xref: /openbmc/qemu/net/vhost-user.c (revision 0b1183e3) 
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 "net/vhost-user.h"
15 #include "chardev/char-fe.h"
16 #include "qemu/config-file.h"
17 #include "qemu/error-report.h"
18 #include "qmp-commands.h"
19 #include "trace.h"
20 
21 typedef struct VhostUserState {
22     NetClientState nc;
23     CharBackend chr; /* only queue index 0 */
24     VHostNetState *vhost_net;
25     guint watch;
26     uint64_t acked_features;
27     bool started;
28 } VhostUserState;
29 
30 VHostNetState *vhost_user_get_vhost_net(NetClientState *nc)
31 {
32     VhostUserState *s = DO_UPCAST(VhostUserState, nc, nc);
33     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_USER);
34     return s->vhost_net;
35 }
36 
37 uint64_t vhost_user_get_acked_features(NetClientState *nc)
38 {
39     VhostUserState *s = DO_UPCAST(VhostUserState, nc, nc);
40     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_USER);
41     return s->acked_features;
42 }
43 
44 static void vhost_user_stop(int queues, NetClientState *ncs[])
45 {
46     VhostUserState *s;
47     int i;
48 
49     for (i = 0; i < queues; i++) {
50         assert(ncs[i]->info->type == NET_CLIENT_DRIVER_VHOST_USER);
51 
52         s = DO_UPCAST(VhostUserState, nc, ncs[i]);
53 
54         if (s->vhost_net) {
55             /* save acked features */
56             uint64_t features = vhost_net_get_acked_features(s->vhost_net);
57             if (features) {
58                 s->acked_features = features;
59             }
60             vhost_net_cleanup(s->vhost_net);
61         }
62     }
63 }
64 
65 static int vhost_user_start(int queues, NetClientState *ncs[], CharBackend *be)
66 {
67     VhostNetOptions options;
68     struct vhost_net *net = NULL;
69     VhostUserState *s;
70     int max_queues;
71     int i;
72 
73     options.backend_type = VHOST_BACKEND_TYPE_USER;
74 
75     for (i = 0; i < queues; i++) {
76         assert(ncs[i]->info->type == NET_CLIENT_DRIVER_VHOST_USER);
77 
78         s = DO_UPCAST(VhostUserState, nc, ncs[i]);
79 
80         options.net_backend = ncs[i];
81         options.opaque      = be;
82         options.busyloop_timeout = 0;
83         net = vhost_net_init(&options);
84         if (!net) {
85             error_report("failed to init vhost_net for queue %d", i);
86             goto err;
87         }
88 
89         if (i == 0) {
90             max_queues = vhost_net_get_max_queues(net);
91             if (queues > max_queues) {
92                 error_report("you are asking more queues than supported: %d",
93                              max_queues);
94                 goto err;
95             }
96         }
97 
98         if (s->vhost_net) {
99             vhost_net_cleanup(s->vhost_net);
100             g_free(s->vhost_net);
101         }
102         s->vhost_net = net;
103     }
104 
105     return 0;
106 
107 err:
108     if (net) {
109         vhost_net_cleanup(net);
110     }
111     vhost_user_stop(i, ncs);
112     return -1;
113 }
114 
115 static ssize_t vhost_user_receive(NetClientState *nc, const uint8_t *buf,
116                                   size_t size)
117 {
118     /* In case of RARP (message size is 60) notify backup to send a fake RARP.
119        This fake RARP will be sent by backend only for guest
120        without GUEST_ANNOUNCE capability.
121      */
122     if (size == 60) {
123         VhostUserState *s = DO_UPCAST(VhostUserState, nc, nc);
124         int r;
125         static int display_rarp_failure = 1;
126         char mac_addr[6];
127 
128         /* extract guest mac address from the RARP message */
129         memcpy(mac_addr, &buf[6], 6);
130 
131         r = vhost_net_notify_migration_done(s->vhost_net, mac_addr);
132 
133         if ((r != 0) && (display_rarp_failure)) {
134             fprintf(stderr,
135                     "Vhost user backend fails to broadcast fake RARP\n");
136             fflush(stderr);
137             display_rarp_failure = 0;
138         }
139     }
140 
141     return size;
142 }
143 
144 static void vhost_user_cleanup(NetClientState *nc)
145 {
146     VhostUserState *s = DO_UPCAST(VhostUserState, nc, nc);
147 
148     if (s->vhost_net) {
149         vhost_net_cleanup(s->vhost_net);
150         g_free(s->vhost_net);
151         s->vhost_net = NULL;
152     }
153     if (nc->queue_index == 0) {
154         qemu_chr_fe_deinit(&s->chr, true);
155     }
156 
157     qemu_purge_queued_packets(nc);
158 }
159 
160 static bool vhost_user_has_vnet_hdr(NetClientState *nc)
161 {
162     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_USER);
163 
164     return true;
165 }
166 
167 static bool vhost_user_has_ufo(NetClientState *nc)
168 {
169     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_USER);
170 
171     return true;
172 }
173 
174 static NetClientInfo net_vhost_user_info = {
175         .type = NET_CLIENT_DRIVER_VHOST_USER,
176         .size = sizeof(VhostUserState),
177         .receive = vhost_user_receive,
178         .cleanup = vhost_user_cleanup,
179         .has_vnet_hdr = vhost_user_has_vnet_hdr,
180         .has_ufo = vhost_user_has_ufo,
181 };
182 
183 static gboolean net_vhost_user_watch(GIOChannel *chan, GIOCondition cond,
184                                            void *opaque)
185 {
186     VhostUserState *s = opaque;
187 
188     qemu_chr_fe_disconnect(&s->chr);
189 
190     return TRUE;
191 }
192 
193 static void net_vhost_user_event(void *opaque, int event);
194 
195 static void chr_closed_bh(void *opaque)
196 {
197     const char *name = opaque;
198     NetClientState *ncs[MAX_QUEUE_NUM];
199     VhostUserState *s;
200     Error *err = NULL;
201     int queues;
202 
203     queues = qemu_find_net_clients_except(name, ncs,
204                                           NET_CLIENT_DRIVER_NIC,
205                                           MAX_QUEUE_NUM);
206     assert(queues < MAX_QUEUE_NUM);
207 
208     s = DO_UPCAST(VhostUserState, nc, ncs[0]);
209 
210     qmp_set_link(name, false, &err);
211     vhost_user_stop(queues, ncs);
212 
213     qemu_chr_fe_set_handlers(&s->chr, NULL, NULL, net_vhost_user_event,
214                              NULL, opaque, NULL, true);
215 
216     if (err) {
217         error_report_err(err);
218     }
219 }
220 
221 static void net_vhost_user_event(void *opaque, int event)
222 {
223     const char *name = opaque;
224     NetClientState *ncs[MAX_QUEUE_NUM];
225     VhostUserState *s;
226     Chardev *chr;
227     Error *err = NULL;
228     int queues;
229 
230     queues = qemu_find_net_clients_except(name, ncs,
231                                           NET_CLIENT_DRIVER_NIC,
232                                           MAX_QUEUE_NUM);
233     assert(queues < MAX_QUEUE_NUM);
234 
235     s = DO_UPCAST(VhostUserState, nc, ncs[0]);
236     chr = qemu_chr_fe_get_driver(&s->chr);
237     trace_vhost_user_event(chr->label, event);
238     switch (event) {
239     case CHR_EVENT_OPENED:
240         if (vhost_user_start(queues, ncs, &s->chr) < 0) {
241             qemu_chr_fe_disconnect(&s->chr);
242             return;
243         }
244         s->watch = qemu_chr_fe_add_watch(&s->chr, G_IO_HUP,
245                                          net_vhost_user_watch, s);
246         qmp_set_link(name, true, &err);
247         s->started = true;
248         break;
249     case CHR_EVENT_CLOSED:
250         /* a close event may happen during a read/write, but vhost
251          * code assumes the vhost_dev remains setup, so delay the
252          * stop & clear to idle.
253          * FIXME: better handle failure in vhost code, remove bh
254          */
255         if (s->watch) {
256             AioContext *ctx = qemu_get_current_aio_context();
257 
258             g_source_remove(s->watch);
259             s->watch = 0;
260             qemu_chr_fe_set_handlers(&s->chr, NULL, NULL, NULL, NULL,
261                                      NULL, NULL, false);
262 
263             aio_bh_schedule_oneshot(ctx, chr_closed_bh, opaque);
264         }
265         break;
266     }
267 
268     if (err) {
269         error_report_err(err);
270     }
271 }
272 
273 static int net_vhost_user_init(NetClientState *peer, const char *device,
274                                const char *name, Chardev *chr,
275                                int queues)
276 {
277     Error *err = NULL;
278     NetClientState *nc, *nc0 = NULL;
279     VhostUserState *s;
280     int i;
281 
282     assert(name);
283     assert(queues > 0);
284 
285     for (i = 0; i < queues; i++) {
286         nc = qemu_new_net_client(&net_vhost_user_info, peer, device, name);
287         snprintf(nc->info_str, sizeof(nc->info_str), "vhost-user%d to %s",
288                  i, chr->label);
289         nc->queue_index = i;
290         if (!nc0) {
291             nc0 = nc;
292             s = DO_UPCAST(VhostUserState, nc, nc);
293             if (!qemu_chr_fe_init(&s->chr, chr, &err)) {
294                 error_report_err(err);
295                 return -1;
296             }
297         }
298 
299     }
300 
301     s = DO_UPCAST(VhostUserState, nc, nc0);
302     do {
303         if (qemu_chr_fe_wait_connected(&s->chr, &err) < 0) {
304             error_report_err(err);
305             return -1;
306         }
307         qemu_chr_fe_set_handlers(&s->chr, NULL, NULL,
308                                  net_vhost_user_event, NULL, nc0->name, NULL,
309                                  true);
310     } while (!s->started);
311 
312     assert(s->vhost_net);
313 
314     return 0;
315 }
316 
317 static Chardev *net_vhost_claim_chardev(
318     const NetdevVhostUserOptions *opts, Error **errp)
319 {
320     Chardev *chr = qemu_chr_find(opts->chardev);
321 
322     if (chr == NULL) {
323         error_setg(errp, "chardev \"%s\" not found", opts->chardev);
324         return NULL;
325     }
326 
327     if (!qemu_chr_has_feature(chr, QEMU_CHAR_FEATURE_RECONNECTABLE)) {
328         error_setg(errp, "chardev \"%s\" is not reconnectable",
329                    opts->chardev);
330         return NULL;
331     }
332     if (!qemu_chr_has_feature(chr, QEMU_CHAR_FEATURE_FD_PASS)) {
333         error_setg(errp, "chardev \"%s\" does not support FD passing",
334                    opts->chardev);
335         return NULL;
336     }
337 
338     return chr;
339 }
340 
341 static int net_vhost_check_net(void *opaque, QemuOpts *opts, Error **errp)
342 {
343     const char *name = opaque;
344     const char *driver, *netdev;
345 
346     driver = qemu_opt_get(opts, "driver");
347     netdev = qemu_opt_get(opts, "netdev");
348 
349     if (!driver || !netdev) {
350         return 0;
351     }
352 
353     if (strcmp(netdev, name) == 0 &&
354         !g_str_has_prefix(driver, "virtio-net-")) {
355         error_setg(errp, "vhost-user requires frontend driver virtio-net-*");
356         return -1;
357     }
358 
359     return 0;
360 }
361 
362 int net_init_vhost_user(const Netdev *netdev, const char *name,
363                         NetClientState *peer, Error **errp)
364 {
365     int queues;
366     const NetdevVhostUserOptions *vhost_user_opts;
367     Chardev *chr;
368 
369     assert(netdev->type == NET_CLIENT_DRIVER_VHOST_USER);
370     vhost_user_opts = &netdev->u.vhost_user;
371 
372     chr = net_vhost_claim_chardev(vhost_user_opts, errp);
373     if (!chr) {
374         return -1;
375     }
376 
377     /* verify net frontend */
378     if (qemu_opts_foreach(qemu_find_opts("device"), net_vhost_check_net,
379                           (char *)name, errp)) {
380         return -1;
381     }
382 
383     queues = vhost_user_opts->has_queues ? vhost_user_opts->queues : 1;
384     if (queues < 1 || queues > MAX_QUEUE_NUM) {
385         error_setg(errp,
386                    "vhost-user number of queues must be in range [1, %d]",
387                    MAX_QUEUE_NUM);
388         return -1;
389     }
390 
391     return net_vhost_user_init(peer, "vhost_user", name, chr, queues);
392 }
393