Merge branch 'acpi-ec'
[linux-drm-fsl-dcu.git] / drivers / vhost / net.c
1 /* Copyright (C) 2009 Red Hat, Inc.
2  * Author: Michael S. Tsirkin <mst@redhat.com>
3  *
4  * This work is licensed under the terms of the GNU GPL, version 2.
5  *
6  * virtio-net server in host kernel.
7  */
8
9 #include <linux/compat.h>
10 #include <linux/eventfd.h>
11 #include <linux/vhost.h>
12 #include <linux/virtio_net.h>
13 #include <linux/miscdevice.h>
14 #include <linux/module.h>
15 #include <linux/moduleparam.h>
16 #include <linux/mutex.h>
17 #include <linux/workqueue.h>
18 #include <linux/file.h>
19 #include <linux/slab.h>
20 #include <linux/vmalloc.h>
21
22 #include <linux/net.h>
23 #include <linux/if_packet.h>
24 #include <linux/if_arp.h>
25 #include <linux/if_tun.h>
26 #include <linux/if_macvlan.h>
27 #include <linux/if_vlan.h>
28
29 #include <net/sock.h>
30
31 #include "vhost.h"
32
33 static int experimental_zcopytx = 1;
34 module_param(experimental_zcopytx, int, 0444);
35 MODULE_PARM_DESC(experimental_zcopytx, "Enable Zero Copy TX;"
36                                        " 1 -Enable; 0 - Disable");
37
38 /* Max number of bytes transferred before requeueing the job.
39  * Using this limit prevents one virtqueue from starving others. */
40 #define VHOST_NET_WEIGHT 0x80000
41
42 /* MAX number of TX used buffers for outstanding zerocopy */
43 #define VHOST_MAX_PEND 128
44 #define VHOST_GOODCOPY_LEN 256
45
46 /*
47  * For transmit, used buffer len is unused; we override it to track buffer
48  * status internally; used for zerocopy tx only.
49  */
50 /* Lower device DMA failed */
51 #define VHOST_DMA_FAILED_LEN    ((__force __virtio32)3)
52 /* Lower device DMA done */
53 #define VHOST_DMA_DONE_LEN      ((__force __virtio32)2)
54 /* Lower device DMA in progress */
55 #define VHOST_DMA_IN_PROGRESS   ((__force __virtio32)1)
56 /* Buffer unused */
57 #define VHOST_DMA_CLEAR_LEN     ((__force __virtio32)0)
58
59 #define VHOST_DMA_IS_DONE(len) ((__force u32)(len) >= (__force u32)VHOST_DMA_DONE_LEN)
60
61 enum {
62         VHOST_NET_FEATURES = VHOST_FEATURES |
63                          (1ULL << VHOST_NET_F_VIRTIO_NET_HDR) |
64                          (1ULL << VIRTIO_NET_F_MRG_RXBUF) |
65                          (1ULL << VIRTIO_F_VERSION_1),
66 };
67
68 enum {
69         VHOST_NET_VQ_RX = 0,
70         VHOST_NET_VQ_TX = 1,
71         VHOST_NET_VQ_MAX = 2,
72 };
73
74 struct vhost_net_ubuf_ref {
75         /* refcount follows semantics similar to kref:
76          *  0: object is released
77          *  1: no outstanding ubufs
78          * >1: outstanding ubufs
79          */
80         atomic_t refcount;
81         wait_queue_head_t wait;
82         struct vhost_virtqueue *vq;
83 };
84
85 struct vhost_net_virtqueue {
86         struct vhost_virtqueue vq;
87         /* hdr is used to store the virtio header.
88          * Since each iovec has >= 1 byte length, we never need more than
89          * header length entries to store the header. */
90         struct iovec hdr[sizeof(struct virtio_net_hdr_mrg_rxbuf)];
91         size_t vhost_hlen;
92         size_t sock_hlen;
93         /* vhost zerocopy support fields below: */
94         /* last used idx for outstanding DMA zerocopy buffers */
95         int upend_idx;
96         /* first used idx for DMA done zerocopy buffers */
97         int done_idx;
98         /* an array of userspace buffers info */
99         struct ubuf_info *ubuf_info;
100         /* Reference counting for outstanding ubufs.
101          * Protected by vq mutex. Writers must also take device mutex. */
102         struct vhost_net_ubuf_ref *ubufs;
103 };
104
105 struct vhost_net {
106         struct vhost_dev dev;
107         struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX];
108         struct vhost_poll poll[VHOST_NET_VQ_MAX];
109         /* Number of TX recently submitted.
110          * Protected by tx vq lock. */
111         unsigned tx_packets;
112         /* Number of times zerocopy TX recently failed.
113          * Protected by tx vq lock. */
114         unsigned tx_zcopy_err;
115         /* Flush in progress. Protected by tx vq lock. */
116         bool tx_flush;
117 };
118
119 static unsigned vhost_net_zcopy_mask __read_mostly;
120
121 static void vhost_net_enable_zcopy(int vq)
122 {
123         vhost_net_zcopy_mask |= 0x1 << vq;
124 }
125
126 static struct vhost_net_ubuf_ref *
127 vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy)
128 {
129         struct vhost_net_ubuf_ref *ubufs;
130         /* No zero copy backend? Nothing to count. */
131         if (!zcopy)
132                 return NULL;
133         ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL);
134         if (!ubufs)
135                 return ERR_PTR(-ENOMEM);
136         atomic_set(&ubufs->refcount, 1);
137         init_waitqueue_head(&ubufs->wait);
138         ubufs->vq = vq;
139         return ubufs;
140 }
141
142 static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs)
143 {
144         int r = atomic_sub_return(1, &ubufs->refcount);
145         if (unlikely(!r))
146                 wake_up(&ubufs->wait);
147         return r;
148 }
149
150 static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs)
151 {
152         vhost_net_ubuf_put(ubufs);
153         wait_event(ubufs->wait, !atomic_read(&ubufs->refcount));
154 }
155
156 static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs)
157 {
158         vhost_net_ubuf_put_and_wait(ubufs);
159         kfree(ubufs);
160 }
161
162 static void vhost_net_clear_ubuf_info(struct vhost_net *n)
163 {
164         int i;
165
166         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
167                 kfree(n->vqs[i].ubuf_info);
168                 n->vqs[i].ubuf_info = NULL;
169         }
170 }
171
172 static int vhost_net_set_ubuf_info(struct vhost_net *n)
173 {
174         bool zcopy;
175         int i;
176
177         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
178                 zcopy = vhost_net_zcopy_mask & (0x1 << i);
179                 if (!zcopy)
180                         continue;
181                 n->vqs[i].ubuf_info = kmalloc(sizeof(*n->vqs[i].ubuf_info) *
182                                               UIO_MAXIOV, GFP_KERNEL);
183                 if  (!n->vqs[i].ubuf_info)
184                         goto err;
185         }
186         return 0;
187
188 err:
189         vhost_net_clear_ubuf_info(n);
190         return -ENOMEM;
191 }
192
193 static void vhost_net_vq_reset(struct vhost_net *n)
194 {
195         int i;
196
197         vhost_net_clear_ubuf_info(n);
198
199         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
200                 n->vqs[i].done_idx = 0;
201                 n->vqs[i].upend_idx = 0;
202                 n->vqs[i].ubufs = NULL;
203                 n->vqs[i].vhost_hlen = 0;
204                 n->vqs[i].sock_hlen = 0;
205         }
206
207 }
208
209 static void vhost_net_tx_packet(struct vhost_net *net)
210 {
211         ++net->tx_packets;
212         if (net->tx_packets < 1024)
213                 return;
214         net->tx_packets = 0;
215         net->tx_zcopy_err = 0;
216 }
217
218 static void vhost_net_tx_err(struct vhost_net *net)
219 {
220         ++net->tx_zcopy_err;
221 }
222
223 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
224 {
225         /* TX flush waits for outstanding DMAs to be done.
226          * Don't start new DMAs.
227          */
228         return !net->tx_flush &&
229                 net->tx_packets / 64 >= net->tx_zcopy_err;
230 }
231
232 static bool vhost_sock_zcopy(struct socket *sock)
233 {
234         return unlikely(experimental_zcopytx) &&
235                 sock_flag(sock->sk, SOCK_ZEROCOPY);
236 }
237
238 /* Pop first len bytes from iovec. Return number of segments used. */
239 static int move_iovec_hdr(struct iovec *from, struct iovec *to,
240                           size_t len, int iov_count)
241 {
242         int seg = 0;
243         size_t size;
244
245         while (len && seg < iov_count) {
246                 size = min(from->iov_len, len);
247                 to->iov_base = from->iov_base;
248                 to->iov_len = size;
249                 from->iov_len -= size;
250                 from->iov_base += size;
251                 len -= size;
252                 ++from;
253                 ++to;
254                 ++seg;
255         }
256         return seg;
257 }
258 /* Copy iovec entries for len bytes from iovec. */
259 static void copy_iovec_hdr(const struct iovec *from, struct iovec *to,
260                            size_t len, int iovcount)
261 {
262         int seg = 0;
263         size_t size;
264
265         while (len && seg < iovcount) {
266                 size = min(from->iov_len, len);
267                 to->iov_base = from->iov_base;
268                 to->iov_len = size;
269                 len -= size;
270                 ++from;
271                 ++to;
272                 ++seg;
273         }
274 }
275
276 /* In case of DMA done not in order in lower device driver for some reason.
277  * upend_idx is used to track end of used idx, done_idx is used to track head
278  * of used idx. Once lower device DMA done contiguously, we will signal KVM
279  * guest used idx.
280  */
281 static void vhost_zerocopy_signal_used(struct vhost_net *net,
282                                        struct vhost_virtqueue *vq)
283 {
284         struct vhost_net_virtqueue *nvq =
285                 container_of(vq, struct vhost_net_virtqueue, vq);
286         int i, add;
287         int j = 0;
288
289         for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
290                 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
291                         vhost_net_tx_err(net);
292                 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
293                         vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
294                         ++j;
295                 } else
296                         break;
297         }
298         while (j) {
299                 add = min(UIO_MAXIOV - nvq->done_idx, j);
300                 vhost_add_used_and_signal_n(vq->dev, vq,
301                                             &vq->heads[nvq->done_idx], add);
302                 nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV;
303                 j -= add;
304         }
305 }
306
307 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
308 {
309         struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
310         struct vhost_virtqueue *vq = ubufs->vq;
311         int cnt;
312
313         rcu_read_lock_bh();
314
315         /* set len to mark this desc buffers done DMA */
316         vq->heads[ubuf->desc].len = success ?
317                 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
318         cnt = vhost_net_ubuf_put(ubufs);
319
320         /*
321          * Trigger polling thread if guest stopped submitting new buffers:
322          * in this case, the refcount after decrement will eventually reach 1.
323          * We also trigger polling periodically after each 16 packets
324          * (the value 16 here is more or less arbitrary, it's tuned to trigger
325          * less than 10% of times).
326          */
327         if (cnt <= 1 || !(cnt % 16))
328                 vhost_poll_queue(&vq->poll);
329
330         rcu_read_unlock_bh();
331 }
332
333 /* Expects to be always run from workqueue - which acts as
334  * read-size critical section for our kind of RCU. */
335 static void handle_tx(struct vhost_net *net)
336 {
337         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
338         struct vhost_virtqueue *vq = &nvq->vq;
339         unsigned out, in, s;
340         int head;
341         struct msghdr msg = {
342                 .msg_name = NULL,
343                 .msg_namelen = 0,
344                 .msg_control = NULL,
345                 .msg_controllen = 0,
346                 .msg_flags = MSG_DONTWAIT,
347         };
348         size_t len, total_len = 0;
349         int err;
350         size_t hdr_size;
351         struct socket *sock;
352         struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
353         bool zcopy, zcopy_used;
354
355         mutex_lock(&vq->mutex);
356         sock = vq->private_data;
357         if (!sock)
358                 goto out;
359
360         vhost_disable_notify(&net->dev, vq);
361
362         hdr_size = nvq->vhost_hlen;
363         zcopy = nvq->ubufs;
364
365         for (;;) {
366                 /* Release DMAs done buffers first */
367                 if (zcopy)
368                         vhost_zerocopy_signal_used(net, vq);
369
370                 /* If more outstanding DMAs, queue the work.
371                  * Handle upend_idx wrap around
372                  */
373                 if (unlikely((nvq->upend_idx + vq->num - VHOST_MAX_PEND)
374                               % UIO_MAXIOV == nvq->done_idx))
375                         break;
376
377                 head = vhost_get_vq_desc(vq, vq->iov,
378                                          ARRAY_SIZE(vq->iov),
379                                          &out, &in,
380                                          NULL, NULL);
381                 /* On error, stop handling until the next kick. */
382                 if (unlikely(head < 0))
383                         break;
384                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
385                 if (head == vq->num) {
386                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
387                                 vhost_disable_notify(&net->dev, vq);
388                                 continue;
389                         }
390                         break;
391                 }
392                 if (in) {
393                         vq_err(vq, "Unexpected descriptor format for TX: "
394                                "out %d, int %d\n", out, in);
395                         break;
396                 }
397                 /* Skip header. TODO: support TSO. */
398                 s = move_iovec_hdr(vq->iov, nvq->hdr, hdr_size, out);
399                 len = iov_length(vq->iov, out);
400                 iov_iter_init(&msg.msg_iter, WRITE, vq->iov, out, len);
401                 /* Sanity check */
402                 if (!len) {
403                         vq_err(vq, "Unexpected header len for TX: "
404                                "%zd expected %zd\n",
405                                iov_length(nvq->hdr, s), hdr_size);
406                         break;
407                 }
408
409                 zcopy_used = zcopy && len >= VHOST_GOODCOPY_LEN
410                                    && (nvq->upend_idx + 1) % UIO_MAXIOV !=
411                                       nvq->done_idx
412                                    && vhost_net_tx_select_zcopy(net);
413
414                 /* use msg_control to pass vhost zerocopy ubuf info to skb */
415                 if (zcopy_used) {
416                         struct ubuf_info *ubuf;
417                         ubuf = nvq->ubuf_info + nvq->upend_idx;
418
419                         vq->heads[nvq->upend_idx].id = cpu_to_vhost32(vq, head);
420                         vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS;
421                         ubuf->callback = vhost_zerocopy_callback;
422                         ubuf->ctx = nvq->ubufs;
423                         ubuf->desc = nvq->upend_idx;
424                         msg.msg_control = ubuf;
425                         msg.msg_controllen = sizeof(ubuf);
426                         ubufs = nvq->ubufs;
427                         atomic_inc(&ubufs->refcount);
428                         nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
429                 } else {
430                         msg.msg_control = NULL;
431                         ubufs = NULL;
432                 }
433                 /* TODO: Check specific error and bomb out unless ENOBUFS? */
434                 err = sock->ops->sendmsg(NULL, sock, &msg, len);
435                 if (unlikely(err < 0)) {
436                         if (zcopy_used) {
437                                 vhost_net_ubuf_put(ubufs);
438                                 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
439                                         % UIO_MAXIOV;
440                         }
441                         vhost_discard_vq_desc(vq, 1);
442                         break;
443                 }
444                 if (err != len)
445                         pr_debug("Truncated TX packet: "
446                                  " len %d != %zd\n", err, len);
447                 if (!zcopy_used)
448                         vhost_add_used_and_signal(&net->dev, vq, head, 0);
449                 else
450                         vhost_zerocopy_signal_used(net, vq);
451                 total_len += len;
452                 vhost_net_tx_packet(net);
453                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
454                         vhost_poll_queue(&vq->poll);
455                         break;
456                 }
457         }
458 out:
459         mutex_unlock(&vq->mutex);
460 }
461
462 static int peek_head_len(struct sock *sk)
463 {
464         struct sk_buff *head;
465         int len = 0;
466         unsigned long flags;
467
468         spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
469         head = skb_peek(&sk->sk_receive_queue);
470         if (likely(head)) {
471                 len = head->len;
472                 if (vlan_tx_tag_present(head))
473                         len += VLAN_HLEN;
474         }
475
476         spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
477         return len;
478 }
479
480 /* This is a multi-buffer version of vhost_get_desc, that works if
481  *      vq has read descriptors only.
482  * @vq          - the relevant virtqueue
483  * @datalen     - data length we'll be reading
484  * @iovcount    - returned count of io vectors we fill
485  * @log         - vhost log
486  * @log_num     - log offset
487  * @quota       - headcount quota, 1 for big buffer
488  *      returns number of buffer heads allocated, negative on error
489  */
490 static int get_rx_bufs(struct vhost_virtqueue *vq,
491                        struct vring_used_elem *heads,
492                        int datalen,
493                        unsigned *iovcount,
494                        struct vhost_log *log,
495                        unsigned *log_num,
496                        unsigned int quota)
497 {
498         unsigned int out, in;
499         int seg = 0;
500         int headcount = 0;
501         unsigned d;
502         int r, nlogs = 0;
503         /* len is always initialized before use since we are always called with
504          * datalen > 0.
505          */
506         u32 uninitialized_var(len);
507
508         while (datalen > 0 && headcount < quota) {
509                 if (unlikely(seg >= UIO_MAXIOV)) {
510                         r = -ENOBUFS;
511                         goto err;
512                 }
513                 r = vhost_get_vq_desc(vq, vq->iov + seg,
514                                       ARRAY_SIZE(vq->iov) - seg, &out,
515                                       &in, log, log_num);
516                 if (unlikely(r < 0))
517                         goto err;
518
519                 d = r;
520                 if (d == vq->num) {
521                         r = 0;
522                         goto err;
523                 }
524                 if (unlikely(out || in <= 0)) {
525                         vq_err(vq, "unexpected descriptor format for RX: "
526                                 "out %d, in %d\n", out, in);
527                         r = -EINVAL;
528                         goto err;
529                 }
530                 if (unlikely(log)) {
531                         nlogs += *log_num;
532                         log += *log_num;
533                 }
534                 heads[headcount].id = cpu_to_vhost32(vq, d);
535                 len = iov_length(vq->iov + seg, in);
536                 heads[headcount].len = cpu_to_vhost32(vq, len);
537                 datalen -= len;
538                 ++headcount;
539                 seg += in;
540         }
541         heads[headcount - 1].len = cpu_to_vhost32(vq, len + datalen);
542         *iovcount = seg;
543         if (unlikely(log))
544                 *log_num = nlogs;
545
546         /* Detect overrun */
547         if (unlikely(datalen > 0)) {
548                 r = UIO_MAXIOV + 1;
549                 goto err;
550         }
551         return headcount;
552 err:
553         vhost_discard_vq_desc(vq, headcount);
554         return r;
555 }
556
557 /* Expects to be always run from workqueue - which acts as
558  * read-size critical section for our kind of RCU. */
559 static void handle_rx(struct vhost_net *net)
560 {
561         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
562         struct vhost_virtqueue *vq = &nvq->vq;
563         unsigned uninitialized_var(in), log;
564         struct vhost_log *vq_log;
565         struct msghdr msg = {
566                 .msg_name = NULL,
567                 .msg_namelen = 0,
568                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
569                 .msg_controllen = 0,
570                 .msg_flags = MSG_DONTWAIT,
571         };
572         struct virtio_net_hdr_mrg_rxbuf hdr = {
573                 .hdr.flags = 0,
574                 .hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
575         };
576         size_t total_len = 0;
577         int err, mergeable;
578         s16 headcount;
579         size_t vhost_hlen, sock_hlen;
580         size_t vhost_len, sock_len;
581         struct socket *sock;
582
583         mutex_lock(&vq->mutex);
584         sock = vq->private_data;
585         if (!sock)
586                 goto out;
587         vhost_disable_notify(&net->dev, vq);
588
589         vhost_hlen = nvq->vhost_hlen;
590         sock_hlen = nvq->sock_hlen;
591
592         vq_log = unlikely(vhost_has_feature(vq, VHOST_F_LOG_ALL)) ?
593                 vq->log : NULL;
594         mergeable = vhost_has_feature(vq, VIRTIO_NET_F_MRG_RXBUF);
595
596         while ((sock_len = peek_head_len(sock->sk))) {
597                 sock_len += sock_hlen;
598                 vhost_len = sock_len + vhost_hlen;
599                 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
600                                         &in, vq_log, &log,
601                                         likely(mergeable) ? UIO_MAXIOV : 1);
602                 /* On error, stop handling until the next kick. */
603                 if (unlikely(headcount < 0))
604                         break;
605                 /* On overrun, truncate and discard */
606                 if (unlikely(headcount > UIO_MAXIOV)) {
607                         iov_iter_init(&msg.msg_iter, READ, vq->iov, 1, 1);
608                         err = sock->ops->recvmsg(NULL, sock, &msg,
609                                                  1, MSG_DONTWAIT | MSG_TRUNC);
610                         pr_debug("Discarded rx packet: len %zd\n", sock_len);
611                         continue;
612                 }
613                 /* OK, now we need to know about added descriptors. */
614                 if (!headcount) {
615                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
616                                 /* They have slipped one in as we were
617                                  * doing that: check again. */
618                                 vhost_disable_notify(&net->dev, vq);
619                                 continue;
620                         }
621                         /* Nothing new?  Wait for eventfd to tell us
622                          * they refilled. */
623                         break;
624                 }
625                 /* We don't need to be notified again. */
626                 if (unlikely((vhost_hlen)))
627                         /* Skip header. TODO: support TSO. */
628                         move_iovec_hdr(vq->iov, nvq->hdr, vhost_hlen, in);
629                 else
630                         /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
631                          * needed because recvmsg can modify msg_iov. */
632                         copy_iovec_hdr(vq->iov, nvq->hdr, sock_hlen, in);
633                 iov_iter_init(&msg.msg_iter, READ, vq->iov, in, sock_len);
634                 err = sock->ops->recvmsg(NULL, sock, &msg,
635                                          sock_len, MSG_DONTWAIT | MSG_TRUNC);
636                 /* Userspace might have consumed the packet meanwhile:
637                  * it's not supposed to do this usually, but might be hard
638                  * to prevent. Discard data we got (if any) and keep going. */
639                 if (unlikely(err != sock_len)) {
640                         pr_debug("Discarded rx packet: "
641                                  " len %d, expected %zd\n", err, sock_len);
642                         vhost_discard_vq_desc(vq, headcount);
643                         continue;
644                 }
645                 if (unlikely(vhost_hlen) &&
646                     memcpy_toiovecend(nvq->hdr, (unsigned char *)&hdr, 0,
647                                       vhost_hlen)) {
648                         vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
649                                vq->iov->iov_base);
650                         break;
651                 }
652                 /* TODO: Should check and handle checksum. */
653
654                 hdr.num_buffers = cpu_to_vhost16(vq, headcount);
655                 if (likely(mergeable) &&
656                     memcpy_toiovecend(nvq->hdr, (void *)&hdr.num_buffers,
657                                       offsetof(typeof(hdr), num_buffers),
658                                       sizeof hdr.num_buffers)) {
659                         vq_err(vq, "Failed num_buffers write");
660                         vhost_discard_vq_desc(vq, headcount);
661                         break;
662                 }
663                 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
664                                             headcount);
665                 if (unlikely(vq_log))
666                         vhost_log_write(vq, vq_log, log, vhost_len);
667                 total_len += vhost_len;
668                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
669                         vhost_poll_queue(&vq->poll);
670                         break;
671                 }
672         }
673 out:
674         mutex_unlock(&vq->mutex);
675 }
676
677 static void handle_tx_kick(struct vhost_work *work)
678 {
679         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
680                                                   poll.work);
681         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
682
683         handle_tx(net);
684 }
685
686 static void handle_rx_kick(struct vhost_work *work)
687 {
688         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
689                                                   poll.work);
690         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
691
692         handle_rx(net);
693 }
694
695 static void handle_tx_net(struct vhost_work *work)
696 {
697         struct vhost_net *net = container_of(work, struct vhost_net,
698                                              poll[VHOST_NET_VQ_TX].work);
699         handle_tx(net);
700 }
701
702 static void handle_rx_net(struct vhost_work *work)
703 {
704         struct vhost_net *net = container_of(work, struct vhost_net,
705                                              poll[VHOST_NET_VQ_RX].work);
706         handle_rx(net);
707 }
708
709 static int vhost_net_open(struct inode *inode, struct file *f)
710 {
711         struct vhost_net *n;
712         struct vhost_dev *dev;
713         struct vhost_virtqueue **vqs;
714         int i;
715
716         n = kmalloc(sizeof *n, GFP_KERNEL | __GFP_NOWARN | __GFP_REPEAT);
717         if (!n) {
718                 n = vmalloc(sizeof *n);
719                 if (!n)
720                         return -ENOMEM;
721         }
722         vqs = kmalloc(VHOST_NET_VQ_MAX * sizeof(*vqs), GFP_KERNEL);
723         if (!vqs) {
724                 kvfree(n);
725                 return -ENOMEM;
726         }
727
728         dev = &n->dev;
729         vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
730         vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
731         n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
732         n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
733         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
734                 n->vqs[i].ubufs = NULL;
735                 n->vqs[i].ubuf_info = NULL;
736                 n->vqs[i].upend_idx = 0;
737                 n->vqs[i].done_idx = 0;
738                 n->vqs[i].vhost_hlen = 0;
739                 n->vqs[i].sock_hlen = 0;
740         }
741         vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX);
742
743         vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
744         vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
745
746         f->private_data = n;
747
748         return 0;
749 }
750
751 static void vhost_net_disable_vq(struct vhost_net *n,
752                                  struct vhost_virtqueue *vq)
753 {
754         struct vhost_net_virtqueue *nvq =
755                 container_of(vq, struct vhost_net_virtqueue, vq);
756         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
757         if (!vq->private_data)
758                 return;
759         vhost_poll_stop(poll);
760 }
761
762 static int vhost_net_enable_vq(struct vhost_net *n,
763                                 struct vhost_virtqueue *vq)
764 {
765         struct vhost_net_virtqueue *nvq =
766                 container_of(vq, struct vhost_net_virtqueue, vq);
767         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
768         struct socket *sock;
769
770         sock = vq->private_data;
771         if (!sock)
772                 return 0;
773
774         return vhost_poll_start(poll, sock->file);
775 }
776
777 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
778                                         struct vhost_virtqueue *vq)
779 {
780         struct socket *sock;
781
782         mutex_lock(&vq->mutex);
783         sock = vq->private_data;
784         vhost_net_disable_vq(n, vq);
785         vq->private_data = NULL;
786         mutex_unlock(&vq->mutex);
787         return sock;
788 }
789
790 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
791                            struct socket **rx_sock)
792 {
793         *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
794         *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
795 }
796
797 static void vhost_net_flush_vq(struct vhost_net *n, int index)
798 {
799         vhost_poll_flush(n->poll + index);
800         vhost_poll_flush(&n->vqs[index].vq.poll);
801 }
802
803 static void vhost_net_flush(struct vhost_net *n)
804 {
805         vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
806         vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
807         if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
808                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
809                 n->tx_flush = true;
810                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
811                 /* Wait for all lower device DMAs done. */
812                 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
813                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
814                 n->tx_flush = false;
815                 atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1);
816                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
817         }
818 }
819
820 static int vhost_net_release(struct inode *inode, struct file *f)
821 {
822         struct vhost_net *n = f->private_data;
823         struct socket *tx_sock;
824         struct socket *rx_sock;
825
826         vhost_net_stop(n, &tx_sock, &rx_sock);
827         vhost_net_flush(n);
828         vhost_dev_stop(&n->dev);
829         vhost_dev_cleanup(&n->dev, false);
830         vhost_net_vq_reset(n);
831         if (tx_sock)
832                 sockfd_put(tx_sock);
833         if (rx_sock)
834                 sockfd_put(rx_sock);
835         /* Make sure no callbacks are outstanding */
836         synchronize_rcu_bh();
837         /* We do an extra flush before freeing memory,
838          * since jobs can re-queue themselves. */
839         vhost_net_flush(n);
840         kfree(n->dev.vqs);
841         kvfree(n);
842         return 0;
843 }
844
845 static struct socket *get_raw_socket(int fd)
846 {
847         struct {
848                 struct sockaddr_ll sa;
849                 char  buf[MAX_ADDR_LEN];
850         } uaddr;
851         int uaddr_len = sizeof uaddr, r;
852         struct socket *sock = sockfd_lookup(fd, &r);
853
854         if (!sock)
855                 return ERR_PTR(-ENOTSOCK);
856
857         /* Parameter checking */
858         if (sock->sk->sk_type != SOCK_RAW) {
859                 r = -ESOCKTNOSUPPORT;
860                 goto err;
861         }
862
863         r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
864                                &uaddr_len, 0);
865         if (r)
866                 goto err;
867
868         if (uaddr.sa.sll_family != AF_PACKET) {
869                 r = -EPFNOSUPPORT;
870                 goto err;
871         }
872         return sock;
873 err:
874         sockfd_put(sock);
875         return ERR_PTR(r);
876 }
877
878 static struct socket *get_tap_socket(int fd)
879 {
880         struct file *file = fget(fd);
881         struct socket *sock;
882
883         if (!file)
884                 return ERR_PTR(-EBADF);
885         sock = tun_get_socket(file);
886         if (!IS_ERR(sock))
887                 return sock;
888         sock = macvtap_get_socket(file);
889         if (IS_ERR(sock))
890                 fput(file);
891         return sock;
892 }
893
894 static struct socket *get_socket(int fd)
895 {
896         struct socket *sock;
897
898         /* special case to disable backend */
899         if (fd == -1)
900                 return NULL;
901         sock = get_raw_socket(fd);
902         if (!IS_ERR(sock))
903                 return sock;
904         sock = get_tap_socket(fd);
905         if (!IS_ERR(sock))
906                 return sock;
907         return ERR_PTR(-ENOTSOCK);
908 }
909
910 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
911 {
912         struct socket *sock, *oldsock;
913         struct vhost_virtqueue *vq;
914         struct vhost_net_virtqueue *nvq;
915         struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
916         int r;
917
918         mutex_lock(&n->dev.mutex);
919         r = vhost_dev_check_owner(&n->dev);
920         if (r)
921                 goto err;
922
923         if (index >= VHOST_NET_VQ_MAX) {
924                 r = -ENOBUFS;
925                 goto err;
926         }
927         vq = &n->vqs[index].vq;
928         nvq = &n->vqs[index];
929         mutex_lock(&vq->mutex);
930
931         /* Verify that ring has been setup correctly. */
932         if (!vhost_vq_access_ok(vq)) {
933                 r = -EFAULT;
934                 goto err_vq;
935         }
936         sock = get_socket(fd);
937         if (IS_ERR(sock)) {
938                 r = PTR_ERR(sock);
939                 goto err_vq;
940         }
941
942         /* start polling new socket */
943         oldsock = vq->private_data;
944         if (sock != oldsock) {
945                 ubufs = vhost_net_ubuf_alloc(vq,
946                                              sock && vhost_sock_zcopy(sock));
947                 if (IS_ERR(ubufs)) {
948                         r = PTR_ERR(ubufs);
949                         goto err_ubufs;
950                 }
951
952                 vhost_net_disable_vq(n, vq);
953                 vq->private_data = sock;
954                 r = vhost_init_used(vq);
955                 if (r)
956                         goto err_used;
957                 r = vhost_net_enable_vq(n, vq);
958                 if (r)
959                         goto err_used;
960
961                 oldubufs = nvq->ubufs;
962                 nvq->ubufs = ubufs;
963
964                 n->tx_packets = 0;
965                 n->tx_zcopy_err = 0;
966                 n->tx_flush = false;
967         }
968
969         mutex_unlock(&vq->mutex);
970
971         if (oldubufs) {
972                 vhost_net_ubuf_put_wait_and_free(oldubufs);
973                 mutex_lock(&vq->mutex);
974                 vhost_zerocopy_signal_used(n, vq);
975                 mutex_unlock(&vq->mutex);
976         }
977
978         if (oldsock) {
979                 vhost_net_flush_vq(n, index);
980                 sockfd_put(oldsock);
981         }
982
983         mutex_unlock(&n->dev.mutex);
984         return 0;
985
986 err_used:
987         vq->private_data = oldsock;
988         vhost_net_enable_vq(n, vq);
989         if (ubufs)
990                 vhost_net_ubuf_put_wait_and_free(ubufs);
991 err_ubufs:
992         sockfd_put(sock);
993 err_vq:
994         mutex_unlock(&vq->mutex);
995 err:
996         mutex_unlock(&n->dev.mutex);
997         return r;
998 }
999
1000 static long vhost_net_reset_owner(struct vhost_net *n)
1001 {
1002         struct socket *tx_sock = NULL;
1003         struct socket *rx_sock = NULL;
1004         long err;
1005         struct vhost_memory *memory;
1006
1007         mutex_lock(&n->dev.mutex);
1008         err = vhost_dev_check_owner(&n->dev);
1009         if (err)
1010                 goto done;
1011         memory = vhost_dev_reset_owner_prepare();
1012         if (!memory) {
1013                 err = -ENOMEM;
1014                 goto done;
1015         }
1016         vhost_net_stop(n, &tx_sock, &rx_sock);
1017         vhost_net_flush(n);
1018         vhost_dev_reset_owner(&n->dev, memory);
1019         vhost_net_vq_reset(n);
1020 done:
1021         mutex_unlock(&n->dev.mutex);
1022         if (tx_sock)
1023                 sockfd_put(tx_sock);
1024         if (rx_sock)
1025                 sockfd_put(rx_sock);
1026         return err;
1027 }
1028
1029 static int vhost_net_set_features(struct vhost_net *n, u64 features)
1030 {
1031         size_t vhost_hlen, sock_hlen, hdr_len;
1032         int i;
1033
1034         hdr_len = (features & ((1ULL << VIRTIO_NET_F_MRG_RXBUF) |
1035                                (1ULL << VIRTIO_F_VERSION_1))) ?
1036                         sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1037                         sizeof(struct virtio_net_hdr);
1038         if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1039                 /* vhost provides vnet_hdr */
1040                 vhost_hlen = hdr_len;
1041                 sock_hlen = 0;
1042         } else {
1043                 /* socket provides vnet_hdr */
1044                 vhost_hlen = 0;
1045                 sock_hlen = hdr_len;
1046         }
1047         mutex_lock(&n->dev.mutex);
1048         if ((features & (1 << VHOST_F_LOG_ALL)) &&
1049             !vhost_log_access_ok(&n->dev)) {
1050                 mutex_unlock(&n->dev.mutex);
1051                 return -EFAULT;
1052         }
1053         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1054                 mutex_lock(&n->vqs[i].vq.mutex);
1055                 n->vqs[i].vq.acked_features = features;
1056                 n->vqs[i].vhost_hlen = vhost_hlen;
1057                 n->vqs[i].sock_hlen = sock_hlen;
1058                 mutex_unlock(&n->vqs[i].vq.mutex);
1059         }
1060         mutex_unlock(&n->dev.mutex);
1061         return 0;
1062 }
1063
1064 static long vhost_net_set_owner(struct vhost_net *n)
1065 {
1066         int r;
1067
1068         mutex_lock(&n->dev.mutex);
1069         if (vhost_dev_has_owner(&n->dev)) {
1070                 r = -EBUSY;
1071                 goto out;
1072         }
1073         r = vhost_net_set_ubuf_info(n);
1074         if (r)
1075                 goto out;
1076         r = vhost_dev_set_owner(&n->dev);
1077         if (r)
1078                 vhost_net_clear_ubuf_info(n);
1079         vhost_net_flush(n);
1080 out:
1081         mutex_unlock(&n->dev.mutex);
1082         return r;
1083 }
1084
1085 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1086                             unsigned long arg)
1087 {
1088         struct vhost_net *n = f->private_data;
1089         void __user *argp = (void __user *)arg;
1090         u64 __user *featurep = argp;
1091         struct vhost_vring_file backend;
1092         u64 features;
1093         int r;
1094
1095         switch (ioctl) {
1096         case VHOST_NET_SET_BACKEND:
1097                 if (copy_from_user(&backend, argp, sizeof backend))
1098                         return -EFAULT;
1099                 return vhost_net_set_backend(n, backend.index, backend.fd);
1100         case VHOST_GET_FEATURES:
1101                 features = VHOST_NET_FEATURES;
1102                 if (copy_to_user(featurep, &features, sizeof features))
1103                         return -EFAULT;
1104                 return 0;
1105         case VHOST_SET_FEATURES:
1106                 if (copy_from_user(&features, featurep, sizeof features))
1107                         return -EFAULT;
1108                 if (features & ~VHOST_NET_FEATURES)
1109                         return -EOPNOTSUPP;
1110                 return vhost_net_set_features(n, features);
1111         case VHOST_RESET_OWNER:
1112                 return vhost_net_reset_owner(n);
1113         case VHOST_SET_OWNER:
1114                 return vhost_net_set_owner(n);
1115         default:
1116                 mutex_lock(&n->dev.mutex);
1117                 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1118                 if (r == -ENOIOCTLCMD)
1119                         r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1120                 else
1121                         vhost_net_flush(n);
1122                 mutex_unlock(&n->dev.mutex);
1123                 return r;
1124         }
1125 }
1126
1127 #ifdef CONFIG_COMPAT
1128 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
1129                                    unsigned long arg)
1130 {
1131         return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
1132 }
1133 #endif
1134
1135 static const struct file_operations vhost_net_fops = {
1136         .owner          = THIS_MODULE,
1137         .release        = vhost_net_release,
1138         .unlocked_ioctl = vhost_net_ioctl,
1139 #ifdef CONFIG_COMPAT
1140         .compat_ioctl   = vhost_net_compat_ioctl,
1141 #endif
1142         .open           = vhost_net_open,
1143         .llseek         = noop_llseek,
1144 };
1145
1146 static struct miscdevice vhost_net_misc = {
1147         .minor = VHOST_NET_MINOR,
1148         .name = "vhost-net",
1149         .fops = &vhost_net_fops,
1150 };
1151
1152 static int vhost_net_init(void)
1153 {
1154         if (experimental_zcopytx)
1155                 vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1156         return misc_register(&vhost_net_misc);
1157 }
1158 module_init(vhost_net_init);
1159
1160 static void vhost_net_exit(void)
1161 {
1162         misc_deregister(&vhost_net_misc);
1163 }
1164 module_exit(vhost_net_exit);
1165
1166 MODULE_VERSION("0.0.1");
1167 MODULE_LICENSE("GPL v2");
1168 MODULE_AUTHOR("Michael S. Tsirkin");
1169 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1170 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1171 MODULE_ALIAS("devname:vhost-net");