net/virtio-user: do not stop stopped device again
[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         if (!dev->started)
188                 goto out;
189
190         for (i = 0; i < dev->max_queue_pairs; ++i)
191                 dev->ops->enable_qp(dev, i, 0);
192
193         if (dev->ops->send_request(dev, VHOST_USER_RESET_OWNER, NULL) < 0) {
194                 PMD_DRV_LOG(INFO, "Failed to reset the device\n");
195                 pthread_mutex_unlock(&dev->mutex);
196                 return -1;
197         }
198         dev->started = false;
199 out:
200         pthread_mutex_unlock(&dev->mutex);
201
202         return 0;
203 }
204
205 static inline void
206 parse_mac(struct virtio_user_dev *dev, const char *mac)
207 {
208         int i, r;
209         uint32_t tmp[ETHER_ADDR_LEN];
210
211         if (!mac)
212                 return;
213
214         r = sscanf(mac, "%x:%x:%x:%x:%x:%x", &tmp[0],
215                         &tmp[1], &tmp[2], &tmp[3], &tmp[4], &tmp[5]);
216         if (r == ETHER_ADDR_LEN) {
217                 for (i = 0; i < ETHER_ADDR_LEN; ++i)
218                         dev->mac_addr[i] = (uint8_t)tmp[i];
219                 dev->mac_specified = 1;
220         } else {
221                 /* ignore the wrong mac, use random mac */
222                 PMD_DRV_LOG(ERR, "wrong format of mac: %s", mac);
223         }
224 }
225
226 static int
227 virtio_user_dev_init_notify(struct virtio_user_dev *dev)
228 {
229         uint32_t i, j;
230         int callfd;
231         int kickfd;
232
233         for (i = 0; i < VIRTIO_MAX_VIRTQUEUES; ++i) {
234                 if (i >= dev->max_queue_pairs * 2) {
235                         dev->kickfds[i] = -1;
236                         dev->callfds[i] = -1;
237                         continue;
238                 }
239
240                 /* May use invalid flag, but some backend uses kickfd and
241                  * callfd as criteria to judge if dev is alive. so finally we
242                  * use real event_fd.
243                  */
244                 callfd = eventfd(0, EFD_CLOEXEC | EFD_NONBLOCK);
245                 if (callfd < 0) {
246                         PMD_DRV_LOG(ERR, "callfd error, %s", strerror(errno));
247                         break;
248                 }
249                 kickfd = eventfd(0, EFD_CLOEXEC | EFD_NONBLOCK);
250                 if (kickfd < 0) {
251                         PMD_DRV_LOG(ERR, "kickfd error, %s", strerror(errno));
252                         break;
253                 }
254                 dev->callfds[i] = callfd;
255                 dev->kickfds[i] = kickfd;
256         }
257
258         if (i < VIRTIO_MAX_VIRTQUEUES) {
259                 for (j = 0; j <= i; ++j) {
260                         close(dev->callfds[j]);
261                         close(dev->kickfds[j]);
262                 }
263
264                 return -1;
265         }
266
267         return 0;
268 }
269
270 static int
271 virtio_user_fill_intr_handle(struct virtio_user_dev *dev)
272 {
273         uint32_t i;
274         struct rte_eth_dev *eth_dev = &rte_eth_devices[dev->port_id];
275
276         if (!eth_dev->intr_handle) {
277                 eth_dev->intr_handle = malloc(sizeof(*eth_dev->intr_handle));
278                 if (!eth_dev->intr_handle) {
279                         PMD_DRV_LOG(ERR, "fail to allocate intr_handle");
280                         return -1;
281                 }
282                 memset(eth_dev->intr_handle, 0, sizeof(*eth_dev->intr_handle));
283         }
284
285         for (i = 0; i < dev->max_queue_pairs; ++i)
286                 eth_dev->intr_handle->efds[i] = dev->callfds[i];
287         eth_dev->intr_handle->nb_efd = dev->max_queue_pairs;
288         eth_dev->intr_handle->max_intr = dev->max_queue_pairs + 1;
289         eth_dev->intr_handle->type = RTE_INTR_HANDLE_VDEV;
290         /* For virtio vdev, no need to read counter for clean */
291         eth_dev->intr_handle->efd_counter_size = 0;
292         eth_dev->intr_handle->fd = -1;
293         if (dev->vhostfd >= 0)
294                 eth_dev->intr_handle->fd = dev->vhostfd;
295         else if (dev->is_server)
296                 eth_dev->intr_handle->fd = dev->listenfd;
297
298         return 0;
299 }
300
301 static void
302 virtio_user_mem_event_cb(enum rte_mem_event type __rte_unused,
303                                                  const void *addr __rte_unused,
304                                                  size_t len __rte_unused,
305                                                  void *arg)
306 {
307         struct virtio_user_dev *dev = arg;
308         struct rte_memseg_list *msl;
309         uint16_t i;
310
311         /* ignore externally allocated memory */
312         msl = rte_mem_virt2memseg_list(addr);
313         if (msl->external)
314                 return;
315
316         pthread_mutex_lock(&dev->mutex);
317
318         if (dev->started == false)
319                 goto exit;
320
321         /* Step 1: pause the active queues */
322         for (i = 0; i < dev->queue_pairs; i++)
323                 dev->ops->enable_qp(dev, i, 0);
324
325         /* Step 2: update memory regions */
326         dev->ops->send_request(dev, VHOST_USER_SET_MEM_TABLE, NULL);
327
328         /* Step 3: resume the active queues */
329         for (i = 0; i < dev->queue_pairs; i++)
330                 dev->ops->enable_qp(dev, i, 1);
331
332 exit:
333         pthread_mutex_unlock(&dev->mutex);
334 }
335
336 static int
337 virtio_user_dev_setup(struct virtio_user_dev *dev)
338 {
339         uint32_t q;
340
341         dev->vhostfd = -1;
342         dev->vhostfds = NULL;
343         dev->tapfds = NULL;
344
345         if (dev->is_server) {
346                 if (access(dev->path, F_OK) == 0 &&
347                     !is_vhost_user_by_type(dev->path)) {
348                         PMD_DRV_LOG(ERR, "Server mode doesn't support vhost-kernel!");
349                         return -1;
350                 }
351                 dev->ops = &virtio_ops_user;
352         } else {
353                 if (is_vhost_user_by_type(dev->path)) {
354                         dev->ops = &virtio_ops_user;
355                 } else {
356                         dev->ops = &virtio_ops_kernel;
357
358                         dev->vhostfds = malloc(dev->max_queue_pairs *
359                                                sizeof(int));
360                         dev->tapfds = malloc(dev->max_queue_pairs *
361                                              sizeof(int));
362                         if (!dev->vhostfds || !dev->tapfds) {
363                                 PMD_INIT_LOG(ERR, "Failed to malloc");
364                                 return -1;
365                         }
366
367                         for (q = 0; q < dev->max_queue_pairs; ++q) {
368                                 dev->vhostfds[q] = -1;
369                                 dev->tapfds[q] = -1;
370                         }
371                 }
372         }
373
374         if (dev->ops->setup(dev) < 0)
375                 return -1;
376
377         if (virtio_user_dev_init_notify(dev) < 0)
378                 return -1;
379
380         if (virtio_user_fill_intr_handle(dev) < 0)
381                 return -1;
382
383         return 0;
384 }
385
386 /* Use below macro to filter features from vhost backend */
387 #define VIRTIO_USER_SUPPORTED_FEATURES                  \
388         (1ULL << VIRTIO_NET_F_MAC               |       \
389          1ULL << VIRTIO_NET_F_STATUS            |       \
390          1ULL << VIRTIO_NET_F_MQ                |       \
391          1ULL << VIRTIO_NET_F_CTRL_MAC_ADDR     |       \
392          1ULL << VIRTIO_NET_F_CTRL_VQ           |       \
393          1ULL << VIRTIO_NET_F_CTRL_RX           |       \
394          1ULL << VIRTIO_NET_F_CTRL_VLAN         |       \
395          1ULL << VIRTIO_NET_F_CSUM              |       \
396          1ULL << VIRTIO_NET_F_HOST_TSO4         |       \
397          1ULL << VIRTIO_NET_F_HOST_TSO6         |       \
398          1ULL << VIRTIO_NET_F_MRG_RXBUF         |       \
399          1ULL << VIRTIO_RING_F_INDIRECT_DESC    |       \
400          1ULL << VIRTIO_NET_F_GUEST_CSUM        |       \
401          1ULL << VIRTIO_NET_F_GUEST_TSO4        |       \
402          1ULL << VIRTIO_NET_F_GUEST_TSO6        |       \
403          1ULL << VIRTIO_F_IN_ORDER              |       \
404          1ULL << VIRTIO_F_VERSION_1)
405
406 int
407 virtio_user_dev_init(struct virtio_user_dev *dev, char *path, int queues,
408                      int cq, int queue_size, const char *mac, char **ifname,
409                      int mrg_rxbuf, int in_order)
410 {
411         pthread_mutex_init(&dev->mutex, NULL);
412         snprintf(dev->path, PATH_MAX, "%s", path);
413         dev->started = 0;
414         dev->max_queue_pairs = queues;
415         dev->queue_pairs = 1; /* mq disabled by default */
416         dev->queue_size = queue_size;
417         dev->mac_specified = 0;
418         dev->unsupported_features = 0;
419         parse_mac(dev, mac);
420
421         if (*ifname) {
422                 dev->ifname = *ifname;
423                 *ifname = NULL;
424         }
425
426         if (virtio_user_dev_setup(dev) < 0) {
427                 PMD_INIT_LOG(ERR, "backend set up fails");
428                 return -1;
429         }
430
431         if (!dev->is_server) {
432                 if (dev->ops->send_request(dev, VHOST_USER_SET_OWNER,
433                                            NULL) < 0) {
434                         PMD_INIT_LOG(ERR, "set_owner fails: %s",
435                                      strerror(errno));
436                         return -1;
437                 }
438
439                 if (dev->ops->send_request(dev, VHOST_USER_GET_FEATURES,
440                                            &dev->device_features) < 0) {
441                         PMD_INIT_LOG(ERR, "get_features failed: %s",
442                                      strerror(errno));
443                         return -1;
444                 }
445         } else {
446                 /* We just pretend vhost-user can support all these features.
447                  * Note that this could be problematic that if some feature is
448                  * negotiated but not supported by the vhost-user which comes
449                  * later.
450                  */
451                 dev->device_features = VIRTIO_USER_SUPPORTED_FEATURES;
452         }
453
454         if (!mrg_rxbuf) {
455                 dev->device_features &= ~(1ull << VIRTIO_NET_F_MRG_RXBUF);
456                 dev->unsupported_features |= (1ull << VIRTIO_NET_F_MRG_RXBUF);
457         }
458
459         if (!in_order) {
460                 dev->device_features &= ~(1ull << VIRTIO_F_IN_ORDER);
461                 dev->unsupported_features |= (1ull << VIRTIO_F_IN_ORDER);
462         }
463
464         if (dev->mac_specified) {
465                 dev->device_features |= (1ull << VIRTIO_NET_F_MAC);
466         } else {
467                 dev->device_features &= ~(1ull << VIRTIO_NET_F_MAC);
468                 dev->unsupported_features |= (1ull << VIRTIO_NET_F_MAC);
469         }
470
471         if (cq) {
472                 /* device does not really need to know anything about CQ,
473                  * so if necessary, we just claim to support CQ
474                  */
475                 dev->device_features |= (1ull << VIRTIO_NET_F_CTRL_VQ);
476         } else {
477                 dev->device_features &= ~(1ull << VIRTIO_NET_F_CTRL_VQ);
478                 /* Also disable features depends on VIRTIO_NET_F_CTRL_VQ */
479                 dev->device_features &= ~(1ull << VIRTIO_NET_F_CTRL_RX);
480                 dev->device_features &= ~(1ull << VIRTIO_NET_F_CTRL_VLAN);
481                 dev->device_features &= ~(1ull << VIRTIO_NET_F_GUEST_ANNOUNCE);
482                 dev->device_features &= ~(1ull << VIRTIO_NET_F_MQ);
483                 dev->device_features &= ~(1ull << VIRTIO_NET_F_CTRL_MAC_ADDR);
484                 dev->unsupported_features |= (1ull << VIRTIO_NET_F_CTRL_VQ);
485                 dev->unsupported_features |= (1ull << VIRTIO_NET_F_CTRL_RX);
486                 dev->unsupported_features |= (1ull << VIRTIO_NET_F_CTRL_VLAN);
487                 dev->unsupported_features |=
488                         (1ull << VIRTIO_NET_F_GUEST_ANNOUNCE);
489                 dev->unsupported_features |= (1ull << VIRTIO_NET_F_MQ);
490                 dev->unsupported_features |=
491                         (1ull << VIRTIO_NET_F_CTRL_MAC_ADDR);
492         }
493
494         /* The backend will not report this feature, we add it explicitly */
495         if (is_vhost_user_by_type(dev->path))
496                 dev->device_features |= (1ull << VIRTIO_NET_F_STATUS);
497
498         dev->device_features &= VIRTIO_USER_SUPPORTED_FEATURES;
499         dev->unsupported_features |= ~VIRTIO_USER_SUPPORTED_FEATURES;
500
501         if (rte_mem_event_callback_register(VIRTIO_USER_MEM_EVENT_CLB_NAME,
502                                 virtio_user_mem_event_cb, dev)) {
503                 if (rte_errno != ENOTSUP) {
504                         PMD_INIT_LOG(ERR, "Failed to register mem event"
505                                         " callback\n");
506                         return -1;
507                 }
508         }
509
510         return 0;
511 }
512
513 void
514 virtio_user_dev_uninit(struct virtio_user_dev *dev)
515 {
516         uint32_t i;
517
518         virtio_user_stop_device(dev);
519
520         rte_mem_event_callback_unregister(VIRTIO_USER_MEM_EVENT_CLB_NAME, dev);
521
522         for (i = 0; i < dev->max_queue_pairs * 2; ++i) {
523                 close(dev->callfds[i]);
524                 close(dev->kickfds[i]);
525         }
526
527         close(dev->vhostfd);
528
529         if (dev->is_server && dev->listenfd >= 0) {
530                 close(dev->listenfd);
531                 dev->listenfd = -1;
532         }
533
534         if (dev->vhostfds) {
535                 for (i = 0; i < dev->max_queue_pairs; ++i)
536                         close(dev->vhostfds[i]);
537                 free(dev->vhostfds);
538                 free(dev->tapfds);
539         }
540
541         free(dev->ifname);
542
543         if (dev->is_server)
544                 unlink(dev->path);
545 }
546
547 uint8_t
548 virtio_user_handle_mq(struct virtio_user_dev *dev, uint16_t q_pairs)
549 {
550         uint16_t i;
551         uint8_t ret = 0;
552
553         if (q_pairs > dev->max_queue_pairs) {
554                 PMD_INIT_LOG(ERR, "multi-q config %u, but only %u supported",
555                              q_pairs, dev->max_queue_pairs);
556                 return -1;
557         }
558
559         /* Server mode can't enable queue pairs if vhostfd is invalid,
560          * always return 0 in this case.
561          */
562         if (!dev->is_server || dev->vhostfd >= 0) {
563                 for (i = 0; i < q_pairs; ++i)
564                         ret |= dev->ops->enable_qp(dev, i, 1);
565                 for (i = q_pairs; i < dev->max_queue_pairs; ++i)
566                         ret |= dev->ops->enable_qp(dev, i, 0);
567         }
568         dev->queue_pairs = q_pairs;
569
570         return ret;
571 }
572
573 static uint32_t
574 virtio_user_handle_ctrl_msg(struct virtio_user_dev *dev, struct vring *vring,
575                             uint16_t idx_hdr)
576 {
577         struct virtio_net_ctrl_hdr *hdr;
578         virtio_net_ctrl_ack status = ~0;
579         uint16_t i, idx_data, idx_status;
580         uint32_t n_descs = 0;
581
582         /* locate desc for header, data, and status */
583         idx_data = vring->desc[idx_hdr].next;
584         n_descs++;
585
586         i = idx_data;
587         while (vring->desc[i].flags == VRING_DESC_F_NEXT) {
588                 i = vring->desc[i].next;
589                 n_descs++;
590         }
591
592         /* locate desc for status */
593         idx_status = i;
594         n_descs++;
595
596         hdr = (void *)(uintptr_t)vring->desc[idx_hdr].addr;
597         if (hdr->class == VIRTIO_NET_CTRL_MQ &&
598             hdr->cmd == VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET) {
599                 uint16_t queues;
600
601                 queues = *(uint16_t *)(uintptr_t)vring->desc[idx_data].addr;
602                 status = virtio_user_handle_mq(dev, queues);
603         }
604
605         /* Update status */
606         *(virtio_net_ctrl_ack *)(uintptr_t)vring->desc[idx_status].addr = status;
607
608         return n_descs;
609 }
610
611 void
612 virtio_user_handle_cq(struct virtio_user_dev *dev, uint16_t queue_idx)
613 {
614         uint16_t avail_idx, desc_idx;
615         struct vring_used_elem *uep;
616         uint32_t n_descs;
617         struct vring *vring = &dev->vrings[queue_idx];
618
619         /* Consume avail ring, using used ring idx as first one */
620         while (vring->used->idx != vring->avail->idx) {
621                 avail_idx = (vring->used->idx) & (vring->num - 1);
622                 desc_idx = vring->avail->ring[avail_idx];
623
624                 n_descs = virtio_user_handle_ctrl_msg(dev, vring, desc_idx);
625
626                 /* Update used ring */
627                 uep = &vring->used->ring[avail_idx];
628                 uep->id = avail_idx;
629                 uep->len = n_descs;
630
631                 vring->used->idx++;
632         }
633 }