blob: f80d3dd41d8c6d420c8fb2d640a81739f2049674 [file] [log] [blame]
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +00001/* 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>
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +000013#include <linux/miscdevice.h>
14#include <linux/module.h>
Michael S. Tsirkinbab632d2011-07-18 03:48:46 +000015#include <linux/moduleparam.h>
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +000016#include <linux/mutex.h>
17#include <linux/workqueue.h>
18#include <linux/rcupdate.h>
19#include <linux/file.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090020#include <linux/slab.h>
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +000021
22#include <linux/net.h>
23#include <linux/if_packet.h>
24#include <linux/if_arp.h>
25#include <linux/if_tun.h>
Arnd Bergmann501c7742010-02-18 05:46:50 +000026#include <linux/if_macvlan.h>
Basil Gorc53cff5e2012-05-03 22:55:23 +000027#include <linux/if_vlan.h>
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +000028
29#include <net/sock.h>
30
31#include "vhost.h"
32
Michael S. Tsirkinf9611c42012-12-06 14:56:00 +020033static int experimental_zcopytx = 1;
Michael S. Tsirkinbab632d2011-07-18 03:48:46 +000034module_param(experimental_zcopytx, int, 0444);
Michael S. Tsirkinf9611c42012-12-06 14:56:00 +020035MODULE_PARM_DESC(experimental_zcopytx, "Enable Zero Copy TX;"
36 " 1 -Enable; 0 - Disable");
Michael S. Tsirkinbab632d2011-07-18 03:48:46 +000037
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +000038/* 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
Michael S. Tsirkinbab632d2011-07-18 03:48:46 +000042/* MAX number of TX used buffers for outstanding zerocopy */
43#define VHOST_MAX_PEND 128
44#define VHOST_GOODCOPY_LEN 256
45
Michael S. Tsirkineaae8132012-11-01 09:16:51 +000046/*
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 3
52/* Lower device DMA done */
53#define VHOST_DMA_DONE_LEN 2
54/* Lower device DMA in progress */
55#define VHOST_DMA_IN_PROGRESS 1
56/* Buffer unused */
57#define VHOST_DMA_CLEAR_LEN 0
58
59#define VHOST_DMA_IS_DONE(len) ((len) >= VHOST_DMA_DONE_LEN)
60
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +000061enum {
Asias He8570a6e2013-05-06 16:38:20 +080062 VHOST_NET_FEATURES = VHOST_FEATURES |
63 (1ULL << VHOST_NET_F_VIRTIO_NET_HDR) |
64 (1ULL << VIRTIO_NET_F_MRG_RXBUF),
65};
66
67enum {
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +000068 VHOST_NET_VQ_RX = 0,
69 VHOST_NET_VQ_TX = 1,
70 VHOST_NET_VQ_MAX = 2,
71};
72
Asias Hefe729a52013-05-06 16:38:24 +080073struct vhost_net_ubuf_ref {
Asias He28394002013-04-27 15:07:46 +080074 struct kref kref;
75 wait_queue_head_t wait;
76 struct vhost_virtqueue *vq;
77};
78
Asias He3ab2e422013-04-27 11:16:48 +080079struct vhost_net_virtqueue {
80 struct vhost_virtqueue vq;
Michael S. Tsirkin81f95a52013-04-28 15:51:40 +030081 /* hdr is used to store the virtio header.
82 * Since each iovec has >= 1 byte length, we never need more than
83 * header length entries to store the header. */
84 struct iovec hdr[sizeof(struct virtio_net_hdr_mrg_rxbuf)];
85 size_t vhost_hlen;
86 size_t sock_hlen;
Asias He28394002013-04-27 15:07:46 +080087 /* vhost zerocopy support fields below: */
88 /* last used idx for outstanding DMA zerocopy buffers */
89 int upend_idx;
90 /* first used idx for DMA done zerocopy buffers */
91 int done_idx;
92 /* an array of userspace buffers info */
93 struct ubuf_info *ubuf_info;
94 /* Reference counting for outstanding ubufs.
95 * Protected by vq mutex. Writers must also take device mutex. */
Asias Hefe729a52013-05-06 16:38:24 +080096 struct vhost_net_ubuf_ref *ubufs;
Asias He3ab2e422013-04-27 11:16:48 +080097};
98
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +000099struct vhost_net {
100 struct vhost_dev dev;
Asias He3ab2e422013-04-27 11:16:48 +0800101 struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX];
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000102 struct vhost_poll poll[VHOST_NET_VQ_MAX];
Michael S. Tsirkineaae8132012-11-01 09:16:51 +0000103 /* Number of TX recently submitted.
104 * Protected by tx vq lock. */
105 unsigned tx_packets;
106 /* Number of times zerocopy TX recently failed.
107 * Protected by tx vq lock. */
108 unsigned tx_zcopy_err;
Michael S. Tsirkin1280c272012-12-04 00:17:14 +0200109 /* Flush in progress. Protected by tx vq lock. */
110 bool tx_flush;
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000111};
112
Asias Hefe729a52013-05-06 16:38:24 +0800113static unsigned vhost_net_zcopy_mask __read_mostly;
Asias He28394002013-04-27 15:07:46 +0800114
Asias Hefe729a52013-05-06 16:38:24 +0800115static void vhost_net_enable_zcopy(int vq)
Asias He28394002013-04-27 15:07:46 +0800116{
Asias Hefe729a52013-05-06 16:38:24 +0800117 vhost_net_zcopy_mask |= 0x1 << vq;
Asias He28394002013-04-27 15:07:46 +0800118}
119
Asias Hefe729a52013-05-06 16:38:24 +0800120static void vhost_net_zerocopy_done_signal(struct kref *kref)
Asias He28394002013-04-27 15:07:46 +0800121{
Asias Hefe729a52013-05-06 16:38:24 +0800122 struct vhost_net_ubuf_ref *ubufs;
123
124 ubufs = container_of(kref, struct vhost_net_ubuf_ref, kref);
Asias He28394002013-04-27 15:07:46 +0800125 wake_up(&ubufs->wait);
126}
127
Asias Hefe729a52013-05-06 16:38:24 +0800128static struct vhost_net_ubuf_ref *
129vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy)
Asias He28394002013-04-27 15:07:46 +0800130{
Asias Hefe729a52013-05-06 16:38:24 +0800131 struct vhost_net_ubuf_ref *ubufs;
Asias He28394002013-04-27 15:07:46 +0800132 /* No zero copy backend? Nothing to count. */
133 if (!zcopy)
134 return NULL;
135 ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL);
136 if (!ubufs)
137 return ERR_PTR(-ENOMEM);
138 kref_init(&ubufs->kref);
139 init_waitqueue_head(&ubufs->wait);
140 ubufs->vq = vq;
141 return ubufs;
142}
143
Asias Hefe729a52013-05-06 16:38:24 +0800144static void vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs)
Asias He28394002013-04-27 15:07:46 +0800145{
Asias Hefe729a52013-05-06 16:38:24 +0800146 kref_put(&ubufs->kref, vhost_net_zerocopy_done_signal);
Asias He28394002013-04-27 15:07:46 +0800147}
148
Asias Hefe729a52013-05-06 16:38:24 +0800149static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs)
Asias He28394002013-04-27 15:07:46 +0800150{
Asias Hefe729a52013-05-06 16:38:24 +0800151 kref_put(&ubufs->kref, vhost_net_zerocopy_done_signal);
Asias He28394002013-04-27 15:07:46 +0800152 wait_event(ubufs->wait, !atomic_read(&ubufs->kref.refcount));
153 kfree(ubufs);
154}
155
Asias Heb1ad8492013-05-06 11:16:00 +0800156static void vhost_net_clear_ubuf_info(struct vhost_net *n)
157{
Asias Heb1ad8492013-05-06 11:16:00 +0800158 int i;
159
Michael S. Tsirkin288cfe72013-06-06 15:20:46 +0300160 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
161 kfree(n->vqs[i].ubuf_info);
162 n->vqs[i].ubuf_info = NULL;
Asias Heb1ad8492013-05-06 11:16:00 +0800163 }
164}
165
Asias He28394002013-04-27 15:07:46 +0800166int vhost_net_set_ubuf_info(struct vhost_net *n)
167{
168 bool zcopy;
169 int i;
170
Michael S. Tsirkin288cfe72013-06-06 15:20:46 +0300171 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
Asias Hefe729a52013-05-06 16:38:24 +0800172 zcopy = vhost_net_zcopy_mask & (0x1 << i);
Asias He28394002013-04-27 15:07:46 +0800173 if (!zcopy)
174 continue;
175 n->vqs[i].ubuf_info = kmalloc(sizeof(*n->vqs[i].ubuf_info) *
176 UIO_MAXIOV, GFP_KERNEL);
177 if (!n->vqs[i].ubuf_info)
178 goto err;
179 }
180 return 0;
181
182err:
Michael S. Tsirkin288cfe72013-06-06 15:20:46 +0300183 vhost_net_clear_ubuf_info(n);
Asias He28394002013-04-27 15:07:46 +0800184 return -ENOMEM;
185}
186
Michael S. Tsirkin81f95a52013-04-28 15:51:40 +0300187void vhost_net_vq_reset(struct vhost_net *n)
Asias He28394002013-04-27 15:07:46 +0800188{
189 int i;
190
Michael S. Tsirkin288cfe72013-06-06 15:20:46 +0300191 vhost_net_clear_ubuf_info(n);
192
Asias He28394002013-04-27 15:07:46 +0800193 for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
194 n->vqs[i].done_idx = 0;
195 n->vqs[i].upend_idx = 0;
196 n->vqs[i].ubufs = NULL;
Michael S. Tsirkin81f95a52013-04-28 15:51:40 +0300197 n->vqs[i].vhost_hlen = 0;
198 n->vqs[i].sock_hlen = 0;
Asias He28394002013-04-27 15:07:46 +0800199 }
200
201}
202
Michael S. Tsirkineaae8132012-11-01 09:16:51 +0000203static void vhost_net_tx_packet(struct vhost_net *net)
204{
205 ++net->tx_packets;
206 if (net->tx_packets < 1024)
207 return;
208 net->tx_packets = 0;
209 net->tx_zcopy_err = 0;
210}
211
212static void vhost_net_tx_err(struct vhost_net *net)
213{
214 ++net->tx_zcopy_err;
215}
216
217static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
218{
Michael S. Tsirkin1280c272012-12-04 00:17:14 +0200219 /* TX flush waits for outstanding DMAs to be done.
220 * Don't start new DMAs.
221 */
222 return !net->tx_flush &&
223 net->tx_packets / 64 >= net->tx_zcopy_err;
Michael S. Tsirkineaae8132012-11-01 09:16:51 +0000224}
225
Michael S. Tsirkinbab632d2011-07-18 03:48:46 +0000226static bool vhost_sock_zcopy(struct socket *sock)
227{
228 return unlikely(experimental_zcopytx) &&
229 sock_flag(sock->sk, SOCK_ZEROCOPY);
230}
231
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000232/* Pop first len bytes from iovec. Return number of segments used. */
233static int move_iovec_hdr(struct iovec *from, struct iovec *to,
234 size_t len, int iov_count)
235{
236 int seg = 0;
237 size_t size;
Krishna Kumard47effe2011-03-01 17:06:37 +0530238
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000239 while (len && seg < iov_count) {
240 size = min(from->iov_len, len);
241 to->iov_base = from->iov_base;
242 to->iov_len = size;
243 from->iov_len -= size;
244 from->iov_base += size;
245 len -= size;
246 ++from;
247 ++to;
248 ++seg;
249 }
250 return seg;
251}
David Stevens8dd014a2010-07-27 18:52:21 +0300252/* Copy iovec entries for len bytes from iovec. */
253static void copy_iovec_hdr(const struct iovec *from, struct iovec *to,
254 size_t len, int iovcount)
255{
256 int seg = 0;
257 size_t size;
Krishna Kumard47effe2011-03-01 17:06:37 +0530258
David Stevens8dd014a2010-07-27 18:52:21 +0300259 while (len && seg < iovcount) {
260 size = min(from->iov_len, len);
261 to->iov_base = from->iov_base;
262 to->iov_len = size;
263 len -= size;
264 ++from;
265 ++to;
266 ++seg;
267 }
268}
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000269
Michael S. Tsirkinb2116162012-11-01 09:16:46 +0000270/* In case of DMA done not in order in lower device driver for some reason.
271 * upend_idx is used to track end of used idx, done_idx is used to track head
272 * of used idx. Once lower device DMA done contiguously, we will signal KVM
273 * guest used idx.
274 */
Michael S. Tsirkineaae8132012-11-01 09:16:51 +0000275static int vhost_zerocopy_signal_used(struct vhost_net *net,
276 struct vhost_virtqueue *vq)
Michael S. Tsirkinb2116162012-11-01 09:16:46 +0000277{
Asias He28394002013-04-27 15:07:46 +0800278 struct vhost_net_virtqueue *nvq =
279 container_of(vq, struct vhost_net_virtqueue, vq);
Michael S. Tsirkinb2116162012-11-01 09:16:46 +0000280 int i;
281 int j = 0;
282
Asias He28394002013-04-27 15:07:46 +0800283 for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
Michael S. Tsirkineaae8132012-11-01 09:16:51 +0000284 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
285 vhost_net_tx_err(net);
Michael S. Tsirkinb2116162012-11-01 09:16:46 +0000286 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
287 vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
288 vhost_add_used_and_signal(vq->dev, vq,
289 vq->heads[i].id, 0);
290 ++j;
291 } else
292 break;
293 }
294 if (j)
Asias He28394002013-04-27 15:07:46 +0800295 nvq->done_idx = i;
Michael S. Tsirkinb2116162012-11-01 09:16:46 +0000296 return j;
297}
298
Michael S. Tsirkineaae8132012-11-01 09:16:51 +0000299static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
Michael S. Tsirkinb2116162012-11-01 09:16:46 +0000300{
Asias Hefe729a52013-05-06 16:38:24 +0800301 struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
Michael S. Tsirkinb2116162012-11-01 09:16:46 +0000302 struct vhost_virtqueue *vq = ubufs->vq;
Michael S. Tsirkin24eb21a2012-11-01 09:16:55 +0000303 int cnt = atomic_read(&ubufs->kref.refcount);
Michael S. Tsirkinb2116162012-11-01 09:16:46 +0000304
Michael S. Tsirkin24eb21a2012-11-01 09:16:55 +0000305 /*
306 * Trigger polling thread if guest stopped submitting new buffers:
307 * in this case, the refcount after decrement will eventually reach 1
308 * so here it is 2.
309 * We also trigger polling periodically after each 16 packets
310 * (the value 16 here is more or less arbitrary, it's tuned to trigger
311 * less than 10% of times).
312 */
313 if (cnt <= 2 || !(cnt % 16))
314 vhost_poll_queue(&vq->poll);
Michael S. Tsirkinb2116162012-11-01 09:16:46 +0000315 /* set len to mark this desc buffers done DMA */
Michael S. Tsirkineaae8132012-11-01 09:16:51 +0000316 vq->heads[ubuf->desc].len = success ?
317 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
Asias Hefe729a52013-05-06 16:38:24 +0800318 vhost_net_ubuf_put(ubufs);
Michael S. Tsirkinb2116162012-11-01 09:16:46 +0000319}
320
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000321/* Expects to be always run from workqueue - which acts as
322 * read-size critical section for our kind of RCU. */
323static void handle_tx(struct vhost_net *net)
324{
Asias He28394002013-04-27 15:07:46 +0800325 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
Michael S. Tsirkin81f95a52013-04-28 15:51:40 +0300326 struct vhost_virtqueue *vq = &nvq->vq;
Michael S. Tsirkind5675bd2010-06-24 16:59:59 +0300327 unsigned out, in, s;
328 int head;
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000329 struct msghdr msg = {
330 .msg_name = NULL,
331 .msg_namelen = 0,
332 .msg_control = NULL,
333 .msg_controllen = 0,
334 .msg_iov = vq->iov,
335 .msg_flags = MSG_DONTWAIT,
336 };
337 size_t len, total_len = 0;
Jason Wang70181d512013-04-10 20:50:48 +0000338 int err;
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000339 size_t hdr_size;
Arnd Bergmann28457ee2010-03-09 19:24:45 +0100340 struct socket *sock;
Asias Hefe729a52013-05-06 16:38:24 +0800341 struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
Michael S. Tsirkincedb9bd2012-12-06 17:00:18 +0200342 bool zcopy, zcopy_used;
Arnd Bergmann28457ee2010-03-09 19:24:45 +0100343
Michael S. Tsirkin5e182472011-01-18 13:04:43 +0200344 /* TODO: check that we are running from vhost_worker? */
Michael S. Tsirkin11cd1a82010-11-14 17:31:52 +0200345 sock = rcu_dereference_check(vq->private_data, 1);
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000346 if (!sock)
347 return;
348
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000349 mutex_lock(&vq->mutex);
Michael S. Tsirkin8ea8cf82011-05-20 02:10:54 +0300350 vhost_disable_notify(&net->dev, vq);
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000351
Michael S. Tsirkin81f95a52013-04-28 15:51:40 +0300352 hdr_size = nvq->vhost_hlen;
Asias He28394002013-04-27 15:07:46 +0800353 zcopy = nvq->ubufs;
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000354
355 for (;;) {
Michael S. Tsirkinbab632d2011-07-18 03:48:46 +0000356 /* Release DMAs done buffers first */
357 if (zcopy)
Michael S. Tsirkineaae8132012-11-01 09:16:51 +0000358 vhost_zerocopy_signal_used(net, vq);
Michael S. Tsirkinbab632d2011-07-18 03:48:46 +0000359
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000360 head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
361 ARRAY_SIZE(vq->iov),
362 &out, &in,
363 NULL, NULL);
Michael S. Tsirkind5675bd2010-06-24 16:59:59 +0300364 /* On error, stop handling until the next kick. */
Michael S. Tsirkin7b3384f2010-07-01 18:40:12 +0300365 if (unlikely(head < 0))
Michael S. Tsirkind5675bd2010-06-24 16:59:59 +0300366 break;
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000367 /* Nothing new? Wait for eventfd to tell us they refilled. */
368 if (head == vq->num) {
Shirley Ma9e380822011-07-20 10:23:12 -0700369 int num_pends;
370
Shirley Ma9e380822011-07-20 10:23:12 -0700371 /* If more outstanding DMAs, queue the work.
372 * Handle upend_idx wrap around
373 */
Asias He28394002013-04-27 15:07:46 +0800374 num_pends = likely(nvq->upend_idx >= nvq->done_idx) ?
375 (nvq->upend_idx - nvq->done_idx) :
376 (nvq->upend_idx + UIO_MAXIOV -
377 nvq->done_idx);
Jason Wang70181d512013-04-10 20:50:48 +0000378 if (unlikely(num_pends > VHOST_MAX_PEND))
Michael S. Tsirkinbab632d2011-07-18 03:48:46 +0000379 break;
Michael S. Tsirkin8ea8cf82011-05-20 02:10:54 +0300380 if (unlikely(vhost_enable_notify(&net->dev, vq))) {
381 vhost_disable_notify(&net->dev, vq);
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000382 continue;
383 }
384 break;
385 }
386 if (in) {
387 vq_err(vq, "Unexpected descriptor format for TX: "
388 "out %d, int %d\n", out, in);
389 break;
390 }
391 /* Skip header. TODO: support TSO. */
Michael S. Tsirkin81f95a52013-04-28 15:51:40 +0300392 s = move_iovec_hdr(vq->iov, nvq->hdr, hdr_size, out);
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000393 msg.msg_iovlen = out;
394 len = iov_length(vq->iov, out);
395 /* Sanity check */
396 if (!len) {
397 vq_err(vq, "Unexpected header len for TX: "
398 "%zd expected %zd\n",
Michael S. Tsirkin81f95a52013-04-28 15:51:40 +0300399 iov_length(nvq->hdr, s), hdr_size);
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000400 break;
401 }
Michael S. Tsirkincedb9bd2012-12-06 17:00:18 +0200402 zcopy_used = zcopy && (len >= VHOST_GOODCOPY_LEN ||
Asias He28394002013-04-27 15:07:46 +0800403 nvq->upend_idx != nvq->done_idx);
Michael S. Tsirkincedb9bd2012-12-06 17:00:18 +0200404
Michael S. Tsirkinbab632d2011-07-18 03:48:46 +0000405 /* use msg_control to pass vhost zerocopy ubuf info to skb */
Michael S. Tsirkincedb9bd2012-12-06 17:00:18 +0200406 if (zcopy_used) {
Asias He28394002013-04-27 15:07:46 +0800407 vq->heads[nvq->upend_idx].id = head;
Michael S. Tsirkineaae8132012-11-01 09:16:51 +0000408 if (!vhost_net_tx_select_zcopy(net) ||
409 len < VHOST_GOODCOPY_LEN) {
Michael S. Tsirkinbab632d2011-07-18 03:48:46 +0000410 /* copy don't need to wait for DMA done */
Asias He28394002013-04-27 15:07:46 +0800411 vq->heads[nvq->upend_idx].len =
Michael S. Tsirkinbab632d2011-07-18 03:48:46 +0000412 VHOST_DMA_DONE_LEN;
413 msg.msg_control = NULL;
414 msg.msg_controllen = 0;
415 ubufs = NULL;
416 } else {
Michael S. Tsirkin46aa92d2013-03-17 02:46:09 +0000417 struct ubuf_info *ubuf;
Asias He28394002013-04-27 15:07:46 +0800418 ubuf = nvq->ubuf_info + nvq->upend_idx;
Michael S. Tsirkinbab632d2011-07-18 03:48:46 +0000419
Asias He28394002013-04-27 15:07:46 +0800420 vq->heads[nvq->upend_idx].len =
Michael S. Tsirkin70e4cb92012-11-01 09:16:37 +0000421 VHOST_DMA_IN_PROGRESS;
Michael S. Tsirkinbab632d2011-07-18 03:48:46 +0000422 ubuf->callback = vhost_zerocopy_callback;
Asias He28394002013-04-27 15:07:46 +0800423 ubuf->ctx = nvq->ubufs;
424 ubuf->desc = nvq->upend_idx;
Michael S. Tsirkinbab632d2011-07-18 03:48:46 +0000425 msg.msg_control = ubuf;
426 msg.msg_controllen = sizeof(ubuf);
Asias He28394002013-04-27 15:07:46 +0800427 ubufs = nvq->ubufs;
Michael S. Tsirkinbab632d2011-07-18 03:48:46 +0000428 kref_get(&ubufs->kref);
429 }
Asias He28394002013-04-27 15:07:46 +0800430 nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
Jason Wang4364d5f2013-06-05 15:40:46 +0800431 } else
432 msg.msg_control = NULL;
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000433 /* TODO: Check specific error and bomb out unless ENOBUFS? */
434 err = sock->ops->sendmsg(NULL, sock, &msg, len);
435 if (unlikely(err < 0)) {
Michael S. Tsirkincedb9bd2012-12-06 17:00:18 +0200436 if (zcopy_used) {
Michael S. Tsirkinbab632d2011-07-18 03:48:46 +0000437 if (ubufs)
Asias Hefe729a52013-05-06 16:38:24 +0800438 vhost_net_ubuf_put(ubufs);
Asias He28394002013-04-27 15:07:46 +0800439 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
440 % UIO_MAXIOV;
Michael S. Tsirkinbab632d2011-07-18 03:48:46 +0000441 }
David Stevens8dd014a2010-07-27 18:52:21 +0300442 vhost_discard_vq_desc(vq, 1);
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000443 break;
444 }
445 if (err != len)
Michael S. Tsirkin95c0ec62010-06-24 17:10:25 +0300446 pr_debug("Truncated TX packet: "
447 " len %d != %zd\n", err, len);
Michael S. Tsirkincedb9bd2012-12-06 17:00:18 +0200448 if (!zcopy_used)
Michael S. Tsirkinbab632d2011-07-18 03:48:46 +0000449 vhost_add_used_and_signal(&net->dev, vq, head, 0);
Jason Wangc8fb2172012-05-02 11:42:41 +0800450 else
Michael S. Tsirkineaae8132012-11-01 09:16:51 +0000451 vhost_zerocopy_signal_used(net, vq);
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000452 total_len += len;
Michael S. Tsirkineaae8132012-11-01 09:16:51 +0000453 vhost_net_tx_packet(net);
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000454 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
455 vhost_poll_queue(&vq->poll);
456 break;
457 }
458 }
459
460 mutex_unlock(&vq->mutex);
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000461}
462
David Stevens8dd014a2010-07-27 18:52:21 +0300463static int peek_head_len(struct sock *sk)
464{
465 struct sk_buff *head;
466 int len = 0;
Jason Wang783e3982011-01-17 16:11:17 +0800467 unsigned long flags;
David Stevens8dd014a2010-07-27 18:52:21 +0300468
Jason Wang783e3982011-01-17 16:11:17 +0800469 spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
David Stevens8dd014a2010-07-27 18:52:21 +0300470 head = skb_peek(&sk->sk_receive_queue);
Basil Gorc53cff5e2012-05-03 22:55:23 +0000471 if (likely(head)) {
David Stevens8dd014a2010-07-27 18:52:21 +0300472 len = head->len;
Basil Gorc53cff5e2012-05-03 22:55:23 +0000473 if (vlan_tx_tag_present(head))
474 len += VLAN_HLEN;
475 }
476
Jason Wang783e3982011-01-17 16:11:17 +0800477 spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
David Stevens8dd014a2010-07-27 18:52:21 +0300478 return len;
479}
480
481/* This is a multi-buffer version of vhost_get_desc, that works if
482 * vq has read descriptors only.
483 * @vq - the relevant virtqueue
484 * @datalen - data length we'll be reading
485 * @iovcount - returned count of io vectors we fill
486 * @log - vhost log
487 * @log_num - log offset
Jason Wang94249362011-01-17 16:11:08 +0800488 * @quota - headcount quota, 1 for big buffer
David Stevens8dd014a2010-07-27 18:52:21 +0300489 * returns number of buffer heads allocated, negative on error
490 */
491static int get_rx_bufs(struct vhost_virtqueue *vq,
492 struct vring_used_elem *heads,
493 int datalen,
494 unsigned *iovcount,
495 struct vhost_log *log,
Jason Wang94249362011-01-17 16:11:08 +0800496 unsigned *log_num,
497 unsigned int quota)
David Stevens8dd014a2010-07-27 18:52:21 +0300498{
499 unsigned int out, in;
500 int seg = 0;
501 int headcount = 0;
502 unsigned d;
503 int r, nlogs = 0;
504
Jason Wang94249362011-01-17 16:11:08 +0800505 while (datalen > 0 && headcount < quota) {
Jason Wange0e9b402010-09-14 23:53:05 +0800506 if (unlikely(seg >= UIO_MAXIOV)) {
David Stevens8dd014a2010-07-27 18:52:21 +0300507 r = -ENOBUFS;
508 goto err;
509 }
510 d = vhost_get_vq_desc(vq->dev, vq, vq->iov + seg,
511 ARRAY_SIZE(vq->iov) - seg, &out,
512 &in, log, log_num);
513 if (d == vq->num) {
514 r = 0;
515 goto err;
516 }
517 if (unlikely(out || in <= 0)) {
518 vq_err(vq, "unexpected descriptor format for RX: "
519 "out %d, in %d\n", out, in);
520 r = -EINVAL;
521 goto err;
522 }
523 if (unlikely(log)) {
524 nlogs += *log_num;
525 log += *log_num;
526 }
527 heads[headcount].id = d;
528 heads[headcount].len = iov_length(vq->iov + seg, in);
529 datalen -= heads[headcount].len;
530 ++headcount;
531 seg += in;
532 }
533 heads[headcount - 1].len += datalen;
534 *iovcount = seg;
535 if (unlikely(log))
536 *log_num = nlogs;
537 return headcount;
538err:
539 vhost_discard_vq_desc(vq, headcount);
540 return r;
541}
542
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000543/* Expects to be always run from workqueue - which acts as
544 * read-size critical section for our kind of RCU. */
Jason Wang94249362011-01-17 16:11:08 +0800545static void handle_rx(struct vhost_net *net)
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000546{
Michael S. Tsirkin81f95a52013-04-28 15:51:40 +0300547 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
548 struct vhost_virtqueue *vq = &nvq->vq;
David Stevens8dd014a2010-07-27 18:52:21 +0300549 unsigned uninitialized_var(in), log;
550 struct vhost_log *vq_log;
551 struct msghdr msg = {
552 .msg_name = NULL,
553 .msg_namelen = 0,
554 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
555 .msg_controllen = 0,
556 .msg_iov = vq->iov,
557 .msg_flags = MSG_DONTWAIT,
558 };
David Stevens8dd014a2010-07-27 18:52:21 +0300559 struct virtio_net_hdr_mrg_rxbuf hdr = {
560 .hdr.flags = 0,
561 .hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
562 };
David Stevens8dd014a2010-07-27 18:52:21 +0300563 size_t total_len = 0;
Michael S. Tsirkin910a5782012-10-24 20:37:51 +0200564 int err, mergeable;
565 s16 headcount;
David Stevens8dd014a2010-07-27 18:52:21 +0300566 size_t vhost_hlen, sock_hlen;
567 size_t vhost_len, sock_len;
Michael S. Tsirkin5e182472011-01-18 13:04:43 +0200568 /* TODO: check that we are running from vhost_worker? */
569 struct socket *sock = rcu_dereference_check(vq->private_data, 1);
Krishna Kumard47effe2011-03-01 17:06:37 +0530570
Michael S. Tsirkinde4d7682011-03-13 23:00:52 +0200571 if (!sock)
David Stevens8dd014a2010-07-27 18:52:21 +0300572 return;
573
David Stevens8dd014a2010-07-27 18:52:21 +0300574 mutex_lock(&vq->mutex);
Michael S. Tsirkin8ea8cf82011-05-20 02:10:54 +0300575 vhost_disable_notify(&net->dev, vq);
Michael S. Tsirkin81f95a52013-04-28 15:51:40 +0300576 vhost_hlen = nvq->vhost_hlen;
577 sock_hlen = nvq->sock_hlen;
David Stevens8dd014a2010-07-27 18:52:21 +0300578
579 vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
580 vq->log : NULL;
Jason Wangcfbdab92011-01-17 16:10:59 +0800581 mergeable = vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF);
David Stevens8dd014a2010-07-27 18:52:21 +0300582
583 while ((sock_len = peek_head_len(sock->sk))) {
584 sock_len += sock_hlen;
585 vhost_len = sock_len + vhost_hlen;
586 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
Jason Wang94249362011-01-17 16:11:08 +0800587 &in, vq_log, &log,
588 likely(mergeable) ? UIO_MAXIOV : 1);
David Stevens8dd014a2010-07-27 18:52:21 +0300589 /* On error, stop handling until the next kick. */
590 if (unlikely(headcount < 0))
591 break;
592 /* OK, now we need to know about added descriptors. */
593 if (!headcount) {
Michael S. Tsirkin8ea8cf82011-05-20 02:10:54 +0300594 if (unlikely(vhost_enable_notify(&net->dev, vq))) {
David Stevens8dd014a2010-07-27 18:52:21 +0300595 /* They have slipped one in as we were
596 * doing that: check again. */
Michael S. Tsirkin8ea8cf82011-05-20 02:10:54 +0300597 vhost_disable_notify(&net->dev, vq);
David Stevens8dd014a2010-07-27 18:52:21 +0300598 continue;
599 }
600 /* Nothing new? Wait for eventfd to tell us
601 * they refilled. */
602 break;
603 }
604 /* We don't need to be notified again. */
605 if (unlikely((vhost_hlen)))
606 /* Skip header. TODO: support TSO. */
Michael S. Tsirkin81f95a52013-04-28 15:51:40 +0300607 move_iovec_hdr(vq->iov, nvq->hdr, vhost_hlen, in);
David Stevens8dd014a2010-07-27 18:52:21 +0300608 else
609 /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
Jason Wanga290aec2010-11-29 13:48:40 +0800610 * needed because recvmsg can modify msg_iov. */
Michael S. Tsirkin81f95a52013-04-28 15:51:40 +0300611 copy_iovec_hdr(vq->iov, nvq->hdr, sock_hlen, in);
David Stevens8dd014a2010-07-27 18:52:21 +0300612 msg.msg_iovlen = in;
613 err = sock->ops->recvmsg(NULL, sock, &msg,
614 sock_len, MSG_DONTWAIT | MSG_TRUNC);
615 /* Userspace might have consumed the packet meanwhile:
616 * it's not supposed to do this usually, but might be hard
617 * to prevent. Discard data we got (if any) and keep going. */
618 if (unlikely(err != sock_len)) {
619 pr_debug("Discarded rx packet: "
620 " len %d, expected %zd\n", err, sock_len);
621 vhost_discard_vq_desc(vq, headcount);
622 continue;
623 }
624 if (unlikely(vhost_hlen) &&
Michael S. Tsirkin81f95a52013-04-28 15:51:40 +0300625 memcpy_toiovecend(nvq->hdr, (unsigned char *)&hdr, 0,
David Stevens8dd014a2010-07-27 18:52:21 +0300626 vhost_hlen)) {
627 vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
628 vq->iov->iov_base);
629 break;
630 }
631 /* TODO: Should check and handle checksum. */
Jason Wangcfbdab92011-01-17 16:10:59 +0800632 if (likely(mergeable) &&
Michael S. Tsirkin81f95a52013-04-28 15:51:40 +0300633 memcpy_toiovecend(nvq->hdr, (unsigned char *)&headcount,
David Stevens8dd014a2010-07-27 18:52:21 +0300634 offsetof(typeof(hdr), num_buffers),
635 sizeof hdr.num_buffers)) {
636 vq_err(vq, "Failed num_buffers write");
637 vhost_discard_vq_desc(vq, headcount);
638 break;
639 }
640 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
641 headcount);
642 if (unlikely(vq_log))
643 vhost_log_write(vq, vq_log, log, vhost_len);
644 total_len += vhost_len;
645 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
646 vhost_poll_queue(&vq->poll);
647 break;
648 }
649 }
650
651 mutex_unlock(&vq->mutex);
David Stevens8dd014a2010-07-27 18:52:21 +0300652}
653
Tejun Heoc23f34452010-06-02 20:40:00 +0200654static void handle_tx_kick(struct vhost_work *work)
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000655{
Tejun Heoc23f34452010-06-02 20:40:00 +0200656 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
657 poll.work);
658 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
659
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000660 handle_tx(net);
661}
662
Tejun Heoc23f34452010-06-02 20:40:00 +0200663static void handle_rx_kick(struct vhost_work *work)
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000664{
Tejun Heoc23f34452010-06-02 20:40:00 +0200665 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
666 poll.work);
667 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
668
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000669 handle_rx(net);
670}
671
Tejun Heoc23f34452010-06-02 20:40:00 +0200672static void handle_tx_net(struct vhost_work *work)
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000673{
Tejun Heoc23f34452010-06-02 20:40:00 +0200674 struct vhost_net *net = container_of(work, struct vhost_net,
675 poll[VHOST_NET_VQ_TX].work);
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000676 handle_tx(net);
677}
678
Tejun Heoc23f34452010-06-02 20:40:00 +0200679static void handle_rx_net(struct vhost_work *work)
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000680{
Tejun Heoc23f34452010-06-02 20:40:00 +0200681 struct vhost_net *net = container_of(work, struct vhost_net,
682 poll[VHOST_NET_VQ_RX].work);
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000683 handle_rx(net);
684}
685
686static int vhost_net_open(struct inode *inode, struct file *f)
687{
688 struct vhost_net *n = kmalloc(sizeof *n, GFP_KERNEL);
Tejun Heoc23f34452010-06-02 20:40:00 +0200689 struct vhost_dev *dev;
Asias He3ab2e422013-04-27 11:16:48 +0800690 struct vhost_virtqueue **vqs;
Asias He28394002013-04-27 15:07:46 +0800691 int r, i;
Tejun Heoc23f34452010-06-02 20:40:00 +0200692
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000693 if (!n)
694 return -ENOMEM;
Asias He3ab2e422013-04-27 11:16:48 +0800695 vqs = kmalloc(VHOST_NET_VQ_MAX * sizeof(*vqs), GFP_KERNEL);
696 if (!vqs) {
697 kfree(n);
698 return -ENOMEM;
699 }
Tejun Heoc23f34452010-06-02 20:40:00 +0200700
701 dev = &n->dev;
Asias He3ab2e422013-04-27 11:16:48 +0800702 vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
703 vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
704 n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
705 n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
Asias He28394002013-04-27 15:07:46 +0800706 for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
707 n->vqs[i].ubufs = NULL;
708 n->vqs[i].ubuf_info = NULL;
709 n->vqs[i].upend_idx = 0;
710 n->vqs[i].done_idx = 0;
Michael S. Tsirkin81f95a52013-04-28 15:51:40 +0300711 n->vqs[i].vhost_hlen = 0;
712 n->vqs[i].sock_hlen = 0;
Asias He28394002013-04-27 15:07:46 +0800713 }
Asias He3ab2e422013-04-27 11:16:48 +0800714 r = vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX);
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000715 if (r < 0) {
716 kfree(n);
Asias He3ab2e422013-04-27 11:16:48 +0800717 kfree(vqs);
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000718 return r;
719 }
720
Tejun Heoc23f34452010-06-02 20:40:00 +0200721 vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
722 vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000723
724 f->private_data = n;
725
726 return 0;
727}
728
729static void vhost_net_disable_vq(struct vhost_net *n,
730 struct vhost_virtqueue *vq)
731{
Asias He3ab2e422013-04-27 11:16:48 +0800732 struct vhost_net_virtqueue *nvq =
733 container_of(vq, struct vhost_net_virtqueue, vq);
734 struct vhost_poll *poll = n->poll + (nvq - n->vqs);
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000735 if (!vq->private_data)
736 return;
Jason Wang70181d512013-04-10 20:50:48 +0000737 vhost_poll_stop(poll);
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000738}
739
Jason Wang2b8b3282013-01-28 01:05:18 +0000740static int vhost_net_enable_vq(struct vhost_net *n,
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000741 struct vhost_virtqueue *vq)
742{
Asias He3ab2e422013-04-27 11:16:48 +0800743 struct vhost_net_virtqueue *nvq =
744 container_of(vq, struct vhost_net_virtqueue, vq);
745 struct vhost_poll *poll = n->poll + (nvq - n->vqs);
Arnd Bergmann28457ee2010-03-09 19:24:45 +0100746 struct socket *sock;
747
748 sock = rcu_dereference_protected(vq->private_data,
749 lockdep_is_held(&vq->mutex));
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000750 if (!sock)
Jason Wang2b8b3282013-01-28 01:05:18 +0000751 return 0;
Jason Wang2b8b3282013-01-28 01:05:18 +0000752
Jason Wang70181d512013-04-10 20:50:48 +0000753 return vhost_poll_start(poll, sock->file);
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000754}
755
756static struct socket *vhost_net_stop_vq(struct vhost_net *n,
757 struct vhost_virtqueue *vq)
758{
759 struct socket *sock;
760
761 mutex_lock(&vq->mutex);
Arnd Bergmann28457ee2010-03-09 19:24:45 +0100762 sock = rcu_dereference_protected(vq->private_data,
763 lockdep_is_held(&vq->mutex));
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000764 vhost_net_disable_vq(n, vq);
765 rcu_assign_pointer(vq->private_data, NULL);
766 mutex_unlock(&vq->mutex);
767 return sock;
768}
769
770static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
771 struct socket **rx_sock)
772{
Asias He3ab2e422013-04-27 11:16:48 +0800773 *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
774 *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000775}
776
777static void vhost_net_flush_vq(struct vhost_net *n, int index)
778{
779 vhost_poll_flush(n->poll + index);
Asias He3ab2e422013-04-27 11:16:48 +0800780 vhost_poll_flush(&n->vqs[index].vq.poll);
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000781}
782
783static void vhost_net_flush(struct vhost_net *n)
784{
785 vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
786 vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
Asias He28394002013-04-27 15:07:46 +0800787 if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
Asias He3ab2e422013-04-27 11:16:48 +0800788 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
Michael S. Tsirkin1280c272012-12-04 00:17:14 +0200789 n->tx_flush = true;
Asias He3ab2e422013-04-27 11:16:48 +0800790 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
Michael S. Tsirkin1280c272012-12-04 00:17:14 +0200791 /* Wait for all lower device DMAs done. */
Asias Hefe729a52013-05-06 16:38:24 +0800792 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
Asias He3ab2e422013-04-27 11:16:48 +0800793 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
Michael S. Tsirkin1280c272012-12-04 00:17:14 +0200794 n->tx_flush = false;
Asias He28394002013-04-27 15:07:46 +0800795 kref_init(&n->vqs[VHOST_NET_VQ_TX].ubufs->kref);
Asias He3ab2e422013-04-27 11:16:48 +0800796 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
Michael S. Tsirkin1280c272012-12-04 00:17:14 +0200797 }
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000798}
799
800static int vhost_net_release(struct inode *inode, struct file *f)
801{
802 struct vhost_net *n = f->private_data;
803 struct socket *tx_sock;
804 struct socket *rx_sock;
805
806 vhost_net_stop(n, &tx_sock, &rx_sock);
807 vhost_net_flush(n);
Michael S. Tsirkinb2116162012-11-01 09:16:46 +0000808 vhost_dev_stop(&n->dev);
Michael S. Tsirkinea5d4042011-11-27 19:05:58 +0200809 vhost_dev_cleanup(&n->dev, false);
Michael S. Tsirkin81f95a52013-04-28 15:51:40 +0300810 vhost_net_vq_reset(n);
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000811 if (tx_sock)
812 fput(tx_sock->file);
813 if (rx_sock)
814 fput(rx_sock->file);
815 /* We do an extra flush before freeing memory,
816 * since jobs can re-queue themselves. */
817 vhost_net_flush(n);
Asias He3ab2e422013-04-27 11:16:48 +0800818 kfree(n->dev.vqs);
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000819 kfree(n);
820 return 0;
821}
822
823static struct socket *get_raw_socket(int fd)
824{
825 struct {
826 struct sockaddr_ll sa;
827 char buf[MAX_ADDR_LEN];
828 } uaddr;
829 int uaddr_len = sizeof uaddr, r;
830 struct socket *sock = sockfd_lookup(fd, &r);
Krishna Kumard47effe2011-03-01 17:06:37 +0530831
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000832 if (!sock)
833 return ERR_PTR(-ENOTSOCK);
834
835 /* Parameter checking */
836 if (sock->sk->sk_type != SOCK_RAW) {
837 r = -ESOCKTNOSUPPORT;
838 goto err;
839 }
840
841 r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
842 &uaddr_len, 0);
843 if (r)
844 goto err;
845
846 if (uaddr.sa.sll_family != AF_PACKET) {
847 r = -EPFNOSUPPORT;
848 goto err;
849 }
850 return sock;
851err:
852 fput(sock->file);
853 return ERR_PTR(r);
854}
855
Arnd Bergmann501c7742010-02-18 05:46:50 +0000856static struct socket *get_tap_socket(int fd)
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000857{
858 struct file *file = fget(fd);
859 struct socket *sock;
Krishna Kumard47effe2011-03-01 17:06:37 +0530860
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000861 if (!file)
862 return ERR_PTR(-EBADF);
863 sock = tun_get_socket(file);
Arnd Bergmann501c7742010-02-18 05:46:50 +0000864 if (!IS_ERR(sock))
865 return sock;
866 sock = macvtap_get_socket(file);
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000867 if (IS_ERR(sock))
868 fput(file);
869 return sock;
870}
871
872static struct socket *get_socket(int fd)
873{
874 struct socket *sock;
Krishna Kumard47effe2011-03-01 17:06:37 +0530875
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000876 /* special case to disable backend */
877 if (fd == -1)
878 return NULL;
879 sock = get_raw_socket(fd);
880 if (!IS_ERR(sock))
881 return sock;
Arnd Bergmann501c7742010-02-18 05:46:50 +0000882 sock = get_tap_socket(fd);
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000883 if (!IS_ERR(sock))
884 return sock;
885 return ERR_PTR(-ENOTSOCK);
886}
887
888static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
889{
890 struct socket *sock, *oldsock;
891 struct vhost_virtqueue *vq;
Asias He28394002013-04-27 15:07:46 +0800892 struct vhost_net_virtqueue *nvq;
Asias Hefe729a52013-05-06 16:38:24 +0800893 struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000894 int r;
895
896 mutex_lock(&n->dev.mutex);
897 r = vhost_dev_check_owner(&n->dev);
898 if (r)
899 goto err;
900
901 if (index >= VHOST_NET_VQ_MAX) {
902 r = -ENOBUFS;
903 goto err;
904 }
Asias He3ab2e422013-04-27 11:16:48 +0800905 vq = &n->vqs[index].vq;
Asias He28394002013-04-27 15:07:46 +0800906 nvq = &n->vqs[index];
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000907 mutex_lock(&vq->mutex);
908
909 /* Verify that ring has been setup correctly. */
910 if (!vhost_vq_access_ok(vq)) {
911 r = -EFAULT;
Jeff Dike1dace8c2010-03-04 16:10:14 -0500912 goto err_vq;
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000913 }
914 sock = get_socket(fd);
915 if (IS_ERR(sock)) {
916 r = PTR_ERR(sock);
Jeff Dike1dace8c2010-03-04 16:10:14 -0500917 goto err_vq;
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000918 }
919
920 /* start polling new socket */
Arnd Bergmann28457ee2010-03-09 19:24:45 +0100921 oldsock = rcu_dereference_protected(vq->private_data,
922 lockdep_is_held(&vq->mutex));
David S. Miller11fe8832010-07-20 18:25:24 -0700923 if (sock != oldsock) {
Asias Hefe729a52013-05-06 16:38:24 +0800924 ubufs = vhost_net_ubuf_alloc(vq,
925 sock && vhost_sock_zcopy(sock));
Michael S. Tsirkinbab632d2011-07-18 03:48:46 +0000926 if (IS_ERR(ubufs)) {
927 r = PTR_ERR(ubufs);
928 goto err_ubufs;
929 }
Jason Wang692a9982013-01-28 01:05:17 +0000930
Krishna Kumard47effe2011-03-01 17:06:37 +0530931 vhost_net_disable_vq(n, vq);
932 rcu_assign_pointer(vq->private_data, sock);
Jason Wangf59281d2011-06-21 18:04:27 +0800933 r = vhost_init_used(vq);
934 if (r)
Jason Wang692a9982013-01-28 01:05:17 +0000935 goto err_used;
Jason Wang2b8b3282013-01-28 01:05:18 +0000936 r = vhost_net_enable_vq(n, vq);
937 if (r)
938 goto err_used;
Jason Wang692a9982013-01-28 01:05:17 +0000939
Asias He28394002013-04-27 15:07:46 +0800940 oldubufs = nvq->ubufs;
941 nvq->ubufs = ubufs;
Michael S. Tsirkin64e9a9b2012-12-03 07:31:51 +0000942
943 n->tx_packets = 0;
944 n->tx_zcopy_err = 0;
Michael S. Tsirkin1280c272012-12-04 00:17:14 +0200945 n->tx_flush = false;
Jeff Dikedd1f4072010-03-04 16:10:14 -0500946 }
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000947
Michael S. Tsirkin1680e902010-07-15 15:19:12 +0300948 mutex_unlock(&vq->mutex);
949
Michael S. Tsirkinc047e5f2011-07-20 13:41:31 +0300950 if (oldubufs) {
Asias Hefe729a52013-05-06 16:38:24 +0800951 vhost_net_ubuf_put_and_wait(oldubufs);
Michael S. Tsirkinc047e5f2011-07-20 13:41:31 +0300952 mutex_lock(&vq->mutex);
Michael S. Tsirkineaae8132012-11-01 09:16:51 +0000953 vhost_zerocopy_signal_used(n, vq);
Michael S. Tsirkinc047e5f2011-07-20 13:41:31 +0300954 mutex_unlock(&vq->mutex);
955 }
Michael S. Tsirkinbab632d2011-07-18 03:48:46 +0000956
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000957 if (oldsock) {
958 vhost_net_flush_vq(n, index);
959 fput(oldsock->file);
960 }
Jeff Dike1dace8c2010-03-04 16:10:14 -0500961
Michael S. Tsirkin1680e902010-07-15 15:19:12 +0300962 mutex_unlock(&n->dev.mutex);
963 return 0;
964
Jason Wang692a9982013-01-28 01:05:17 +0000965err_used:
966 rcu_assign_pointer(vq->private_data, oldsock);
967 vhost_net_enable_vq(n, vq);
968 if (ubufs)
Asias Hefe729a52013-05-06 16:38:24 +0800969 vhost_net_ubuf_put_and_wait(ubufs);
Michael S. Tsirkinbab632d2011-07-18 03:48:46 +0000970err_ubufs:
971 fput(sock->file);
Jeff Dike1dace8c2010-03-04 16:10:14 -0500972err_vq:
973 mutex_unlock(&vq->mutex);
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000974err:
975 mutex_unlock(&n->dev.mutex);
976 return r;
977}
978
979static long vhost_net_reset_owner(struct vhost_net *n)
980{
981 struct socket *tx_sock = NULL;
982 struct socket *rx_sock = NULL;
983 long err;
Michael S. Tsirkin150b9e52013-04-28 17:12:08 +0300984 struct vhost_memory *memory;
Krishna Kumard47effe2011-03-01 17:06:37 +0530985
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000986 mutex_lock(&n->dev.mutex);
987 err = vhost_dev_check_owner(&n->dev);
988 if (err)
989 goto done;
Michael S. Tsirkin150b9e52013-04-28 17:12:08 +0300990 memory = vhost_dev_reset_owner_prepare();
991 if (!memory) {
992 err = -ENOMEM;
993 goto done;
994 }
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000995 vhost_net_stop(n, &tx_sock, &rx_sock);
996 vhost_net_flush(n);
Michael S. Tsirkin150b9e52013-04-28 17:12:08 +0300997 vhost_dev_reset_owner(&n->dev, memory);
Michael S. Tsirkin81f95a52013-04-28 15:51:40 +0300998 vhost_net_vq_reset(n);
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +0000999done:
1000 mutex_unlock(&n->dev.mutex);
1001 if (tx_sock)
1002 fput(tx_sock->file);
1003 if (rx_sock)
1004 fput(rx_sock->file);
1005 return err;
1006}
1007
1008static int vhost_net_set_features(struct vhost_net *n, u64 features)
1009{
David Stevens8dd014a2010-07-27 18:52:21 +03001010 size_t vhost_hlen, sock_hlen, hdr_len;
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +00001011 int i;
David Stevens8dd014a2010-07-27 18:52:21 +03001012
1013 hdr_len = (features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ?
1014 sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1015 sizeof(struct virtio_net_hdr);
1016 if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1017 /* vhost provides vnet_hdr */
1018 vhost_hlen = hdr_len;
1019 sock_hlen = 0;
1020 } else {
1021 /* socket provides vnet_hdr */
1022 vhost_hlen = 0;
1023 sock_hlen = hdr_len;
1024 }
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +00001025 mutex_lock(&n->dev.mutex);
1026 if ((features & (1 << VHOST_F_LOG_ALL)) &&
1027 !vhost_log_access_ok(&n->dev)) {
1028 mutex_unlock(&n->dev.mutex);
1029 return -EFAULT;
1030 }
1031 n->dev.acked_features = features;
1032 smp_wmb();
1033 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
Asias He3ab2e422013-04-27 11:16:48 +08001034 mutex_lock(&n->vqs[i].vq.mutex);
Michael S. Tsirkin81f95a52013-04-28 15:51:40 +03001035 n->vqs[i].vhost_hlen = vhost_hlen;
1036 n->vqs[i].sock_hlen = sock_hlen;
Asias He3ab2e422013-04-27 11:16:48 +08001037 mutex_unlock(&n->vqs[i].vq.mutex);
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +00001038 }
1039 vhost_net_flush(n);
1040 mutex_unlock(&n->dev.mutex);
1041 return 0;
1042}
1043
Asias Heb1ad8492013-05-06 11:16:00 +08001044static long vhost_net_set_owner(struct vhost_net *n)
1045{
1046 int r;
1047
1048 mutex_lock(&n->dev.mutex);
Michael S. Tsirkin05c05352013-06-06 15:20:39 +03001049 if (vhost_dev_has_owner(&n->dev)) {
1050 r = -EBUSY;
1051 goto out;
1052 }
Asias Heb1ad8492013-05-06 11:16:00 +08001053 r = vhost_net_set_ubuf_info(n);
1054 if (r)
1055 goto out;
1056 r = vhost_dev_set_owner(&n->dev);
1057 if (r)
1058 vhost_net_clear_ubuf_info(n);
1059 vhost_net_flush(n);
1060out:
1061 mutex_unlock(&n->dev.mutex);
1062 return r;
1063}
1064
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +00001065static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1066 unsigned long arg)
1067{
1068 struct vhost_net *n = f->private_data;
1069 void __user *argp = (void __user *)arg;
1070 u64 __user *featurep = argp;
1071 struct vhost_vring_file backend;
1072 u64 features;
1073 int r;
Krishna Kumard47effe2011-03-01 17:06:37 +05301074
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +00001075 switch (ioctl) {
1076 case VHOST_NET_SET_BACKEND:
Takuya Yoshikawad3553a52010-05-27 19:01:58 +09001077 if (copy_from_user(&backend, argp, sizeof backend))
1078 return -EFAULT;
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +00001079 return vhost_net_set_backend(n, backend.index, backend.fd);
1080 case VHOST_GET_FEATURES:
Stefan Hajnoczi0dd05a32012-07-21 06:55:36 +00001081 features = VHOST_NET_FEATURES;
Takuya Yoshikawad3553a52010-05-27 19:01:58 +09001082 if (copy_to_user(featurep, &features, sizeof features))
1083 return -EFAULT;
1084 return 0;
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +00001085 case VHOST_SET_FEATURES:
Takuya Yoshikawad3553a52010-05-27 19:01:58 +09001086 if (copy_from_user(&features, featurep, sizeof features))
1087 return -EFAULT;
Stefan Hajnoczi0dd05a32012-07-21 06:55:36 +00001088 if (features & ~VHOST_NET_FEATURES)
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +00001089 return -EOPNOTSUPP;
1090 return vhost_net_set_features(n, features);
1091 case VHOST_RESET_OWNER:
1092 return vhost_net_reset_owner(n);
Asias Heb1ad8492013-05-06 11:16:00 +08001093 case VHOST_SET_OWNER:
1094 return vhost_net_set_owner(n);
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +00001095 default:
1096 mutex_lock(&n->dev.mutex);
Michael S. Tsirkin935cdee2012-12-06 14:03:34 +02001097 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1098 if (r == -ENOIOCTLCMD)
1099 r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1100 else
1101 vhost_net_flush(n);
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +00001102 mutex_unlock(&n->dev.mutex);
1103 return r;
1104 }
1105}
1106
1107#ifdef CONFIG_COMPAT
1108static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
1109 unsigned long arg)
1110{
1111 return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
1112}
1113#endif
1114
Tobias Klauser373a83a2010-05-17 15:12:49 +02001115static const struct file_operations vhost_net_fops = {
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +00001116 .owner = THIS_MODULE,
1117 .release = vhost_net_release,
1118 .unlocked_ioctl = vhost_net_ioctl,
1119#ifdef CONFIG_COMPAT
1120 .compat_ioctl = vhost_net_compat_ioctl,
1121#endif
1122 .open = vhost_net_open,
Arnd Bergmann6038f372010-08-15 18:52:59 +02001123 .llseek = noop_llseek,
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +00001124};
1125
1126static struct miscdevice vhost_net_misc = {
stephen hemminger7c7c7f02012-01-11 19:30:38 +00001127 .minor = VHOST_NET_MINOR,
1128 .name = "vhost-net",
1129 .fops = &vhost_net_fops,
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +00001130};
1131
Christoph Hellwiga8d37822010-04-13 14:11:25 -04001132static int vhost_net_init(void)
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +00001133{
Michael S. Tsirkinbab632d2011-07-18 03:48:46 +00001134 if (experimental_zcopytx)
Asias Hefe729a52013-05-06 16:38:24 +08001135 vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
Tejun Heoc23f34452010-06-02 20:40:00 +02001136 return misc_register(&vhost_net_misc);
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +00001137}
1138module_init(vhost_net_init);
1139
Christoph Hellwiga8d37822010-04-13 14:11:25 -04001140static void vhost_net_exit(void)
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +00001141{
1142 misc_deregister(&vhost_net_misc);
Michael S. Tsirkin3a4d5c94e2010-01-14 06:17:27 +00001143}
1144module_exit(vhost_net_exit);
1145
1146MODULE_VERSION("0.0.1");
1147MODULE_LICENSE("GPL v2");
1148MODULE_AUTHOR("Michael S. Tsirkin");
1149MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
stephen hemminger7c7c7f02012-01-11 19:30:38 +00001150MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1151MODULE_ALIAS("devname:vhost-net");