqueue associated with this receive_queue */
97         struct virtqueue *vq;
98
99         struct napi_struct napi;
100
101         struct bpf_prog __rcu *xdp_prog;
102
103         /* Chain pages by the private ptr. */
104         struct page *pages;
105
106         /* Average packet length for mergeable receive buffers. */
107         struct ewma_pkt_len mrg_avg_pkt_len;
108
109         /* Page frag for packet buffer allocation. */
110         struct page_frag alloc_frag;
111
112         /* RX: fragments + linear part + virtio header */
113         struct scatterlist sg[MAX_SKB_FRAGS + 2];
114
115         /* Name of this receive queue: input.$index */
116         char name[40];
117 };
118
119 struct virtnet_info {
120         struct virtio_device *vdev;
121         struct virtqueue *cvq;
122         struct net_device *dev;
123         struct send_queue *sq;
124         struct receive_queue *rq;
125         unsigned int status;
126
127         /* Max # of queue pairs supported by the device */
128         u16 max_queue_pairs;
129
130         /* # of queue pairs currently used by the driver */
131         u16 curr_queue_pairs;
132
133         /* # of XDP queue pairs currently used by the driver */
134         u16 xdp_queue_pairs;
135
136         /* I like... big packets and I cannot lie! */
137         bool big_packets;
138
139         /* Host will merge rx buffers for big packets (shake it! shake it!) */
140         bool mergeable_rx_bufs;
141
142         /* Has control virtqueue */
143         bool has_cvq;
144
145         /* Host can handle any s/g split between our header and packet data */
146         bool any_header_sg;
147
148         /* Packet virtio header size */
149         u8 hdr_len;
150
151         /* Active statistics */
152         struct virtnet_stats __percpu *stats;
153
154         /* Work struct for refilling if we run low on memory. */
155         struct delayed_work refill;
156
157         /* Work struct for config space updates */
158         struct work_struct config_work;
159
160         /* Does the affinity hint is set for virtqueues? */
161         bool affinity_hint_set;
162
163         /* CPU hotplug instances for online & dead */
164         struct hlist_node node;
165         struct hlist_node node_dead;
166
167         /* Control VQ buffers: protected by the rtnl lock */
168         struct virtio_net_ctrl_hdr ctrl_hdr;
169         virtio_net_ctrl_ack ctrl_status;
170         struct virtio_net_ctrl_mq ctrl_mq;
171         u8 ctrl_promisc;
172         u8 ctrl_allmulti;
173         u16 ctrl_vid;
174
175         /* Ethtool settings */
176         u8 duplex;
177         u32 speed;
178 };
179
180 struct padded_vnet_hdr {
181         struct virtio_net_hdr_mrg_rxbuf hdr;
182         /*
183          * hdr is in a separate sg buffer, and data sg buffer shares same page
184          * with this header sg. This padding makes next sg 16 byte aligned
185          * after the header.
186          */
187         char padding[4];
188 };
189
190 /* Converting between virtqueue no. and kernel tx/rx queue no.
191  * 0:rx0 1:tx0 2:rx1 3:tx1 ... 2N:rxN 2N+1:txN 2N+2:cvq
192  */
193 static int vq2txq(struct virtqueue *vq)
194 {
195         return (vq->index - 1) / 2;
196 }
197
198 static int txq2vq(int txq)
199 {
200         return txq * 2 + 1;
201 }
202
203 static int vq2rxq(struct virtqueue *vq)
204 {
205         return vq->index / 2;
206 }
207
208 static int rxq2vq(int rxq)
209 {
210         return rxq * 2;
211 }
212
213 static inline struct virtio_net_hdr_mrg_rxbuf *skb_vnet_hdr(struct sk_buff *skb)
214 {
215         return (struct virtio_net_hdr_mrg_rxbuf *)skb->cb;
216 }
217
218 /*
219  * private is used to chain pages for big packets, put the whole
220  * most recent used list in the beginning for reuse
221  */
222 static void give_pages(struct receive_queue *rq, struct page *page)
223 {
224         struct page *end;
225
226         /* Find end of list, sew whole thing into vi->rq.pages. */
227         for (end = page; end->private; end = (struct page *)end->private);
228         end->private = (unsigned long)rq->pages;
229         rq->pages = page;
230 }
231
232 static struct page *get_a_page(struct receive_queue *rq, gfp_t gfp_mask)
233 {
234         struct page *p = rq->pages;
235
236         if (p) {
237                 rq->pages = (struct page *)p->private;
238                 /* clear private here, it is used to chain pages */
239                 p->private = 0;
240         } else
241                 p = alloc_page(gfp_mask);
242         return p;
243 }
244
245 static void virtqueue_napi_schedule(struct napi_struct *napi,
246                                     struct virtqueue *vq)
247 {
248         if (napi_schedule_prep(napi)) {
249                 virtqueue_disable_cb(vq);
250                 __napi_schedule(napi);
251         }
252 }
253
254 static void virtqueue_napi_complete(struct napi_struct *napi,
255                                     struct virtqueue *vq, int processed)
256 {
257         int opaque;
258
259         opaque = virtqueue_enable_cb_prepare(vq);
260         if (napi_complete_done(napi, processed) &&
261             unlikely(virtqueue_poll(vq, opaque)))
262                 virtqueue_napi_schedule(napi, vq);
263 }
264
265 static void skb_xmit_done(struct virtqueue *vq)
266 {
267         struct virtnet_info *vi = vq->vdev->priv;
268         struct napi_struct *napi = &vi->sq[vq2txq(vq)].napi;
269
270         /* Suppress further interrupts. */
271         virtqueue_disable_cb(vq);
272
273         if (napi->weight)
274                 virtqueue_napi_schedule(napi, vq);
275         else
276                 /* We were probably waiting for more output buffers. */
277                 netif_wake_subqueue(vi->dev, vq2txq(vq));
278 }
279
280 static unsigned int mergeable_ctx_to_buf_truesize(unsigned long mrg_ctx)
281 {
282         unsigned int truesize = mrg_ctx & (MERGEABLE_BUFFER_ALIGN - 1);
283         return (truesize + 1) * MERGEABLE_BUFFER_ALIGN;
284 }
285
286 static void *mergeable_ctx_to_buf_address(unsigned long mrg_ctx)
287 {
288         return (void *)(mrg_ctx & -MERGEABLE_BUFFER_ALIGN);
289
290 }
291
292 static unsigned long mergeable_buf_to_ctx(void *buf, unsigned int truesize)
293 {
294         unsigned int size = truesize / MERGEABLE_BUFFER_ALIGN;
295         return (unsigned long)buf | (size - 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         memcpy(skb_put(skb, copy), 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 static struct sk_buff *receive_small(struct net_device *dev,
419                                      struct virtnet_info *vi,
420                                      struct receive_queue *rq,
421                                      void *buf, unsigned int len)
422 {
423         struct sk_buff *skb;
424         struct bpf_prog *xdp_prog;
425         unsigned int xdp_headroom = virtnet_get_headroom(vi);
426         unsigned int header_offset = VIRTNET_RX_PAD + xdp_headroom;
427         unsigned int headroom = vi->hdr_len + header_offset;
428         unsigned int buflen = SKB_DATA_ALIGN(GOOD_PACKET_LEN + headroom) +
429                               SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
430         unsigned int delta = 0;
431         len -= vi->hdr_len;
432
433         rcu_read_lock();
434         xdp_prog = rcu_dereference(rq->xdp_prog);
435         if (xdp_prog) {
436                 struct virtio_net_hdr_mrg_rxbuf *hdr = buf + header_offset;
437                 struct xdp_buff xdp;
438                 void *orig_data;
439                 u32 act;
440
441                 if (unlikely(hdr->hdr.gso_type || hdr->hdr.flags))
442                         goto err_xdp;
443
444                 xdp.data_hard_start = buf + VIRTNET_RX_PAD + vi->hdr_len;
445                 xdp.data = xdp.data_hard_start + xdp_headroom;
446                 xdp.data_end = xdp.data + len;
447                 orig_data = xdp.data;
448                 act = bpf_prog_run_xdp(xdp_prog, &xdp);
449
450                 switch (act) {
451                 case XDP_PASS:
452                         /* Recalculate length in case bpf program changed it */
453                         delta = orig_data - xdp.data;
454                         break;
455                 case XDP_TX:
456                         if (unlikely(!virtnet_xdp_xmit(vi, rq, &xdp)))
457                                 trace_xdp_exception(vi->dev, xdp_prog, act);
458                         rcu_read_unlock();
459                         goto xdp_xmit;
460                 default:
461                         bpf_warn_invalid_xdp_action(act);
462                 case XDP_ABORTED:
463                         trace_xdp_exception(vi->dev, xdp_prog, act);
464                 case XDP_DROP:
465                         goto err_xdp;
466                 }
467         }
468         rcu_read_unlock();
469
470         skb = build_skb(buf, buflen);
471         if (!skb) {
472                 put_page(virt_to_head_page(buf));
473                 goto err;
474         }
475         skb_reserve(skb, headroom - delta);
476         skb_put(skb, len + delta);
477         if (!delta) {
478                 buf += header_offset;
479                 memcpy(skb_vnet_hdr(skb), buf, vi->hdr_len);
480         } /* keep zeroed vnet hdr since packet was changed by bpf */
481
482 err:
483         return skb;
484
485 err_xdp:
486         rcu_read_unlock();
487         dev->stats.rx_dropped++;
488         put_page(virt_to_head_page(buf));
489 xdp_xmit:
490         return NULL;
491 }
492
493 static struct sk_buff *receive_big(struct net_device *dev,
494                                    struct virtnet_info *vi,
495                                    struct receive_queue *rq,
496                                    void *buf,
497                                    unsigned int len)
498 {
499         struct page *page = buf;
500         struct sk_buff *skb = page_to_skb(vi, rq, page, 0, len, PAGE_SIZE);
501
502         if (unlikely(!skb))
503                 goto err;
504
505         return skb;
506
507 err:
508         dev->stats.rx_dropped++;
509         give_pages(rq, page);
510         return NULL;
511 }
512
513 /* The conditions to enable XDP should preclude the underlying device from
514  * sending packets across multiple buffers (num_buf > 1). However per spec
515  * it does not appear to be illegal to do so but rather just against convention.
516  * So in order to avoid making a system unresponsive the packets are pushed
517  * into a page and the XDP program is run. This will be extremely slow and we
518  * push a warning to the user to fix this as soon as possible. Fixing this may
519  * require resolving the underlying hardware to determine why multiple buffers
520  * are being received or simply loading the XDP program in the ingress stack
521  * after the skb is built because there is no advantage to running it here
522  * anymore.
523  */
524 static struct page *xdp_linearize_page(struct receive_queue *rq,
525                                        u16 *num_buf,
526                                        struct page *p,
527                                        int offset,
528                                        unsigned int *len)
529 {
530         struct page *page = alloc_page(GFP_ATOMIC);
531         unsigned int page_off = VIRTIO_XDP_HEADROOM;
532
533         if (!page)
534                 return NULL;
535
536         memcpy(page_address(page) + page_off, page_address(p) + offset, *len);
537         page_off += *len;
538
539         while (--*num_buf) {
540                 unsigned int buflen;
541                 unsigned long ctx;
542                 void *buf;
543                 int off;
544
545                 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &buflen);
546                 if (unlikely(!ctx))
547                         goto err_buf;
548
549                 buf = mergeable_ctx_to_buf_address(ctx);
550                 p = virt_to_head_page(buf);
551                 off = buf - page_address(p);
552
553                 /* guard against a misconfigured or uncooperative backend that
554                  * is sending packet larger than the MTU.
555                  */
556                 if ((page_off + buflen) > PAGE_SIZE) {
557                         put_page(p);
558                         goto err_buf;
559                 }
560
561                 memcpy(page_address(page) + page_off,
562                        page_address(p) + off, buflen);
563                 page_off += buflen;
564                 put_page(p);
565         }
566
567         /* Headroom does not contribute to packet length */
568         *len = page_off - VIRTIO_XDP_HEADROOM;
569         return page;
570 err_buf:
571         __free_pages(page, 0);
572         return NULL;
573 }
574
575 static struct sk_buff *receive_mergeable(struct net_device *dev,
576                                          struct virtnet_info *vi,
577                                          struct receive_queue *rq,
578                                          unsigned long ctx,
579                                          unsigned int len)
580 {
581         void *buf = mergeable_ctx_to_buf_address(ctx);
582         struct virtio_net_hdr_mrg_rxbuf *hdr = buf;
583         u16 num_buf = virtio16_to_cpu(vi->vdev, hdr->num_buffers);
584         struct page *page = virt_to_head_page(buf);
585         int offset = buf - page_address(page);
586         struct sk_buff *head_skb, *curr_skb;
587         struct bpf_prog *xdp_prog;
588         unsigned int truesize;
589
590         head_skb = NULL;
591
592         rcu_read_lock();
593         xdp_prog = rcu_dereference(rq->xdp_prog);
594         if (xdp_prog) {
595                 struct page *xdp_page;
596                 struct xdp_buff xdp;
597                 void *data;
598                 u32 act;
599
600                 /* This happens when rx buffer size is underestimated */
601                 if (unlikely(num_buf > 1)) {
602                         /* linearize data for XDP */
603                         xdp_page = xdp_linearize_page(rq, &num_buf,
604                                                       page, offset, &len);
605                         if (!xdp_page)
606                                 goto err_xdp;
607                         offset = VIRTIO_XDP_HEADROOM;
608                 } else {
609                         xdp_page = page;
610                 }
611
612                 /* Transient failure which in theory could occur if
613                  * in-flight packets from before XDP was enabled reach
614                  * the receive path after XDP is loaded. In practice I
615                  * was not able to create this condition.
616                  */
617                 if (unlikely(hdr->hdr.gso_type))
618                         goto err_xdp;
619
620                 /* Allow consuming headroom but reserve enough space to push
621                  * the descriptor on if we get an XDP_TX return code.
622                  */
623                 data = page_address(xdp_page) + offset;
624                 xdp.data_hard_start = data - VIRTIO_XDP_HEADROOM + vi->hdr_len;
625                 xdp.data = data + vi->hdr_len;
626                 xdp.data_end = xdp.data + (len - vi->hdr_len);
627                 act = bpf_prog_run_xdp(xdp_prog, &xdp);
628
629                 switch (act) {
630                 case XDP_PASS:
631                         /* recalculate offset to account for any header
632                          * adjustments. Note other cases do not build an
633                          * skb and avoid using offset
634                          */
635                         offset = xdp.data -
636                                         page_address(xdp_page) - vi->hdr_len;
637
638                         /* We can only create skb based on xdp_page. */
639                         if (unlikely(xdp_page != page)) {
640                                 rcu_read_unlock();
641                                 put_page(page);
642                                 head_skb = page_to_skb(vi, rq, xdp_page,
643                                                        offset, len, PAGE_SIZE);
644                                 ewma_pkt_len_add(&rq->mrg_avg_pkt_len, len);
645                                 return head_skb;
646                         }
647                         break;
648                 case XDP_TX:
649                         if (unlikely(!virtnet_xdp_xmit(vi, rq, &xdp)))
650                                 trace_xdp_exception(vi->dev, xdp_prog, act);
651                         ewma_pkt_len_add(&rq->mrg_avg_pkt_len, len);
652                         if (unlikely(xdp_page != page))
653                                 goto err_xdp;
654                         rcu_read_unlock();
655                         goto xdp_xmit;
656                 default:
657                         bpf_warn_invalid_xdp_action(act);
658                 case XDP_ABORTED:
659                         trace_xdp_exception(vi->dev, xdp_prog, act);
660                 case XDP_DROP:
661                         if (unlikely(xdp_page != page))
662                                 __free_pages(xdp_page, 0);
663                         ewma_pkt_len_add(&rq->mrg_avg_pkt_len, len);
664                         goto err_xdp;
665                 }
666         }
667         rcu_read_unlock();
668
669         truesize = max(len, mergeable_ctx_to_buf_truesize(ctx));
670         head_skb = page_to_skb(vi, rq, page, offset, len, truesize);
671         curr_skb = head_skb;
672
673         if (unlikely(!curr_skb))
674                 goto err_skb;
675         while (--num_buf) {
676                 int num_skb_frags;
677
678                 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &len);
679                 if (unlikely(!ctx)) {
680                         pr_debug("%s: rx error: %d buffers out of %d missing\n",
681                                  dev->name, num_buf,
682                                  virtio16_to_cpu(vi->vdev,
683                                                  hdr->num_buffers));
684                         dev->stats.rx_length_errors++;
685                         goto err_buf;
686                 }
687
688                 buf = mergeable_ctx_to_buf_address(ctx);
689                 page = virt_to_head_page(buf);
690
691                 num_skb_frags = skb_shinfo(curr_skb)->nr_frags;
692                 if (unlikely(num_skb_frags == MAX_SKB_FRAGS)) {
693                         struct sk_buff *nskb = alloc_skb(0, GFP_ATOMIC);
694
695                         if (unlikely(!nskb))
696                                 goto err_skb;
697                         if (curr_skb == head_skb)
698                                 skb_shinfo(curr_skb)->frag_list = nskb;
699                         else
700                                 curr_skb->next = nskb;
701                         curr_skb = nskb;
702                         head_skb->truesize += nskb->truesize;
703                         num_skb_frags = 0;
704                 }
705                 truesize = max(len, mergeable_ctx_to_buf_truesize(ctx));
706                 if (curr_skb != head_skb) {
707                         head_skb->data_len += len;
708                         head_skb->len += len;
709                         head_skb->truesize += truesize;
710                 }
711                 offset = buf - page_address(page);
712                 if (skb_can_coalesce(curr_skb, num_skb_frags, page, offset)) {
713                         put_page(page);
714                         skb_coalesce_rx_frag(curr_skb, num_skb_frags - 1,
715                                              len, truesize);
716                 } else {
717                         skb_add_rx_frag(curr_skb, num_skb_frags, page,
718                                         offset, len, truesize);
719                 }
720         }
721
722         ewma_pkt_len_add(&rq->mrg_avg_pkt_len, head_skb->len);
723         return head_skb;
724
725 err_xdp:
726         rcu_read_unlock();
727 err_skb:
728         put_page(page);
729         while (--num_buf) {
730                 ctx = (unsigned long)virtqueue_get_buf(rq->vq, &len);
731                 if (unlikely(!ctx)) {
732                         pr_debug("%s: rx error: %d buffers missing\n",
733                                  dev->name, num_buf);
734                         dev->stats.rx_length_errors++;
735                         break;
736                 }
737                 page = virt_to_head_page(mergeable_ctx_to_buf_address(ctx));
738                 put_page(page);
739         }
740 err_buf:
741         dev->stats.rx_dropped++;
742         dev_kfree_skb(head_skb);
743 xdp_xmit:
744         return NULL;
745 }
746
747 static int receive_buf(struct virtnet_info *vi, struct receive_queue *rq,
748                        void *buf, unsigned int len)
749 {
750         struct net_device *dev = vi->dev;
751         struct sk_buff *skb;
752         struct virtio_net_hdr_mrg_rxbuf *hdr;
753         int ret;
754
755         if (unlikely(len < vi->hdr_len + ETH_HLEN)) {
756                 pr_debug("%s: short packet %i\n", dev->name, len);
757                 dev->stats.rx_length_errors++;
758                 if (vi->mergeable_rx_bufs) {
759                         unsigned long ctx = (unsigned long)buf;
760                         void *base = mergeable_ctx_to_buf_address(ctx);
761                         put_page(virt_to_head_page(base));
762                 } else if (vi->big_packets) {
763                         give_pages(rq, buf);
764                 } else {
765                         put_page(virt_to_head_page(buf));
766                 }
767                 return 0;
768         }
769
770         if (vi->mergeable_rx_bufs)
771                 skb = receive_mergeable(dev, vi, rq, (unsigned long)buf, len);
772         else if (vi->big_packets)
773                 skb = receive_big(dev, vi, rq, buf, len);
774         else
775                 skb = receive_small(dev, vi, rq, buf, len);
776
777         if (unlikely(!skb))
778                 return 0;
779
780         hdr = skb_vnet_hdr(skb);
781
782         ret = skb->len;
783
784         if (hdr->hdr.flags & VIRTIO_NET_HDR_F_DATA_VALID)
785                 skb->ip_summed = CHECKSUM_UNNECESSARY;
786
787         if (virtio_net_hdr_to_skb(skb, &hdr->hdr,
788                                   virtio_is_little_endian(vi->vdev))) {
789                 net_warn_ratelimited("%s: bad gso: type: %u, size: %u\n",
790                                      dev->name, hdr->hdr.gso_type,
791                                      hdr->hdr.gso_size);
792                 goto frame_err;
793         }
794
795         skb->protocol = eth_type_trans(skb, dev);
796         pr_debug("Receiving skb proto 0x%04x len %i type %i\n",
797                  ntohs(skb->protocol), skb->len, skb->pkt_type);
798
799         napi_gro_receive(&rq->napi, skb);
800         return ret;
801
802 frame_err:
803         dev->stats.rx_frame_errors++;
804         dev_kfree_skb(skb);
805         return 0;
806 }
807
808 static int add_recvbuf_small(struct virtnet_info *vi, struct receive_queue *rq,
809                              gfp_t gfp)
810 {
811         struct page_frag *alloc_frag = &rq->alloc_frag;
812         char *buf;
813         unsigned int xdp_headroom = virtnet_get_headroom(vi);
814         int len = vi->hdr_len + VIRTNET_RX_PAD + GOOD_PACKET_LEN + xdp_headroom;
815         int err;
816
817         len = SKB_DATA_ALIGN(len) +
818               SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
819         if (unlikely(!skb_page_frag_refill(len, alloc_frag, gfp)))
820                 return -ENOMEM;
821
822         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
823         get_page(alloc_frag->page);
824         alloc_frag->offset += len;
825         sg_init_one(rq->sg, buf + VIRTNET_RX_PAD + xdp_headroom,
826                     vi->hdr_len + GOOD_PACKET_LEN);
827         err = virtqueue_add_inbuf(rq->vq, rq->sg, 1, buf, gfp);
828         if (err < 0)
829                 put_page(virt_to_head_page(buf));
830
831         return err;
832 }
833
834 static int add_recvbuf_big(struct virtnet_info *vi, struct receive_queue *rq,
835                            gfp_t gfp)
836 {
837         struct page *first, *list = NULL;
838         char *p;
839         int i, err, offset;
840
841         sg_init_table(rq->sg, MAX_SKB_FRAGS + 2);
842
843         /* page in rq->sg[MAX_SKB_FRAGS + 1] is list tail */
844         for (i = MAX_SKB_FRAGS + 1; i > 1; --i) {
845                 first = get_a_page(rq, gfp);
846                 if (!first) {
847                         if (list)
848                                 give_pages(rq, list);
849                         return -ENOMEM;
850                 }
851                 sg_set_buf(&rq->sg[i], page_address(first), PAGE_SIZE);
852
853                 /* chain new page in list head to match sg */
854                 first->private = (unsigned long)list;
855                 list = first;
856         }
857
858         first = get_a_page(rq, gfp);
859         if (!first) {
860                 give_pages(rq, list);
861                 return -ENOMEM;
862         }
863         p = page_address(first);
864
865         /* rq->sg[0], rq->sg[1] share the same page */
866         /* a separated rq->sg[0] for header - required in case !any_header_sg */
867         sg_set_buf(&rq->sg[0], p, vi->hdr_len);
868
869         /* rq->sg[1] for data packet, from offset */
870         offset = sizeof(struct padded_vnet_hdr);
871         sg_set_buf(&rq->sg[1], p + offset, PAGE_SIZE - offset);
872
873         /* chain first in list head */
874         first->private = (unsigned long)list;
875         err = virtqueue_add_inbuf(rq->vq, rq->sg, MAX_SKB_FRAGS + 2,
876                                   first, gfp);
877         if (err < 0)
878                 give_pages(rq, first);
879
880         return err;
881 }
882
883 static unsigned int get_mergeable_buf_len(struct ewma_pkt_len *avg_pkt_len)
884 {
885         const size_t hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
886         unsigned int len;
887
888         len = hdr_len + clamp_t(unsigned int, ewma_pkt_len_read(avg_pkt_len),
889                         GOOD_PACKET_LEN, PAGE_SIZE - hdr_len);
890         return ALIGN(len, MERGEABLE_BUFFER_ALIGN);
891 }
892
893 static int add_recvbuf_mergeable(struct virtnet_info *vi,
894                                  struct receive_queue *rq, gfp_t gfp)
895 {
896         struct page_frag *alloc_frag = &rq->alloc_frag;
897         unsigned int headroom = virtnet_get_headroom(vi);
898         char *buf;
899         unsigned long ctx;
900         int err;
901         unsigned int len, hole;
902
903         len = get_mergeable_buf_len(&rq->mrg_avg_pkt_len);
904         if (unlikely(!skb_page_frag_refill(len + headroom, alloc_frag, gfp)))
905                 return -ENOMEM;
906
907         buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
908         buf += headroom; /* advance address leaving hole at front of pkt */
909         ctx = mergeable_buf_to_ctx(buf, len);
910         get_page(alloc_frag->page);
911         alloc_frag->offset += len + headroom;
912         hole = alloc_frag->size - alloc_frag->offset;
913         if (hole < len + headroom) {
914                 /* To avoid internal fragmentation, if there is very likely not
915                  * enough space for another buffer, add the remaining space to
916                  * the current buffer. This extra space is not included in
917                  * the truesize stored in ctx.
918                  */
919                 len += hole;
920                 alloc_frag->offset += hole;
921         }
922
923         sg_init_one(rq->sg, buf, len);
924         err = virtqueue_add_inbuf(rq->vq, rq->sg, 1, (void *)ctx, gfp);
925         if (err < 0)
926                 put_page(virt_to_head_page(buf));
927
928         return err;
929 }
930
931 /*
932  * Returns false if we couldn't fill entirely (OOM).
933  *
934  * Normally run in the receive path, but can also be run from ndo_open
935  * before we're receiving packets, or from refill_work which is
936  * careful to disable receiving (using napi_disable).
937  */
938 static bool try_fill_recv(struct virtnet_info *vi, struct receive_queue *rq,
939                           gfp_t gfp)
940 {
941         int err;
942         bool oom;
943
944         gfp |= __GFP_COLD;
945         do {
946                 if (vi->mergeable_rx_bufs)
947                         err = add_recvbuf_mergeable(vi, rq, gfp);
948                 else if (vi->big_packets)
949                         err = add_recvbuf_big(vi, rq, gfp);
950                 else
951                         err = add_recvbuf_small(vi, rq, gfp);
952
953                 oom = err == -ENOMEM;
954                 if (err)
955                         break;
956         } while (rq->vq->num_free);
957         virtqueue_kick(rq->vq);
958         return !oom;
959 }
960
961 static void skb_recv_done(struct virtqueue *rvq)
962 {
963         struct virtnet_info *vi = rvq->vdev->priv;
964         struct receive_queue *rq = &vi->rq[vq2rxq(rvq)];
965
966         virtqueue_napi_schedule(&rq->napi, rvq);
967 }
968
969 static void virtnet_napi_enable(struct virtqueue *vq, struct napi_struct *napi)
970 {
971         napi_enable(napi);
972
973         /* If all buffers were filled by other side before we napi_enabled, we
974          * won't get another interrupt, so process any outstanding packets now.
975          * Call local_bh_enable after to trigger softIRQ processing.
976          */
977         local_bh_disable();
978         virtqueue_napi_schedule(napi, vq);
979         local_bh_enable();
980 }
981
982 static void virtnet_napi_tx_enable(struct virtnet_info *vi,
983                                    struct virtqueue *vq,
984                                    struct napi_struct *napi)
985 {
986         if (!napi->weight)
987                 return;
988
989         /* Tx napi touches cachelines on the cpu handling tx interrupts. Only
990          * enable the feature if this is likely affine with the transmit path.
991          */
992         if (!vi->affinity_hint_set) {
993                 napi->weight = 0;
994                 return;
995         }
996
997         return virtnet_napi_enable(vq, napi);
998 }
999
1000 static void virtnet_napi_tx_disable(struct napi_struct *napi)
1001 {
1002         if (napi->weight)
1003                 napi_disable(napi);
1004 }
1005
1006 static void refill_work(struct work_struct *work)
1007 {
1008         struct virtnet_info *vi =
1009                 container_of(work, struct virtnet_info, refill.work);
1010         bool still_empty;
1011         int i;
1012
1013         for (i = 0; i < vi->curr_queue_pairs; i++) {
1014                 struct receive_queue *rq = &vi->rq[i];
1015
1016                 napi_disable(&rq->napi);
1017                 still_empty = !try_fill_recv(vi, rq, GFP_KERNEL);
1018                 virtnet_napi_enable(rq->vq, &rq->napi);
1019
1020                 /* In theory, this can happen: if we don't get any buffers in
1021                  * we will *never* try to fill again.
1022                  */
1023                 if (still_empty)
1024                         schedule_delayed_work(&vi->refill, HZ/2);
1025         }
1026 }
1027
1028 static int virtnet_receive(struct receive_queue *rq, int budget)
1029 {
1030         struct virtnet_info *vi = rq->vq->vdev->priv;
1031         unsigned int len, received = 0, bytes = 0;
1032         void *buf;
1033         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
1034
1035         while (received < budget &&
1036                (buf = virtqueue_get_buf(rq->vq, &len)) != NULL) {
1037                 bytes += receive_buf(vi, rq, buf, len);
1038                 received++;
1039         }
1040
1041         if (rq->vq->num_free > virtqueue_get_vring_size(rq->vq) / 2) {
1042                 if (!try_fill_recv(vi, rq, GFP_ATOMIC))
1043                         schedule_delayed_work(&vi->refill, 0);
1044         }
1045
1046         u64_stats_update_begin(&stats->rx_syncp);
1047         stats->rx_bytes += bytes;
1048         stats->rx_packets += received;
1049         u64_stats_update_end(&stats->rx_syncp);
1050
1051         return received;
1052 }
1053
1054 static void free_old_xmit_skbs(struct send_queue *sq)
1055 {
1056         struct sk_buff *skb;
1057         unsigned int len;
1058         struct virtnet_info *vi = sq->vq->vdev->priv;
1059         struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
1060         unsigned int packets = 0;
1061         unsigned int bytes = 0;
1062
1063         while ((skb = virtqueue_get_buf(sq->vq, &len)) != NULL) {
1064                 pr_debug("Sent skb %p\n", skb);
1065
1066                 bytes += skb->len;
1067                 packets++;
1068
1069                 dev_kfree_skb_any(skb);
1070         }
1071
1072         /* Avoid overhead when no packets have been processed
1073          * happens when called speculatively from start_xmit.
1074          */
1075         if (!packets)
1076                 return;
1077
1078         u64_stats_update_begin(&stats->tx_syncp);
1079         stats->tx_bytes += bytes;
1080         stats->tx_packets += packets;
1081         u64_stats_update_end(&stats->tx_syncp);
1082 }
1083
1084 static void virtnet_poll_cleantx(struct receive_queue *rq)
1085 {
1086         struct virtnet_info *vi = rq->vq->vdev->priv;
1087         unsigned int index = vq2rxq(rq->vq);
1088         struct send_queue *sq = &vi->sq[index];
1089         struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, index);
1090
1091         if (!sq->napi.weight)
1092                 return;
1093
1094         if (__netif_tx_trylock(txq)) {
1095                 free_old_xmit_skbs(sq);
1096                 __netif_tx_unlock(txq);
1097         }
1098
1099         if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1100                 netif_tx_wake_queue(txq);
1101 }
1102
1103 static int virtnet_poll(struct napi_struct *napi, int budget)
1104 {
1105         struct receive_queue *rq =
1106                 container_of(napi, struct receive_queue, napi);
1107         unsigned int received;
1108
1109         virtnet_poll_cleantx(rq);
1110
1111         received = virtnet_receive(rq, budget);
1112
1113         /* Out of packets? */
1114         if (received < budget)
1115                 virtqueue_napi_complete(napi, rq->vq, received);
1116
1117         return received;
1118 }
1119
1120 static int virtnet_open(struct net_device *dev)
1121 {
1122         struct virtnet_info *vi = netdev_priv(dev);
1123         int i;
1124
1125         for (i = 0; i < vi->max_queue_pairs; i++) {
1126                 if (i < vi->curr_queue_pairs)
1127                         /* Make sure we have some buffers: if oom use wq. */
1128                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1129                                 schedule_delayed_work(&vi->refill, 0);
1130                 virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1131                 virtnet_napi_tx_enable(vi, vi->sq[i].vq, &vi->sq[i].napi);
1132         }
1133
1134         return 0;
1135 }
1136
1137 static int virtnet_poll_tx(struct napi_struct *napi, int budget)
1138 {
1139         struct send_queue *sq = container_of(napi, struct send_queue, napi);
1140         struct virtnet_info *vi = sq->vq->vdev->priv;
1141         struct netdev_queue *txq = netdev_get_tx_queue(vi->dev, vq2txq(sq->vq));
1142
1143         __netif_tx_lock(txq, raw_smp_processor_id());
1144         free_old_xmit_skbs(sq);
1145         __netif_tx_unlock(txq);
1146
1147         virtqueue_napi_complete(napi, sq->vq, 0);
1148
1149         if (sq->vq->num_free >= 2 + MAX_SKB_FRAGS)
1150                 netif_tx_wake_queue(txq);
1151
1152         return 0;
1153 }
1154
1155 static int xmit_skb(struct send_queue *sq, struct sk_buff *skb)
1156 {
1157         struct virtio_net_hdr_mrg_rxbuf *hdr;
1158         const unsigned char *dest = ((struct ethhdr *)skb->data)->h_dest;
1159         struct virtnet_info *vi = sq->vq->vdev->priv;
1160         unsigned num_sg;
1161         unsigned hdr_len = vi->hdr_len;
1162         bool can_push;
1163
1164         pr_debug("%s: xmit %p %pM\n", vi->dev->name, skb, dest);
1165
1166         can_push = vi->any_header_sg &&
1167                 !((unsigned long)skb->data & (__alignof__(*hdr) - 1)) &&
1168                 !skb_header_cloned(skb) && skb_headroom(skb) >= hdr_len;
1169         /* Even if we can, don't push here yet as this would skew
1170          * csum_start offset below. */
1171         if (can_push)
1172                 hdr = (struct virtio_net_hdr_mrg_rxbuf *)(skb->data - hdr_len);
1173         else
1174                 hdr = skb_vnet_hdr(skb);
1175
1176         if (virtio_net_hdr_from_skb(skb, &hdr->hdr,
1177                                     virtio_is_little_endian(vi->vdev), false))
1178                 BUG();
1179
1180         if (vi->mergeable_rx_bufs)
1181                 hdr->num_buffers = 0;
1182
1183         sg_init_table(sq->sg, skb_shinfo(skb)->nr_frags + (can_push ? 1 : 2));
1184         if (can_push) {
1185                 __skb_push(skb, hdr_len);
1186                 num_sg = skb_to_sgvec(skb, sq->sg, 0, skb->len);
1187                 /* Pull header back to avoid skew in tx bytes calculations. */
1188                 __skb_pull(skb, hdr_len);
1189         } else {
1190                 sg_set_buf(sq->sg, hdr, hdr_len);
1191                 num_sg = skb_to_sgvec(skb, sq->sg + 1, 0, skb->len) + 1;
1192         }
1193         return virtqueue_add_outbuf(sq->vq, sq->sg, num_sg, skb, GFP_ATOMIC);
1194 }
1195
1196 static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
1197 {
1198         struct virtnet_info *vi = netdev_priv(dev);
1199         int qnum = skb_get_queue_mapping(skb);
1200         struct send_queue *sq = &vi->sq[qnum];
1201         int err;
1202         struct netdev_queue *txq = netdev_get_tx_queue(dev, qnum);
1203         bool kick = !skb->xmit_more;
1204         bool use_napi = sq->napi.weight;
1205
1206         /* Free up any pending old buffers before queueing new ones. */
1207         free_old_xmit_skbs(sq);
1208
1209         if (use_napi && kick)
1210                 virtqueue_enable_cb_delayed(sq->vq);
1211
1212         /* timestamp packet in software */
1213         skb_tx_timestamp(skb);
1214
1215         /* Try to transmit */
1216         err = xmit_skb(sq, skb);
1217
1218         /* This should not happen! */
1219         if (unlikely(err)) {
1220                 dev->stats.tx_fifo_errors++;
1221                 if (net_ratelimit())
1222                         dev_warn(&dev->dev,
1223                                  "Unexpected TXQ (%d) queue failure: %d\n", qnum, err);
1224                 dev->stats.tx_dropped++;
1225                 dev_kfree_skb_any(skb);
1226                 return NETDEV_TX_OK;
1227         }
1228
1229         /* Don't wait up for transmitted skbs to be freed. */
1230         if (!use_napi) {
1231                 skb_orphan(skb);
1232                 nf_reset(skb);
1233         }
1234
1235         /* If running out of space, stop queue to avoid getting packets that we
1236          * are then unable to transmit.
1237          * An alternative would be to force queuing layer to requeue the skb by
1238          * returning NETDEV_TX_BUSY. However, NETDEV_TX_BUSY should not be
1239          * returned in a normal path of operation: it means that driver is not
1240          * maintaining the TX queue stop/start state properly, and causes
1241          * the stack to do a non-trivial amount of useless work.
1242          * Since most packets only take 1 or 2 ring slots, stopping the queue
1243          * early means 16 slots are typically wasted.
1244          */
1245         if (sq->vq->num_free < 2+MAX_SKB_FRAGS) {
1246                 netif_stop_subqueue(dev, qnum);
1247                 if (!use_napi &&
1248                     unlikely(!virtqueue_enable_cb_delayed(sq->vq))) {
1249                         /* More just got used, free them then recheck. */
1250                         free_old_xmit_skbs(sq);
1251                         if (sq->vq->num_free >= 2+MAX_SKB_FRAGS) {
1252                                 netif_start_subqueue(dev, qnum);
1253                                 virtqueue_disable_cb(sq->vq);
1254                         }
1255                 }
1256         }
1257
1258         if (kick || netif_xmit_stopped(txq))
1259                 virtqueue_kick(sq->vq);
1260
1261         return NETDEV_TX_OK;
1262 }
1263
1264 /*
1265  * Send command via the control virtqueue and check status.  Commands
1266  * supported by the hypervisor, as indicated by feature bits, should
1267  * never fail unless improperly formatted.
1268  */
1269 static bool virtnet_send_command(struct virtnet_info *vi, u8 class, u8 cmd,
1270                                  struct scatterlist *out)
1271 {
1272         struct scatterlist *sgs[4], hdr, stat;
1273         unsigned out_num = 0, tmp;
1274
1275         /* Caller should know better */
1276         BUG_ON(!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ));
1277
1278         vi->ctrl_status = ~0;
1279         vi->ctrl_hdr.class = class;
1280         vi->ctrl_hdr.cmd = cmd;
1281         /* Add header */
1282         sg_init_one(&hdr, &vi->ctrl_hdr, sizeof(vi->ctrl_hdr));
1283         sgs[out_num++] = &hdr;
1284
1285         if (out)
1286                 sgs[out_num++] = out;
1287
1288         /* Add return status. */
1289         sg_init_one(&stat, &vi->ctrl_status, sizeof(vi->ctrl_status));
1290         sgs[out_num] = &stat;
1291
1292         BUG_ON(out_num + 1 > ARRAY_SIZE(sgs));
1293         virtqueue_add_sgs(vi->cvq, sgs, out_num, 1, vi, GFP_ATOMIC);
1294
1295         if (unlikely(!virtqueue_kick(vi->cvq)))
1296                 return vi->ctrl_status == VIRTIO_NET_OK;
1297
1298         /* Spin for a response, the kick causes an ioport write, trapping
1299          * into the hypervisor, so the request should be handled immediately.
1300          */
1301         while (!virtqueue_get_buf(vi->cvq, &tmp) &&
1302                !virtqueue_is_broken(vi->cvq))
1303                 cpu_relax();
1304
1305         return vi->ctrl_status == VIRTIO_NET_OK;
1306 }
1307
1308 static int virtnet_set_mac_address(struct net_device *dev, void *p)
1309 {
1310         struct virtnet_info *vi = netdev_priv(dev);
1311         struct virtio_device *vdev = vi->vdev;
1312         int ret;
1313         struct sockaddr *addr;
1314         struct scatterlist sg;
1315
1316         addr = kmemdup(p, sizeof(*addr), GFP_KERNEL);
1317         if (!addr)
1318                 return -ENOMEM;
1319
1320         ret = eth_prepare_mac_addr_change(dev, addr);
1321         if (ret)
1322                 goto out;
1323
1324         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
1325                 sg_init_one(&sg, addr->sa_data, dev->addr_len);
1326                 if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1327                                           VIRTIO_NET_CTRL_MAC_ADDR_SET, &sg)) {
1328                         dev_warn(&vdev->dev,
1329                                  "Failed to set mac address by vq command.\n");
1330                         ret = -EINVAL;
1331                         goto out;
1332                 }
1333         } else if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC) &&
1334                    !virtio_has_feature(vdev, VIRTIO_F_VERSION_1)) {
1335                 unsigned int i;
1336
1337                 /* Naturally, this has an atomicity problem. */
1338                 for (i = 0; i < dev->addr_len; i++)
1339                         virtio_cwrite8(vdev,
1340                                        offsetof(struct virtio_net_config, mac) +
1341                                        i, addr->sa_data[i]);
1342         }
1343
1344         eth_commit_mac_addr_change(dev, p);
1345         ret = 0;
1346
1347 out:
1348         kfree(addr);
1349         return ret;
1350 }
1351
1352 static void virtnet_stats(struct net_device *dev,
1353                           struct rtnl_link_stats64 *tot)
1354 {
1355         struct virtnet_info *vi = netdev_priv(dev);
1356         int cpu;
1357         unsigned int start;
1358
1359         for_each_possible_cpu(cpu) {
1360                 struct virtnet_stats *stats = per_cpu_ptr(vi->stats, cpu);
1361                 u64 tpackets, tbytes, rpackets, rbytes;
1362
1363                 do {
1364                         start = u64_stats_fetch_begin_irq(&stats->tx_syncp);
1365                         tpackets = stats->tx_packets;
1366                         tbytes   = stats->tx_bytes;
1367                 } while (u64_stats_fetch_retry_irq(&stats->tx_syncp, start));
1368
1369                 do {
1370                         start = u64_stats_fetch_begin_irq(&stats->rx_syncp);
1371                         rpackets = stats->rx_packets;
1372                         rbytes   = stats->rx_bytes;
1373                 } while (u64_stats_fetch_retry_irq(&stats->rx_syncp, start));
1374
1375                 tot->rx_packets += rpackets;
1376                 tot->tx_packets += tpackets;
1377                 tot->rx_bytes   += rbytes;
1378                 tot->tx_bytes   += tbytes;
1379         }
1380
1381         tot->tx_dropped = dev->stats.tx_dropped;
1382         tot->tx_fifo_errors = dev->stats.tx_fifo_errors;
1383         tot->rx_dropped = dev->stats.rx_dropped;
1384         tot->rx_length_errors = dev->stats.rx_length_errors;
1385         tot->rx_frame_errors = dev->stats.rx_frame_errors;
1386 }
1387
1388 #ifdef CONFIG_NET_POLL_CONTROLLER
1389 static void virtnet_netpoll(struct net_device *dev)
1390 {
1391         struct virtnet_info *vi = netdev_priv(dev);
1392         int i;
1393
1394         for (i = 0; i < vi->curr_queue_pairs; i++)
1395                 napi_schedule(&vi->rq[i].napi);
1396 }
1397 #endif
1398
1399 static void virtnet_ack_link_announce(struct virtnet_info *vi)
1400 {
1401         rtnl_lock();
1402         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_ANNOUNCE,
1403                                   VIRTIO_NET_CTRL_ANNOUNCE_ACK, NULL))
1404                 dev_warn(&vi->dev->dev, "Failed to ack link announce.\n");
1405         rtnl_unlock();
1406 }
1407
1408 static int _virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1409 {
1410         struct scatterlist sg;
1411         struct net_device *dev = vi->dev;
1412
1413         if (!vi->has_cvq || !virtio_has_feature(vi->vdev, VIRTIO_NET_F_MQ))
1414                 return 0;
1415
1416         vi->ctrl_mq.virtqueue_pairs = cpu_to_virtio16(vi->vdev, queue_pairs);
1417         sg_init_one(&sg, &vi->ctrl_mq, sizeof(vi->ctrl_mq));
1418
1419         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MQ,
1420                                   VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &sg)) {
1421                 dev_warn(&dev->dev, "Fail to set num of queue pairs to %d\n",
1422                          queue_pairs);
1423                 return -EINVAL;
1424         } else {
1425                 vi->curr_queue_pairs = queue_pairs;
1426                 /* virtnet_open() will refill when device is going to up. */
1427                 if (dev->flags & IFF_UP)
1428                         schedule_delayed_work(&vi->refill, 0);
1429         }
1430
1431         return 0;
1432 }
1433
1434 static int virtnet_set_queues(struct virtnet_info *vi, u16 queue_pairs)
1435 {
1436         int err;
1437
1438         rtnl_lock();
1439         err = _virtnet_set_queues(vi, queue_pairs);
1440         rtnl_unlock();
1441         return err;
1442 }
1443
1444 static int virtnet_close(struct net_device *dev)
1445 {
1446         struct virtnet_info *vi = netdev_priv(dev);
1447         int i;
1448
1449         /* Make sure refill_work doesn't re-enable napi! */
1450         cancel_delayed_work_sync(&vi->refill);
1451
1452         for (i = 0; i < vi->max_queue_pairs; i++) {
1453                 napi_disable(&vi->rq[i].napi);
1454                 virtnet_napi_tx_disable(&vi->sq[i].napi);
1455         }
1456
1457         return 0;
1458 }
1459
1460 static void virtnet_set_rx_mode(struct net_device *dev)
1461 {
1462         struct virtnet_info *vi = netdev_priv(dev);
1463         struct scatterlist sg[2];
1464         struct virtio_net_ctrl_mac *mac_data;
1465         struct netdev_hw_addr *ha;
1466         int uc_count;
1467         int mc_count;
1468         void *buf;
1469         int i;
1470
1471         /* We can't dynamically set ndo_set_rx_mode, so return gracefully */
1472         if (!virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_RX))
1473                 return;
1474
1475         vi->ctrl_promisc = ((dev->flags & IFF_PROMISC) != 0);
1476         vi->ctrl_allmulti = ((dev->flags & IFF_ALLMULTI) != 0);
1477
1478         sg_init_one(sg, &vi->ctrl_promisc, sizeof(vi->ctrl_promisc));
1479
1480         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1481                                   VIRTIO_NET_CTRL_RX_PROMISC, sg))
1482                 dev_warn(&dev->dev, "Failed to %sable promisc mode.\n",
1483                          vi->ctrl_promisc ? "en" : "dis");
1484
1485         sg_init_one(sg, &vi->ctrl_allmulti, sizeof(vi->ctrl_allmulti));
1486
1487         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_RX,
1488                                   VIRTIO_NET_CTRL_RX_ALLMULTI, sg))
1489                 dev_warn(&dev->dev, "Failed to %sable allmulti mode.\n",
1490                          vi->ctrl_allmulti ? "en" : "dis");
1491
1492         uc_count = netdev_uc_count(dev);
1493         mc_count = netdev_mc_count(dev);
1494         /* MAC filter - use one buffer for both lists */
1495         buf = kzalloc(((uc_count + mc_count) * ETH_ALEN) +
1496                       (2 * sizeof(mac_data->entries)), GFP_ATOMIC);
1497         mac_data = buf;
1498         if (!buf)
1499                 return;
1500
1501         sg_init_table(sg, 2);
1502
1503         /* Store the unicast list and count in the front of the buffer */
1504         mac_data->entries = cpu_to_virtio32(vi->vdev, uc_count);
1505         i = 0;
1506         netdev_for_each_uc_addr(ha, dev)
1507                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1508
1509         sg_set_buf(&sg[0], mac_data,
1510                    sizeof(mac_data->entries) + (uc_count * ETH_ALEN));
1511
1512         /* multicast list and count fill the end */
1513         mac_data = (void *)&mac_data->macs[uc_count][0];
1514
1515         mac_data->entries = cpu_to_virtio32(vi->vdev, mc_count);
1516         i = 0;
1517         netdev_for_each_mc_addr(ha, dev)
1518                 memcpy(&mac_data->macs[i++][0], ha->addr, ETH_ALEN);
1519
1520         sg_set_buf(&sg[1], mac_data,
1521                    sizeof(mac_data->entries) + (mc_count * ETH_ALEN));
1522
1523         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_MAC,
1524                                   VIRTIO_NET_CTRL_MAC_TABLE_SET, sg))
1525                 dev_warn(&dev->dev, "Failed to set MAC filter table.\n");
1526
1527         kfree(buf);
1528 }
1529
1530 static int virtnet_vlan_rx_add_vid(struct net_device *dev,
1531                                    __be16 proto, u16 vid)
1532 {
1533         struct virtnet_info *vi = netdev_priv(dev);
1534         struct scatterlist sg;
1535
1536         vi->ctrl_vid = vid;
1537         sg_init_one(&sg, &vi->ctrl_vid, sizeof(vi->ctrl_vid));
1538
1539         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1540                                   VIRTIO_NET_CTRL_VLAN_ADD, &sg))
1541                 dev_warn(&dev->dev, "Failed to add VLAN ID %d.\n", vid);
1542         return 0;
1543 }
1544
1545 static int virtnet_vlan_rx_kill_vid(struct net_device *dev,
1546                                     __be16 proto, u16 vid)
1547 {
1548         struct virtnet_info *vi = netdev_priv(dev);
1549         struct scatterlist sg;
1550
1551         vi->ctrl_vid = vid;
1552         sg_init_one(&sg, &vi->ctrl_vid, sizeof(vi->ctrl_vid));
1553
1554         if (!virtnet_send_command(vi, VIRTIO_NET_CTRL_VLAN,
1555                                   VIRTIO_NET_CTRL_VLAN_DEL, &sg))
1556                 dev_warn(&dev->dev, "Failed to kill VLAN ID %d.\n", vid);
1557         return 0;
1558 }
1559
1560 static void virtnet_clean_affinity(struct virtnet_info *vi, long hcpu)
1561 {
1562         int i;
1563
1564         if (vi->affinity_hint_set) {
1565                 for (i = 0; i < vi->max_queue_pairs; i++) {
1566                         virtqueue_set_affinity(vi->rq[i].vq, -1);
1567                         virtqueue_set_affinity(vi->sq[i].vq, -1);
1568                 }
1569
1570                 vi->affinity_hint_set = false;
1571         }
1572 }
1573
1574 static void virtnet_set_affinity(struct virtnet_info *vi)
1575 {
1576         int i;
1577         int cpu;
1578
1579         /* In multiqueue mode, when the number of cpu is equal to the number of
1580          * queue pairs, we let the queue pairs to be private to one cpu by
1581          * setting the affinity hint to eliminate the contention.
1582          */
1583         if (vi->curr_queue_pairs == 1 ||
1584             vi->max_queue_pairs != num_online_cpus()) {
1585                 virtnet_clean_affinity(vi, -1);
1586                 return;
1587         }
1588
1589         i = 0;
1590         for_each_online_cpu(cpu) {
1591                 virtqueue_set_affinity(vi->rq[i].vq, cpu);
1592                 virtqueue_set_affinity(vi->sq[i].vq, cpu);
1593                 netif_set_xps_queue(vi->dev, cpumask_of(cpu), i);
1594                 i++;
1595         }
1596
1597         vi->affinity_hint_set = true;
1598 }
1599
1600 static int virtnet_cpu_online(unsigned int cpu, struct hlist_node *node)
1601 {
1602         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1603                                                    node);
1604         virtnet_set_affinity(vi);
1605         return 0;
1606 }
1607
1608 static int virtnet_cpu_dead(unsigned int cpu, struct hlist_node *node)
1609 {
1610         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1611                                                    node_dead);
1612         virtnet_set_affinity(vi);
1613         return 0;
1614 }
1615
1616 static int virtnet_cpu_down_prep(unsigned int cpu, struct hlist_node *node)
1617 {
1618         struct virtnet_info *vi = hlist_entry_safe(node, struct virtnet_info,
1619                                                    node);
1620
1621         virtnet_clean_affinity(vi, cpu);
1622         return 0;
1623 }
1624
1625 static enum cpuhp_state virtionet_online;
1626
1627 static int virtnet_cpu_notif_add(struct virtnet_info *vi)
1628 {
1629         int ret;
1630
1631         ret = cpuhp_state_add_instance_nocalls(virtionet_online, &vi->node);
1632         if (ret)
1633                 return ret;
1634         ret = cpuhp_state_add_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1635                                                &vi->node_dead);
1636         if (!ret)
1637                 return ret;
1638         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1639         return ret;
1640 }
1641
1642 static void virtnet_cpu_notif_remove(struct virtnet_info *vi)
1643 {
1644         cpuhp_state_remove_instance_nocalls(virtionet_online, &vi->node);
1645         cpuhp_state_remove_instance_nocalls(CPUHP_VIRT_NET_DEAD,
1646                                             &vi->node_dead);
1647 }
1648
1649 static void virtnet_get_ringparam(struct net_device *dev,
1650                                 struct ethtool_ringparam *ring)
1651 {
1652         struct virtnet_info *vi = netdev_priv(dev);
1653
1654         ring->rx_max_pending = virtqueue_get_vring_size(vi->rq[0].vq);
1655         ring->tx_max_pending = virtqueue_get_vring_size(vi->sq[0].vq);
1656         ring->rx_pending = ring->rx_max_pending;
1657         ring->tx_pending = ring->tx_max_pending;
1658 }
1659
1660
1661 static void virtnet_get_drvinfo(struct net_device *dev,
1662                                 struct ethtool_drvinfo *info)
1663 {
1664         struct virtnet_info *vi = netdev_priv(dev);
1665         struct virtio_device *vdev = vi->vdev;
1666
1667         strlcpy(info->driver, KBUILD_MODNAME, sizeof(info->driver));
1668         strlcpy(info->version, VIRTNET_DRIVER_VERSION, sizeof(info->version));
1669         strlcpy(info->bus_info, virtio_bus_name(vdev), sizeof(info->bus_info));
1670
1671 }
1672
1673 /* TODO: Eliminate OOO packets during switching */
1674 static int virtnet_set_channels(struct net_device *dev,
1675                                 struct ethtool_channels *channels)
1676 {
1677         struct virtnet_info *vi = netdev_priv(dev);
1678         u16 queue_pairs = channels->combined_count;
1679         int err;
1680
1681         /* We don't support separate rx/tx channels.
1682          * We don't allow setting 'other' channels.
1683          */
1684         if (channels->rx_count || channels->tx_count || channels->other_count)
1685                 return -EINVAL;
1686
1687         if (queue_pairs > vi->max_queue_pairs || queue_pairs == 0)
1688                 return -EINVAL;
1689
1690         /* For now we don't support modifying channels while XDP is loaded
1691          * also when XDP is loaded all RX queues have XDP programs so we only
1692          * need to check a single RX queue.
1693          */
1694         if (vi->rq[0].xdp_prog)
1695                 return -EINVAL;
1696
1697         get_online_cpus();
1698         err = _virtnet_set_queues(vi, queue_pairs);
1699         if (!err) {
1700                 netif_set_real_num_tx_queues(dev, queue_pairs);
1701                 netif_set_real_num_rx_queues(dev, queue_pairs);
1702
1703                 virtnet_set_affinity(vi);
1704         }
1705         put_online_cpus();
1706
1707         return err;
1708 }
1709
1710 static void virtnet_get_channels(struct net_device *dev,
1711                                  struct ethtool_channels *channels)
1712 {
1713         struct virtnet_info *vi = netdev_priv(dev);
1714
1715         channels->combined_count = vi->curr_queue_pairs;
1716         channels->max_combined = vi->max_queue_pairs;
1717         channels->max_other = 0;
1718         channels->rx_count = 0;
1719         channels->tx_count = 0;
1720         channels->other_count = 0;
1721 }
1722
1723 /* Check if the user is trying to change anything besides speed/duplex */
1724 static bool
1725 virtnet_validate_ethtool_cmd(const struct ethtool_link_ksettings *cmd)
1726 {
1727         struct ethtool_link_ksettings diff1 = *cmd;
1728         struct ethtool_link_ksettings diff2 = {};
1729
1730         /* cmd is always set so we need to clear it, validate the port type
1731          * and also without autonegotiation we can ignore advertising
1732          */
1733         diff1.base.speed = 0;
1734         diff2.base.port = PORT_OTHER;
1735         ethtool_link_ksettings_zero_link_mode(&diff1, advertising);
1736         diff1.base.duplex = 0;
1737         diff1.base.cmd = 0;
1738         diff1.base.link_mode_masks_nwords = 0;
1739
1740         return !memcmp(&diff1.base, &diff2.base, sizeof(diff1.base)) &&
1741                 bitmap_empty(diff1.link_modes.supported,
1742                              __ETHTOOL_LINK_MODE_MASK_NBITS) &&
1743                 bitmap_empty(diff1.link_modes.advertising,
1744                              __ETHTOOL_LINK_MODE_MASK_NBITS) &&
1745                 bitmap_empty(diff1.link_modes.lp_advertising,
1746                              __ETHTOOL_LINK_MODE_MASK_NBITS);
1747 }
1748
1749 static int virtnet_set_link_ksettings(struct net_device *dev,
1750                                       const struct ethtool_link_ksettings *cmd)
1751 {
1752         struct virtnet_info *vi = netdev_priv(dev);
1753         u32 speed;
1754
1755         speed = cmd->base.speed;
1756         /* don't allow custom speed and duplex */
1757         if (!ethtool_validate_speed(speed) ||
1758             !ethtool_validate_duplex(cmd->base.duplex) ||
1759             !virtnet_validate_ethtool_cmd(cmd))
1760                 return -EINVAL;
1761         vi->speed = speed;
1762         vi->duplex = cmd->base.duplex;
1763
1764         return 0;
1765 }
1766
1767 static int virtnet_get_link_ksettings(struct net_device *dev,
1768                                       struct ethtool_link_ksettings *cmd)
1769 {
1770         struct virtnet_info *vi = netdev_priv(dev);
1771
1772         cmd->base.speed = vi->speed;
1773         cmd->base.duplex = vi->duplex;
1774         cmd->base.port = PORT_OTHER;
1775
1776         return 0;
1777 }
1778
1779 static void virtnet_init_settings(struct net_device *dev)
1780 {
1781         struct virtnet_info *vi = netdev_priv(dev);
1782
1783         vi->speed = SPEED_UNKNOWN;
1784         vi->duplex = DUPLEX_UNKNOWN;
1785 }
1786
1787 static const struct ethtool_ops virtnet_ethtool_ops = {
1788         .get_drvinfo = virtnet_get_drvinfo,
1789         .get_link = ethtool_op_get_link,
1790         .get_ringparam = virtnet_get_ringparam,
1791         .set_channels = virtnet_set_channels,
1792         .get_channels = virtnet_get_channels,
1793         .get_ts_info = ethtool_op_get_ts_info,
1794         .get_link_ksettings = virtnet_get_link_ksettings,
1795         .set_link_ksettings = virtnet_set_link_ksettings,
1796 };
1797
1798 static void virtnet_freeze_down(struct virtio_device *vdev)
1799 {
1800         struct virtnet_info *vi = vdev->priv;
1801         int i;
1802
1803         /* Make sure no work handler is accessing the device */
1804         flush_work(&vi->config_work);
1805
1806         netif_device_detach(vi->dev);
1807         cancel_delayed_work_sync(&vi->refill);
1808
1809         if (netif_running(vi->dev)) {
1810                 for (i = 0; i < vi->max_queue_pairs; i++) {
1811                         napi_disable(&vi->rq[i].napi);
1812                         virtnet_napi_tx_disable(&vi->sq[i].napi);
1813                 }
1814         }
1815 }
1816
1817 static int init_vqs(struct virtnet_info *vi);
1818 static void _remove_vq_common(struct virtnet_info *vi);
1819
1820 static int virtnet_restore_up(struct virtio_device *vdev)
1821 {
1822         struct virtnet_info *vi = vdev->priv;
1823         int err, i;
1824
1825         err = init_vqs(vi);
1826         if (err)
1827                 return err;
1828
1829         virtio_device_ready(vdev);
1830
1831         if (netif_running(vi->dev)) {
1832                 for (i = 0; i < vi->curr_queue_pairs; i++)
1833                         if (!try_fill_recv(vi, &vi->rq[i], GFP_KERNEL))
1834                                 schedule_delayed_work(&vi->refill, 0);
1835
1836                 for (i = 0; i < vi->max_queue_pairs; i++) {
1837                         virtnet_napi_enable(vi->rq[i].vq, &vi->rq[i].napi);
1838                         virtnet_napi_tx_enable(vi, vi->sq[i].vq,
1839                                                &vi->sq[i].napi);
1840                 }
1841         }
1842
1843         netif_device_attach(vi->dev);
1844         return err;
1845 }
1846
1847 static int virtnet_reset(struct virtnet_info *vi, int curr_qp, int xdp_qp)
1848 {
1849         struct virtio_device *dev = vi->vdev;
1850         int ret;
1851
1852         virtio_config_disable(dev);
1853         dev->failed = dev->config->get_status(dev) & VIRTIO_CONFIG_S_FAILED;
1854         virtnet_freeze_down(dev);
1855         _remove_vq_common(vi);
1856
1857         dev->config->reset(dev);
1858         virtio_add_status(dev, VIRTIO_CONFIG_S_ACKNOWLEDGE);
1859         virtio_add_status(dev, VIRTIO_CONFIG_S_DRIVER);
1860
1861         ret = virtio_finalize_features(dev);
1862         if (ret)
1863                 goto err;
1864
1865         vi->xdp_queue_pairs = xdp_qp;
1866         ret = virtnet_restore_up(dev);
1867         if (ret)
1868                 goto err;
1869         ret = _virtnet_set_queues(vi, curr_qp);
1870         if (ret)
1871                 goto err;
1872
1873         virtio_add_status(dev, VIRTIO_CONFIG_S_DRIVER_OK);
1874         virtio_config_enable(dev);
1875         return 0;
1876 err:
1877         virtio_add_status(dev, VIRTIO_CONFIG_S_FAILED);
1878         return ret;
1879 }
1880
1881 static int virtnet_xdp_set(struct net_device *dev, struct bpf_prog *prog,
1882                            struct netlink_ext_ack *extack)
1883 {
1884         unsigned long int max_sz = PAGE_SIZE - sizeof(struct padded_vnet_hdr);
1885         struct virtnet_info *vi = netdev_priv(dev);
1886         struct bpf_prog *old_prog;
1887         u16 xdp_qp = 0, curr_qp;
1888         int i, err;
1889
1890         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO4) ||
1891             virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_TSO6) ||
1892             virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_ECN) ||
1893             virtio_has_feature(vi->vdev, VIRTIO_NET_F_GUEST_UFO)) {
1894                 NL_SET_ERR_MSG_MOD(extack, "Can't set XDP while host is implementing LRO, disable LRO first");
1895                 return -EOPNOTSUPP;
1896         }
1897
1898         if (vi->mergeable_rx_bufs && !vi->any_header_sg) {
1899                 NL_SET_ERR_MSG_MOD(extack, "XDP expects header/data in single page, any_header_sg required");
1900                 return -EINVAL;
1901         }
1902
1903         if (dev->mtu > max_sz) {
1904                 NL_SET_ERR_MSG_MOD(extack, "MTU too large to enable XDP");
1905                 netdev_warn(dev, "XDP requires MTU less than %lu\n", max_sz);
1906                 return -EINVAL;
1907         }
1908
1909         curr_qp = vi->curr_queue_pairs - vi->xdp_queue_pairs;
1910         if (prog)
1911                 xdp_qp = nr_cpu_ids;
1912
1913         /* XDP requires extra queues for XDP_TX */
1914         if (curr_qp + xdp_qp > vi->max_queue_pairs) {
1915                 NL_SET_ERR_MSG_MOD(extack, "Too few free TX rings available");
1916                 netdev_warn(dev, "request %i queues but max is %i\n",
1917                             curr_qp + xdp_qp, vi->max_queue_pairs);
1918                 return -ENOMEM;
1919         }
1920
1921         if (prog) {
1922                 prog = bpf_prog_add(prog, vi->max_queue_pairs - 1);
1923                 if (IS_ERR(prog))
1924                         return PTR_ERR(prog);
1925         }
1926
1927         /* Changing the headroom in buffers is a disruptive operation because
1928          * existing buffers must be flushed and reallocated. This will happen
1929          * when a xdp program is initially added or xdp is disabled by removing
1930          * the xdp program resulting in number of XDP queues changing.
1931          */
1932         if (vi->xdp_queue_pairs != xdp_qp) {
1933                 err = virtnet_reset(vi, curr_qp + xdp_qp, xdp_qp);
1934                 if (err) {
1935                         dev_warn(&dev->dev, "XDP reset failure.\n");
1936                         goto virtio_reset_err;
1937                 }
1938         }
1939
1940         netif_set_real_num_rx_queues(dev, curr_qp + xdp_qp);
1941
1942         for (i = 0; i < vi->max_queue_pairs; i++) {
1943                 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
1944                 rcu_assign_pointer(vi->rq[i].xdp_prog, prog);
1945                 if (old_prog)
1946                         bpf_prog_put(old_prog);
1947         }
1948
1949         return 0;
1950
1951 virtio_reset_err:
1952         /* On reset error do our best to unwind XDP changes inflight and return
1953          * error up to user space for resolution. The underlying reset hung on
1954          * us so not much we can do here.
1955          */
1956         if (prog)
1957                 bpf_prog_sub(prog, vi->max_queue_pairs - 1);
1958         return err;
1959 }
1960
1961 static bool virtnet_xdp_query(struct net_device *dev)
1962 {
1963         struct virtnet_info *vi = netdev_priv(dev);
1964         int i;
1965
1966         for (i = 0; i < vi->max_queue_pairs; i++) {
1967                 if (vi->rq[i].xdp_prog)
1968                         return true;
1969         }
1970         return false;
1971 }
1972
1973 static int virtnet_xdp(struct net_device *dev, struct netdev_xdp *xdp)
1974 {
1975         switch (xdp->command) {
1976         case XDP_SETUP_PROG:
1977                 return virtnet_xdp_set(dev, xdp->prog, xdp->extack);
1978         case XDP_QUERY_PROG:
1979                 xdp->prog_attached = virtnet_xdp_query(dev);
1980                 return 0;
1981         default:
1982                 return -EINVAL;
1983         }
1984 }
1985
1986 static const struct net_device_ops virtnet_netdev = {
1987         .ndo_open            = virtnet_open,
1988         .ndo_stop            = virtnet_close,
1989         .ndo_start_xmit      = start_xmit,
1990         .ndo_validate_addr   = eth_validate_addr,
1991         .ndo_set_mac_address = virtnet_set_mac_address,
1992         .ndo_set_rx_mode     = virtnet_set_rx_mode,
1993         .ndo_get_stats64     = virtnet_stats,
1994         .ndo_vlan_rx_add_vid = virtnet_vlan_rx_add_vid,
1995         .ndo_vlan_rx_kill_vid = virtnet_vlan_rx_kill_vid,
1996 #ifdef CONFIG_NET_POLL_CONTROLLER
1997         .ndo_poll_controller = virtnet_netpoll,
1998 #endif
1999         .ndo_xdp                = virtnet_xdp,
2000 };
2001
2002 static void virtnet_config_changed_work(struct work_struct *work)
2003 {
2004         struct virtnet_info *vi =
2005                 container_of(work, struct virtnet_info, config_work);
2006         u16 v;
2007
2008         if (virtio_cread_feature(vi->vdev, VIRTIO_NET_F_STATUS,
2009                                  struct virtio_net_config, status, &v) < 0)
2010                 return;
2011
2012         if (v & VIRTIO_NET_S_ANNOUNCE) {
2013                 netdev_notify_peers(vi->dev);
2014                 virtnet_ack_link_announce(vi);
2015         }
2016
2017         /* Ignore unknown (future) status bits */
2018         v &= VIRTIO_NET_S_LINK_UP;
2019
2020         if (vi->status == v)
2021                 return;
2022
2023         vi->status = v;
2024
2025         if (vi->status & VIRTIO_NET_S_LINK_UP) {
2026                 netif_carrier_on(vi->dev);
2027                 netif_tx_wake_all_queues(vi->dev);
2028         } else {
2029                 netif_carrier_off(vi->dev);
2030                 netif_tx_stop_all_queues(vi->dev);
2031         }
2032 }
2033
2034 static void virtnet_config_changed(struct virtio_device *vdev)
2035 {
2036         struct virtnet_info *vi = vdev->priv;
2037
2038         schedule_work(&vi->config_work);
2039 }
2040
2041 static void virtnet_free_queues(struct virtnet_info *vi)
2042 {
2043         int i;
2044
2045         for (i = 0; i < vi->max_queue_pairs; i++) {
2046                 napi_hash_del(&vi->rq[i].napi);
2047                 netif_napi_del(&vi->rq[i].napi);
2048                 netif_napi_del(&vi->sq[i].napi);
2049         }
2050
2051         /* We called napi_hash_del() before netif_napi_del(),
2052          * we need to respect an RCU grace period before freeing vi->rq
2053          */
2054         synchronize_net();
2055
2056         kfree(vi->rq);
2057         kfree(vi->sq);
2058 }
2059
2060 static void _free_receive_bufs(struct virtnet_info *vi)
2061 {
2062         struct bpf_prog *old_prog;
2063         int i;
2064
2065         for (i = 0; i < vi->max_queue_pairs; i++) {
2066                 while (vi->rq[i].pages)
2067                         __free_pages(get_a_page(&vi->rq[i], GFP_KERNEL), 0);
2068
2069                 old_prog = rtnl_dereference(vi->rq[i].xdp_prog);
2070                 RCU_INIT_POINTER(vi->rq[i].xdp_prog, NULL);
2071                 if (old_prog)
2072                         bpf_prog_put(old_prog);
2073         }
2074 }
2075
2076 static void free_receive_bufs(struct virtnet_info *vi)
2077 {
2078         rtnl_lock();
2079         _free_receive_bufs(vi);
2080         rtnl_unlock();
2081 }
2082
2083 static void free_receive_page_frags(struct virtnet_info *vi)
2084 {
2085         int i;
2086         for (i = 0; i < vi->max_queue_pairs; i++)
2087                 if (vi->rq[i].alloc_frag.page)
2088                         put_page(vi->rq[i].alloc_frag.page);
2089 }
2090
2091 static bool is_xdp_raw_buffer_queue(struct virtnet_info *vi, int q)
2092 {
2093         if (q < (vi->curr_queue_pairs - vi->xdp_queue_pairs))
2094                 return false;
2095         else if (q < vi->curr_queue_pairs)
2096                 return true;
2097         else
2098                 return false;
2099 }
2100
2101 static void free_unused_bufs(struct virtnet_info *vi)
2102 {
2103         void *buf;
2104         int i;
2105
2106         for (i = 0; i < vi->max_queue_pairs; i++) {
2107                 struct virtqueue *vq = vi->sq[i].vq;
2108                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2109                         if (!is_xdp_raw_buffer_queue(vi, i))
2110                                 dev_kfree_skb(buf);
2111                         else
2112                                 put_page(virt_to_head_page(buf));
2113                 }
2114         }
2115
2116         for (i = 0; i < vi->max_queue_pairs; i++) {
2117                 struct virtqueue *vq = vi->rq[i].vq;
2118
2119                 while ((buf = virtqueue_detach_unused_buf(vq)) != NULL) {
2120                         if (vi->mergeable_rx_bufs) {
2121                                 unsigned long ctx = (unsigned long)buf;
2122                                 void *base = mergeable_ctx_to_buf_address(ctx);
2123                                 put_page(virt_to_head_page(base));
2124                         } else if (vi->big_packets) {
2125                                 give_pages(&vi->rq[i], buf);
2126                         } else {
2127                                 put_page(virt_to_head_page(buf));
2128                         }
2129                 }
2130         }
2131 }
2132
2133 static void virtnet_del_vqs(struct virtnet_info *vi)
2134 {
2135         struct virtio_device *vdev = vi->vdev;
2136
2137         virtnet_clean_affinity(vi, -1);
2138
2139         vdev->config->del_vqs(vdev);
2140
2141         virtnet_free_queues(vi);
2142 }
2143
2144 static int virtnet_find_vqs(struct virtnet_info *vi)
2145 {
2146         vq_callback_t **callbacks;
2147         struct virtqueue **vqs;
2148         int ret = -ENOMEM;
2149         int i, total_vqs;
2150         const char **names;
2151
2152         /* We expect 1 RX virtqueue followed by 1 TX virtqueue, followed by
2153          * possible N-1 RX/TX queue pairs used in multiqueue mode, followed by
2154          * possible control vq.
2155          */
2156         total_vqs = vi->max_queue_pairs * 2 +
2157                     virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VQ);
2158
2159         /* Allocate space for find_vqs parameters */
2160         vqs = kzalloc(total_vqs * sizeof(*vqs), GFP_KERNEL);
2161         if (!vqs)
2162                 goto err_vq;
2163         callbacks = kmalloc(total_vqs * sizeof(*callbacks), GFP_KERNEL);
2164         if (!callbacks)
2165                 goto err_callback;
2166         names = kmalloc(total_vqs * sizeof(*names), GFP_KERNEL);
2167         if (!names)
2168                 goto err_names;
2169
2170         /* Parameters for control virtqueue, if any */
2171         if (vi->has_cvq) {
2172                 callbacks[total_vqs - 1] = NULL;
2173                 names[total_vqs - 1] = "control";
2174         }
2175
2176         /* Allocate/initialize parameters for send/receive virtqueues */
2177         for (i = 0; i < vi->max_queue_pairs; i++) {
2178                 callbacks[rxq2vq(i)] = skb_recv_done;
2179                 callbacks[txq2vq(i)] = skb_xmit_done;
2180                 sprintf(vi->rq[i].name, "input.%d", i);
2181                 sprintf(vi->sq[i].name, "output.%d", i);
2182                 names[rxq2vq(i)] = vi->rq[i].name;
2183                 names[txq2vq(i)] = vi->sq[i].name;
2184         }
2185
2186         ret = vi->vdev->config->find_vqs(vi->vdev, total_vqs, vqs, callbacks,
2187                                          names, NULL);
2188         if (ret)
2189                 goto err_find;
2190
2191         if (vi->has_cvq) {
2192                 vi->cvq = vqs[total_vqs - 1];
2193                 if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_CTRL_VLAN))
2194                         vi->dev->features |= NETIF_F_HW_VLAN_CTAG_FILTER;
2195         }
2196
2197         for (i = 0; i < vi->max_queue_pairs; i++) {
2198                 vi->rq[i].vq = vqs[rxq2vq(i)];
2199                 vi->sq[i].vq = vqs[txq2vq(i)];
2200         }
2201
2202         kfree(names);
2203         kfree(callbacks);
2204         kfree(vqs);
2205
2206         return 0;
2207
2208 err_find:
2209         kfree(names);
2210 err_names:
2211         kfree(callbacks);
2212 err_callback:
2213         kfree(vqs);
2214 err_vq:
2215         return ret;
2216 }
2217
2218 static int virtnet_alloc_queues(struct virtnet_info *vi)
2219 {
2220         int i;
2221
2222         vi->sq = kzalloc(sizeof(*vi->sq) * vi->max_queue_pairs, GFP_KERNEL);
2223         if (!vi->sq)
2224                 goto err_sq;
2225         vi->rq = kzalloc(sizeof(*vi->rq) * vi->max_queue_pairs, GFP_KERNEL);
2226         if (!vi->rq)
2227                 goto err_rq;
2228
2229         INIT_DELAYED_WORK(&vi->refill, refill_work);
2230         for (i = 0; i < vi->max_queue_pairs; i++) {
2231                 vi->rq[i].pages = NULL;
2232                 netif_napi_add(vi->dev, &vi->rq[i].napi, virtnet_poll,
2233                                napi_weight);
2234                 netif_tx_napi_add(vi->dev, &vi->sq[i].napi, virtnet_poll_tx,
2235                                   napi_tx ? napi_weight : 0);
2236
2237                 sg_init_table(vi->rq[i].sg, ARRAY_SIZE(vi->rq[i].sg));
2238                 ewma_pkt_len_init(&vi->rq[i].mrg_avg_pkt_len);
2239                 sg_init_table(vi->sq[i].sg, ARRAY_SIZE(vi->sq[i].sg));
2240         }
2241
2242         return 0;
2243
2244 err_rq:
2245         kfree(vi->sq);
2246 err_sq:
2247         return -ENOMEM;
2248 }
2249
2250 static int init_vqs(struct virtnet_info *vi)
2251 {
2252         int ret;
2253
2254         /* Allocate send & receive queues */
2255         ret = virtnet_alloc_queues(vi);
2256         if (ret)
2257                 goto err;
2258
2259         ret = virtnet_find_vqs(vi);
2260         if (ret)
2261                 goto err_free;
2262
2263         get_online_cpus();
2264         virtnet_set_affinity(vi);
2265         put_online_cpus();
2266
2267         return 0;
2268
2269 err_free:
2270         virtnet_free_queues(vi);
2271 err:
2272         return ret;
2273 }
2274
2275 #ifdef CONFIG_SYSFS
2276 static ssize_t mergeable_rx_buffer_size_show(struct netdev_rx_queue *queue,
2277                 struct rx_queue_attribute *attribute, char *buf)
2278 {
2279         struct virtnet_info *vi = netdev_priv(queue->dev);
2280         unsigned int queue_index = get_netdev_rx_queue_index(queue);
2281         struct ewma_pkt_len *avg;
2282
2283         BUG_ON(queue_index >= vi->max_queue_pairs);
2284         avg = &vi->rq[queue_index].mrg_avg_pkt_len;
2285         return sprintf(buf, "%u\n", get_mergeable_buf_len(avg));
2286 }
2287
2288 static struct rx_queue_attribute mergeable_rx_buffer_size_attribute =
2289         __ATTR_RO(mergeable_rx_buffer_size);
2290
2291 static struct attribute *virtio_net_mrg_rx_attrs[] = {
2292         &mergeable_rx_buffer_size_attribute.attr,
2293         NULL
2294 };
2295
2296 static const struct attribute_group virtio_net_mrg_rx_group = {
2297         .name = "virtio_net",
2298         .attrs = virtio_net_mrg_rx_attrs
2299 };
2300 #endif
2301
2302 static bool virtnet_fail_on_feature(struct virtio_device *vdev,
2303                                     unsigned int fbit,
2304                                     const char *fname, const char *dname)
2305 {
2306         if (!virtio_has_feature(vdev, fbit))
2307                 return false;
2308
2309         dev_err(&vdev->dev, "device advertises feature %s but not %s",
2310                 fname, dname);
2311
2312         return true;
2313 }
2314
2315 #define VIRTNET_FAIL_ON(vdev, fbit, dbit)                       \
2316         virtnet_fail_on_feature(vdev, fbit, #fbit, dbit)
2317
2318 static bool virtnet_validate_features(struct virtio_device *vdev)
2319 {
2320         if (!virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) &&
2321             (VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_RX,
2322                              "VIRTIO_NET_F_CTRL_VQ") ||
2323              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_VLAN,
2324                              "VIRTIO_NET_F_CTRL_VQ") ||
2325              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_GUEST_ANNOUNCE,
2326                              "VIRTIO_NET_F_CTRL_VQ") ||
2327              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_MQ, "VIRTIO_NET_F_CTRL_VQ") ||
2328              VIRTNET_FAIL_ON(vdev, VIRTIO_NET_F_CTRL_MAC_ADDR,
2329                              "VIRTIO_NET_F_CTRL_VQ"))) {
2330                 return false;
2331         }
2332
2333         return true;
2334 }
2335
2336 #define MIN_MTU ETH_MIN_MTU
2337 #define MAX_MTU ETH_MAX_MTU
2338
2339 static int virtnet_validate(struct virtio_device *vdev)
2340 {
2341         if (!vdev->config->get) {
2342                 dev_err(&vdev->dev, "%s failure: config access disabled\n",
2343                         __func__);
2344                 return -EINVAL;
2345         }
2346
2347         if (!virtnet_validate_features(vdev))
2348                 return -EINVAL;
2349
2350         if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2351                 int mtu = virtio_cread16(vdev,
2352                                          offsetof(struct virtio_net_config,
2353                                                   mtu));
2354                 if (mtu < MIN_MTU)
2355                         __virtio_clear_bit(vdev, VIRTIO_NET_F_MTU);
2356         }
2357
2358         return 0;
2359 }
2360
2361 static int virtnet_probe(struct virtio_device *vdev)
2362 {
2363         int i, err;
2364         struct net_device *dev;
2365         struct virtnet_info *vi;
2366         u16 max_queue_pairs;
2367         int mtu;
2368
2369         /* Find if host supports multiqueue virtio_net device */
2370         err = virtio_cread_feature(vdev, VIRTIO_NET_F_MQ,
2371                                    struct virtio_net_config,
2372                                    max_virtqueue_pairs, &max_queue_pairs);
2373
2374         /* We need at least 2 queue's */
2375         if (err || max_queue_pairs < VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN ||
2376             max_queue_pairs > VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX ||
2377             !virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2378                 max_queue_pairs = 1;
2379
2380         /* Allocate ourselves a network device with room for our info */
2381         dev = alloc_etherdev_mq(sizeof(struct virtnet_info), max_queue_pairs);
2382         if (!dev)
2383                 return -ENOMEM;
2384
2385         /* Set up network device as normal. */
2386         dev->priv_flags |= IFF_UNICAST_FLT | IFF_LIVE_ADDR_CHANGE;
2387         dev->netdev_ops = &virtnet_netdev;
2388         dev->features = NETIF_F_HIGHDMA;
2389
2390         dev->ethtool_ops = &virtnet_ethtool_ops;
2391         SET_NETDEV_DEV(dev, &vdev->dev);
2392
2393         /* Do we support "hardware" checksums? */
2394         if (virtio_has_feature(vdev, VIRTIO_NET_F_CSUM)) {
2395                 /* This opens up the world of extra features. */
2396                 dev->hw_features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2397                 if (csum)
2398                         dev->features |= NETIF_F_HW_CSUM | NETIF_F_SG;
2399
2400                 if (virtio_has_feature(vdev, VIRTIO_NET_F_GSO)) {
2401                         dev->hw_features |= NETIF_F_TSO | NETIF_F_UFO
2402                                 | NETIF_F_TSO_ECN | NETIF_F_TSO6;
2403                 }
2404                 /* Individual feature bits: what can host handle? */
2405                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO4))
2406                         dev->hw_features |= NETIF_F_TSO;
2407                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_TSO6))
2408                         dev->hw_features |= NETIF_F_TSO6;
2409                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_ECN))
2410                         dev->hw_features |= NETIF_F_TSO_ECN;
2411                 if (virtio_has_feature(vdev, VIRTIO_NET_F_HOST_UFO))
2412                         dev->hw_features |= NETIF_F_UFO;
2413
2414                 dev->features |= NETIF_F_GSO_ROBUST;
2415
2416                 if (gso)
2417                         dev->features |= dev->hw_features & (NETIF_F_ALL_TSO|NETIF_F_UFO);
2418                 /* (!csum && gso) case will be fixed by register_netdev() */
2419         }
2420         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_CSUM))
2421                 dev->features |= NETIF_F_RXCSUM;
2422
2423         dev->vlan_features = dev->features;
2424
2425         /* MTU range: 68 - 65535 */
2426         dev->min_mtu = MIN_MTU;
2427         dev->max_mtu = MAX_MTU;
2428
2429         /* Configuration may specify what MAC to use.  Otherwise random. */
2430         if (virtio_has_feature(vdev, VIRTIO_NET_F_MAC))
2431                 virtio_cread_bytes(vdev,
2432                                    offsetof(struct virtio_net_config, mac),
2433                                    dev->dev_addr, dev->addr_len);
2434         else
2435                 eth_hw_addr_random(dev);
2436
2437         /* Set up our device-specific information */
2438         vi = netdev_priv(dev);
2439         vi->dev = dev;
2440         vi->vdev = vdev;
2441         vdev->priv = vi;
2442         vi->stats = alloc_percpu(struct virtnet_stats);
2443         err = -ENOMEM;
2444         if (vi->stats == NULL)
2445                 goto free;
2446
2447         for_each_possible_cpu(i) {
2448                 struct virtnet_stats *virtnet_stats;
2449                 virtnet_stats = per_cpu_ptr(vi->stats, i);
2450                 u64_stats_init(&virtnet_stats->tx_syncp);
2451                 u64_stats_init(&virtnet_stats->rx_syncp);
2452         }
2453
2454         INIT_WORK(&vi->config_work, virtnet_config_changed_work);
2455
2456         /* If we can receive ANY GSO packets, we must allocate large ones. */
2457         if (virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO4) ||
2458             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_TSO6) ||
2459             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_ECN) ||
2460             virtio_has_feature(vdev, VIRTIO_NET_F_GUEST_UFO))
2461                 vi->big_packets = true;
2462
2463         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF))
2464                 vi->mergeable_rx_bufs = true;
2465
2466         if (virtio_has_feature(vdev, VIRTIO_NET_F_MRG_RXBUF) ||
2467             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2468                 vi->hdr_len = sizeof(struct virtio_net_hdr_mrg_rxbuf);
2469         else
2470                 vi->hdr_len = sizeof(struct virtio_net_hdr);
2471
2472         if (virtio_has_feature(vdev, VIRTIO_F_ANY_LAYOUT) ||
2473             virtio_has_feature(vdev, VIRTIO_F_VERSION_1))
2474                 vi->any_header_sg = true;
2475
2476         if (virtio_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ))
2477                 vi->has_cvq = true;
2478
2479         if (virtio_has_feature(vdev, VIRTIO_NET_F_MTU)) {
2480                 mtu = virtio_cread16(vdev,
2481                                      offsetof(struct virtio_net_config,
2482                                               mtu));
2483                 if (mtu < dev->min_mtu) {
2484                         /* Should never trigger: MTU was previously validated
2485                          * in virtnet_validate.
2486                          */
2487                         dev_err(&vdev->dev, "device MTU appears to have changed "
2488                                 "it is now %d < %d", mtu, dev->min_mtu);
2489                         goto free_stats;
2490                 }
2491
2492                 dev->mtu = mtu;
2493                 dev->max_mtu = mtu;
2494
2495                 /* TODO: size buffers correctly in this case. */
2496                 if (dev->mtu > ETH_DATA_LEN)
2497                         vi->big_packets = true;
2498         }
2499
2500         if (vi->any_header_sg)
2501                 dev->needed_headroom = vi->hdr_len;
2502
2503         /* Enable multiqueue by default */
2504         if (num_online_cpus() >= max_queue_pairs)
2505                 vi->curr_queue_pairs = max_queue_pairs;
2506         else
2507                 vi->curr_queue_pairs = num_online_cpus();
2508         vi->max_queue_pairs = max_queue_pairs;
2509
2510         /* Allocate/initialize the rx/tx queues, and invoke find_vqs */
2511         err = init_vqs(vi);
2512         if (err)
2513                 goto free_stats;
2514
2515 #ifdef CONFIG_SYSFS
2516         if (vi->mergeable_rx_bufs)
2517                 dev->sysfs_rx_queue_group = &virtio_net_mrg_rx_group;
2518 #endif
2519         netif_set_real_num_tx_queues(dev, vi->curr_queue_pairs);
2520         netif_set_real_num_rx_queues(dev, vi->curr_queue_pairs);
2521
2522         virtnet_init_settings(dev);
2523
2524         err = register_netdev(dev);
2525         if (err) {
2526                 pr_debug("virtio_net: registering device failed\n");
2527                 goto free_vqs;
2528         }
2529
2530         virtio_device_ready(vdev);
2531
2532         err = virtnet_cpu_notif_add(vi);
2533         if (err) {
2534                 pr_debug("virtio_net: registering cpu notifier failed\n");
2535                 goto free_unregister_netdev;
2536         }
2537
2538         virtnet_set_queues(vi, vi->curr_queue_pairs);
2539
2540         /* Assume link up if device can't report link status,
2541            otherwise get link status from config. */
2542         if (virtio_has_feature(vi->vdev, VIRTIO_NET_F_STATUS)) {
2543                 netif_carrier_off(dev);
2544                 schedule_work(&vi->config_work);
2545         } else {
2546                 vi->status = VIRTIO_NET_S_LINK_UP;
2547                 netif_carrier_on(dev);
2548         }
2549
2550         pr_debug("virtnet: registered device %s with %d RX and TX vq's\n",
2551                  dev->name, max_queue_pairs);
2552
2553         return 0;
2554
2555 free_unregister_netdev:
2556         vi->vdev->config->reset(vdev);
2557
2558         unregister_netdev(dev);
2559 free_vqs:
2560         cancel_delayed_work_sync(&vi->refill);
2561         free_receive_page_frags(vi);
2562         virtnet_del_vqs(vi);
2563 free_stats:
2564         free_percpu(vi->stats);
2565 free:
2566         free_netdev(dev);
2567         return err;
2568 }
2569
2570 static void _remove_vq_common(struct virtnet_info *vi)
2571 {
2572         vi->vdev->config->reset(vi->vdev);
2573         free_unused_bufs(vi);
2574         _free_receive_bufs(vi);
2575         free_receive_page_frags(vi);
2576         virtnet_del_vqs(vi);
2577 }
2578
2579 static void remove_vq_common(struct virtnet_info *vi)
2580 {
2581         vi->vdev->config->reset(vi->vdev);
2582
2583         /* Free unused buffers in both send and recv, if any. */
2584         free_unused_bufs(vi);
2585
2586         free_receive_bufs(vi);
2587
2588         free_receive_page_frags(vi);
2589
2590         virtnet_del_vqs(vi);
2591 }
2592
2593 static void virtnet_remove(struct virtio_device *vdev)
2594 {
2595         struct virtnet_info *vi = vdev->priv;
2596
2597         virtnet_cpu_notif_remove(vi);
2598
2599         /* Make sure no work handler is accessing the device. */
2600         flush_work(&vi->config_work);
2601
2602         unregister_netdev(vi->dev);
2603
2604         remove_vq_common(vi);
2605
2606         free_percpu(vi->stats);
2607         free_netdev(vi->dev);
2608 }
2609
2610 #ifdef CONFIG_PM_SLEEP
2611 static int virtnet_freeze(struct virtio_device *vdev)
2612 {
2613         struct virtnet_info *vi = vdev->priv;
2614
2615         virtnet_cpu_notif_remove(vi);
2616         virtnet_freeze_down(vdev);
2617         remove_vq_common(vi);
2618
2619         return 0;
2620 }
2621
2622 static int virtnet_restore(struct virtio_device *vdev)
2623 {
2624         struct virtnet_info *vi = vdev->priv;
2625         int err;
2626
2627         err = virtnet_restore_up(vdev);
2628         if (err)
2629                 return err;
2630         virtnet_set_queues(vi, vi->curr_queue_pairs);
2631
2632         err = virtnet_cpu_notif_add(vi);
2633         if (err)
2634                 return err;
2635
2636         return 0;
2637 }
2638 #endif
2639
2640 static struct virtio_device_id id_table[] = {
2641         { VIRTIO_ID_NET, VIRTIO_DEV_ANY_ID },
2642         { 0 },
2643 };
2644
2645 #define VIRTNET_FEATURES \
2646         VIRTIO_NET_F_CSUM, VIRTIO_NET_F_GUEST_CSUM, \
2647         VIRTIO_NET_F_MAC, \
2648         VIRTIO_NET_F_HOST_TSO4, VIRTIO_NET_F_HOST_UFO, VIRTIO_NET_F_HOST_TSO6, \
2649         VIRTIO_NET_F_HOST_ECN, VIRTIO_NET_F_GUEST_TSO4, VIRTIO_NET_F_GUEST_TSO6, \
2650         VIRTIO_NET_F_GUEST_ECN, VIRTIO_NET_F_GUEST_UFO, \
2651         VIRTIO_NET_F_MRG_RXBUF, VIRTIO_NET_F_STATUS, VIRTIO_NET_F_CTRL_VQ, \
2652         VIRTIO_NET_F_CTRL_RX, VIRTIO_NET_F_CTRL_VLAN, \
2653         VIRTIO_NET_F_GUEST_ANNOUNCE, VIRTIO_NET_F_MQ, \
2654         VIRTIO_NET_F_CTRL_MAC_ADDR, \
2655         VIRTIO_NET_F_MTU
2656
2657 static unsigned int features[] = {
2658         VIRTNET_FEATURES,
2659 };
2660
2661 static unsigned int features_legacy[] = {
2662         VIRTNET_FEATURES,
2663         VIRTIO_NET_F_GSO,
2664         VIRTIO_F_ANY_LAYOUT,
2665 };
2666
2667 static struct virtio_driver virtio_net_driver = {
2668         .feature_table = features,
2669         .feature_table_size = ARRAY_SIZE(features),
2670         .feature_table_legacy = features_legacy,
2671         .feature_table_size_legacy = ARRAY_SIZE(features_legacy),
2672         .driver.name =  KBUILD_MODNAME,
2673         .driver.owner = THIS_MODULE,
2674         .id_table =     id_table,
2675         .validate =     virtnet_validate,
2676         .probe =        virtnet_probe,
2677         .remove =       virtnet_remove,
2678         .config_changed = virtnet_config_changed,
2679 #ifdef CONFIG_PM_SLEEP
2680         .freeze =       virtnet_freeze,
2681         .restore =      virtnet_restore,
2682 #endif
2683 };
2684
2685 static __init int virtio_net_driver_init(void)
2686 {
2687         int ret;
2688
2689         ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "virtio/net:online",
2690                                       virtnet_cpu_online,
2691                                       virtnet_cpu_down_prep);
2692         if (ret < 0)
2693                 goto out;
2694         virtionet_online = ret;
2695         ret = cpuhp_setup_state_multi(CPUHP_VIRT_NET_DEAD, "virtio/net:dead",
2696                                       NULL, virtnet_cpu_dead);
2697         if (ret)
2698                 goto err_dead;
2699
2700         ret = register_virtio_driver(&virtio_net_driver);
2701         if (ret)
2702                 goto err_virtio;
2703         return 0;
2704 err_virtio:
2705         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
2706 err_dead:
2707         cpuhp_remove_multi_state(virtionet_online);
2708 out:
2709         return ret;
2710 }
2711 module_init(virtio_net_driver_init);
2712
2713 static __exit void virtio_net_driver_exit(void)
2714 {
2715         cpuhp_remove_multi_state(CPUHP_VIRT_NET_DEAD);
2716         cpuhp_remove_multi_state(virtionet_online);
2717         unregister_virtio_driver(&virtio_net_driver);
2718 }
2719 module_exit(virtio_net_driver_exit);
2720
2721 MODULE_DEVICE_TABLE(virtio, id_table);
2722 MODULE_DESCRIPTION("Virtio network driver");
2723 MODULE_LICENSE("GPL");