Commit | Line | Data |
---|---|---|
3a4d5c94 MT |
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> | |
3a4d5c94 MT |
13 | #include <linux/miscdevice.h> |
14 | #include <linux/module.h> | |
bab632d6 | 15 | #include <linux/moduleparam.h> |
3a4d5c94 MT |
16 | #include <linux/mutex.h> |
17 | #include <linux/workqueue.h> | |
18 | #include <linux/rcupdate.h> | |
19 | #include <linux/file.h> | |
5a0e3ad6 | 20 | #include <linux/slab.h> |
3a4d5c94 MT |
21 | |
22 | #include <linux/net.h> | |
23 | #include <linux/if_packet.h> | |
24 | #include <linux/if_arp.h> | |
25 | #include <linux/if_tun.h> | |
501c774c | 26 | #include <linux/if_macvlan.h> |
c53cff5e | 27 | #include <linux/if_vlan.h> |
3a4d5c94 MT |
28 | |
29 | #include <net/sock.h> | |
30 | ||
31 | #include "vhost.h" | |
32 | ||
bab632d6 MT |
33 | static int experimental_zcopytx; |
34 | module_param(experimental_zcopytx, int, 0444); | |
35 | MODULE_PARM_DESC(experimental_zcopytx, "Enable Experimental Zero Copy TX"); | |
36 | ||
3a4d5c94 MT |
37 | /* Max number of bytes transferred before requeueing the job. |
38 | * Using this limit prevents one virtqueue from starving others. */ | |
39 | #define VHOST_NET_WEIGHT 0x80000 | |
40 | ||
bab632d6 MT |
41 | /* MAX number of TX used buffers for outstanding zerocopy */ |
42 | #define VHOST_MAX_PEND 128 | |
43 | #define VHOST_GOODCOPY_LEN 256 | |
44 | ||
3a4d5c94 MT |
45 | enum { |
46 | VHOST_NET_VQ_RX = 0, | |
47 | VHOST_NET_VQ_TX = 1, | |
48 | VHOST_NET_VQ_MAX = 2, | |
49 | }; | |
50 | ||
51 | enum vhost_net_poll_state { | |
52 | VHOST_NET_POLL_DISABLED = 0, | |
53 | VHOST_NET_POLL_STARTED = 1, | |
54 | VHOST_NET_POLL_STOPPED = 2, | |
55 | }; | |
56 | ||
57 | struct vhost_net { | |
58 | struct vhost_dev dev; | |
59 | struct vhost_virtqueue vqs[VHOST_NET_VQ_MAX]; | |
60 | struct vhost_poll poll[VHOST_NET_VQ_MAX]; | |
61 | /* Tells us whether we are polling a socket for TX. | |
62 | * We only do this when socket buffer fills up. | |
63 | * Protected by tx vq lock. */ | |
64 | enum vhost_net_poll_state tx_poll_state; | |
65 | }; | |
66 | ||
bab632d6 MT |
67 | static bool vhost_sock_zcopy(struct socket *sock) |
68 | { | |
69 | return unlikely(experimental_zcopytx) && | |
70 | sock_flag(sock->sk, SOCK_ZEROCOPY); | |
71 | } | |
72 | ||
3a4d5c94 MT |
73 | /* Pop first len bytes from iovec. Return number of segments used. */ |
74 | static int move_iovec_hdr(struct iovec *from, struct iovec *to, | |
75 | size_t len, int iov_count) | |
76 | { | |
77 | int seg = 0; | |
78 | size_t size; | |
d47effe1 | 79 | |
3a4d5c94 MT |
80 | while (len && seg < iov_count) { |
81 | size = min(from->iov_len, len); | |
82 | to->iov_base = from->iov_base; | |
83 | to->iov_len = size; | |
84 | from->iov_len -= size; | |
85 | from->iov_base += size; | |
86 | len -= size; | |
87 | ++from; | |
88 | ++to; | |
89 | ++seg; | |
90 | } | |
91 | return seg; | |
92 | } | |
8dd014ad DS |
93 | /* Copy iovec entries for len bytes from iovec. */ |
94 | static void copy_iovec_hdr(const struct iovec *from, struct iovec *to, | |
95 | size_t len, int iovcount) | |
96 | { | |
97 | int seg = 0; | |
98 | size_t size; | |
d47effe1 | 99 | |
8dd014ad DS |
100 | while (len && seg < iovcount) { |
101 | size = min(from->iov_len, len); | |
102 | to->iov_base = from->iov_base; | |
103 | to->iov_len = size; | |
104 | len -= size; | |
105 | ++from; | |
106 | ++to; | |
107 | ++seg; | |
108 | } | |
109 | } | |
3a4d5c94 MT |
110 | |
111 | /* Caller must have TX VQ lock */ | |
112 | static void tx_poll_stop(struct vhost_net *net) | |
113 | { | |
114 | if (likely(net->tx_poll_state != VHOST_NET_POLL_STARTED)) | |
115 | return; | |
116 | vhost_poll_stop(net->poll + VHOST_NET_VQ_TX); | |
117 | net->tx_poll_state = VHOST_NET_POLL_STOPPED; | |
118 | } | |
119 | ||
120 | /* Caller must have TX VQ lock */ | |
121 | static void tx_poll_start(struct vhost_net *net, struct socket *sock) | |
122 | { | |
123 | if (unlikely(net->tx_poll_state != VHOST_NET_POLL_STOPPED)) | |
124 | return; | |
125 | vhost_poll_start(net->poll + VHOST_NET_VQ_TX, sock->file); | |
126 | net->tx_poll_state = VHOST_NET_POLL_STARTED; | |
127 | } | |
128 | ||
b211616d MT |
129 | /* In case of DMA done not in order in lower device driver for some reason. |
130 | * upend_idx is used to track end of used idx, done_idx is used to track head | |
131 | * of used idx. Once lower device DMA done contiguously, we will signal KVM | |
132 | * guest used idx. | |
133 | */ | |
134 | int vhost_zerocopy_signal_used(struct vhost_virtqueue *vq) | |
135 | { | |
136 | int i; | |
137 | int j = 0; | |
138 | ||
139 | for (i = vq->done_idx; i != vq->upend_idx; i = (i + 1) % UIO_MAXIOV) { | |
140 | if (VHOST_DMA_IS_DONE(vq->heads[i].len)) { | |
141 | vq->heads[i].len = VHOST_DMA_CLEAR_LEN; | |
142 | vhost_add_used_and_signal(vq->dev, vq, | |
143 | vq->heads[i].id, 0); | |
144 | ++j; | |
145 | } else | |
146 | break; | |
147 | } | |
148 | if (j) | |
149 | vq->done_idx = i; | |
150 | return j; | |
151 | } | |
152 | ||
153 | static void vhost_zerocopy_callback(struct ubuf_info *ubuf, int status) | |
154 | { | |
155 | struct vhost_ubuf_ref *ubufs = ubuf->ctx; | |
156 | struct vhost_virtqueue *vq = ubufs->vq; | |
157 | ||
158 | vhost_poll_queue(&vq->poll); | |
159 | /* set len to mark this desc buffers done DMA */ | |
160 | vq->heads[ubuf->desc].len = status ? | |
161 | VHOST_DMA_FAILED_LEN : VHOST_DMA_DONE_LEN; | |
162 | vhost_ubuf_put(ubufs); | |
163 | } | |
164 | ||
3a4d5c94 MT |
165 | /* Expects to be always run from workqueue - which acts as |
166 | * read-size critical section for our kind of RCU. */ | |
167 | static void handle_tx(struct vhost_net *net) | |
168 | { | |
169 | struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_TX]; | |
d5675bd2 MT |
170 | unsigned out, in, s; |
171 | int head; | |
3a4d5c94 MT |
172 | struct msghdr msg = { |
173 | .msg_name = NULL, | |
174 | .msg_namelen = 0, | |
175 | .msg_control = NULL, | |
176 | .msg_controllen = 0, | |
177 | .msg_iov = vq->iov, | |
178 | .msg_flags = MSG_DONTWAIT, | |
179 | }; | |
180 | size_t len, total_len = 0; | |
181 | int err, wmem; | |
182 | size_t hdr_size; | |
28457ee6 | 183 | struct socket *sock; |
bab632d6 MT |
184 | struct vhost_ubuf_ref *uninitialized_var(ubufs); |
185 | bool zcopy; | |
28457ee6 | 186 | |
5e18247b | 187 | /* TODO: check that we are running from vhost_worker? */ |
11cd1a8b | 188 | sock = rcu_dereference_check(vq->private_data, 1); |
3a4d5c94 MT |
189 | if (!sock) |
190 | return; | |
191 | ||
192 | wmem = atomic_read(&sock->sk->sk_wmem_alloc); | |
39286fa4 SS |
193 | if (wmem >= sock->sk->sk_sndbuf) { |
194 | mutex_lock(&vq->mutex); | |
195 | tx_poll_start(net, sock); | |
196 | mutex_unlock(&vq->mutex); | |
3a4d5c94 | 197 | return; |
39286fa4 | 198 | } |
3a4d5c94 | 199 | |
3a4d5c94 | 200 | mutex_lock(&vq->mutex); |
8ea8cf89 | 201 | vhost_disable_notify(&net->dev, vq); |
3a4d5c94 | 202 | |
0e255572 | 203 | if (wmem < sock->sk->sk_sndbuf / 2) |
3a4d5c94 | 204 | tx_poll_stop(net); |
8dd014ad | 205 | hdr_size = vq->vhost_hlen; |
c460f057 | 206 | zcopy = vq->ubufs; |
3a4d5c94 MT |
207 | |
208 | for (;;) { | |
bab632d6 MT |
209 | /* Release DMAs done buffers first */ |
210 | if (zcopy) | |
211 | vhost_zerocopy_signal_used(vq); | |
212 | ||
3a4d5c94 MT |
213 | head = vhost_get_vq_desc(&net->dev, vq, vq->iov, |
214 | ARRAY_SIZE(vq->iov), | |
215 | &out, &in, | |
216 | NULL, NULL); | |
d5675bd2 | 217 | /* On error, stop handling until the next kick. */ |
7b3384fc | 218 | if (unlikely(head < 0)) |
d5675bd2 | 219 | break; |
3a4d5c94 MT |
220 | /* Nothing new? Wait for eventfd to tell us they refilled. */ |
221 | if (head == vq->num) { | |
9e380825 SM |
222 | int num_pends; |
223 | ||
3a4d5c94 MT |
224 | wmem = atomic_read(&sock->sk->sk_wmem_alloc); |
225 | if (wmem >= sock->sk->sk_sndbuf * 3 / 4) { | |
226 | tx_poll_start(net, sock); | |
227 | set_bit(SOCK_ASYNC_NOSPACE, &sock->flags); | |
228 | break; | |
229 | } | |
9e380825 SM |
230 | /* If more outstanding DMAs, queue the work. |
231 | * Handle upend_idx wrap around | |
232 | */ | |
233 | num_pends = likely(vq->upend_idx >= vq->done_idx) ? | |
234 | (vq->upend_idx - vq->done_idx) : | |
235 | (vq->upend_idx + UIO_MAXIOV - vq->done_idx); | |
236 | if (unlikely(num_pends > VHOST_MAX_PEND)) { | |
bab632d6 MT |
237 | tx_poll_start(net, sock); |
238 | set_bit(SOCK_ASYNC_NOSPACE, &sock->flags); | |
239 | break; | |
240 | } | |
8ea8cf89 MT |
241 | if (unlikely(vhost_enable_notify(&net->dev, vq))) { |
242 | vhost_disable_notify(&net->dev, vq); | |
3a4d5c94 MT |
243 | continue; |
244 | } | |
245 | break; | |
246 | } | |
247 | if (in) { | |
248 | vq_err(vq, "Unexpected descriptor format for TX: " | |
249 | "out %d, int %d\n", out, in); | |
250 | break; | |
251 | } | |
252 | /* Skip header. TODO: support TSO. */ | |
253 | s = move_iovec_hdr(vq->iov, vq->hdr, hdr_size, out); | |
254 | msg.msg_iovlen = out; | |
255 | len = iov_length(vq->iov, out); | |
256 | /* Sanity check */ | |
257 | if (!len) { | |
258 | vq_err(vq, "Unexpected header len for TX: " | |
259 | "%zd expected %zd\n", | |
260 | iov_length(vq->hdr, s), hdr_size); | |
261 | break; | |
262 | } | |
bab632d6 MT |
263 | /* use msg_control to pass vhost zerocopy ubuf info to skb */ |
264 | if (zcopy) { | |
265 | vq->heads[vq->upend_idx].id = head; | |
266 | if (len < VHOST_GOODCOPY_LEN) { | |
267 | /* copy don't need to wait for DMA done */ | |
268 | vq->heads[vq->upend_idx].len = | |
269 | VHOST_DMA_DONE_LEN; | |
270 | msg.msg_control = NULL; | |
271 | msg.msg_controllen = 0; | |
272 | ubufs = NULL; | |
273 | } else { | |
274 | struct ubuf_info *ubuf = &vq->ubuf_info[head]; | |
275 | ||
70e4cb9a MT |
276 | vq->heads[vq->upend_idx].len = |
277 | VHOST_DMA_IN_PROGRESS; | |
bab632d6 | 278 | ubuf->callback = vhost_zerocopy_callback; |
ca8f4fb2 | 279 | ubuf->ctx = vq->ubufs; |
bab632d6 MT |
280 | ubuf->desc = vq->upend_idx; |
281 | msg.msg_control = ubuf; | |
282 | msg.msg_controllen = sizeof(ubuf); | |
283 | ubufs = vq->ubufs; | |
284 | kref_get(&ubufs->kref); | |
285 | } | |
286 | vq->upend_idx = (vq->upend_idx + 1) % UIO_MAXIOV; | |
287 | } | |
3a4d5c94 MT |
288 | /* TODO: Check specific error and bomb out unless ENOBUFS? */ |
289 | err = sock->ops->sendmsg(NULL, sock, &msg, len); | |
290 | if (unlikely(err < 0)) { | |
bab632d6 MT |
291 | if (zcopy) { |
292 | if (ubufs) | |
293 | vhost_ubuf_put(ubufs); | |
294 | vq->upend_idx = ((unsigned)vq->upend_idx - 1) % | |
295 | UIO_MAXIOV; | |
296 | } | |
8dd014ad | 297 | vhost_discard_vq_desc(vq, 1); |
dbf34207 JW |
298 | if (err == -EAGAIN || err == -ENOBUFS) |
299 | tx_poll_start(net, sock); | |
3a4d5c94 MT |
300 | break; |
301 | } | |
302 | if (err != len) | |
95c0ec6a MT |
303 | pr_debug("Truncated TX packet: " |
304 | " len %d != %zd\n", err, len); | |
bab632d6 MT |
305 | if (!zcopy) |
306 | vhost_add_used_and_signal(&net->dev, vq, head, 0); | |
c8fb217a JW |
307 | else |
308 | vhost_zerocopy_signal_used(vq); | |
3a4d5c94 MT |
309 | total_len += len; |
310 | if (unlikely(total_len >= VHOST_NET_WEIGHT)) { | |
311 | vhost_poll_queue(&vq->poll); | |
312 | break; | |
313 | } | |
314 | } | |
315 | ||
316 | mutex_unlock(&vq->mutex); | |
3a4d5c94 MT |
317 | } |
318 | ||
8dd014ad DS |
319 | static int peek_head_len(struct sock *sk) |
320 | { | |
321 | struct sk_buff *head; | |
322 | int len = 0; | |
783e3988 | 323 | unsigned long flags; |
8dd014ad | 324 | |
783e3988 | 325 | spin_lock_irqsave(&sk->sk_receive_queue.lock, flags); |
8dd014ad | 326 | head = skb_peek(&sk->sk_receive_queue); |
c53cff5e | 327 | if (likely(head)) { |
8dd014ad | 328 | len = head->len; |
c53cff5e BG |
329 | if (vlan_tx_tag_present(head)) |
330 | len += VLAN_HLEN; | |
331 | } | |
332 | ||
783e3988 | 333 | spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags); |
8dd014ad DS |
334 | return len; |
335 | } | |
336 | ||
337 | /* This is a multi-buffer version of vhost_get_desc, that works if | |
338 | * vq has read descriptors only. | |
339 | * @vq - the relevant virtqueue | |
340 | * @datalen - data length we'll be reading | |
341 | * @iovcount - returned count of io vectors we fill | |
342 | * @log - vhost log | |
343 | * @log_num - log offset | |
94249369 | 344 | * @quota - headcount quota, 1 for big buffer |
8dd014ad DS |
345 | * returns number of buffer heads allocated, negative on error |
346 | */ | |
347 | static int get_rx_bufs(struct vhost_virtqueue *vq, | |
348 | struct vring_used_elem *heads, | |
349 | int datalen, | |
350 | unsigned *iovcount, | |
351 | struct vhost_log *log, | |
94249369 JW |
352 | unsigned *log_num, |
353 | unsigned int quota) | |
8dd014ad DS |
354 | { |
355 | unsigned int out, in; | |
356 | int seg = 0; | |
357 | int headcount = 0; | |
358 | unsigned d; | |
359 | int r, nlogs = 0; | |
360 | ||
94249369 | 361 | while (datalen > 0 && headcount < quota) { |
e0e9b406 | 362 | if (unlikely(seg >= UIO_MAXIOV)) { |
8dd014ad DS |
363 | r = -ENOBUFS; |
364 | goto err; | |
365 | } | |
366 | d = vhost_get_vq_desc(vq->dev, vq, vq->iov + seg, | |
367 | ARRAY_SIZE(vq->iov) - seg, &out, | |
368 | &in, log, log_num); | |
369 | if (d == vq->num) { | |
370 | r = 0; | |
371 | goto err; | |
372 | } | |
373 | if (unlikely(out || in <= 0)) { | |
374 | vq_err(vq, "unexpected descriptor format for RX: " | |
375 | "out %d, in %d\n", out, in); | |
376 | r = -EINVAL; | |
377 | goto err; | |
378 | } | |
379 | if (unlikely(log)) { | |
380 | nlogs += *log_num; | |
381 | log += *log_num; | |
382 | } | |
383 | heads[headcount].id = d; | |
384 | heads[headcount].len = iov_length(vq->iov + seg, in); | |
385 | datalen -= heads[headcount].len; | |
386 | ++headcount; | |
387 | seg += in; | |
388 | } | |
389 | heads[headcount - 1].len += datalen; | |
390 | *iovcount = seg; | |
391 | if (unlikely(log)) | |
392 | *log_num = nlogs; | |
393 | return headcount; | |
394 | err: | |
395 | vhost_discard_vq_desc(vq, headcount); | |
396 | return r; | |
397 | } | |
398 | ||
3a4d5c94 MT |
399 | /* Expects to be always run from workqueue - which acts as |
400 | * read-size critical section for our kind of RCU. */ | |
94249369 | 401 | static void handle_rx(struct vhost_net *net) |
3a4d5c94 | 402 | { |
8dd014ad DS |
403 | struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_RX]; |
404 | unsigned uninitialized_var(in), log; | |
405 | struct vhost_log *vq_log; | |
406 | struct msghdr msg = { | |
407 | .msg_name = NULL, | |
408 | .msg_namelen = 0, | |
409 | .msg_control = NULL, /* FIXME: get and handle RX aux data. */ | |
410 | .msg_controllen = 0, | |
411 | .msg_iov = vq->iov, | |
412 | .msg_flags = MSG_DONTWAIT, | |
413 | }; | |
8dd014ad DS |
414 | struct virtio_net_hdr_mrg_rxbuf hdr = { |
415 | .hdr.flags = 0, | |
416 | .hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE | |
417 | }; | |
8dd014ad | 418 | size_t total_len = 0; |
cfbdab95 | 419 | int err, headcount, mergeable; |
8dd014ad DS |
420 | size_t vhost_hlen, sock_hlen; |
421 | size_t vhost_len, sock_len; | |
5e18247b MT |
422 | /* TODO: check that we are running from vhost_worker? */ |
423 | struct socket *sock = rcu_dereference_check(vq->private_data, 1); | |
d47effe1 | 424 | |
de4d768a | 425 | if (!sock) |
8dd014ad DS |
426 | return; |
427 | ||
8dd014ad | 428 | mutex_lock(&vq->mutex); |
8ea8cf89 | 429 | vhost_disable_notify(&net->dev, vq); |
8dd014ad DS |
430 | vhost_hlen = vq->vhost_hlen; |
431 | sock_hlen = vq->sock_hlen; | |
432 | ||
433 | vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ? | |
434 | vq->log : NULL; | |
cfbdab95 | 435 | mergeable = vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF); |
8dd014ad DS |
436 | |
437 | while ((sock_len = peek_head_len(sock->sk))) { | |
438 | sock_len += sock_hlen; | |
439 | vhost_len = sock_len + vhost_hlen; | |
440 | headcount = get_rx_bufs(vq, vq->heads, vhost_len, | |
94249369 JW |
441 | &in, vq_log, &log, |
442 | likely(mergeable) ? UIO_MAXIOV : 1); | |
8dd014ad DS |
443 | /* On error, stop handling until the next kick. */ |
444 | if (unlikely(headcount < 0)) | |
445 | break; | |
446 | /* OK, now we need to know about added descriptors. */ | |
447 | if (!headcount) { | |
8ea8cf89 | 448 | if (unlikely(vhost_enable_notify(&net->dev, vq))) { |
8dd014ad DS |
449 | /* They have slipped one in as we were |
450 | * doing that: check again. */ | |
8ea8cf89 | 451 | vhost_disable_notify(&net->dev, vq); |
8dd014ad DS |
452 | continue; |
453 | } | |
454 | /* Nothing new? Wait for eventfd to tell us | |
455 | * they refilled. */ | |
456 | break; | |
457 | } | |
458 | /* We don't need to be notified again. */ | |
459 | if (unlikely((vhost_hlen))) | |
460 | /* Skip header. TODO: support TSO. */ | |
461 | move_iovec_hdr(vq->iov, vq->hdr, vhost_hlen, in); | |
462 | else | |
463 | /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF: | |
a290aec8 | 464 | * needed because recvmsg can modify msg_iov. */ |
8dd014ad DS |
465 | copy_iovec_hdr(vq->iov, vq->hdr, sock_hlen, in); |
466 | msg.msg_iovlen = in; | |
467 | err = sock->ops->recvmsg(NULL, sock, &msg, | |
468 | sock_len, MSG_DONTWAIT | MSG_TRUNC); | |
469 | /* Userspace might have consumed the packet meanwhile: | |
470 | * it's not supposed to do this usually, but might be hard | |
471 | * to prevent. Discard data we got (if any) and keep going. */ | |
472 | if (unlikely(err != sock_len)) { | |
473 | pr_debug("Discarded rx packet: " | |
474 | " len %d, expected %zd\n", err, sock_len); | |
475 | vhost_discard_vq_desc(vq, headcount); | |
476 | continue; | |
477 | } | |
478 | if (unlikely(vhost_hlen) && | |
479 | memcpy_toiovecend(vq->hdr, (unsigned char *)&hdr, 0, | |
480 | vhost_hlen)) { | |
481 | vq_err(vq, "Unable to write vnet_hdr at addr %p\n", | |
482 | vq->iov->iov_base); | |
483 | break; | |
484 | } | |
485 | /* TODO: Should check and handle checksum. */ | |
cfbdab95 | 486 | if (likely(mergeable) && |
8dd014ad DS |
487 | memcpy_toiovecend(vq->hdr, (unsigned char *)&headcount, |
488 | offsetof(typeof(hdr), num_buffers), | |
489 | sizeof hdr.num_buffers)) { | |
490 | vq_err(vq, "Failed num_buffers write"); | |
491 | vhost_discard_vq_desc(vq, headcount); | |
492 | break; | |
493 | } | |
494 | vhost_add_used_and_signal_n(&net->dev, vq, vq->heads, | |
495 | headcount); | |
496 | if (unlikely(vq_log)) | |
497 | vhost_log_write(vq, vq_log, log, vhost_len); | |
498 | total_len += vhost_len; | |
499 | if (unlikely(total_len >= VHOST_NET_WEIGHT)) { | |
500 | vhost_poll_queue(&vq->poll); | |
501 | break; | |
502 | } | |
503 | } | |
504 | ||
505 | mutex_unlock(&vq->mutex); | |
8dd014ad DS |
506 | } |
507 | ||
c23f3445 | 508 | static void handle_tx_kick(struct vhost_work *work) |
3a4d5c94 | 509 | { |
c23f3445 TH |
510 | struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue, |
511 | poll.work); | |
512 | struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev); | |
513 | ||
3a4d5c94 MT |
514 | handle_tx(net); |
515 | } | |
516 | ||
c23f3445 | 517 | static void handle_rx_kick(struct vhost_work *work) |
3a4d5c94 | 518 | { |
c23f3445 TH |
519 | struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue, |
520 | poll.work); | |
521 | struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev); | |
522 | ||
3a4d5c94 MT |
523 | handle_rx(net); |
524 | } | |
525 | ||
c23f3445 | 526 | static void handle_tx_net(struct vhost_work *work) |
3a4d5c94 | 527 | { |
c23f3445 TH |
528 | struct vhost_net *net = container_of(work, struct vhost_net, |
529 | poll[VHOST_NET_VQ_TX].work); | |
3a4d5c94 MT |
530 | handle_tx(net); |
531 | } | |
532 | ||
c23f3445 | 533 | static void handle_rx_net(struct vhost_work *work) |
3a4d5c94 | 534 | { |
c23f3445 TH |
535 | struct vhost_net *net = container_of(work, struct vhost_net, |
536 | poll[VHOST_NET_VQ_RX].work); | |
3a4d5c94 MT |
537 | handle_rx(net); |
538 | } | |
539 | ||
540 | static int vhost_net_open(struct inode *inode, struct file *f) | |
541 | { | |
542 | struct vhost_net *n = kmalloc(sizeof *n, GFP_KERNEL); | |
c23f3445 | 543 | struct vhost_dev *dev; |
3a4d5c94 | 544 | int r; |
c23f3445 | 545 | |
3a4d5c94 MT |
546 | if (!n) |
547 | return -ENOMEM; | |
c23f3445 TH |
548 | |
549 | dev = &n->dev; | |
3a4d5c94 MT |
550 | n->vqs[VHOST_NET_VQ_TX].handle_kick = handle_tx_kick; |
551 | n->vqs[VHOST_NET_VQ_RX].handle_kick = handle_rx_kick; | |
c23f3445 | 552 | r = vhost_dev_init(dev, n->vqs, VHOST_NET_VQ_MAX); |
3a4d5c94 MT |
553 | if (r < 0) { |
554 | kfree(n); | |
555 | return r; | |
556 | } | |
557 | ||
c23f3445 TH |
558 | vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev); |
559 | vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev); | |
3a4d5c94 MT |
560 | n->tx_poll_state = VHOST_NET_POLL_DISABLED; |
561 | ||
562 | f->private_data = n; | |
563 | ||
564 | return 0; | |
565 | } | |
566 | ||
567 | static void vhost_net_disable_vq(struct vhost_net *n, | |
568 | struct vhost_virtqueue *vq) | |
569 | { | |
570 | if (!vq->private_data) | |
571 | return; | |
572 | if (vq == n->vqs + VHOST_NET_VQ_TX) { | |
573 | tx_poll_stop(n); | |
574 | n->tx_poll_state = VHOST_NET_POLL_DISABLED; | |
575 | } else | |
576 | vhost_poll_stop(n->poll + VHOST_NET_VQ_RX); | |
577 | } | |
578 | ||
579 | static void vhost_net_enable_vq(struct vhost_net *n, | |
580 | struct vhost_virtqueue *vq) | |
581 | { | |
28457ee6 AB |
582 | struct socket *sock; |
583 | ||
584 | sock = rcu_dereference_protected(vq->private_data, | |
585 | lockdep_is_held(&vq->mutex)); | |
3a4d5c94 MT |
586 | if (!sock) |
587 | return; | |
588 | if (vq == n->vqs + VHOST_NET_VQ_TX) { | |
589 | n->tx_poll_state = VHOST_NET_POLL_STOPPED; | |
590 | tx_poll_start(n, sock); | |
591 | } else | |
592 | vhost_poll_start(n->poll + VHOST_NET_VQ_RX, sock->file); | |
593 | } | |
594 | ||
595 | static struct socket *vhost_net_stop_vq(struct vhost_net *n, | |
596 | struct vhost_virtqueue *vq) | |
597 | { | |
598 | struct socket *sock; | |
599 | ||
600 | mutex_lock(&vq->mutex); | |
28457ee6 AB |
601 | sock = rcu_dereference_protected(vq->private_data, |
602 | lockdep_is_held(&vq->mutex)); | |
3a4d5c94 MT |
603 | vhost_net_disable_vq(n, vq); |
604 | rcu_assign_pointer(vq->private_data, NULL); | |
605 | mutex_unlock(&vq->mutex); | |
606 | return sock; | |
607 | } | |
608 | ||
609 | static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock, | |
610 | struct socket **rx_sock) | |
611 | { | |
612 | *tx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_TX); | |
613 | *rx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_RX); | |
614 | } | |
615 | ||
616 | static void vhost_net_flush_vq(struct vhost_net *n, int index) | |
617 | { | |
618 | vhost_poll_flush(n->poll + index); | |
619 | vhost_poll_flush(&n->dev.vqs[index].poll); | |
620 | } | |
621 | ||
622 | static void vhost_net_flush(struct vhost_net *n) | |
623 | { | |
624 | vhost_net_flush_vq(n, VHOST_NET_VQ_TX); | |
625 | vhost_net_flush_vq(n, VHOST_NET_VQ_RX); | |
626 | } | |
627 | ||
628 | static int vhost_net_release(struct inode *inode, struct file *f) | |
629 | { | |
630 | struct vhost_net *n = f->private_data; | |
631 | struct socket *tx_sock; | |
632 | struct socket *rx_sock; | |
b211616d | 633 | int i; |
3a4d5c94 MT |
634 | |
635 | vhost_net_stop(n, &tx_sock, &rx_sock); | |
636 | vhost_net_flush(n); | |
b211616d MT |
637 | vhost_dev_stop(&n->dev); |
638 | for (i = 0; i < n->dev.nvqs; ++i) { | |
639 | /* Wait for all lower device DMAs done. */ | |
640 | if (n->dev.vqs[i].ubufs) | |
641 | vhost_ubuf_put_and_wait(n->dev.vqs[i].ubufs); | |
642 | ||
643 | vhost_zerocopy_signal_used(n, &n->dev.vqs[i]); | |
644 | } | |
ea5d4046 | 645 | vhost_dev_cleanup(&n->dev, false); |
3a4d5c94 MT |
646 | if (tx_sock) |
647 | fput(tx_sock->file); | |
648 | if (rx_sock) | |
649 | fput(rx_sock->file); | |
650 | /* We do an extra flush before freeing memory, | |
651 | * since jobs can re-queue themselves. */ | |
652 | vhost_net_flush(n); | |
653 | kfree(n); | |
654 | return 0; | |
655 | } | |
656 | ||
657 | static struct socket *get_raw_socket(int fd) | |
658 | { | |
659 | struct { | |
660 | struct sockaddr_ll sa; | |
661 | char buf[MAX_ADDR_LEN]; | |
662 | } uaddr; | |
663 | int uaddr_len = sizeof uaddr, r; | |
664 | struct socket *sock = sockfd_lookup(fd, &r); | |
d47effe1 | 665 | |
3a4d5c94 MT |
666 | if (!sock) |
667 | return ERR_PTR(-ENOTSOCK); | |
668 | ||
669 | /* Parameter checking */ | |
670 | if (sock->sk->sk_type != SOCK_RAW) { | |
671 | r = -ESOCKTNOSUPPORT; | |
672 | goto err; | |
673 | } | |
674 | ||
675 | r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa, | |
676 | &uaddr_len, 0); | |
677 | if (r) | |
678 | goto err; | |
679 | ||
680 | if (uaddr.sa.sll_family != AF_PACKET) { | |
681 | r = -EPFNOSUPPORT; | |
682 | goto err; | |
683 | } | |
684 | return sock; | |
685 | err: | |
686 | fput(sock->file); | |
687 | return ERR_PTR(r); | |
688 | } | |
689 | ||
501c774c | 690 | static struct socket *get_tap_socket(int fd) |
3a4d5c94 MT |
691 | { |
692 | struct file *file = fget(fd); | |
693 | struct socket *sock; | |
d47effe1 | 694 | |
3a4d5c94 MT |
695 | if (!file) |
696 | return ERR_PTR(-EBADF); | |
697 | sock = tun_get_socket(file); | |
501c774c AB |
698 | if (!IS_ERR(sock)) |
699 | return sock; | |
700 | sock = macvtap_get_socket(file); | |
3a4d5c94 MT |
701 | if (IS_ERR(sock)) |
702 | fput(file); | |
703 | return sock; | |
704 | } | |
705 | ||
706 | static struct socket *get_socket(int fd) | |
707 | { | |
708 | struct socket *sock; | |
d47effe1 | 709 | |
3a4d5c94 MT |
710 | /* special case to disable backend */ |
711 | if (fd == -1) | |
712 | return NULL; | |
713 | sock = get_raw_socket(fd); | |
714 | if (!IS_ERR(sock)) | |
715 | return sock; | |
501c774c | 716 | sock = get_tap_socket(fd); |
3a4d5c94 MT |
717 | if (!IS_ERR(sock)) |
718 | return sock; | |
719 | return ERR_PTR(-ENOTSOCK); | |
720 | } | |
721 | ||
722 | static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd) | |
723 | { | |
724 | struct socket *sock, *oldsock; | |
725 | struct vhost_virtqueue *vq; | |
bab632d6 | 726 | struct vhost_ubuf_ref *ubufs, *oldubufs = NULL; |
3a4d5c94 MT |
727 | int r; |
728 | ||
729 | mutex_lock(&n->dev.mutex); | |
730 | r = vhost_dev_check_owner(&n->dev); | |
731 | if (r) | |
732 | goto err; | |
733 | ||
734 | if (index >= VHOST_NET_VQ_MAX) { | |
735 | r = -ENOBUFS; | |
736 | goto err; | |
737 | } | |
738 | vq = n->vqs + index; | |
739 | mutex_lock(&vq->mutex); | |
740 | ||
741 | /* Verify that ring has been setup correctly. */ | |
742 | if (!vhost_vq_access_ok(vq)) { | |
743 | r = -EFAULT; | |
1dace8c8 | 744 | goto err_vq; |
3a4d5c94 MT |
745 | } |
746 | sock = get_socket(fd); | |
747 | if (IS_ERR(sock)) { | |
748 | r = PTR_ERR(sock); | |
1dace8c8 | 749 | goto err_vq; |
3a4d5c94 MT |
750 | } |
751 | ||
752 | /* start polling new socket */ | |
28457ee6 AB |
753 | oldsock = rcu_dereference_protected(vq->private_data, |
754 | lockdep_is_held(&vq->mutex)); | |
11fe8839 | 755 | if (sock != oldsock) { |
bab632d6 MT |
756 | ubufs = vhost_ubuf_alloc(vq, sock && vhost_sock_zcopy(sock)); |
757 | if (IS_ERR(ubufs)) { | |
758 | r = PTR_ERR(ubufs); | |
759 | goto err_ubufs; | |
760 | } | |
761 | oldubufs = vq->ubufs; | |
762 | vq->ubufs = ubufs; | |
d47effe1 KK |
763 | vhost_net_disable_vq(n, vq); |
764 | rcu_assign_pointer(vq->private_data, sock); | |
765 | vhost_net_enable_vq(n, vq); | |
f59281da JW |
766 | |
767 | r = vhost_init_used(vq); | |
768 | if (r) | |
769 | goto err_vq; | |
dd1f4078 | 770 | } |
3a4d5c94 | 771 | |
1680e906 MT |
772 | mutex_unlock(&vq->mutex); |
773 | ||
c047e5f3 | 774 | if (oldubufs) { |
bab632d6 | 775 | vhost_ubuf_put_and_wait(oldubufs); |
c047e5f3 MT |
776 | mutex_lock(&vq->mutex); |
777 | vhost_zerocopy_signal_used(vq); | |
778 | mutex_unlock(&vq->mutex); | |
779 | } | |
bab632d6 | 780 | |
3a4d5c94 MT |
781 | if (oldsock) { |
782 | vhost_net_flush_vq(n, index); | |
783 | fput(oldsock->file); | |
784 | } | |
1dace8c8 | 785 | |
1680e906 MT |
786 | mutex_unlock(&n->dev.mutex); |
787 | return 0; | |
788 | ||
bab632d6 MT |
789 | err_ubufs: |
790 | fput(sock->file); | |
1dace8c8 JD |
791 | err_vq: |
792 | mutex_unlock(&vq->mutex); | |
3a4d5c94 MT |
793 | err: |
794 | mutex_unlock(&n->dev.mutex); | |
795 | return r; | |
796 | } | |
797 | ||
798 | static long vhost_net_reset_owner(struct vhost_net *n) | |
799 | { | |
800 | struct socket *tx_sock = NULL; | |
801 | struct socket *rx_sock = NULL; | |
802 | long err; | |
d47effe1 | 803 | |
3a4d5c94 MT |
804 | mutex_lock(&n->dev.mutex); |
805 | err = vhost_dev_check_owner(&n->dev); | |
806 | if (err) | |
807 | goto done; | |
808 | vhost_net_stop(n, &tx_sock, &rx_sock); | |
809 | vhost_net_flush(n); | |
810 | err = vhost_dev_reset_owner(&n->dev); | |
811 | done: | |
812 | mutex_unlock(&n->dev.mutex); | |
813 | if (tx_sock) | |
814 | fput(tx_sock->file); | |
815 | if (rx_sock) | |
816 | fput(rx_sock->file); | |
817 | return err; | |
818 | } | |
819 | ||
820 | static int vhost_net_set_features(struct vhost_net *n, u64 features) | |
821 | { | |
8dd014ad | 822 | size_t vhost_hlen, sock_hlen, hdr_len; |
3a4d5c94 | 823 | int i; |
8dd014ad DS |
824 | |
825 | hdr_len = (features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ? | |
826 | sizeof(struct virtio_net_hdr_mrg_rxbuf) : | |
827 | sizeof(struct virtio_net_hdr); | |
828 | if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) { | |
829 | /* vhost provides vnet_hdr */ | |
830 | vhost_hlen = hdr_len; | |
831 | sock_hlen = 0; | |
832 | } else { | |
833 | /* socket provides vnet_hdr */ | |
834 | vhost_hlen = 0; | |
835 | sock_hlen = hdr_len; | |
836 | } | |
3a4d5c94 MT |
837 | mutex_lock(&n->dev.mutex); |
838 | if ((features & (1 << VHOST_F_LOG_ALL)) && | |
839 | !vhost_log_access_ok(&n->dev)) { | |
840 | mutex_unlock(&n->dev.mutex); | |
841 | return -EFAULT; | |
842 | } | |
843 | n->dev.acked_features = features; | |
844 | smp_wmb(); | |
845 | for (i = 0; i < VHOST_NET_VQ_MAX; ++i) { | |
846 | mutex_lock(&n->vqs[i].mutex); | |
8dd014ad DS |
847 | n->vqs[i].vhost_hlen = vhost_hlen; |
848 | n->vqs[i].sock_hlen = sock_hlen; | |
3a4d5c94 MT |
849 | mutex_unlock(&n->vqs[i].mutex); |
850 | } | |
851 | vhost_net_flush(n); | |
852 | mutex_unlock(&n->dev.mutex); | |
853 | return 0; | |
854 | } | |
855 | ||
856 | static long vhost_net_ioctl(struct file *f, unsigned int ioctl, | |
857 | unsigned long arg) | |
858 | { | |
859 | struct vhost_net *n = f->private_data; | |
860 | void __user *argp = (void __user *)arg; | |
861 | u64 __user *featurep = argp; | |
862 | struct vhost_vring_file backend; | |
863 | u64 features; | |
864 | int r; | |
d47effe1 | 865 | |
3a4d5c94 MT |
866 | switch (ioctl) { |
867 | case VHOST_NET_SET_BACKEND: | |
d3553a52 TY |
868 | if (copy_from_user(&backend, argp, sizeof backend)) |
869 | return -EFAULT; | |
3a4d5c94 MT |
870 | return vhost_net_set_backend(n, backend.index, backend.fd); |
871 | case VHOST_GET_FEATURES: | |
0dd05a3b | 872 | features = VHOST_NET_FEATURES; |
d3553a52 TY |
873 | if (copy_to_user(featurep, &features, sizeof features)) |
874 | return -EFAULT; | |
875 | return 0; | |
3a4d5c94 | 876 | case VHOST_SET_FEATURES: |
d3553a52 TY |
877 | if (copy_from_user(&features, featurep, sizeof features)) |
878 | return -EFAULT; | |
0dd05a3b | 879 | if (features & ~VHOST_NET_FEATURES) |
3a4d5c94 MT |
880 | return -EOPNOTSUPP; |
881 | return vhost_net_set_features(n, features); | |
882 | case VHOST_RESET_OWNER: | |
883 | return vhost_net_reset_owner(n); | |
884 | default: | |
885 | mutex_lock(&n->dev.mutex); | |
886 | r = vhost_dev_ioctl(&n->dev, ioctl, arg); | |
887 | vhost_net_flush(n); | |
888 | mutex_unlock(&n->dev.mutex); | |
889 | return r; | |
890 | } | |
891 | } | |
892 | ||
893 | #ifdef CONFIG_COMPAT | |
894 | static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl, | |
895 | unsigned long arg) | |
896 | { | |
897 | return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg)); | |
898 | } | |
899 | #endif | |
900 | ||
373a83a6 | 901 | static const struct file_operations vhost_net_fops = { |
3a4d5c94 MT |
902 | .owner = THIS_MODULE, |
903 | .release = vhost_net_release, | |
904 | .unlocked_ioctl = vhost_net_ioctl, | |
905 | #ifdef CONFIG_COMPAT | |
906 | .compat_ioctl = vhost_net_compat_ioctl, | |
907 | #endif | |
908 | .open = vhost_net_open, | |
6038f373 | 909 | .llseek = noop_llseek, |
3a4d5c94 MT |
910 | }; |
911 | ||
912 | static struct miscdevice vhost_net_misc = { | |
7c7c7f01 | 913 | .minor = VHOST_NET_MINOR, |
914 | .name = "vhost-net", | |
915 | .fops = &vhost_net_fops, | |
3a4d5c94 MT |
916 | }; |
917 | ||
a8d3782f | 918 | static int vhost_net_init(void) |
3a4d5c94 | 919 | { |
bab632d6 MT |
920 | if (experimental_zcopytx) |
921 | vhost_enable_zcopy(VHOST_NET_VQ_TX); | |
c23f3445 | 922 | return misc_register(&vhost_net_misc); |
3a4d5c94 MT |
923 | } |
924 | module_init(vhost_net_init); | |
925 | ||
a8d3782f | 926 | static void vhost_net_exit(void) |
3a4d5c94 MT |
927 | { |
928 | misc_deregister(&vhost_net_misc); | |
3a4d5c94 MT |
929 | } |
930 | module_exit(vhost_net_exit); | |
931 | ||
932 | MODULE_VERSION("0.0.1"); | |
933 | MODULE_LICENSE("GPL v2"); | |
934 | MODULE_AUTHOR("Michael S. Tsirkin"); | |
935 | MODULE_DESCRIPTION("Host kernel accelerator for virtio net"); | |
7c7c7f01 | 936 | MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR); |
937 | MODULE_ALIAS("devname:vhost-net"); |