xref: /openbmc/qemu/net/vhost-user.c (revision 2f44bea9)
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 "hw/virtio/vhost-user.h"
16 #include "chardev/char-fe.h"
17 #include "qapi/error.h"
18 #include "qapi/qapi-commands-net.h"
19 #include "qemu/config-file.h"
20 #include "qemu/error-report.h"
21 #include "qemu/option.h"
22 #include "trace.h"
23 
24 typedef struct NetVhostUserState {
25     NetClientState nc;
26     CharBackend chr; /* only queue index 0 */
27     VhostUserState *vhost_user;
28     VHostNetState *vhost_net;
29     guint watch;
30     uint64_t acked_features;
31     bool started;
32 } NetVhostUserState;
33 
34 VHostNetState *vhost_user_get_vhost_net(NetClientState *nc)
35 {
36     NetVhostUserState *s = DO_UPCAST(NetVhostUserState, nc, nc);
37     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_USER);
38     return s->vhost_net;
39 }
40 
41 uint64_t vhost_user_get_acked_features(NetClientState *nc)
42 {
43     NetVhostUserState *s = DO_UPCAST(NetVhostUserState, nc, nc);
44     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_USER);
45     return s->acked_features;
46 }
47 
48 static void vhost_user_stop(int queues, NetClientState *ncs[])
49 {
50     NetVhostUserState *s;
51     int i;
52 
53     for (i = 0; i < queues; i++) {
54         assert(ncs[i]->info->type == NET_CLIENT_DRIVER_VHOST_USER);
55 
56         s = DO_UPCAST(NetVhostUserState, nc, ncs[i]);
57 
58         if (s->vhost_net) {
59             /* save acked features */
60             uint64_t features = vhost_net_get_acked_features(s->vhost_net);
61             if (features) {
62                 s->acked_features = features;
63             }
64             vhost_net_cleanup(s->vhost_net);
65         }
66     }
67 }
68 
69 static int vhost_user_start(int queues, NetClientState *ncs[],
70                             VhostUserState *be)
71 {
72     VhostNetOptions options;
73     struct vhost_net *net = NULL;
74     NetVhostUserState *s;
75     int max_queues;
76     int i;
77 
78     options.backend_type = VHOST_BACKEND_TYPE_USER;
79 
80     for (i = 0; i < queues; i++) {
81         assert(ncs[i]->info->type == NET_CLIENT_DRIVER_VHOST_USER);
82 
83         s = DO_UPCAST(NetVhostUserState, nc, ncs[i]);
84 
85         options.net_backend = ncs[i];
86         options.opaque      = be;
87         options.busyloop_timeout = 0;
88         options.nvqs = 2;
89         net = vhost_net_init(&options);
90         if (!net) {
91             error_report("failed to init vhost_net for queue %d", i);
92             goto err;
93         }
94 
95         if (i == 0) {
96             max_queues = vhost_net_get_max_queues(net);
97             if (queues > max_queues) {
98                 error_report("you are asking more queues than supported: %d",
99                              max_queues);
100                 goto err;
101             }
102         }
103 
104         if (s->vhost_net) {
105             vhost_net_cleanup(s->vhost_net);
106             g_free(s->vhost_net);
107         }
108         s->vhost_net = net;
109     }
110 
111     return 0;
112 
113 err:
114     if (net) {
115         vhost_net_cleanup(net);
116         g_free(net);
117     }
118     vhost_user_stop(i, ncs);
119     return -1;
120 }
121 
122 static ssize_t vhost_user_receive(NetClientState *nc, const uint8_t *buf,
123                                   size_t size)
124 {
125     /* In case of RARP (message size is 60) notify backup to send a fake RARP.
126        This fake RARP will be sent by backend only for guest
127        without GUEST_ANNOUNCE capability.
128      */
129     if (size == 60) {
130         NetVhostUserState *s = DO_UPCAST(NetVhostUserState, nc, nc);
131         int r;
132         static int display_rarp_failure = 1;
133         char mac_addr[6];
134 
135         /* extract guest mac address from the RARP message */
136         memcpy(mac_addr, &buf[6], 6);
137 
138         r = vhost_net_notify_migration_done(s->vhost_net, mac_addr);
139 
140         if ((r != 0) && (display_rarp_failure)) {
141             fprintf(stderr,
142                     "Vhost user backend fails to broadcast fake RARP\n");
143             fflush(stderr);
144             display_rarp_failure = 0;
145         }
146     }
147 
148     return size;
149 }
150 
151 static void net_vhost_user_cleanup(NetClientState *nc)
152 {
153     NetVhostUserState *s = DO_UPCAST(NetVhostUserState, nc, nc);
154 
155     if (s->vhost_net) {
156         vhost_net_cleanup(s->vhost_net);
157         g_free(s->vhost_net);
158         s->vhost_net = NULL;
159     }
160     if (nc->queue_index == 0) {
161         if (s->watch) {
162             g_source_remove(s->watch);
163             s->watch = 0;
164         }
165         qemu_chr_fe_deinit(&s->chr, true);
166         if (s->vhost_user) {
167             vhost_user_cleanup(s->vhost_user);
168             g_free(s->vhost_user);
169             s->vhost_user = NULL;
170         }
171     }
172 
173     qemu_purge_queued_packets(nc);
174 }
175 
176 static int vhost_user_set_vnet_endianness(NetClientState *nc,
177                                           bool enable)
178 {
179     /* Nothing to do.  If the server supports
180      * VHOST_USER_PROTOCOL_F_CROSS_ENDIAN, it will get the
181      * vnet header endianness from there.  If it doesn't, negotiation
182      * fails.
183      */
184     return 0;
185 }
186 
187 static bool vhost_user_has_vnet_hdr(NetClientState *nc)
188 {
189     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_USER);
190 
191     return true;
192 }
193 
194 static bool vhost_user_has_ufo(NetClientState *nc)
195 {
196     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_USER);
197 
198     return true;
199 }
200 
201 static NetClientInfo net_vhost_user_info = {
202         .type = NET_CLIENT_DRIVER_VHOST_USER,
203         .size = sizeof(NetVhostUserState),
204         .receive = vhost_user_receive,
205         .cleanup = net_vhost_user_cleanup,
206         .has_vnet_hdr = vhost_user_has_vnet_hdr,
207         .has_ufo = vhost_user_has_ufo,
208         .set_vnet_be = vhost_user_set_vnet_endianness,
209         .set_vnet_le = vhost_user_set_vnet_endianness,
210 };
211 
212 static gboolean net_vhost_user_watch(void *do_not_use, GIOCondition cond,
213                                      void *opaque)
214 {
215     NetVhostUserState *s = opaque;
216 
217     qemu_chr_fe_disconnect(&s->chr);
218 
219     return TRUE;
220 }
221 
222 static void net_vhost_user_event(void *opaque, QEMUChrEvent event);
223 
224 static void chr_closed_bh(void *opaque)
225 {
226     const char *name = opaque;
227     NetClientState *ncs[MAX_QUEUE_NUM];
228     NetVhostUserState *s;
229     Error *err = NULL;
230     int queues, i;
231 
232     queues = qemu_find_net_clients_except(name, ncs,
233                                           NET_CLIENT_DRIVER_NIC,
234                                           MAX_QUEUE_NUM);
235     assert(queues < MAX_QUEUE_NUM);
236 
237     s = DO_UPCAST(NetVhostUserState, nc, ncs[0]);
238 
239     for (i = queues -1; i >= 0; i--) {
240         s = DO_UPCAST(NetVhostUserState, nc, ncs[i]);
241 
242         if (s->vhost_net) {
243             s->acked_features = vhost_net_get_acked_features(s->vhost_net);
244         }
245     }
246 
247     qmp_set_link(name, false, &err);
248 
249     qemu_chr_fe_set_handlers(&s->chr, NULL, NULL, net_vhost_user_event,
250                              NULL, opaque, NULL, true);
251 
252     if (err) {
253         error_report_err(err);
254     }
255 }
256 
257 static void net_vhost_user_event(void *opaque, QEMUChrEvent event)
258 {
259     const char *name = opaque;
260     NetClientState *ncs[MAX_QUEUE_NUM];
261     NetVhostUserState *s;
262     Chardev *chr;
263     Error *err = NULL;
264     int queues;
265 
266     queues = qemu_find_net_clients_except(name, ncs,
267                                           NET_CLIENT_DRIVER_NIC,
268                                           MAX_QUEUE_NUM);
269     assert(queues < MAX_QUEUE_NUM);
270 
271     s = DO_UPCAST(NetVhostUserState, nc, ncs[0]);
272     chr = qemu_chr_fe_get_driver(&s->chr);
273     trace_vhost_user_event(chr->label, event);
274     switch (event) {
275     case CHR_EVENT_OPENED:
276         if (vhost_user_start(queues, ncs, s->vhost_user) < 0) {
277             qemu_chr_fe_disconnect(&s->chr);
278             return;
279         }
280         s->watch = qemu_chr_fe_add_watch(&s->chr, G_IO_HUP,
281                                          net_vhost_user_watch, s);
282         qmp_set_link(name, true, &err);
283         s->started = true;
284         break;
285     case CHR_EVENT_CLOSED:
286         /* a close event may happen during a read/write, but vhost
287          * code assumes the vhost_dev remains setup, so delay the
288          * stop & clear to idle.
289          * FIXME: better handle failure in vhost code, remove bh
290          */
291         if (s->watch) {
292             AioContext *ctx = qemu_get_current_aio_context();
293 
294             g_source_remove(s->watch);
295             s->watch = 0;
296             qemu_chr_fe_set_handlers(&s->chr, NULL, NULL, NULL, NULL,
297                                      NULL, NULL, false);
298 
299             aio_bh_schedule_oneshot(ctx, chr_closed_bh, opaque);
300         }
301         break;
302     case CHR_EVENT_BREAK:
303     case CHR_EVENT_MUX_IN:
304     case CHR_EVENT_MUX_OUT:
305         /* Ignore */
306         break;
307     }
308 
309     if (err) {
310         error_report_err(err);
311     }
312 }
313 
314 static int net_vhost_user_init(NetClientState *peer, const char *device,
315                                const char *name, Chardev *chr,
316                                int queues)
317 {
318     Error *err = NULL;
319     NetClientState *nc, *nc0 = NULL;
320     NetVhostUserState *s = NULL;
321     VhostUserState *user;
322     int i;
323 
324     assert(name);
325     assert(queues > 0);
326 
327     user = g_new0(struct VhostUserState, 1);
328     for (i = 0; i < queues; i++) {
329         nc = qemu_new_net_client(&net_vhost_user_info, peer, device, name);
330         snprintf(nc->info_str, sizeof(nc->info_str), "vhost-user%d to %s",
331                  i, chr->label);
332         nc->queue_index = i;
333         if (!nc0) {
334             nc0 = nc;
335             s = DO_UPCAST(NetVhostUserState, nc, nc);
336             if (!qemu_chr_fe_init(&s->chr, chr, &err) ||
337                 !vhost_user_init(user, &s->chr, &err)) {
338                 error_report_err(err);
339                 goto err;
340             }
341         }
342         s = DO_UPCAST(NetVhostUserState, nc, nc);
343         s->vhost_user = user;
344     }
345 
346     s = DO_UPCAST(NetVhostUserState, nc, nc0);
347     do {
348         if (qemu_chr_fe_wait_connected(&s->chr, &err) < 0) {
349             error_report_err(err);
350             goto err;
351         }
352         qemu_chr_fe_set_handlers(&s->chr, NULL, NULL,
353                                  net_vhost_user_event, NULL, nc0->name, NULL,
354                                  true);
355     } while (!s->started);
356 
357     assert(s->vhost_net);
358 
359     return 0;
360 
361 err:
362     if (user) {
363         vhost_user_cleanup(user);
364         g_free(user);
365         if (s) {
366             s->vhost_user = NULL;
367         }
368     }
369     if (nc0) {
370         qemu_del_net_client(nc0);
371     }
372 
373     return -1;
374 }
375 
376 static Chardev *net_vhost_claim_chardev(
377     const NetdevVhostUserOptions *opts, Error **errp)
378 {
379     Chardev *chr = qemu_chr_find(opts->chardev);
380 
381     if (chr == NULL) {
382         error_setg(errp, "chardev \"%s\" not found", opts->chardev);
383         return NULL;
384     }
385 
386     if (!qemu_chr_has_feature(chr, QEMU_CHAR_FEATURE_RECONNECTABLE)) {
387         error_setg(errp, "chardev \"%s\" is not reconnectable",
388                    opts->chardev);
389         return NULL;
390     }
391     if (!qemu_chr_has_feature(chr, QEMU_CHAR_FEATURE_FD_PASS)) {
392         error_setg(errp, "chardev \"%s\" does not support FD passing",
393                    opts->chardev);
394         return NULL;
395     }
396 
397     return chr;
398 }
399 
400 static int net_vhost_check_net(void *opaque, QemuOpts *opts, Error **errp)
401 {
402     const char *name = opaque;
403     const char *driver, *netdev;
404 
405     driver = qemu_opt_get(opts, "driver");
406     netdev = qemu_opt_get(opts, "netdev");
407 
408     if (!driver || !netdev) {
409         return 0;
410     }
411 
412     if (strcmp(netdev, name) == 0 &&
413         !g_str_has_prefix(driver, "virtio-net-")) {
414         error_setg(errp, "vhost-user requires frontend driver virtio-net-*");
415         return -1;
416     }
417 
418     return 0;
419 }
420 
421 int net_init_vhost_user(const Netdev *netdev, const char *name,
422                         NetClientState *peer, Error **errp)
423 {
424     int queues;
425     const NetdevVhostUserOptions *vhost_user_opts;
426     Chardev *chr;
427 
428     assert(netdev->type == NET_CLIENT_DRIVER_VHOST_USER);
429     vhost_user_opts = &netdev->u.vhost_user;
430 
431     chr = net_vhost_claim_chardev(vhost_user_opts, errp);
432     if (!chr) {
433         return -1;
434     }
435 
436     /* verify net frontend */
437     if (qemu_opts_foreach(qemu_find_opts("device"), net_vhost_check_net,
438                           (char *)name, errp)) {
439         return -1;
440     }
441 
442     queues = vhost_user_opts->has_queues ? vhost_user_opts->queues : 1;
443     if (queues < 1 || queues > MAX_QUEUE_NUM) {
444         error_setg(errp,
445                    "vhost-user number of queues must be in range [1, %d]",
446                    MAX_QUEUE_NUM);
447         return -1;
448     }
449 
450     return net_vhost_user_init(peer, "vhost_user", name, chr, queues);
451 }
452