net/virtio-user: fix multiple queue for vhost-kernel
[dpdk.git] / drivers / net / virtio / virtio_user / virtio_user_dev.c
1 /* SPDX-License-Identifier: BSD-3-Clause
2  * Copyright(c) 2010-2016 Intel Corporation
3  */
4
5 #include <stdint.h>
6 #include <stdio.h>
7 #include <fcntl.h>
8 #include <string.h>
9 #include <errno.h>
10 #include <sys/mman.h>
11 #include <unistd.h>
12 #include <sys/eventfd.h>
13 #include <sys/types.h>
14 #include <sys/stat.h>
15
16 #include <rte_eal_memconfig.h>
17
18 #include "vhost.h"
19 #include "virtio_user_dev.h"
20 #include "../virtio_ethdev.h"
21
22 #define VIRTIO_USER_MEM_EVENT_CLB_NAME "virtio_user_mem_event_clb"
23
24 static int
25 virtio_user_create_queue(struct virtio_user_dev *dev, uint32_t queue_sel)
26 {
27         /* Of all per virtqueue MSGs, make sure VHOST_SET_VRING_CALL come
28          * firstly because vhost depends on this msg to allocate virtqueue
29          * pair.
30          */
31         struct vhost_vring_file file;
32
33         file.index = queue_sel;
34         file.fd = dev->callfds[queue_sel];
35         dev->ops->send_request(dev, VHOST_USER_SET_VRING_CALL, &file);
36
37         return 0;
38 }
39
40 static int
41 virtio_user_kick_queue(struct virtio_user_dev *dev, uint32_t queue_sel)
42 {
43         struct vhost_vring_file file;
44         struct vhost_vring_state state;
45         struct vring *vring = &dev->vrings[queue_sel];
46         struct vhost_vring_addr addr = {
47                 .index = queue_sel,
48                 .desc_user_addr = (uint64_t)(uintptr_t)vring->desc,
49                 .avail_user_addr = (uint64_t)(uintptr_t)vring->avail,
50                 .used_user_addr = (uint64_t)(uintptr_t)vring->used,
51                 .log_guest_addr = 0,
52                 .flags = 0, /* disable log */
53         };
54
55         state.index = queue_sel;
56         state.num = vring->num;
57         dev->ops->send_request(dev, VHOST_USER_SET_VRING_NUM, &state);
58
59         state.index = queue_sel;
60         state.num = 0; /* no reservation */
61         dev->ops->send_request(dev, VHOST_USER_SET_VRING_BASE, &state);
62
63         dev->ops->send_request(dev, VHOST_USER_SET_VRING_ADDR, &addr);
64
65         /* Of all per virtqueue MSGs, make sure VHOST_USER_SET_VRING_KICK comes
66          * lastly because vhost depends on this msg to judge if
67          * virtio is ready.
68          */
69         file.index = queue_sel;
70         file.fd = dev->kickfds[queue_sel];
71         dev->ops->send_request(dev, VHOST_USER_SET_VRING_KICK, &file);
72
73         return 0;
74 }
75
76 static int
77 virtio_user_queue_setup(struct virtio_user_dev *dev,
78                         int (*fn)(struct virtio_user_dev *, uint32_t))
79 {
80         uint32_t i, queue_sel;
81
82         for (i = 0; i < dev->max_queue_pairs; ++i) {
83                 queue_sel = 2 * i + VTNET_SQ_RQ_QUEUE_IDX;
84                 if (fn(dev, queue_sel) < 0) {
85                         PMD_DRV_LOG(INFO, "setup rx vq fails: %u", i);
86                         return -1;
87                 }
88         }
89         for (i = 0; i < dev->max_queue_pairs; ++i) {
90                 queue_sel = 2 * i + VTNET_SQ_TQ_QUEUE_IDX;
91                 if (fn(dev, queue_sel) < 0) {
92                         PMD_DRV_LOG(INFO, "setup tx vq fails: %u", i);
93                         return -1;
94                 }
95         }
96
97         return 0;
98 }
99
100 int
101 is_vhost_user_by_type(const char *path)
102 {
103         struct stat sb;
104
105         if (stat(path, &sb) == -1)
106                 return 0;
107
108         return S_ISSOCK(sb.st_mode);
109 }
110
111 int
112 virtio_user_start_device(struct virtio_user_dev *dev)
113 {
114         struct rte_mem_config *mcfg = rte_eal_get_configuration()->mem_config;
115         uint64_t features;
116         int ret;
117
118         /*
119          * XXX workaround!
120          *
121          * We need to make sure that the locks will be
122          * taken in the correct order to avoid deadlocks.
123          *
124          * Before releasing this lock, this thread should
125          * not trigger any memory hotplug events.
126          *
127          * This is a temporary workaround, and should be
128          * replaced when we get proper supports from the
129          * memory subsystem in the future.
130          */
131         rte_rwlock_read_lock(&mcfg->memory_hotplug_lock);
132         pthread_mutex_lock(&dev->mutex);
133
134         if (is_vhost_user_by_type(dev->path) && dev->vhostfd < 0)
135                 goto error;
136
137         /* Do not check return as already done in init, or reset in stop */
138         dev->ops->send_request(dev, VHOST_USER_SET_OWNER, NULL);
139
140         /* Step 0: tell vhost to create queues */
141         if (virtio_user_queue_setup(dev, virtio_user_create_queue) < 0)
142                 goto error;
143
144         /* Step 1: set features */
145         features = dev->features;
146         /* Strip VIRTIO_NET_F_MAC, as MAC address is handled in vdev init */
147         features &= ~(1ull << VIRTIO_NET_F_MAC);
148         /* Strip VIRTIO_NET_F_CTRL_VQ, as devices do not really need to know */
149         features &= ~(1ull << VIRTIO_NET_F_CTRL_VQ);
150         features &= ~(1ull << VIRTIO_NET_F_STATUS);
151         ret = dev->ops->send_request(dev, VHOST_USER_SET_FEATURES, &features);
152         if (ret < 0)
153                 goto error;
154         PMD_DRV_LOG(INFO, "set features: %" PRIx64, features);
155
156         /* Step 2: share memory regions */
157         ret = dev->ops->send_request(dev, VHOST_USER_SET_MEM_TABLE, NULL);
158         if (ret < 0)
159                 goto error;
160
161         /* Step 3: kick queues */
162         if (virtio_user_queue_setup(dev, virtio_user_kick_queue) < 0)
163                 goto error;
164
165         /* Step 4: enable queues
166          * we enable the 1st queue pair by default.
167          */
168         dev->ops->enable_qp(dev, 0, 1);
169
170         dev->started = true;
171         pthread_mutex_unlock(&dev->mutex);
172         rte_rwlock_read_unlock(&mcfg->memory_hotplug_lock);
173
174         return 0;
175 error:
176         pthread_mutex_unlock(&dev->mutex);
177         rte_rwlock_read_unlock(&mcfg->memory_hotplug_lock);
178         /* TODO: free resource here or caller to check */
179         return -1;
180 }
181
182 int virtio_user_stop_device(struct virtio_user_dev *dev)
183 {
184         uint32_t i;
185
186         pthread_mutex_lock(&dev->mutex);
187         for (i = 0; i < dev->max_queue_pairs; ++i)
188                 dev->ops->enable_qp(dev, i, 0);
189
190         if (dev->ops->send_request(dev, VHOST_USER_RESET_OWNER, NULL) < 0) {
191                 PMD_DRV_LOG(INFO, "Failed to reset the device\n");
192                 pthread_mutex_unlock(&dev->mutex);
193                 return -1;
194         }
195         dev->started = false;
196         pthread_mutex_unlock(&dev->mutex);
197
198         return 0;
199 }
200
201 static inline void
202 parse_mac(struct virtio_user_dev *dev, const char *mac)
203 {
204         int i, r;
205         uint32_t tmp[ETHER_ADDR_LEN];
206
207         if (!mac)
208                 return;
209
210         r = sscanf(mac, "%x:%x:%x:%x:%x:%x", &tmp[0],
211                         &tmp[1], &tmp[2], &tmp[3], &tmp[4], &tmp[5]);
212         if (r == ETHER_ADDR_LEN) {
213                 for (i = 0; i < ETHER_ADDR_LEN; ++i)
214                         dev->mac_addr[i] = (uint8_t)tmp[i];
215                 dev->mac_specified = 1;
216         } else {
217                 /* ignore the wrong mac, use random mac */
218                 PMD_DRV_LOG(ERR, "wrong format of mac: %s", mac);
219         }
220 }
221
222 static int
223 virtio_user_dev_init_notify(struct virtio_user_dev *dev)
224 {
225         uint32_t i, j;
226         int callfd;
227         int kickfd;
228
229         for (i = 0; i < VIRTIO_MAX_VIRTQUEUES; ++i) {
230                 if (i >= dev->max_queue_pairs * 2) {
231                         dev->kickfds[i] = -1;
232                         dev->callfds[i] = -1;
233                         continue;
234                 }
235
236                 /* May use invalid flag, but some backend uses kickfd and
237                  * callfd as criteria to judge if dev is alive. so finally we
238                  * use real event_fd.
239                  */
240                 callfd = eventfd(0, EFD_CLOEXEC | EFD_NONBLOCK);
241                 if (callfd < 0) {
242                         PMD_DRV_LOG(ERR, "callfd error, %s", strerror(errno));
243                         break;
244                 }
245                 kickfd = eventfd(0, EFD_CLOEXEC | EFD_NONBLOCK);
246                 if (kickfd < 0) {
247                         PMD_DRV_LOG(ERR, "kickfd error, %s", strerror(errno));
248                         break;
249                 }
250                 dev->callfds[i] = callfd;
251                 dev->kickfds[i] = kickfd;
252         }
253
254         if (i < VIRTIO_MAX_VIRTQUEUES) {
255                 for (j = 0; j <= i; ++j) {
256                         close(dev->callfds[j]);
257                         close(dev->kickfds[j]);
258                 }
259
260                 return -1;
261         }
262
263         return 0;
264 }
265
266 static int
267 virtio_user_fill_intr_handle(struct virtio_user_dev *dev)
268 {
269         uint32_t i;
270         struct rte_eth_dev *eth_dev = &rte_eth_devices[dev->port_id];
271
272         if (!eth_dev->intr_handle) {
273                 eth_dev->intr_handle = malloc(sizeof(*eth_dev->intr_handle));
274                 if (!eth_dev->intr_handle) {
275                         PMD_DRV_LOG(ERR, "fail to allocate intr_handle");
276                         return -1;
277                 }
278                 memset(eth_dev->intr_handle, 0, sizeof(*eth_dev->intr_handle));
279         }
280
281         for (i = 0; i < dev->max_queue_pairs; ++i)
282                 eth_dev->intr_handle->efds[i] = dev->callfds[i];
283         eth_dev->intr_handle->nb_efd = dev->max_queue_pairs;
284         eth_dev->intr_handle->max_intr = dev->max_queue_pairs + 1;
285         eth_dev->intr_handle->type = RTE_INTR_HANDLE_VDEV;
286         /* For virtio vdev, no need to read counter for clean */
287         eth_dev->intr_handle->efd_counter_size = 0;
288         eth_dev->intr_handle->fd = -1;
289         if (dev->vhostfd >= 0)
290                 eth_dev->intr_handle->fd = dev->vhostfd;
291         else if (dev->is_server)
292                 eth_dev->intr_handle->fd = dev->listenfd;
293
294         return 0;
295 }
296
297 static void
298 virtio_user_mem_event_cb(enum rte_mem_event type __rte_unused,
299                                                  const void *addr __rte_unused,
300                                                  size_t len __rte_unused,
301                                                  void *arg)
302 {
303         struct virtio_user_dev *dev = arg;
304         uint16_t i;
305
306         pthread_mutex_lock(&dev->mutex);
307
308         if (dev->started == false)
309                 goto exit;
310
311         /* Step 1: pause the active queues */
312         for (i = 0; i < dev->queue_pairs; i++)
313                 dev->ops->enable_qp(dev, i, 0);
314
315         /* Step 2: update memory regions */
316         dev->ops->send_request(dev, VHOST_USER_SET_MEM_TABLE, NULL);
317
318         /* Step 3: resume the active queues */
319         for (i = 0; i < dev->queue_pairs; i++)
320                 dev->ops->enable_qp(dev, i, 1);
321
322 exit:
323         pthread_mutex_unlock(&dev->mutex);
324 }
325
326 static int
327 virtio_user_dev_setup(struct virtio_user_dev *dev)
328 {
329         uint32_t q;
330
331         dev->vhostfd = -1;
332         dev->vhostfds = NULL;
333         dev->tapfds = NULL;
334
335         if (dev->is_server) {
336                 if (access(dev->path, F_OK) == 0 &&
337                     !is_vhost_user_by_type(dev->path)) {
338                         PMD_DRV_LOG(ERR, "Server mode doesn't support vhost-kernel!");
339                         return -1;
340                 }
341                 dev->ops = &ops_user;
342         } else {
343                 if (is_vhost_user_by_type(dev->path)) {
344                         dev->ops = &ops_user;
345                 } else {
346                         dev->ops = &ops_kernel;
347
348                         dev->vhostfds = malloc(dev->max_queue_pairs *
349                                                sizeof(int));
350                         dev->tapfds = malloc(dev->max_queue_pairs *
351                                              sizeof(int));
352                         if (!dev->vhostfds || !dev->tapfds) {
353                                 PMD_INIT_LOG(ERR, "Failed to malloc");
354                                 return -1;
355                         }
356
357                         for (q = 0; q < dev->max_queue_pairs; ++q) {
358                                 dev->vhostfds[q] = -1;
359                                 dev->tapfds[q] = -1;
360                         }
361                 }
362         }
363
364         if (dev->ops->setup(dev) < 0)
365                 return -1;
366
367         if (virtio_user_dev_init_notify(dev) < 0)
368                 return -1;
369
370         if (virtio_user_fill_intr_handle(dev) < 0)
371                 return -1;
372
373         return 0;
374 }
375
376 /* Use below macro to filter features from vhost backend */
377 #define VIRTIO_USER_SUPPORTED_FEATURES                  \
378         (1ULL << VIRTIO_NET_F_MAC               |       \
379          1ULL << VIRTIO_NET_F_STATUS            |       \
380          1ULL << VIRTIO_NET_F_MQ                |       \
381          1ULL << VIRTIO_NET_F_CTRL_MAC_ADDR     |       \
382          1ULL << VIRTIO_NET_F_CTRL_VQ           |       \
383          1ULL << VIRTIO_NET_F_CTRL_RX           |       \
384          1ULL << VIRTIO_NET_F_CTRL_VLAN         |       \
385          1ULL << VIRTIO_NET_F_CSUM              |       \
386          1ULL << VIRTIO_NET_F_HOST_TSO4         |       \
387          1ULL << VIRTIO_NET_F_HOST_TSO6         |       \
388          1ULL << VIRTIO_NET_F_MRG_RXBUF         |       \
389          1ULL << VIRTIO_RING_F_INDIRECT_DESC    |       \
390          1ULL << VIRTIO_NET_F_GUEST_CSUM        |       \
391          1ULL << VIRTIO_NET_F_GUEST_TSO4        |       \
392          1ULL << VIRTIO_NET_F_GUEST_TSO6        |       \
393          1ULL << VIRTIO_F_IN_ORDER              |       \
394          1ULL << VIRTIO_F_VERSION_1)
395
396 int
397 virtio_user_dev_init(struct virtio_user_dev *dev, char *path, int queues,
398                      int cq, int queue_size, const char *mac, char **ifname,
399                      int mrg_rxbuf, int in_order)
400 {
401         pthread_mutex_init(&dev->mutex, NULL);
402         snprintf(dev->path, PATH_MAX, "%s", path);
403         dev->started = 0;
404         dev->max_queue_pairs = queues;
405         dev->queue_pairs = 1; /* mq disabled by default */
406         dev->queue_size = queue_size;
407         dev->mac_specified = 0;
408         dev->unsupported_features = 0;
409         parse_mac(dev, mac);
410
411         if (*ifname) {
412                 dev->ifname = *ifname;
413                 *ifname = NULL;
414         }
415
416         if (virtio_user_dev_setup(dev) < 0) {
417                 PMD_INIT_LOG(ERR, "backend set up fails");
418                 return -1;
419         }
420
421         if (!dev->is_server) {
422                 if (dev->ops->send_request(dev, VHOST_USER_SET_OWNER,
423                                            NULL) < 0) {
424                         PMD_INIT_LOG(ERR, "set_owner fails: %s",
425                                      strerror(errno));
426                         return -1;
427                 }
428
429                 if (dev->ops->send_request(dev, VHOST_USER_GET_FEATURES,
430                                            &dev->device_features) < 0) {
431                         PMD_INIT_LOG(ERR, "get_features failed: %s",
432                                      strerror(errno));
433                         return -1;
434                 }
435         } else {
436                 /* We just pretend vhost-user can support all these features.
437                  * Note that this could be problematic that if some feature is
438                  * negotiated but not supported by the vhost-user which comes
439                  * later.
440                  */
441                 dev->device_features = VIRTIO_USER_SUPPORTED_FEATURES;
442         }
443
444         if (!mrg_rxbuf) {
445                 dev->device_features &= ~(1ull << VIRTIO_NET_F_MRG_RXBUF);
446                 dev->unsupported_features |= (1ull << VIRTIO_NET_F_MRG_RXBUF);
447         }
448
449         if (!in_order) {
450                 dev->device_features &= ~(1ull << VIRTIO_F_IN_ORDER);
451                 dev->unsupported_features |= (1ull << VIRTIO_F_IN_ORDER);
452         }
453
454         if (dev->mac_specified) {
455                 dev->device_features |= (1ull << VIRTIO_NET_F_MAC);
456         } else {
457                 dev->device_features &= ~(1ull << VIRTIO_NET_F_MAC);
458                 dev->unsupported_features |= (1ull << VIRTIO_NET_F_MAC);
459         }
460
461         if (cq) {
462                 /* device does not really need to know anything about CQ,
463                  * so if necessary, we just claim to support CQ
464                  */
465                 dev->device_features |= (1ull << VIRTIO_NET_F_CTRL_VQ);
466         } else {
467                 dev->device_features &= ~(1ull << VIRTIO_NET_F_CTRL_VQ);
468                 /* Also disable features depends on VIRTIO_NET_F_CTRL_VQ */
469                 dev->device_features &= ~(1ull << VIRTIO_NET_F_CTRL_RX);
470                 dev->device_features &= ~(1ull << VIRTIO_NET_F_CTRL_VLAN);
471                 dev->device_features &= ~(1ull << VIRTIO_NET_F_GUEST_ANNOUNCE);
472                 dev->device_features &= ~(1ull << VIRTIO_NET_F_MQ);
473                 dev->device_features &= ~(1ull << VIRTIO_NET_F_CTRL_MAC_ADDR);
474                 dev->unsupported_features |= (1ull << VIRTIO_NET_F_CTRL_VQ);
475                 dev->unsupported_features |= (1ull << VIRTIO_NET_F_CTRL_RX);
476                 dev->unsupported_features |= (1ull << VIRTIO_NET_F_CTRL_VLAN);
477                 dev->unsupported_features |=
478                         (1ull << VIRTIO_NET_F_GUEST_ANNOUNCE);
479                 dev->unsupported_features |= (1ull << VIRTIO_NET_F_MQ);
480                 dev->unsupported_features |=
481                         (1ull << VIRTIO_NET_F_CTRL_MAC_ADDR);
482         }
483
484         /* The backend will not report this feature, we add it explicitly */
485         if (is_vhost_user_by_type(dev->path))
486                 dev->device_features |= (1ull << VIRTIO_NET_F_STATUS);
487
488         dev->device_features &= VIRTIO_USER_SUPPORTED_FEATURES;
489         dev->unsupported_features |= ~VIRTIO_USER_SUPPORTED_FEATURES;
490
491         if (rte_mem_event_callback_register(VIRTIO_USER_MEM_EVENT_CLB_NAME,
492                                 virtio_user_mem_event_cb, dev)) {
493                 if (rte_errno != ENOTSUP) {
494                         PMD_INIT_LOG(ERR, "Failed to register mem event"
495                                         " callback\n");
496                         return -1;
497                 }
498         }
499
500         return 0;
501 }
502
503 void
504 virtio_user_dev_uninit(struct virtio_user_dev *dev)
505 {
506         uint32_t i;
507
508         virtio_user_stop_device(dev);
509
510         rte_mem_event_callback_unregister(VIRTIO_USER_MEM_EVENT_CLB_NAME, dev);
511
512         for (i = 0; i < dev->max_queue_pairs * 2; ++i) {
513                 close(dev->callfds[i]);
514                 close(dev->kickfds[i]);
515         }
516
517         close(dev->vhostfd);
518
519         if (dev->is_server && dev->listenfd >= 0) {
520                 close(dev->listenfd);
521                 dev->listenfd = -1;
522         }
523
524         if (dev->vhostfds) {
525                 for (i = 0; i < dev->max_queue_pairs; ++i)
526                         close(dev->vhostfds[i]);
527                 free(dev->vhostfds);
528                 free(dev->tapfds);
529         }
530
531         free(dev->ifname);
532
533         if (dev->is_server)
534                 unlink(dev->path);
535 }
536
537 uint8_t
538 virtio_user_handle_mq(struct virtio_user_dev *dev, uint16_t q_pairs)
539 {
540         uint16_t i;
541         uint8_t ret = 0;
542
543         if (q_pairs > dev->max_queue_pairs) {
544                 PMD_INIT_LOG(ERR, "multi-q config %u, but only %u supported",
545                              q_pairs, dev->max_queue_pairs);
546                 return -1;
547         }
548
549         /* Server mode can't enable queue pairs if vhostfd is invalid,
550          * always return 0 in this case.
551          */
552         if (!dev->is_server || dev->vhostfd >= 0) {
553                 for (i = 0; i < q_pairs; ++i)
554                         ret |= dev->ops->enable_qp(dev, i, 1);
555                 for (i = q_pairs; i < dev->max_queue_pairs; ++i)
556                         ret |= dev->ops->enable_qp(dev, i, 0);
557         }
558         dev->queue_pairs = q_pairs;
559
560         return ret;
561 }
562
563 static uint32_t
564 virtio_user_handle_ctrl_msg(struct virtio_user_dev *dev, struct vring *vring,
565                             uint16_t idx_hdr)
566 {
567         struct virtio_net_ctrl_hdr *hdr;
568         virtio_net_ctrl_ack status = ~0;
569         uint16_t i, idx_data, idx_status;
570         uint32_t n_descs = 0;
571
572         /* locate desc for header, data, and status */
573         idx_data = vring->desc[idx_hdr].next;
574         n_descs++;
575
576         i = idx_data;
577         while (vring->desc[i].flags == VRING_DESC_F_NEXT) {
578                 i = vring->desc[i].next;
579                 n_descs++;
580         }
581
582         /* locate desc for status */
583         idx_status = i;
584         n_descs++;
585
586         hdr = (void *)(uintptr_t)vring->desc[idx_hdr].addr;
587         if (hdr->class == VIRTIO_NET_CTRL_MQ &&
588             hdr->cmd == VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET) {
589                 uint16_t queues;
590
591                 queues = *(uint16_t *)(uintptr_t)vring->desc[idx_data].addr;
592                 status = virtio_user_handle_mq(dev, queues);
593         }
594
595         /* Update status */
596         *(virtio_net_ctrl_ack *)(uintptr_t)vring->desc[idx_status].addr = status;
597
598         return n_descs;
599 }
600
601 void
602 virtio_user_handle_cq(struct virtio_user_dev *dev, uint16_t queue_idx)
603 {
604         uint16_t avail_idx, desc_idx;
605         struct vring_used_elem *uep;
606         uint32_t n_descs;
607         struct vring *vring = &dev->vrings[queue_idx];
608
609         /* Consume avail ring, using used ring idx as first one */
610         while (vring->used->idx != vring->avail->idx) {
611                 avail_idx = (vring->used->idx) & (vring->num - 1);
612                 desc_idx = vring->avail->ring[avail_idx];
613
614                 n_descs = virtio_user_handle_ctrl_msg(dev, vring, desc_idx);
615
616                 /* Update used ring */
617                 uep = &vring->used->ring[avail_idx];
618                 uep->id = avail_idx;
619                 uep->len = n_descs;
620
621                 vring->used->idx++;
622         }
623 }