xref: /openbmc/qemu/util/vhost-user-server.c (revision 55abfc1f)
1 /*
2  * Sharing QEMU devices via vhost-user protocol
3  *
4  * Copyright (c) Coiby Xu <coiby.xu@gmail.com>.
5  * Copyright (c) 2020 Red Hat, Inc.
6  *
7  * This work is licensed under the terms of the GNU GPL, version 2 or
8  * later.  See the COPYING file in the top-level directory.
9  */
10 #include "qemu/osdep.h"
11 #include "qemu/error-report.h"
12 #include "qemu/main-loop.h"
13 #include "qemu/vhost-user-server.h"
14 #include "block/aio-wait.h"
15 
16 /*
17  * Theory of operation:
18  *
19  * VuServer is started and stopped by vhost_user_server_start() and
20  * vhost_user_server_stop() from the main loop thread. Starting the server
21  * opens a vhost-user UNIX domain socket and listens for incoming connections.
22  * Only one connection is allowed at a time.
23  *
24  * The connection is handled by the vu_client_trip() coroutine in the
25  * VuServer->ctx AioContext. The coroutine consists of a vu_dispatch() loop
26  * where libvhost-user calls vu_message_read() to receive the next vhost-user
27  * protocol messages over the UNIX domain socket.
28  *
29  * When virtqueues are set up libvhost-user calls set_watch() to monitor kick
30  * fds. These fds are also handled in the VuServer->ctx AioContext.
31  *
32  * Both vu_client_trip() and kick fd monitoring can be stopped by shutting down
33  * the socket connection. Shutting down the socket connection causes
34  * vu_message_read() to fail since no more data can be received from the socket.
35  * After vu_dispatch() fails, vu_client_trip() calls vu_deinit() to stop
36  * libvhost-user before terminating the coroutine. vu_deinit() calls
37  * remove_watch() to stop monitoring kick fds and this stops virtqueue
38  * processing.
39  *
40  * When vu_client_trip() has finished cleaning up it schedules a BH in the main
41  * loop thread to accept the next client connection.
42  *
43  * When libvhost-user detects an error it calls panic_cb() and sets the
44  * dev->broken flag. Both vu_client_trip() and kick fd processing stop when
45  * the dev->broken flag is set.
46  *
47  * It is possible to switch AioContexts using
48  * vhost_user_server_detach_aio_context() and
49  * vhost_user_server_attach_aio_context(). They stop monitoring fds in the old
50  * AioContext and resume monitoring in the new AioContext. The vu_client_trip()
51  * coroutine remains in a yielded state during the switch. This is made
52  * possible by QIOChannel's support for spurious coroutine re-entry in
53  * qio_channel_yield(). The coroutine will restart I/O when re-entered from the
54  * new AioContext.
55  */
56 
57 static void vmsg_close_fds(VhostUserMsg *vmsg)
58 {
59     int i;
60     for (i = 0; i < vmsg->fd_num; i++) {
61         close(vmsg->fds[i]);
62     }
63 }
64 
65 static void vmsg_unblock_fds(VhostUserMsg *vmsg)
66 {
67     int i;
68     for (i = 0; i < vmsg->fd_num; i++) {
69         qemu_socket_set_nonblock(vmsg->fds[i]);
70     }
71 }
72 
73 static void panic_cb(VuDev *vu_dev, const char *buf)
74 {
75     error_report("vu_panic: %s", buf);
76 }
77 
78 void vhost_user_server_inc_in_flight(VuServer *server)
79 {
80     assert(!server->wait_idle);
81     qatomic_inc(&server->in_flight);
82 }
83 
84 void vhost_user_server_dec_in_flight(VuServer *server)
85 {
86     if (qatomic_fetch_dec(&server->in_flight) == 1) {
87         if (server->wait_idle) {
88             aio_co_wake(server->co_trip);
89         }
90     }
91 }
92 
93 bool vhost_user_server_has_in_flight(VuServer *server)
94 {
95     return qatomic_load_acquire(&server->in_flight) > 0;
96 }
97 
98 static bool coroutine_fn
99 vu_message_read(VuDev *vu_dev, int conn_fd, VhostUserMsg *vmsg)
100 {
101     struct iovec iov = {
102         .iov_base = (char *)vmsg,
103         .iov_len = VHOST_USER_HDR_SIZE,
104     };
105     int rc, read_bytes = 0;
106     Error *local_err = NULL;
107     const size_t max_fds = G_N_ELEMENTS(vmsg->fds);
108     VuServer *server = container_of(vu_dev, VuServer, vu_dev);
109     QIOChannel *ioc = server->ioc;
110 
111     vmsg->fd_num = 0;
112     if (!ioc) {
113         error_report_err(local_err);
114         goto fail;
115     }
116 
117     assert(qemu_in_coroutine());
118     do {
119         size_t nfds = 0;
120         int *fds = NULL;
121 
122         /*
123          * qio_channel_readv_full may have short reads, keeping calling it
124          * until getting VHOST_USER_HDR_SIZE or 0 bytes in total
125          */
126         rc = qio_channel_readv_full(ioc, &iov, 1, &fds, &nfds, 0, &local_err);
127         if (rc < 0) {
128             if (rc == QIO_CHANNEL_ERR_BLOCK) {
129                 assert(local_err == NULL);
130                 if (server->ctx) {
131                     server->in_qio_channel_yield = true;
132                     qio_channel_yield(ioc, G_IO_IN);
133                     server->in_qio_channel_yield = false;
134                 } else {
135                     return false;
136                 }
137                 continue;
138             } else {
139                 error_report_err(local_err);
140                 goto fail;
141             }
142         }
143 
144         if (nfds > 0) {
145             if (vmsg->fd_num + nfds > max_fds) {
146                 error_report("A maximum of %zu fds are allowed, "
147                              "however got %zu fds now",
148                              max_fds, vmsg->fd_num + nfds);
149                 g_free(fds);
150                 goto fail;
151             }
152             memcpy(vmsg->fds + vmsg->fd_num, fds, nfds * sizeof(vmsg->fds[0]));
153             vmsg->fd_num += nfds;
154             g_free(fds);
155         }
156 
157         if (rc == 0) { /* socket closed */
158             goto fail;
159         }
160 
161         iov.iov_base += rc;
162         iov.iov_len -= rc;
163         read_bytes += rc;
164     } while (read_bytes != VHOST_USER_HDR_SIZE);
165 
166     /* qio_channel_readv_full will make socket fds blocking, unblock them */
167     vmsg_unblock_fds(vmsg);
168     if (vmsg->size > sizeof(vmsg->payload)) {
169         error_report("Error: too big message request: %d, "
170                      "size: vmsg->size: %u, "
171                      "while sizeof(vmsg->payload) = %zu",
172                      vmsg->request, vmsg->size, sizeof(vmsg->payload));
173         goto fail;
174     }
175 
176     struct iovec iov_payload = {
177         .iov_base = (char *)&vmsg->payload,
178         .iov_len = vmsg->size,
179     };
180     if (vmsg->size) {
181         rc = qio_channel_readv_all_eof(ioc, &iov_payload, 1, &local_err);
182         if (rc != 1) {
183             if (local_err) {
184                 error_report_err(local_err);
185             }
186             goto fail;
187         }
188     }
189 
190     return true;
191 
192 fail:
193     vmsg_close_fds(vmsg);
194 
195     return false;
196 }
197 
198 static coroutine_fn void vu_client_trip(void *opaque)
199 {
200     VuServer *server = opaque;
201     VuDev *vu_dev = &server->vu_dev;
202 
203     while (!vu_dev->broken) {
204         if (server->quiescing) {
205             server->co_trip = NULL;
206             aio_wait_kick();
207             return;
208         }
209         /* vu_dispatch() returns false if server->ctx went away */
210         if (!vu_dispatch(vu_dev) && server->ctx) {
211             break;
212         }
213     }
214 
215     if (vhost_user_server_has_in_flight(server)) {
216         /* Wait for requests to complete before we can unmap the memory */
217         server->wait_idle = true;
218         qemu_coroutine_yield();
219         server->wait_idle = false;
220     }
221     assert(!vhost_user_server_has_in_flight(server));
222 
223     vu_deinit(vu_dev);
224 
225     /* vu_deinit() should have called remove_watch() */
226     assert(QTAILQ_EMPTY(&server->vu_fd_watches));
227 
228     object_unref(OBJECT(server->sioc));
229     server->sioc = NULL;
230 
231     object_unref(OBJECT(server->ioc));
232     server->ioc = NULL;
233 
234     server->co_trip = NULL;
235     if (server->restart_listener_bh) {
236         qemu_bh_schedule(server->restart_listener_bh);
237     }
238     aio_wait_kick();
239 }
240 
241 /*
242  * a wrapper for vu_kick_cb
243  *
244  * since aio_dispatch can only pass one user data pointer to the
245  * callback function, pack VuDev and pvt into a struct. Then unpack it
246  * and pass them to vu_kick_cb
247  */
248 static void kick_handler(void *opaque)
249 {
250     VuFdWatch *vu_fd_watch = opaque;
251     VuDev *vu_dev = vu_fd_watch->vu_dev;
252 
253     vu_fd_watch->cb(vu_dev, 0, vu_fd_watch->pvt);
254 
255     /* Stop vu_client_trip() if an error occurred in vu_fd_watch->cb() */
256     if (vu_dev->broken) {
257         VuServer *server = container_of(vu_dev, VuServer, vu_dev);
258 
259         qio_channel_shutdown(server->ioc, QIO_CHANNEL_SHUTDOWN_BOTH, NULL);
260     }
261 }
262 
263 static VuFdWatch *find_vu_fd_watch(VuServer *server, int fd)
264 {
265 
266     VuFdWatch *vu_fd_watch, *next;
267     QTAILQ_FOREACH_SAFE(vu_fd_watch, &server->vu_fd_watches, next, next) {
268         if (vu_fd_watch->fd == fd) {
269             return vu_fd_watch;
270         }
271     }
272     return NULL;
273 }
274 
275 static void
276 set_watch(VuDev *vu_dev, int fd, int vu_evt,
277           vu_watch_cb cb, void *pvt)
278 {
279 
280     VuServer *server = container_of(vu_dev, VuServer, vu_dev);
281     g_assert(vu_dev);
282     g_assert(fd >= 0);
283     g_assert(cb);
284 
285     VuFdWatch *vu_fd_watch = find_vu_fd_watch(server, fd);
286 
287     if (!vu_fd_watch) {
288         vu_fd_watch = g_new0(VuFdWatch, 1);
289 
290         QTAILQ_INSERT_TAIL(&server->vu_fd_watches, vu_fd_watch, next);
291 
292         vu_fd_watch->fd = fd;
293         vu_fd_watch->cb = cb;
294         qemu_socket_set_nonblock(fd);
295         aio_set_fd_handler(server->ctx, fd, kick_handler,
296                            NULL, NULL, NULL, vu_fd_watch);
297         vu_fd_watch->vu_dev = vu_dev;
298         vu_fd_watch->pvt = pvt;
299     }
300 }
301 
302 
303 static void remove_watch(VuDev *vu_dev, int fd)
304 {
305     VuServer *server;
306     g_assert(vu_dev);
307     g_assert(fd >= 0);
308 
309     server = container_of(vu_dev, VuServer, vu_dev);
310 
311     VuFdWatch *vu_fd_watch = find_vu_fd_watch(server, fd);
312 
313     if (!vu_fd_watch) {
314         return;
315     }
316     aio_set_fd_handler(server->ctx, fd, NULL, NULL, NULL, NULL, NULL);
317 
318     QTAILQ_REMOVE(&server->vu_fd_watches, vu_fd_watch, next);
319     g_free(vu_fd_watch);
320 }
321 
322 
323 static void vu_accept(QIONetListener *listener, QIOChannelSocket *sioc,
324                       gpointer opaque)
325 {
326     VuServer *server = opaque;
327 
328     if (server->sioc) {
329         warn_report("Only one vhost-user client is allowed to "
330                     "connect the server one time");
331         return;
332     }
333 
334     if (!vu_init(&server->vu_dev, server->max_queues, sioc->fd, panic_cb,
335                  vu_message_read, set_watch, remove_watch, server->vu_iface)) {
336         error_report("Failed to initialize libvhost-user");
337         return;
338     }
339 
340     /*
341      * Unset the callback function for network listener to make another
342      * vhost-user client keeping waiting until this client disconnects
343      */
344     qio_net_listener_set_client_func(server->listener,
345                                      NULL,
346                                      NULL,
347                                      NULL);
348     server->sioc = sioc;
349     /*
350      * Increase the object reference, so sioc will not freed by
351      * qio_net_listener_channel_func which will call object_unref(OBJECT(sioc))
352      */
353     object_ref(OBJECT(server->sioc));
354     qio_channel_set_name(QIO_CHANNEL(sioc), "vhost-user client");
355     server->ioc = QIO_CHANNEL(sioc);
356     object_ref(OBJECT(server->ioc));
357 
358     /* TODO vu_message_write() spins if non-blocking! */
359     qio_channel_set_blocking(server->ioc, false, NULL);
360 
361     qio_channel_set_follow_coroutine_ctx(server->ioc, true);
362 
363     /* Attaching the AioContext starts the vu_client_trip coroutine */
364     aio_context_acquire(server->ctx);
365     vhost_user_server_attach_aio_context(server, server->ctx);
366     aio_context_release(server->ctx);
367 }
368 
369 /* server->ctx acquired by caller */
370 void vhost_user_server_stop(VuServer *server)
371 {
372     qemu_bh_delete(server->restart_listener_bh);
373     server->restart_listener_bh = NULL;
374 
375     if (server->sioc) {
376         VuFdWatch *vu_fd_watch;
377 
378         QTAILQ_FOREACH(vu_fd_watch, &server->vu_fd_watches, next) {
379             aio_set_fd_handler(server->ctx, vu_fd_watch->fd,
380                                NULL, NULL, NULL, NULL, vu_fd_watch);
381         }
382 
383         qio_channel_shutdown(server->ioc, QIO_CHANNEL_SHUTDOWN_BOTH, NULL);
384 
385         AIO_WAIT_WHILE(server->ctx, server->co_trip);
386     }
387 
388     if (server->listener) {
389         qio_net_listener_disconnect(server->listener);
390         object_unref(OBJECT(server->listener));
391     }
392 }
393 
394 /*
395  * Allow the next client to connect to the server. Called from a BH in the main
396  * loop.
397  */
398 static void restart_listener_bh(void *opaque)
399 {
400     VuServer *server = opaque;
401 
402     qio_net_listener_set_client_func(server->listener, vu_accept, server,
403                                      NULL);
404 }
405 
406 /* Called with ctx acquired */
407 void vhost_user_server_attach_aio_context(VuServer *server, AioContext *ctx)
408 {
409     VuFdWatch *vu_fd_watch;
410 
411     server->ctx = ctx;
412 
413     if (!server->sioc) {
414         return;
415     }
416 
417     QTAILQ_FOREACH(vu_fd_watch, &server->vu_fd_watches, next) {
418         aio_set_fd_handler(ctx, vu_fd_watch->fd, kick_handler, NULL,
419                            NULL, NULL, vu_fd_watch);
420     }
421 
422     if (server->co_trip) {
423         /*
424          * The caller didn't fully shut down co_trip (this can happen on
425          * non-polling drains like in bdrv_graph_wrlock()). This is okay as long
426          * as it no longer tries to shut it down and we're guaranteed to still
427          * be in the same AioContext as before.
428          *
429          * co_ctx can still be NULL if we get multiple calls and only just
430          * scheduled a new coroutine in the else branch.
431          */
432         AioContext *co_ctx = qemu_coroutine_get_aio_context(server->co_trip);
433 
434         assert(!server->quiescing);
435         assert(!co_ctx || co_ctx == ctx);
436     } else {
437         server->co_trip = qemu_coroutine_create(vu_client_trip, server);
438         assert(!server->in_qio_channel_yield);
439         aio_co_schedule(ctx, server->co_trip);
440     }
441 }
442 
443 /* Called with server->ctx acquired */
444 void vhost_user_server_detach_aio_context(VuServer *server)
445 {
446     if (server->sioc) {
447         VuFdWatch *vu_fd_watch;
448 
449         QTAILQ_FOREACH(vu_fd_watch, &server->vu_fd_watches, next) {
450             aio_set_fd_handler(server->ctx, vu_fd_watch->fd,
451                                NULL, NULL, NULL, NULL, vu_fd_watch);
452         }
453     }
454 
455     server->ctx = NULL;
456 
457     if (server->ioc) {
458         if (server->in_qio_channel_yield) {
459             /* Stop receiving the next vhost-user message */
460             qio_channel_wake_read(server->ioc);
461         }
462     }
463 }
464 
465 bool vhost_user_server_start(VuServer *server,
466                              SocketAddress *socket_addr,
467                              AioContext *ctx,
468                              uint16_t max_queues,
469                              const VuDevIface *vu_iface,
470                              Error **errp)
471 {
472     QEMUBH *bh;
473     QIONetListener *listener;
474 
475     if (socket_addr->type != SOCKET_ADDRESS_TYPE_UNIX &&
476         socket_addr->type != SOCKET_ADDRESS_TYPE_FD) {
477         error_setg(errp, "Only socket address types 'unix' and 'fd' are supported");
478         return false;
479     }
480 
481     listener = qio_net_listener_new();
482     if (qio_net_listener_open_sync(listener, socket_addr, 1,
483                                    errp) < 0) {
484         object_unref(OBJECT(listener));
485         return false;
486     }
487 
488     bh = qemu_bh_new(restart_listener_bh, server);
489 
490     /* zero out unspecified fields */
491     *server = (VuServer) {
492         .listener              = listener,
493         .restart_listener_bh   = bh,
494         .vu_iface              = vu_iface,
495         .max_queues            = max_queues,
496         .ctx                   = ctx,
497     };
498 
499     qio_net_listener_set_name(server->listener, "vhost-user-backend-listener");
500 
501     qio_net_listener_set_client_func(server->listener,
502                                      vu_accept,
503                                      server,
504                                      NULL);
505 
506     QTAILQ_INIT(&server->vu_fd_watches);
507     return true;
508 }
509