xref: /openbmc/qemu/net/vhost-user.c (revision 0d70c5aa1bbfb0f5099d53d6e084337a8246cc0c) 
1 /*
2  * vhost-user.c
3  *
4  * Copyright (c) 2013 Virtual Open Systems Sarl.
5  *
6  * This work is licensed under the terms of the GNU GPL, version 2 or later.
7  * See the COPYING file in the top-level directory.
8  *
9  */
10 
11 #include "qemu/osdep.h"
12 #include "clients.h"
13 #include "net/vhost_net.h"
14 #include "hw/virtio/vhost.h"
15 #include "hw/virtio/vhost-user.h"
16 #include "standard-headers/linux/virtio_net.h"
17 #include "chardev/char-fe.h"
18 #include "qapi/error.h"
19 #include "qapi/qapi-commands-net.h"
20 #include "qapi/qapi-events-net.h"
21 #include "qemu/config-file.h"
22 #include "qemu/error-report.h"
23 #include "qemu/option.h"
24 #include "trace.h"
25 
26 static const int user_feature_bits[] = {
27     VIRTIO_F_NOTIFY_ON_EMPTY,
28     VIRTIO_F_NOTIFICATION_DATA,
29     VIRTIO_RING_F_INDIRECT_DESC,
30     VIRTIO_RING_F_EVENT_IDX,
31 
32     VIRTIO_F_ANY_LAYOUT,
33     VIRTIO_F_VERSION_1,
34     VIRTIO_NET_F_CSUM,
35     VIRTIO_NET_F_GUEST_CSUM,
36     VIRTIO_NET_F_GSO,
37     VIRTIO_NET_F_GUEST_TSO4,
38     VIRTIO_NET_F_GUEST_TSO6,
39     VIRTIO_NET_F_GUEST_ECN,
40     VIRTIO_NET_F_GUEST_UFO,
41     VIRTIO_NET_F_HOST_TSO4,
42     VIRTIO_NET_F_HOST_TSO6,
43     VIRTIO_NET_F_HOST_ECN,
44     VIRTIO_NET_F_HOST_UFO,
45     VIRTIO_NET_F_MRG_RXBUF,
46     VIRTIO_NET_F_MTU,
47     VIRTIO_F_IOMMU_PLATFORM,
48     VIRTIO_F_RING_PACKED,
49     VIRTIO_F_RING_RESET,
50     VIRTIO_F_IN_ORDER,
51     VIRTIO_NET_F_RSS,
52     VIRTIO_NET_F_RSC_EXT,
53     VIRTIO_NET_F_HASH_REPORT,
54     VIRTIO_NET_F_GUEST_USO4,
55     VIRTIO_NET_F_GUEST_USO6,
56     VIRTIO_NET_F_HOST_USO,
57 
58     /* This bit implies RARP isn't sent by QEMU out of band */
59     VIRTIO_NET_F_GUEST_ANNOUNCE,
60 
61     VIRTIO_NET_F_MQ,
62 
63     VHOST_INVALID_FEATURE_BIT
64 };
65 
66 typedef struct NetVhostUserState {
67     NetClientState nc;
68     CharBackend chr; /* only queue index 0 */
69     VhostUserState *vhost_user;
70     VHostNetState *vhost_net;
71     guint watch;
72     uint64_t acked_features;
73     bool started;
74 } NetVhostUserState;
75 
76 static struct vhost_net *vhost_user_get_vhost_net(NetClientState *nc)
77 {
78     NetVhostUserState *s = DO_UPCAST(NetVhostUserState, nc, nc);
79     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_USER);
80     return s->vhost_net;
81 }
82 
83 static uint64_t vhost_user_get_acked_features(NetClientState *nc)
84 {
85     NetVhostUserState *s = DO_UPCAST(NetVhostUserState, nc, nc);
86     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_USER);
87     return s->acked_features;
88 }
89 
90 static void vhost_user_save_acked_features(NetClientState *nc)
91 {
92     NetVhostUserState *s;
93 
94     s = DO_UPCAST(NetVhostUserState, nc, nc);
95     if (s->vhost_net) {
96         uint64_t features = vhost_net_get_acked_features(s->vhost_net);
97         if (features) {
98             s->acked_features = features;
99         }
100     }
101 }
102 
103 static void vhost_user_stop(int queues, NetClientState *ncs[])
104 {
105     int i;
106     NetVhostUserState *s;
107 
108     for (i = 0; i < queues; i++) {
109         assert(ncs[i]->info->type == NET_CLIENT_DRIVER_VHOST_USER);
110 
111         s = DO_UPCAST(NetVhostUserState, nc, ncs[i]);
112 
113         if (s->vhost_net) {
114             vhost_user_save_acked_features(ncs[i]);
115             vhost_net_cleanup(s->vhost_net);
116         }
117     }
118 }
119 
120 static int vhost_user_start(int queues, NetClientState *ncs[],
121                             VhostUserState *be)
122 {
123     VhostNetOptions options;
124     struct vhost_net *net = NULL;
125     NetVhostUserState *s;
126     int max_queues;
127     int i;
128 
129     options.backend_type = VHOST_BACKEND_TYPE_USER;
130 
131     for (i = 0; i < queues; i++) {
132         assert(ncs[i]->info->type == NET_CLIENT_DRIVER_VHOST_USER);
133 
134         s = DO_UPCAST(NetVhostUserState, nc, ncs[i]);
135 
136         options.net_backend = ncs[i];
137         options.opaque      = be;
138         options.busyloop_timeout = 0;
139         options.nvqs = 2;
140         options.feature_bits = user_feature_bits;
141         options.max_tx_queue_size = VIRTQUEUE_MAX_SIZE;
142         options.get_acked_features = vhost_user_get_acked_features;
143         options.save_acked_features = vhost_user_save_acked_features;
144         options.is_vhost_user = true;
145 
146         net = vhost_net_init(&options);
147         if (!net) {
148             error_report("failed to init vhost_net for queue %d", i);
149             goto err;
150         }
151 
152         if (i == 0) {
153             max_queues = vhost_net_get_max_queues(net);
154             if (queues > max_queues) {
155                 error_report("you are asking more queues than supported: %d",
156                              max_queues);
157                 goto err;
158             }
159         }
160 
161         if (s->vhost_net) {
162             vhost_net_cleanup(s->vhost_net);
163             g_free(s->vhost_net);
164         }
165         s->vhost_net = net;
166     }
167 
168     return 0;
169 
170 err:
171     if (net) {
172         vhost_net_cleanup(net);
173         g_free(net);
174     }
175     vhost_user_stop(i, ncs);
176     return -1;
177 }
178 
179 static ssize_t vhost_user_receive(NetClientState *nc, const uint8_t *buf,
180                                   size_t size)
181 {
182     /* In case of RARP (message size is 60) notify backup to send a fake RARP.
183        This fake RARP will be sent by backend only for guest
184        without GUEST_ANNOUNCE capability.
185      */
186     if (size == 60) {
187         NetVhostUserState *s = DO_UPCAST(NetVhostUserState, nc, nc);
188         int r;
189         static int display_rarp_failure = 1;
190         char mac_addr[6];
191 
192         /* extract guest mac address from the RARP message */
193         memcpy(mac_addr, &buf[6], 6);
194 
195         r = vhost_net_notify_migration_done(s->vhost_net, mac_addr);
196 
197         if ((r != 0) && (display_rarp_failure)) {
198             fprintf(stderr,
199                     "Vhost user backend fails to broadcast fake RARP\n");
200             fflush(stderr);
201             display_rarp_failure = 0;
202         }
203     }
204 
205     return size;
206 }
207 
208 static void net_vhost_user_cleanup(NetClientState *nc)
209 {
210     NetVhostUserState *s = DO_UPCAST(NetVhostUserState, nc, nc);
211 
212     if (s->vhost_net) {
213         vhost_net_cleanup(s->vhost_net);
214         g_free(s->vhost_net);
215         s->vhost_net = NULL;
216     }
217     if (nc->queue_index == 0) {
218         if (s->watch) {
219             g_source_remove(s->watch);
220             s->watch = 0;
221         }
222         qemu_chr_fe_deinit(&s->chr, true);
223         if (s->vhost_user) {
224             vhost_user_cleanup(s->vhost_user);
225             g_free(s->vhost_user);
226             s->vhost_user = NULL;
227         }
228     }
229 
230     qemu_purge_queued_packets(nc);
231 }
232 
233 static int vhost_user_set_vnet_endianness(NetClientState *nc,
234                                           bool enable)
235 {
236     /* Nothing to do.  If the server supports
237      * VHOST_USER_PROTOCOL_F_CROSS_ENDIAN, it will get the
238      * vnet header endianness from there.  If it doesn't, negotiation
239      * fails.
240      */
241     return 0;
242 }
243 
244 static bool vhost_user_has_vnet_hdr(NetClientState *nc)
245 {
246     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_USER);
247 
248     return true;
249 }
250 
251 static bool vhost_user_has_ufo(NetClientState *nc)
252 {
253     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_USER);
254 
255     return true;
256 }
257 
258 static bool vhost_user_check_peer_type(NetClientState *nc, ObjectClass *oc,
259                                        Error **errp)
260 {
261     const char *driver = object_class_get_name(oc);
262 
263     if (!g_str_has_prefix(driver, "virtio-net-")) {
264         error_setg(errp, "vhost-user requires frontend driver virtio-net-*");
265         return false;
266     }
267 
268     return true;
269 }
270 
271 static NetClientInfo net_vhost_user_info = {
272         .type = NET_CLIENT_DRIVER_VHOST_USER,
273         .size = sizeof(NetVhostUserState),
274         .receive = vhost_user_receive,
275         .cleanup = net_vhost_user_cleanup,
276         .has_vnet_hdr = vhost_user_has_vnet_hdr,
277         .has_ufo = vhost_user_has_ufo,
278         .set_vnet_be = vhost_user_set_vnet_endianness,
279         .set_vnet_le = vhost_user_set_vnet_endianness,
280         .check_peer_type = vhost_user_check_peer_type,
281         .get_vhost_net = vhost_user_get_vhost_net,
282 };
283 
284 static gboolean net_vhost_user_watch(void *do_not_use, GIOCondition cond,
285                                      void *opaque)
286 {
287     NetVhostUserState *s = opaque;
288 
289     qemu_chr_fe_disconnect(&s->chr);
290 
291     return G_SOURCE_CONTINUE;
292 }
293 
294 static void net_vhost_user_event(void *opaque, QEMUChrEvent event);
295 
296 static void chr_closed_bh(void *opaque)
297 {
298     const char *name = opaque;
299     NetClientState *ncs[MAX_QUEUE_NUM];
300     NetVhostUserState *s;
301     int queues, i;
302 
303     queues = qemu_find_net_clients_except(name, ncs,
304                                           NET_CLIENT_DRIVER_NIC,
305                                           MAX_QUEUE_NUM);
306     assert(queues < MAX_QUEUE_NUM);
307 
308     s = DO_UPCAST(NetVhostUserState, nc, ncs[0]);
309 
310     for (i = queues -1; i >= 0; i--) {
311         vhost_user_save_acked_features(ncs[i]);
312     }
313 
314     net_client_set_link(ncs, queues, false);
315 
316     qemu_chr_fe_set_handlers(&s->chr, NULL, NULL, net_vhost_user_event,
317                              NULL, opaque, NULL, true);
318 
319     qapi_event_send_netdev_vhost_user_disconnected(name);
320 }
321 
322 static void net_vhost_user_event(void *opaque, QEMUChrEvent event)
323 {
324     const char *name = opaque;
325     NetClientState *ncs[MAX_QUEUE_NUM];
326     NetVhostUserState *s;
327     Chardev *chr;
328     int queues;
329 
330     queues = qemu_find_net_clients_except(name, ncs,
331                                           NET_CLIENT_DRIVER_NIC,
332                                           MAX_QUEUE_NUM);
333     assert(queues < MAX_QUEUE_NUM);
334 
335     s = DO_UPCAST(NetVhostUserState, nc, ncs[0]);
336     chr = qemu_chr_fe_get_driver(&s->chr);
337     trace_vhost_user_event(chr->label, event);
338     switch (event) {
339     case CHR_EVENT_OPENED:
340         if (vhost_user_start(queues, ncs, s->vhost_user) < 0) {
341             qemu_chr_fe_disconnect(&s->chr);
342             return;
343         }
344         s->watch = qemu_chr_fe_add_watch(&s->chr, G_IO_HUP,
345                                          net_vhost_user_watch, s);
346         net_client_set_link(ncs, queues, true);
347         s->started = true;
348         qapi_event_send_netdev_vhost_user_connected(name, chr->label);
349         break;
350     case CHR_EVENT_CLOSED:
351         /* a close event may happen during a read/write, but vhost
352          * code assumes the vhost_dev remains setup, so delay the
353          * stop & clear to idle.
354          * FIXME: better handle failure in vhost code, remove bh
355          */
356         if (s->watch) {
357             AioContext *ctx = qemu_get_current_aio_context();
358 
359             g_source_remove(s->watch);
360             s->watch = 0;
361             qemu_chr_fe_set_handlers(&s->chr, NULL, NULL, NULL, NULL,
362                                      NULL, NULL, false);
363 
364             aio_bh_schedule_oneshot(ctx, chr_closed_bh, opaque);
365         }
366         break;
367     case CHR_EVENT_BREAK:
368     case CHR_EVENT_MUX_IN:
369     case CHR_EVENT_MUX_OUT:
370         /* Ignore */
371         break;
372     }
373 }
374 
375 static int net_vhost_user_init(NetClientState *peer, const char *device,
376                                const char *name, Chardev *chr,
377                                int queues)
378 {
379     Error *err = NULL;
380     NetClientState *nc, *nc0 = NULL;
381     NetVhostUserState *s = NULL;
382     VhostUserState *user;
383     int i;
384 
385     assert(name);
386     assert(queues > 0);
387 
388     user = g_new0(struct VhostUserState, 1);
389     for (i = 0; i < queues; i++) {
390         nc = qemu_new_net_client(&net_vhost_user_info, peer, device, name);
391         qemu_set_info_str(nc, "vhost-user%d to %s", i, chr->label);
392         nc->queue_index = i;
393         if (!nc0) {
394             nc0 = nc;
395             s = DO_UPCAST(NetVhostUserState, nc, nc);
396             if (!qemu_chr_fe_init(&s->chr, chr, &err) ||
397                 !vhost_user_init(user, &s->chr, &err)) {
398                 error_report_err(err);
399                 goto err;
400             }
401         }
402         s = DO_UPCAST(NetVhostUserState, nc, nc);
403         s->vhost_user = user;
404     }
405 
406     s = DO_UPCAST(NetVhostUserState, nc, nc0);
407     do {
408         if (qemu_chr_fe_wait_connected(&s->chr, &err) < 0) {
409             error_report_err(err);
410             goto err;
411         }
412         qemu_chr_fe_set_handlers(&s->chr, NULL, NULL,
413                                  net_vhost_user_event, NULL, nc0->name, NULL,
414                                  true);
415     } while (!s->started);
416 
417     assert(s->vhost_net);
418 
419     return 0;
420 
421 err:
422     if (user) {
423         vhost_user_cleanup(user);
424         g_free(user);
425         if (s) {
426             s->vhost_user = NULL;
427         }
428     }
429     if (nc0) {
430         qemu_del_net_client(nc0);
431     }
432 
433     return -1;
434 }
435 
436 static Chardev *net_vhost_claim_chardev(
437     const NetdevVhostUserOptions *opts, Error **errp)
438 {
439     Chardev *chr = qemu_chr_find(opts->chardev);
440 
441     if (chr == NULL) {
442         error_setg(errp, "chardev \"%s\" not found", opts->chardev);
443         return NULL;
444     }
445 
446     if (!qemu_chr_has_feature(chr, QEMU_CHAR_FEATURE_RECONNECTABLE)) {
447         error_setg(errp, "chardev \"%s\" is not reconnectable",
448                    opts->chardev);
449         return NULL;
450     }
451     if (!qemu_chr_has_feature(chr, QEMU_CHAR_FEATURE_FD_PASS)) {
452         error_setg(errp, "chardev \"%s\" does not support FD passing",
453                    opts->chardev);
454         return NULL;
455     }
456 
457     return chr;
458 }
459 
460 int net_init_vhost_user(const Netdev *netdev, const char *name,
461                         NetClientState *peer, Error **errp)
462 {
463     int queues;
464     const NetdevVhostUserOptions *vhost_user_opts;
465     Chardev *chr;
466 
467     assert(netdev->type == NET_CLIENT_DRIVER_VHOST_USER);
468     vhost_user_opts = &netdev->u.vhost_user;
469 
470     chr = net_vhost_claim_chardev(vhost_user_opts, errp);
471     if (!chr) {
472         return -1;
473     }
474 
475     queues = vhost_user_opts->has_queues ? vhost_user_opts->queues : 1;
476     if (queues < 1 || queues > MAX_QUEUE_NUM) {
477         error_setg(errp,
478                    "vhost-user number of queues must be in range [1, %d]",
479                    MAX_QUEUE_NUM);
480         return -1;
481     }
482 
483     return net_vhost_user_init(peer, "vhost_user", name, chr, queues);
484 }
485