xref: /openbmc/qemu/block/nbd.c (revision 19f4ed36)
1 /*
2  * QEMU Block driver for  NBD
3  *
4  * Copyright (c) 2019 Virtuozzo International GmbH.
5  * Copyright (C) 2016 Red Hat, Inc.
6  * Copyright (C) 2008 Bull S.A.S.
7  *     Author: Laurent Vivier <Laurent.Vivier@bull.net>
8  *
9  * Some parts:
10  *    Copyright (C) 2007 Anthony Liguori <anthony@codemonkey.ws>
11  *
12  * Permission is hereby granted, free of charge, to any person obtaining a copy
13  * of this software and associated documentation files (the "Software"), to deal
14  * in the Software without restriction, including without limitation the rights
15  * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
16  * copies of the Software, and to permit persons to whom the Software is
17  * furnished to do so, subject to the following conditions:
18  *
19  * The above copyright notice and this permission notice shall be included in
20  * all copies or substantial portions of the Software.
21  *
22  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
23  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
24  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
25  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
26  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
27  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
28  * THE SOFTWARE.
29  */
30 
31 #include "qemu/osdep.h"
32 
33 #include "trace.h"
34 #include "qemu/uri.h"
35 #include "qemu/option.h"
36 #include "qemu/cutils.h"
37 #include "qemu/main-loop.h"
38 #include "qemu/atomic.h"
39 
40 #include "qapi/qapi-visit-sockets.h"
41 #include "qapi/qmp/qstring.h"
42 #include "qapi/clone-visitor.h"
43 
44 #include "block/qdict.h"
45 #include "block/nbd.h"
46 #include "block/block_int.h"
47 
48 #include "qemu/yank.h"
49 
50 #define EN_OPTSTR ":exportname="
51 #define MAX_NBD_REQUESTS    16
52 
53 #define HANDLE_TO_INDEX(bs, handle) ((handle) ^ (uint64_t)(intptr_t)(bs))
54 #define INDEX_TO_HANDLE(bs, index)  ((index)  ^ (uint64_t)(intptr_t)(bs))
55 
56 typedef struct {
57     Coroutine *coroutine;
58     uint64_t offset;        /* original offset of the request */
59     bool receiving;         /* waiting for connection_co? */
60 } NBDClientRequest;
61 
62 typedef enum NBDClientState {
63     NBD_CLIENT_CONNECTING_WAIT,
64     NBD_CLIENT_CONNECTING_NOWAIT,
65     NBD_CLIENT_CONNECTED,
66     NBD_CLIENT_QUIT
67 } NBDClientState;
68 
69 typedef enum NBDConnectThreadState {
70     /* No thread, no pending results */
71     CONNECT_THREAD_NONE,
72 
73     /* Thread is running, no results for now */
74     CONNECT_THREAD_RUNNING,
75 
76     /*
77      * Thread is running, but requestor exited. Thread should close
78      * the new socket and free the connect state on exit.
79      */
80     CONNECT_THREAD_RUNNING_DETACHED,
81 
82     /* Thread finished, results are stored in a state */
83     CONNECT_THREAD_FAIL,
84     CONNECT_THREAD_SUCCESS
85 } NBDConnectThreadState;
86 
87 typedef struct NBDConnectThread {
88     /* Initialization constants */
89     SocketAddress *saddr; /* address to connect to */
90     /*
91      * Bottom half to schedule on completion. Scheduled only if bh_ctx is not
92      * NULL
93      */
94     QEMUBHFunc *bh_func;
95     void *bh_opaque;
96 
97     /*
98      * Result of last attempt. Valid in FAIL and SUCCESS states.
99      * If you want to steal error, don't forget to set pointer to NULL.
100      */
101     QIOChannelSocket *sioc;
102     Error *err;
103 
104     /* state and bh_ctx are protected by mutex */
105     QemuMutex mutex;
106     NBDConnectThreadState state; /* current state of the thread */
107     AioContext *bh_ctx; /* where to schedule bh (NULL means don't schedule) */
108 } NBDConnectThread;
109 
110 typedef struct BDRVNBDState {
111     QIOChannelSocket *sioc; /* The master data channel */
112     QIOChannel *ioc; /* The current I/O channel which may differ (eg TLS) */
113     NBDExportInfo info;
114 
115     CoMutex send_mutex;
116     CoQueue free_sema;
117     Coroutine *connection_co;
118     Coroutine *teardown_co;
119     QemuCoSleepState *connection_co_sleep_ns_state;
120     bool drained;
121     bool wait_drained_end;
122     int in_flight;
123     NBDClientState state;
124     int connect_status;
125     Error *connect_err;
126     bool wait_in_flight;
127 
128     QEMUTimer *reconnect_delay_timer;
129 
130     NBDClientRequest requests[MAX_NBD_REQUESTS];
131     NBDReply reply;
132     BlockDriverState *bs;
133 
134     /* Connection parameters */
135     uint32_t reconnect_delay;
136     SocketAddress *saddr;
137     char *export, *tlscredsid;
138     QCryptoTLSCreds *tlscreds;
139     const char *hostname;
140     char *x_dirty_bitmap;
141     bool alloc_depth;
142 
143     bool wait_connect;
144     NBDConnectThread *connect_thread;
145 } BDRVNBDState;
146 
147 static int nbd_establish_connection(BlockDriverState *bs, SocketAddress *saddr,
148                                     Error **errp);
149 static int nbd_co_establish_connection(BlockDriverState *bs, Error **errp);
150 static void nbd_co_establish_connection_cancel(BlockDriverState *bs,
151                                                bool detach);
152 static int nbd_client_handshake(BlockDriverState *bs, Error **errp);
153 static void nbd_yank(void *opaque);
154 
155 static void nbd_clear_bdrvstate(BDRVNBDState *s)
156 {
157     object_unref(OBJECT(s->tlscreds));
158     qapi_free_SocketAddress(s->saddr);
159     s->saddr = NULL;
160     g_free(s->export);
161     s->export = NULL;
162     g_free(s->tlscredsid);
163     s->tlscredsid = NULL;
164     g_free(s->x_dirty_bitmap);
165     s->x_dirty_bitmap = NULL;
166 }
167 
168 static void nbd_channel_error(BDRVNBDState *s, int ret)
169 {
170     if (ret == -EIO) {
171         if (qatomic_load_acquire(&s->state) == NBD_CLIENT_CONNECTED) {
172             s->state = s->reconnect_delay ? NBD_CLIENT_CONNECTING_WAIT :
173                                             NBD_CLIENT_CONNECTING_NOWAIT;
174         }
175     } else {
176         if (qatomic_load_acquire(&s->state) == NBD_CLIENT_CONNECTED) {
177             qio_channel_shutdown(s->ioc, QIO_CHANNEL_SHUTDOWN_BOTH, NULL);
178         }
179         s->state = NBD_CLIENT_QUIT;
180     }
181 }
182 
183 static void nbd_recv_coroutines_wake_all(BDRVNBDState *s)
184 {
185     int i;
186 
187     for (i = 0; i < MAX_NBD_REQUESTS; i++) {
188         NBDClientRequest *req = &s->requests[i];
189 
190         if (req->coroutine && req->receiving) {
191             aio_co_wake(req->coroutine);
192         }
193     }
194 }
195 
196 static void reconnect_delay_timer_del(BDRVNBDState *s)
197 {
198     if (s->reconnect_delay_timer) {
199         timer_free(s->reconnect_delay_timer);
200         s->reconnect_delay_timer = NULL;
201     }
202 }
203 
204 static void reconnect_delay_timer_cb(void *opaque)
205 {
206     BDRVNBDState *s = opaque;
207 
208     if (qatomic_load_acquire(&s->state) == NBD_CLIENT_CONNECTING_WAIT) {
209         s->state = NBD_CLIENT_CONNECTING_NOWAIT;
210         while (qemu_co_enter_next(&s->free_sema, NULL)) {
211             /* Resume all queued requests */
212         }
213     }
214 
215     reconnect_delay_timer_del(s);
216 }
217 
218 static void reconnect_delay_timer_init(BDRVNBDState *s, uint64_t expire_time_ns)
219 {
220     if (qatomic_load_acquire(&s->state) != NBD_CLIENT_CONNECTING_WAIT) {
221         return;
222     }
223 
224     assert(!s->reconnect_delay_timer);
225     s->reconnect_delay_timer = aio_timer_new(bdrv_get_aio_context(s->bs),
226                                              QEMU_CLOCK_REALTIME,
227                                              SCALE_NS,
228                                              reconnect_delay_timer_cb, s);
229     timer_mod(s->reconnect_delay_timer, expire_time_ns);
230 }
231 
232 static void nbd_client_detach_aio_context(BlockDriverState *bs)
233 {
234     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
235 
236     /* Timer is deleted in nbd_client_co_drain_begin() */
237     assert(!s->reconnect_delay_timer);
238     /*
239      * If reconnect is in progress we may have no ->ioc.  It will be
240      * re-instantiated in the proper aio context once the connection is
241      * reestablished.
242      */
243     if (s->ioc) {
244         qio_channel_detach_aio_context(QIO_CHANNEL(s->ioc));
245     }
246 }
247 
248 static void nbd_client_attach_aio_context_bh(void *opaque)
249 {
250     BlockDriverState *bs = opaque;
251     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
252 
253     if (s->connection_co) {
254         /*
255          * The node is still drained, so we know the coroutine has yielded in
256          * nbd_read_eof(), the only place where bs->in_flight can reach 0, or
257          * it is entered for the first time. Both places are safe for entering
258          * the coroutine.
259          */
260         qemu_aio_coroutine_enter(bs->aio_context, s->connection_co);
261     }
262     bdrv_dec_in_flight(bs);
263 }
264 
265 static void nbd_client_attach_aio_context(BlockDriverState *bs,
266                                           AioContext *new_context)
267 {
268     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
269 
270     /*
271      * s->connection_co is either yielded from nbd_receive_reply or from
272      * nbd_co_reconnect_loop()
273      */
274     if (qatomic_load_acquire(&s->state) == NBD_CLIENT_CONNECTED) {
275         qio_channel_attach_aio_context(QIO_CHANNEL(s->ioc), new_context);
276     }
277 
278     bdrv_inc_in_flight(bs);
279 
280     /*
281      * Need to wait here for the BH to run because the BH must run while the
282      * node is still drained.
283      */
284     aio_wait_bh_oneshot(new_context, nbd_client_attach_aio_context_bh, bs);
285 }
286 
287 static void coroutine_fn nbd_client_co_drain_begin(BlockDriverState *bs)
288 {
289     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
290 
291     s->drained = true;
292     if (s->connection_co_sleep_ns_state) {
293         qemu_co_sleep_wake(s->connection_co_sleep_ns_state);
294     }
295 
296     nbd_co_establish_connection_cancel(bs, false);
297 
298     reconnect_delay_timer_del(s);
299 
300     if (qatomic_load_acquire(&s->state) == NBD_CLIENT_CONNECTING_WAIT) {
301         s->state = NBD_CLIENT_CONNECTING_NOWAIT;
302         qemu_co_queue_restart_all(&s->free_sema);
303     }
304 }
305 
306 static void coroutine_fn nbd_client_co_drain_end(BlockDriverState *bs)
307 {
308     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
309 
310     s->drained = false;
311     if (s->wait_drained_end) {
312         s->wait_drained_end = false;
313         aio_co_wake(s->connection_co);
314     }
315 }
316 
317 
318 static void nbd_teardown_connection(BlockDriverState *bs)
319 {
320     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
321 
322     if (s->ioc) {
323         /* finish any pending coroutines */
324         qio_channel_shutdown(s->ioc, QIO_CHANNEL_SHUTDOWN_BOTH, NULL);
325     } else if (s->sioc) {
326         /* abort negotiation */
327         qio_channel_shutdown(QIO_CHANNEL(s->sioc), QIO_CHANNEL_SHUTDOWN_BOTH,
328                              NULL);
329     }
330 
331     s->state = NBD_CLIENT_QUIT;
332     if (s->connection_co) {
333         if (s->connection_co_sleep_ns_state) {
334             qemu_co_sleep_wake(s->connection_co_sleep_ns_state);
335         }
336         nbd_co_establish_connection_cancel(bs, true);
337     }
338     if (qemu_in_coroutine()) {
339         s->teardown_co = qemu_coroutine_self();
340         /* connection_co resumes us when it terminates */
341         qemu_coroutine_yield();
342         s->teardown_co = NULL;
343     } else {
344         BDRV_POLL_WHILE(bs, s->connection_co);
345     }
346     assert(!s->connection_co);
347 }
348 
349 static bool nbd_client_connecting(BDRVNBDState *s)
350 {
351     NBDClientState state = qatomic_load_acquire(&s->state);
352     return state == NBD_CLIENT_CONNECTING_WAIT ||
353         state == NBD_CLIENT_CONNECTING_NOWAIT;
354 }
355 
356 static bool nbd_client_connecting_wait(BDRVNBDState *s)
357 {
358     return qatomic_load_acquire(&s->state) == NBD_CLIENT_CONNECTING_WAIT;
359 }
360 
361 static void connect_bh(void *opaque)
362 {
363     BDRVNBDState *state = opaque;
364 
365     assert(state->wait_connect);
366     state->wait_connect = false;
367     aio_co_wake(state->connection_co);
368 }
369 
370 static void nbd_init_connect_thread(BDRVNBDState *s)
371 {
372     s->connect_thread = g_new(NBDConnectThread, 1);
373 
374     *s->connect_thread = (NBDConnectThread) {
375         .saddr = QAPI_CLONE(SocketAddress, s->saddr),
376         .state = CONNECT_THREAD_NONE,
377         .bh_func = connect_bh,
378         .bh_opaque = s,
379     };
380 
381     qemu_mutex_init(&s->connect_thread->mutex);
382 }
383 
384 static void nbd_free_connect_thread(NBDConnectThread *thr)
385 {
386     if (thr->sioc) {
387         qio_channel_close(QIO_CHANNEL(thr->sioc), NULL);
388     }
389     error_free(thr->err);
390     qapi_free_SocketAddress(thr->saddr);
391     g_free(thr);
392 }
393 
394 static void *connect_thread_func(void *opaque)
395 {
396     NBDConnectThread *thr = opaque;
397     int ret;
398     bool do_free = false;
399 
400     thr->sioc = qio_channel_socket_new();
401 
402     error_free(thr->err);
403     thr->err = NULL;
404     ret = qio_channel_socket_connect_sync(thr->sioc, thr->saddr, &thr->err);
405     if (ret < 0) {
406         object_unref(OBJECT(thr->sioc));
407         thr->sioc = NULL;
408     }
409 
410     qemu_mutex_lock(&thr->mutex);
411 
412     switch (thr->state) {
413     case CONNECT_THREAD_RUNNING:
414         thr->state = ret < 0 ? CONNECT_THREAD_FAIL : CONNECT_THREAD_SUCCESS;
415         if (thr->bh_ctx) {
416             aio_bh_schedule_oneshot(thr->bh_ctx, thr->bh_func, thr->bh_opaque);
417 
418             /* play safe, don't reuse bh_ctx on further connection attempts */
419             thr->bh_ctx = NULL;
420         }
421         break;
422     case CONNECT_THREAD_RUNNING_DETACHED:
423         do_free = true;
424         break;
425     default:
426         abort();
427     }
428 
429     qemu_mutex_unlock(&thr->mutex);
430 
431     if (do_free) {
432         nbd_free_connect_thread(thr);
433     }
434 
435     return NULL;
436 }
437 
438 static int coroutine_fn
439 nbd_co_establish_connection(BlockDriverState *bs, Error **errp)
440 {
441     int ret;
442     QemuThread thread;
443     BDRVNBDState *s = bs->opaque;
444     NBDConnectThread *thr = s->connect_thread;
445 
446     if (!thr) {
447         /* detached */
448         return -1;
449     }
450 
451     qemu_mutex_lock(&thr->mutex);
452 
453     switch (thr->state) {
454     case CONNECT_THREAD_FAIL:
455     case CONNECT_THREAD_NONE:
456         error_free(thr->err);
457         thr->err = NULL;
458         thr->state = CONNECT_THREAD_RUNNING;
459         qemu_thread_create(&thread, "nbd-connect",
460                            connect_thread_func, thr, QEMU_THREAD_DETACHED);
461         break;
462     case CONNECT_THREAD_SUCCESS:
463         /* Previous attempt finally succeeded in background */
464         thr->state = CONNECT_THREAD_NONE;
465         s->sioc = thr->sioc;
466         thr->sioc = NULL;
467         yank_register_function(BLOCKDEV_YANK_INSTANCE(bs->node_name),
468                                nbd_yank, bs);
469         qemu_mutex_unlock(&thr->mutex);
470         return 0;
471     case CONNECT_THREAD_RUNNING:
472         /* Already running, will wait */
473         break;
474     default:
475         abort();
476     }
477 
478     thr->bh_ctx = qemu_get_current_aio_context();
479 
480     qemu_mutex_unlock(&thr->mutex);
481 
482 
483     /*
484      * We are going to wait for connect-thread finish, but
485      * nbd_client_co_drain_begin() can interrupt.
486      *
487      * Note that wait_connect variable is not visible for connect-thread. It
488      * doesn't need mutex protection, it used only inside home aio context of
489      * bs.
490      */
491     s->wait_connect = true;
492     qemu_coroutine_yield();
493 
494     if (!s->connect_thread) {
495         /* detached */
496         return -1;
497     }
498     assert(thr == s->connect_thread);
499 
500     qemu_mutex_lock(&thr->mutex);
501 
502     switch (thr->state) {
503     case CONNECT_THREAD_SUCCESS:
504     case CONNECT_THREAD_FAIL:
505         thr->state = CONNECT_THREAD_NONE;
506         error_propagate(errp, thr->err);
507         thr->err = NULL;
508         s->sioc = thr->sioc;
509         thr->sioc = NULL;
510         if (s->sioc) {
511             yank_register_function(BLOCKDEV_YANK_INSTANCE(bs->node_name),
512                                    nbd_yank, bs);
513         }
514         ret = (s->sioc ? 0 : -1);
515         break;
516     case CONNECT_THREAD_RUNNING:
517     case CONNECT_THREAD_RUNNING_DETACHED:
518         /*
519          * Obviously, drained section wants to start. Report the attempt as
520          * failed. Still connect thread is executing in background, and its
521          * result may be used for next connection attempt.
522          */
523         ret = -1;
524         error_setg(errp, "Connection attempt cancelled by other operation");
525         break;
526 
527     case CONNECT_THREAD_NONE:
528         /*
529          * Impossible. We've seen this thread running. So it should be
530          * running or at least give some results.
531          */
532         abort();
533 
534     default:
535         abort();
536     }
537 
538     qemu_mutex_unlock(&thr->mutex);
539 
540     return ret;
541 }
542 
543 /*
544  * nbd_co_establish_connection_cancel
545  * Cancel nbd_co_establish_connection asynchronously: it will finish soon, to
546  * allow drained section to begin.
547  *
548  * If detach is true, also cleanup the state (or if thread is running, move it
549  * to CONNECT_THREAD_RUNNING_DETACHED state). s->connect_thread becomes NULL if
550  * detach is true.
551  */
552 static void nbd_co_establish_connection_cancel(BlockDriverState *bs,
553                                                bool detach)
554 {
555     BDRVNBDState *s = bs->opaque;
556     NBDConnectThread *thr = s->connect_thread;
557     bool wake = false;
558     bool do_free = false;
559 
560     qemu_mutex_lock(&thr->mutex);
561 
562     if (thr->state == CONNECT_THREAD_RUNNING) {
563         /* We can cancel only in running state, when bh is not yet scheduled */
564         thr->bh_ctx = NULL;
565         if (s->wait_connect) {
566             s->wait_connect = false;
567             wake = true;
568         }
569         if (detach) {
570             thr->state = CONNECT_THREAD_RUNNING_DETACHED;
571             s->connect_thread = NULL;
572         }
573     } else if (detach) {
574         do_free = true;
575     }
576 
577     qemu_mutex_unlock(&thr->mutex);
578 
579     if (do_free) {
580         nbd_free_connect_thread(thr);
581         s->connect_thread = NULL;
582     }
583 
584     if (wake) {
585         aio_co_wake(s->connection_co);
586     }
587 }
588 
589 static coroutine_fn void nbd_reconnect_attempt(BDRVNBDState *s)
590 {
591     int ret;
592     Error *local_err = NULL;
593 
594     if (!nbd_client_connecting(s)) {
595         return;
596     }
597 
598     /* Wait for completion of all in-flight requests */
599 
600     qemu_co_mutex_lock(&s->send_mutex);
601 
602     while (s->in_flight > 0) {
603         qemu_co_mutex_unlock(&s->send_mutex);
604         nbd_recv_coroutines_wake_all(s);
605         s->wait_in_flight = true;
606         qemu_coroutine_yield();
607         s->wait_in_flight = false;
608         qemu_co_mutex_lock(&s->send_mutex);
609     }
610 
611     qemu_co_mutex_unlock(&s->send_mutex);
612 
613     if (!nbd_client_connecting(s)) {
614         return;
615     }
616 
617     /*
618      * Now we are sure that nobody is accessing the channel, and no one will
619      * try until we set the state to CONNECTED.
620      */
621 
622     /* Finalize previous connection if any */
623     if (s->ioc) {
624         qio_channel_detach_aio_context(QIO_CHANNEL(s->ioc));
625         yank_unregister_function(BLOCKDEV_YANK_INSTANCE(s->bs->node_name),
626                                  nbd_yank, s->bs);
627         object_unref(OBJECT(s->sioc));
628         s->sioc = NULL;
629         object_unref(OBJECT(s->ioc));
630         s->ioc = NULL;
631     }
632 
633     if (nbd_co_establish_connection(s->bs, &local_err) < 0) {
634         ret = -ECONNREFUSED;
635         goto out;
636     }
637 
638     bdrv_dec_in_flight(s->bs);
639 
640     ret = nbd_client_handshake(s->bs, &local_err);
641 
642     if (s->drained) {
643         s->wait_drained_end = true;
644         while (s->drained) {
645             /*
646              * We may be entered once from nbd_client_attach_aio_context_bh
647              * and then from nbd_client_co_drain_end. So here is a loop.
648              */
649             qemu_coroutine_yield();
650         }
651     }
652     bdrv_inc_in_flight(s->bs);
653 
654 out:
655     s->connect_status = ret;
656     error_free(s->connect_err);
657     s->connect_err = NULL;
658     error_propagate(&s->connect_err, local_err);
659 
660     if (ret >= 0) {
661         /* successfully connected */
662         s->state = NBD_CLIENT_CONNECTED;
663         qemu_co_queue_restart_all(&s->free_sema);
664     }
665 }
666 
667 static coroutine_fn void nbd_co_reconnect_loop(BDRVNBDState *s)
668 {
669     uint64_t timeout = 1 * NANOSECONDS_PER_SECOND;
670     uint64_t max_timeout = 16 * NANOSECONDS_PER_SECOND;
671 
672     if (qatomic_load_acquire(&s->state) == NBD_CLIENT_CONNECTING_WAIT) {
673         reconnect_delay_timer_init(s, qemu_clock_get_ns(QEMU_CLOCK_REALTIME) +
674                                    s->reconnect_delay * NANOSECONDS_PER_SECOND);
675     }
676 
677     nbd_reconnect_attempt(s);
678 
679     while (nbd_client_connecting(s)) {
680         if (s->drained) {
681             bdrv_dec_in_flight(s->bs);
682             s->wait_drained_end = true;
683             while (s->drained) {
684                 /*
685                  * We may be entered once from nbd_client_attach_aio_context_bh
686                  * and then from nbd_client_co_drain_end. So here is a loop.
687                  */
688                 qemu_coroutine_yield();
689             }
690             bdrv_inc_in_flight(s->bs);
691         } else {
692             qemu_co_sleep_ns_wakeable(QEMU_CLOCK_REALTIME, timeout,
693                                       &s->connection_co_sleep_ns_state);
694             if (s->drained) {
695                 continue;
696             }
697             if (timeout < max_timeout) {
698                 timeout *= 2;
699             }
700         }
701 
702         nbd_reconnect_attempt(s);
703     }
704 
705     reconnect_delay_timer_del(s);
706 }
707 
708 static coroutine_fn void nbd_connection_entry(void *opaque)
709 {
710     BDRVNBDState *s = opaque;
711     uint64_t i;
712     int ret = 0;
713     Error *local_err = NULL;
714 
715     while (qatomic_load_acquire(&s->state) != NBD_CLIENT_QUIT) {
716         /*
717          * The NBD client can only really be considered idle when it has
718          * yielded from qio_channel_readv_all_eof(), waiting for data. This is
719          * the point where the additional scheduled coroutine entry happens
720          * after nbd_client_attach_aio_context().
721          *
722          * Therefore we keep an additional in_flight reference all the time and
723          * only drop it temporarily here.
724          */
725 
726         if (nbd_client_connecting(s)) {
727             nbd_co_reconnect_loop(s);
728         }
729 
730         if (qatomic_load_acquire(&s->state) != NBD_CLIENT_CONNECTED) {
731             continue;
732         }
733 
734         assert(s->reply.handle == 0);
735         ret = nbd_receive_reply(s->bs, s->ioc, &s->reply, &local_err);
736 
737         if (local_err) {
738             trace_nbd_read_reply_entry_fail(ret, error_get_pretty(local_err));
739             error_free(local_err);
740             local_err = NULL;
741         }
742         if (ret <= 0) {
743             nbd_channel_error(s, ret ? ret : -EIO);
744             continue;
745         }
746 
747         /*
748          * There's no need for a mutex on the receive side, because the
749          * handler acts as a synchronization point and ensures that only
750          * one coroutine is called until the reply finishes.
751          */
752         i = HANDLE_TO_INDEX(s, s->reply.handle);
753         if (i >= MAX_NBD_REQUESTS ||
754             !s->requests[i].coroutine ||
755             !s->requests[i].receiving ||
756             (nbd_reply_is_structured(&s->reply) && !s->info.structured_reply))
757         {
758             nbd_channel_error(s, -EINVAL);
759             continue;
760         }
761 
762         /*
763          * We're woken up again by the request itself.  Note that there
764          * is no race between yielding and reentering connection_co.  This
765          * is because:
766          *
767          * - if the request runs on the same AioContext, it is only
768          *   entered after we yield
769          *
770          * - if the request runs on a different AioContext, reentering
771          *   connection_co happens through a bottom half, which can only
772          *   run after we yield.
773          */
774         aio_co_wake(s->requests[i].coroutine);
775         qemu_coroutine_yield();
776     }
777 
778     qemu_co_queue_restart_all(&s->free_sema);
779     nbd_recv_coroutines_wake_all(s);
780     bdrv_dec_in_flight(s->bs);
781 
782     s->connection_co = NULL;
783     if (s->ioc) {
784         qio_channel_detach_aio_context(QIO_CHANNEL(s->ioc));
785         yank_unregister_function(BLOCKDEV_YANK_INSTANCE(s->bs->node_name),
786                                  nbd_yank, s->bs);
787         object_unref(OBJECT(s->sioc));
788         s->sioc = NULL;
789         object_unref(OBJECT(s->ioc));
790         s->ioc = NULL;
791     }
792 
793     if (s->teardown_co) {
794         aio_co_wake(s->teardown_co);
795     }
796     aio_wait_kick();
797 }
798 
799 static int nbd_co_send_request(BlockDriverState *bs,
800                                NBDRequest *request,
801                                QEMUIOVector *qiov)
802 {
803     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
804     int rc, i = -1;
805 
806     qemu_co_mutex_lock(&s->send_mutex);
807     while (s->in_flight == MAX_NBD_REQUESTS || nbd_client_connecting_wait(s)) {
808         qemu_co_queue_wait(&s->free_sema, &s->send_mutex);
809     }
810 
811     if (qatomic_load_acquire(&s->state) != NBD_CLIENT_CONNECTED) {
812         rc = -EIO;
813         goto err;
814     }
815 
816     s->in_flight++;
817 
818     for (i = 0; i < MAX_NBD_REQUESTS; i++) {
819         if (s->requests[i].coroutine == NULL) {
820             break;
821         }
822     }
823 
824     g_assert(qemu_in_coroutine());
825     assert(i < MAX_NBD_REQUESTS);
826 
827     s->requests[i].coroutine = qemu_coroutine_self();
828     s->requests[i].offset = request->from;
829     s->requests[i].receiving = false;
830 
831     request->handle = INDEX_TO_HANDLE(s, i);
832 
833     assert(s->ioc);
834 
835     if (qiov) {
836         qio_channel_set_cork(s->ioc, true);
837         rc = nbd_send_request(s->ioc, request);
838         if (qatomic_load_acquire(&s->state) == NBD_CLIENT_CONNECTED &&
839             rc >= 0) {
840             if (qio_channel_writev_all(s->ioc, qiov->iov, qiov->niov,
841                                        NULL) < 0) {
842                 rc = -EIO;
843             }
844         } else if (rc >= 0) {
845             rc = -EIO;
846         }
847         qio_channel_set_cork(s->ioc, false);
848     } else {
849         rc = nbd_send_request(s->ioc, request);
850     }
851 
852 err:
853     if (rc < 0) {
854         nbd_channel_error(s, rc);
855         if (i != -1) {
856             s->requests[i].coroutine = NULL;
857             s->in_flight--;
858         }
859         if (s->in_flight == 0 && s->wait_in_flight) {
860             aio_co_wake(s->connection_co);
861         } else {
862             qemu_co_queue_next(&s->free_sema);
863         }
864     }
865     qemu_co_mutex_unlock(&s->send_mutex);
866     return rc;
867 }
868 
869 static inline uint16_t payload_advance16(uint8_t **payload)
870 {
871     *payload += 2;
872     return lduw_be_p(*payload - 2);
873 }
874 
875 static inline uint32_t payload_advance32(uint8_t **payload)
876 {
877     *payload += 4;
878     return ldl_be_p(*payload - 4);
879 }
880 
881 static inline uint64_t payload_advance64(uint8_t **payload)
882 {
883     *payload += 8;
884     return ldq_be_p(*payload - 8);
885 }
886 
887 static int nbd_parse_offset_hole_payload(BDRVNBDState *s,
888                                          NBDStructuredReplyChunk *chunk,
889                                          uint8_t *payload, uint64_t orig_offset,
890                                          QEMUIOVector *qiov, Error **errp)
891 {
892     uint64_t offset;
893     uint32_t hole_size;
894 
895     if (chunk->length != sizeof(offset) + sizeof(hole_size)) {
896         error_setg(errp, "Protocol error: invalid payload for "
897                          "NBD_REPLY_TYPE_OFFSET_HOLE");
898         return -EINVAL;
899     }
900 
901     offset = payload_advance64(&payload);
902     hole_size = payload_advance32(&payload);
903 
904     if (!hole_size || offset < orig_offset || hole_size > qiov->size ||
905         offset > orig_offset + qiov->size - hole_size) {
906         error_setg(errp, "Protocol error: server sent chunk exceeding requested"
907                          " region");
908         return -EINVAL;
909     }
910     if (s->info.min_block &&
911         !QEMU_IS_ALIGNED(hole_size, s->info.min_block)) {
912         trace_nbd_structured_read_compliance("hole");
913     }
914 
915     qemu_iovec_memset(qiov, offset - orig_offset, 0, hole_size);
916 
917     return 0;
918 }
919 
920 /*
921  * nbd_parse_blockstatus_payload
922  * Based on our request, we expect only one extent in reply, for the
923  * base:allocation context.
924  */
925 static int nbd_parse_blockstatus_payload(BDRVNBDState *s,
926                                          NBDStructuredReplyChunk *chunk,
927                                          uint8_t *payload, uint64_t orig_length,
928                                          NBDExtent *extent, Error **errp)
929 {
930     uint32_t context_id;
931 
932     /* The server succeeded, so it must have sent [at least] one extent */
933     if (chunk->length < sizeof(context_id) + sizeof(*extent)) {
934         error_setg(errp, "Protocol error: invalid payload for "
935                          "NBD_REPLY_TYPE_BLOCK_STATUS");
936         return -EINVAL;
937     }
938 
939     context_id = payload_advance32(&payload);
940     if (s->info.context_id != context_id) {
941         error_setg(errp, "Protocol error: unexpected context id %d for "
942                          "NBD_REPLY_TYPE_BLOCK_STATUS, when negotiated context "
943                          "id is %d", context_id,
944                          s->info.context_id);
945         return -EINVAL;
946     }
947 
948     extent->length = payload_advance32(&payload);
949     extent->flags = payload_advance32(&payload);
950 
951     if (extent->length == 0) {
952         error_setg(errp, "Protocol error: server sent status chunk with "
953                    "zero length");
954         return -EINVAL;
955     }
956 
957     /*
958      * A server sending unaligned block status is in violation of the
959      * protocol, but as qemu-nbd 3.1 is such a server (at least for
960      * POSIX files that are not a multiple of 512 bytes, since qemu
961      * rounds files up to 512-byte multiples but lseek(SEEK_HOLE)
962      * still sees an implicit hole beyond the real EOF), it's nicer to
963      * work around the misbehaving server. If the request included
964      * more than the final unaligned block, truncate it back to an
965      * aligned result; if the request was only the final block, round
966      * up to the full block and change the status to fully-allocated
967      * (always a safe status, even if it loses information).
968      */
969     if (s->info.min_block && !QEMU_IS_ALIGNED(extent->length,
970                                                    s->info.min_block)) {
971         trace_nbd_parse_blockstatus_compliance("extent length is unaligned");
972         if (extent->length > s->info.min_block) {
973             extent->length = QEMU_ALIGN_DOWN(extent->length,
974                                              s->info.min_block);
975         } else {
976             extent->length = s->info.min_block;
977             extent->flags = 0;
978         }
979     }
980 
981     /*
982      * We used NBD_CMD_FLAG_REQ_ONE, so the server should not have
983      * sent us any more than one extent, nor should it have included
984      * status beyond our request in that extent. However, it's easy
985      * enough to ignore the server's noncompliance without killing the
986      * connection; just ignore trailing extents, and clamp things to
987      * the length of our request.
988      */
989     if (chunk->length > sizeof(context_id) + sizeof(*extent)) {
990         trace_nbd_parse_blockstatus_compliance("more than one extent");
991     }
992     if (extent->length > orig_length) {
993         extent->length = orig_length;
994         trace_nbd_parse_blockstatus_compliance("extent length too large");
995     }
996 
997     /*
998      * HACK: if we are using x-dirty-bitmaps to access
999      * qemu:allocation-depth, treat all depths > 2 the same as 2,
1000      * since nbd_client_co_block_status is only expecting the low two
1001      * bits to be set.
1002      */
1003     if (s->alloc_depth && extent->flags > 2) {
1004         extent->flags = 2;
1005     }
1006 
1007     return 0;
1008 }
1009 
1010 /*
1011  * nbd_parse_error_payload
1012  * on success @errp contains message describing nbd error reply
1013  */
1014 static int nbd_parse_error_payload(NBDStructuredReplyChunk *chunk,
1015                                    uint8_t *payload, int *request_ret,
1016                                    Error **errp)
1017 {
1018     uint32_t error;
1019     uint16_t message_size;
1020 
1021     assert(chunk->type & (1 << 15));
1022 
1023     if (chunk->length < sizeof(error) + sizeof(message_size)) {
1024         error_setg(errp,
1025                    "Protocol error: invalid payload for structured error");
1026         return -EINVAL;
1027     }
1028 
1029     error = nbd_errno_to_system_errno(payload_advance32(&payload));
1030     if (error == 0) {
1031         error_setg(errp, "Protocol error: server sent structured error chunk "
1032                          "with error = 0");
1033         return -EINVAL;
1034     }
1035 
1036     *request_ret = -error;
1037     message_size = payload_advance16(&payload);
1038 
1039     if (message_size > chunk->length - sizeof(error) - sizeof(message_size)) {
1040         error_setg(errp, "Protocol error: server sent structured error chunk "
1041                          "with incorrect message size");
1042         return -EINVAL;
1043     }
1044 
1045     /* TODO: Add a trace point to mention the server complaint */
1046 
1047     /* TODO handle ERROR_OFFSET */
1048 
1049     return 0;
1050 }
1051 
1052 static int nbd_co_receive_offset_data_payload(BDRVNBDState *s,
1053                                               uint64_t orig_offset,
1054                                               QEMUIOVector *qiov, Error **errp)
1055 {
1056     QEMUIOVector sub_qiov;
1057     uint64_t offset;
1058     size_t data_size;
1059     int ret;
1060     NBDStructuredReplyChunk *chunk = &s->reply.structured;
1061 
1062     assert(nbd_reply_is_structured(&s->reply));
1063 
1064     /* The NBD spec requires at least one byte of payload */
1065     if (chunk->length <= sizeof(offset)) {
1066         error_setg(errp, "Protocol error: invalid payload for "
1067                          "NBD_REPLY_TYPE_OFFSET_DATA");
1068         return -EINVAL;
1069     }
1070 
1071     if (nbd_read64(s->ioc, &offset, "OFFSET_DATA offset", errp) < 0) {
1072         return -EIO;
1073     }
1074 
1075     data_size = chunk->length - sizeof(offset);
1076     assert(data_size);
1077     if (offset < orig_offset || data_size > qiov->size ||
1078         offset > orig_offset + qiov->size - data_size) {
1079         error_setg(errp, "Protocol error: server sent chunk exceeding requested"
1080                          " region");
1081         return -EINVAL;
1082     }
1083     if (s->info.min_block && !QEMU_IS_ALIGNED(data_size, s->info.min_block)) {
1084         trace_nbd_structured_read_compliance("data");
1085     }
1086 
1087     qemu_iovec_init(&sub_qiov, qiov->niov);
1088     qemu_iovec_concat(&sub_qiov, qiov, offset - orig_offset, data_size);
1089     ret = qio_channel_readv_all(s->ioc, sub_qiov.iov, sub_qiov.niov, errp);
1090     qemu_iovec_destroy(&sub_qiov);
1091 
1092     return ret < 0 ? -EIO : 0;
1093 }
1094 
1095 #define NBD_MAX_MALLOC_PAYLOAD 1000
1096 static coroutine_fn int nbd_co_receive_structured_payload(
1097         BDRVNBDState *s, void **payload, Error **errp)
1098 {
1099     int ret;
1100     uint32_t len;
1101 
1102     assert(nbd_reply_is_structured(&s->reply));
1103 
1104     len = s->reply.structured.length;
1105 
1106     if (len == 0) {
1107         return 0;
1108     }
1109 
1110     if (payload == NULL) {
1111         error_setg(errp, "Unexpected structured payload");
1112         return -EINVAL;
1113     }
1114 
1115     if (len > NBD_MAX_MALLOC_PAYLOAD) {
1116         error_setg(errp, "Payload too large");
1117         return -EINVAL;
1118     }
1119 
1120     *payload = g_new(char, len);
1121     ret = nbd_read(s->ioc, *payload, len, "structured payload", errp);
1122     if (ret < 0) {
1123         g_free(*payload);
1124         *payload = NULL;
1125         return ret;
1126     }
1127 
1128     return 0;
1129 }
1130 
1131 /*
1132  * nbd_co_do_receive_one_chunk
1133  * for simple reply:
1134  *   set request_ret to received reply error
1135  *   if qiov is not NULL: read payload to @qiov
1136  * for structured reply chunk:
1137  *   if error chunk: read payload, set @request_ret, do not set @payload
1138  *   else if offset_data chunk: read payload data to @qiov, do not set @payload
1139  *   else: read payload to @payload
1140  *
1141  * If function fails, @errp contains corresponding error message, and the
1142  * connection with the server is suspect.  If it returns 0, then the
1143  * transaction succeeded (although @request_ret may be a negative errno
1144  * corresponding to the server's error reply), and errp is unchanged.
1145  */
1146 static coroutine_fn int nbd_co_do_receive_one_chunk(
1147         BDRVNBDState *s, uint64_t handle, bool only_structured,
1148         int *request_ret, QEMUIOVector *qiov, void **payload, Error **errp)
1149 {
1150     int ret;
1151     int i = HANDLE_TO_INDEX(s, handle);
1152     void *local_payload = NULL;
1153     NBDStructuredReplyChunk *chunk;
1154 
1155     if (payload) {
1156         *payload = NULL;
1157     }
1158     *request_ret = 0;
1159 
1160     /* Wait until we're woken up by nbd_connection_entry.  */
1161     s->requests[i].receiving = true;
1162     qemu_coroutine_yield();
1163     s->requests[i].receiving = false;
1164     if (qatomic_load_acquire(&s->state) != NBD_CLIENT_CONNECTED) {
1165         error_setg(errp, "Connection closed");
1166         return -EIO;
1167     }
1168     assert(s->ioc);
1169 
1170     assert(s->reply.handle == handle);
1171 
1172     if (nbd_reply_is_simple(&s->reply)) {
1173         if (only_structured) {
1174             error_setg(errp, "Protocol error: simple reply when structured "
1175                              "reply chunk was expected");
1176             return -EINVAL;
1177         }
1178 
1179         *request_ret = -nbd_errno_to_system_errno(s->reply.simple.error);
1180         if (*request_ret < 0 || !qiov) {
1181             return 0;
1182         }
1183 
1184         return qio_channel_readv_all(s->ioc, qiov->iov, qiov->niov,
1185                                      errp) < 0 ? -EIO : 0;
1186     }
1187 
1188     /* handle structured reply chunk */
1189     assert(s->info.structured_reply);
1190     chunk = &s->reply.structured;
1191 
1192     if (chunk->type == NBD_REPLY_TYPE_NONE) {
1193         if (!(chunk->flags & NBD_REPLY_FLAG_DONE)) {
1194             error_setg(errp, "Protocol error: NBD_REPLY_TYPE_NONE chunk without"
1195                        " NBD_REPLY_FLAG_DONE flag set");
1196             return -EINVAL;
1197         }
1198         if (chunk->length) {
1199             error_setg(errp, "Protocol error: NBD_REPLY_TYPE_NONE chunk with"
1200                        " nonzero length");
1201             return -EINVAL;
1202         }
1203         return 0;
1204     }
1205 
1206     if (chunk->type == NBD_REPLY_TYPE_OFFSET_DATA) {
1207         if (!qiov) {
1208             error_setg(errp, "Unexpected NBD_REPLY_TYPE_OFFSET_DATA chunk");
1209             return -EINVAL;
1210         }
1211 
1212         return nbd_co_receive_offset_data_payload(s, s->requests[i].offset,
1213                                                   qiov, errp);
1214     }
1215 
1216     if (nbd_reply_type_is_error(chunk->type)) {
1217         payload = &local_payload;
1218     }
1219 
1220     ret = nbd_co_receive_structured_payload(s, payload, errp);
1221     if (ret < 0) {
1222         return ret;
1223     }
1224 
1225     if (nbd_reply_type_is_error(chunk->type)) {
1226         ret = nbd_parse_error_payload(chunk, local_payload, request_ret, errp);
1227         g_free(local_payload);
1228         return ret;
1229     }
1230 
1231     return 0;
1232 }
1233 
1234 /*
1235  * nbd_co_receive_one_chunk
1236  * Read reply, wake up connection_co and set s->quit if needed.
1237  * Return value is a fatal error code or normal nbd reply error code
1238  */
1239 static coroutine_fn int nbd_co_receive_one_chunk(
1240         BDRVNBDState *s, uint64_t handle, bool only_structured,
1241         int *request_ret, QEMUIOVector *qiov, NBDReply *reply, void **payload,
1242         Error **errp)
1243 {
1244     int ret = nbd_co_do_receive_one_chunk(s, handle, only_structured,
1245                                           request_ret, qiov, payload, errp);
1246 
1247     if (ret < 0) {
1248         memset(reply, 0, sizeof(*reply));
1249         nbd_channel_error(s, ret);
1250     } else {
1251         /* For assert at loop start in nbd_connection_entry */
1252         *reply = s->reply;
1253     }
1254     s->reply.handle = 0;
1255 
1256     if (s->connection_co && !s->wait_in_flight) {
1257         /*
1258          * We must check s->wait_in_flight, because we may entered by
1259          * nbd_recv_coroutines_wake_all(), in this case we should not
1260          * wake connection_co here, it will woken by last request.
1261          */
1262         aio_co_wake(s->connection_co);
1263     }
1264 
1265     return ret;
1266 }
1267 
1268 typedef struct NBDReplyChunkIter {
1269     int ret;
1270     int request_ret;
1271     Error *err;
1272     bool done, only_structured;
1273 } NBDReplyChunkIter;
1274 
1275 static void nbd_iter_channel_error(NBDReplyChunkIter *iter,
1276                                    int ret, Error **local_err)
1277 {
1278     assert(local_err && *local_err);
1279     assert(ret < 0);
1280 
1281     if (!iter->ret) {
1282         iter->ret = ret;
1283         error_propagate(&iter->err, *local_err);
1284     } else {
1285         error_free(*local_err);
1286     }
1287 
1288     *local_err = NULL;
1289 }
1290 
1291 static void nbd_iter_request_error(NBDReplyChunkIter *iter, int ret)
1292 {
1293     assert(ret < 0);
1294 
1295     if (!iter->request_ret) {
1296         iter->request_ret = ret;
1297     }
1298 }
1299 
1300 /*
1301  * NBD_FOREACH_REPLY_CHUNK
1302  * The pointer stored in @payload requires g_free() to free it.
1303  */
1304 #define NBD_FOREACH_REPLY_CHUNK(s, iter, handle, structured, \
1305                                 qiov, reply, payload) \
1306     for (iter = (NBDReplyChunkIter) { .only_structured = structured }; \
1307          nbd_reply_chunk_iter_receive(s, &iter, handle, qiov, reply, payload);)
1308 
1309 /*
1310  * nbd_reply_chunk_iter_receive
1311  * The pointer stored in @payload requires g_free() to free it.
1312  */
1313 static bool nbd_reply_chunk_iter_receive(BDRVNBDState *s,
1314                                          NBDReplyChunkIter *iter,
1315                                          uint64_t handle,
1316                                          QEMUIOVector *qiov, NBDReply *reply,
1317                                          void **payload)
1318 {
1319     int ret, request_ret;
1320     NBDReply local_reply;
1321     NBDStructuredReplyChunk *chunk;
1322     Error *local_err = NULL;
1323     if (qatomic_load_acquire(&s->state) != NBD_CLIENT_CONNECTED) {
1324         error_setg(&local_err, "Connection closed");
1325         nbd_iter_channel_error(iter, -EIO, &local_err);
1326         goto break_loop;
1327     }
1328 
1329     if (iter->done) {
1330         /* Previous iteration was last. */
1331         goto break_loop;
1332     }
1333 
1334     if (reply == NULL) {
1335         reply = &local_reply;
1336     }
1337 
1338     ret = nbd_co_receive_one_chunk(s, handle, iter->only_structured,
1339                                    &request_ret, qiov, reply, payload,
1340                                    &local_err);
1341     if (ret < 0) {
1342         nbd_iter_channel_error(iter, ret, &local_err);
1343     } else if (request_ret < 0) {
1344         nbd_iter_request_error(iter, request_ret);
1345     }
1346 
1347     /* Do not execute the body of NBD_FOREACH_REPLY_CHUNK for simple reply. */
1348     if (nbd_reply_is_simple(reply) ||
1349         qatomic_load_acquire(&s->state) != NBD_CLIENT_CONNECTED) {
1350         goto break_loop;
1351     }
1352 
1353     chunk = &reply->structured;
1354     iter->only_structured = true;
1355 
1356     if (chunk->type == NBD_REPLY_TYPE_NONE) {
1357         /* NBD_REPLY_FLAG_DONE is already checked in nbd_co_receive_one_chunk */
1358         assert(chunk->flags & NBD_REPLY_FLAG_DONE);
1359         goto break_loop;
1360     }
1361 
1362     if (chunk->flags & NBD_REPLY_FLAG_DONE) {
1363         /* This iteration is last. */
1364         iter->done = true;
1365     }
1366 
1367     /* Execute the loop body */
1368     return true;
1369 
1370 break_loop:
1371     s->requests[HANDLE_TO_INDEX(s, handle)].coroutine = NULL;
1372 
1373     qemu_co_mutex_lock(&s->send_mutex);
1374     s->in_flight--;
1375     if (s->in_flight == 0 && s->wait_in_flight) {
1376         aio_co_wake(s->connection_co);
1377     } else {
1378         qemu_co_queue_next(&s->free_sema);
1379     }
1380     qemu_co_mutex_unlock(&s->send_mutex);
1381 
1382     return false;
1383 }
1384 
1385 static int nbd_co_receive_return_code(BDRVNBDState *s, uint64_t handle,
1386                                       int *request_ret, Error **errp)
1387 {
1388     NBDReplyChunkIter iter;
1389 
1390     NBD_FOREACH_REPLY_CHUNK(s, iter, handle, false, NULL, NULL, NULL) {
1391         /* nbd_reply_chunk_iter_receive does all the work */
1392     }
1393 
1394     error_propagate(errp, iter.err);
1395     *request_ret = iter.request_ret;
1396     return iter.ret;
1397 }
1398 
1399 static int nbd_co_receive_cmdread_reply(BDRVNBDState *s, uint64_t handle,
1400                                         uint64_t offset, QEMUIOVector *qiov,
1401                                         int *request_ret, Error **errp)
1402 {
1403     NBDReplyChunkIter iter;
1404     NBDReply reply;
1405     void *payload = NULL;
1406     Error *local_err = NULL;
1407 
1408     NBD_FOREACH_REPLY_CHUNK(s, iter, handle, s->info.structured_reply,
1409                             qiov, &reply, &payload)
1410     {
1411         int ret;
1412         NBDStructuredReplyChunk *chunk = &reply.structured;
1413 
1414         assert(nbd_reply_is_structured(&reply));
1415 
1416         switch (chunk->type) {
1417         case NBD_REPLY_TYPE_OFFSET_DATA:
1418             /*
1419              * special cased in nbd_co_receive_one_chunk, data is already
1420              * in qiov
1421              */
1422             break;
1423         case NBD_REPLY_TYPE_OFFSET_HOLE:
1424             ret = nbd_parse_offset_hole_payload(s, &reply.structured, payload,
1425                                                 offset, qiov, &local_err);
1426             if (ret < 0) {
1427                 nbd_channel_error(s, ret);
1428                 nbd_iter_channel_error(&iter, ret, &local_err);
1429             }
1430             break;
1431         default:
1432             if (!nbd_reply_type_is_error(chunk->type)) {
1433                 /* not allowed reply type */
1434                 nbd_channel_error(s, -EINVAL);
1435                 error_setg(&local_err,
1436                            "Unexpected reply type: %d (%s) for CMD_READ",
1437                            chunk->type, nbd_reply_type_lookup(chunk->type));
1438                 nbd_iter_channel_error(&iter, -EINVAL, &local_err);
1439             }
1440         }
1441 
1442         g_free(payload);
1443         payload = NULL;
1444     }
1445 
1446     error_propagate(errp, iter.err);
1447     *request_ret = iter.request_ret;
1448     return iter.ret;
1449 }
1450 
1451 static int nbd_co_receive_blockstatus_reply(BDRVNBDState *s,
1452                                             uint64_t handle, uint64_t length,
1453                                             NBDExtent *extent,
1454                                             int *request_ret, Error **errp)
1455 {
1456     NBDReplyChunkIter iter;
1457     NBDReply reply;
1458     void *payload = NULL;
1459     Error *local_err = NULL;
1460     bool received = false;
1461 
1462     assert(!extent->length);
1463     NBD_FOREACH_REPLY_CHUNK(s, iter, handle, false, NULL, &reply, &payload) {
1464         int ret;
1465         NBDStructuredReplyChunk *chunk = &reply.structured;
1466 
1467         assert(nbd_reply_is_structured(&reply));
1468 
1469         switch (chunk->type) {
1470         case NBD_REPLY_TYPE_BLOCK_STATUS:
1471             if (received) {
1472                 nbd_channel_error(s, -EINVAL);
1473                 error_setg(&local_err, "Several BLOCK_STATUS chunks in reply");
1474                 nbd_iter_channel_error(&iter, -EINVAL, &local_err);
1475             }
1476             received = true;
1477 
1478             ret = nbd_parse_blockstatus_payload(s, &reply.structured,
1479                                                 payload, length, extent,
1480                                                 &local_err);
1481             if (ret < 0) {
1482                 nbd_channel_error(s, ret);
1483                 nbd_iter_channel_error(&iter, ret, &local_err);
1484             }
1485             break;
1486         default:
1487             if (!nbd_reply_type_is_error(chunk->type)) {
1488                 nbd_channel_error(s, -EINVAL);
1489                 error_setg(&local_err,
1490                            "Unexpected reply type: %d (%s) "
1491                            "for CMD_BLOCK_STATUS",
1492                            chunk->type, nbd_reply_type_lookup(chunk->type));
1493                 nbd_iter_channel_error(&iter, -EINVAL, &local_err);
1494             }
1495         }
1496 
1497         g_free(payload);
1498         payload = NULL;
1499     }
1500 
1501     if (!extent->length && !iter.request_ret) {
1502         error_setg(&local_err, "Server did not reply with any status extents");
1503         nbd_iter_channel_error(&iter, -EIO, &local_err);
1504     }
1505 
1506     error_propagate(errp, iter.err);
1507     *request_ret = iter.request_ret;
1508     return iter.ret;
1509 }
1510 
1511 static int nbd_co_request(BlockDriverState *bs, NBDRequest *request,
1512                           QEMUIOVector *write_qiov)
1513 {
1514     int ret, request_ret;
1515     Error *local_err = NULL;
1516     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
1517 
1518     assert(request->type != NBD_CMD_READ);
1519     if (write_qiov) {
1520         assert(request->type == NBD_CMD_WRITE);
1521         assert(request->len == iov_size(write_qiov->iov, write_qiov->niov));
1522     } else {
1523         assert(request->type != NBD_CMD_WRITE);
1524     }
1525 
1526     do {
1527         ret = nbd_co_send_request(bs, request, write_qiov);
1528         if (ret < 0) {
1529             continue;
1530         }
1531 
1532         ret = nbd_co_receive_return_code(s, request->handle,
1533                                          &request_ret, &local_err);
1534         if (local_err) {
1535             trace_nbd_co_request_fail(request->from, request->len,
1536                                       request->handle, request->flags,
1537                                       request->type,
1538                                       nbd_cmd_lookup(request->type),
1539                                       ret, error_get_pretty(local_err));
1540             error_free(local_err);
1541             local_err = NULL;
1542         }
1543     } while (ret < 0 && nbd_client_connecting_wait(s));
1544 
1545     return ret ? ret : request_ret;
1546 }
1547 
1548 static int nbd_client_co_preadv(BlockDriverState *bs, uint64_t offset,
1549                                 uint64_t bytes, QEMUIOVector *qiov, int flags)
1550 {
1551     int ret, request_ret;
1552     Error *local_err = NULL;
1553     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
1554     NBDRequest request = {
1555         .type = NBD_CMD_READ,
1556         .from = offset,
1557         .len = bytes,
1558     };
1559 
1560     assert(bytes <= NBD_MAX_BUFFER_SIZE);
1561     assert(!flags);
1562 
1563     if (!bytes) {
1564         return 0;
1565     }
1566     /*
1567      * Work around the fact that the block layer doesn't do
1568      * byte-accurate sizing yet - if the read exceeds the server's
1569      * advertised size because the block layer rounded size up, then
1570      * truncate the request to the server and tail-pad with zero.
1571      */
1572     if (offset >= s->info.size) {
1573         assert(bytes < BDRV_SECTOR_SIZE);
1574         qemu_iovec_memset(qiov, 0, 0, bytes);
1575         return 0;
1576     }
1577     if (offset + bytes > s->info.size) {
1578         uint64_t slop = offset + bytes - s->info.size;
1579 
1580         assert(slop < BDRV_SECTOR_SIZE);
1581         qemu_iovec_memset(qiov, bytes - slop, 0, slop);
1582         request.len -= slop;
1583     }
1584 
1585     do {
1586         ret = nbd_co_send_request(bs, &request, NULL);
1587         if (ret < 0) {
1588             continue;
1589         }
1590 
1591         ret = nbd_co_receive_cmdread_reply(s, request.handle, offset, qiov,
1592                                            &request_ret, &local_err);
1593         if (local_err) {
1594             trace_nbd_co_request_fail(request.from, request.len, request.handle,
1595                                       request.flags, request.type,
1596                                       nbd_cmd_lookup(request.type),
1597                                       ret, error_get_pretty(local_err));
1598             error_free(local_err);
1599             local_err = NULL;
1600         }
1601     } while (ret < 0 && nbd_client_connecting_wait(s));
1602 
1603     return ret ? ret : request_ret;
1604 }
1605 
1606 static int nbd_client_co_pwritev(BlockDriverState *bs, uint64_t offset,
1607                                  uint64_t bytes, QEMUIOVector *qiov, int flags)
1608 {
1609     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
1610     NBDRequest request = {
1611         .type = NBD_CMD_WRITE,
1612         .from = offset,
1613         .len = bytes,
1614     };
1615 
1616     assert(!(s->info.flags & NBD_FLAG_READ_ONLY));
1617     if (flags & BDRV_REQ_FUA) {
1618         assert(s->info.flags & NBD_FLAG_SEND_FUA);
1619         request.flags |= NBD_CMD_FLAG_FUA;
1620     }
1621 
1622     assert(bytes <= NBD_MAX_BUFFER_SIZE);
1623 
1624     if (!bytes) {
1625         return 0;
1626     }
1627     return nbd_co_request(bs, &request, qiov);
1628 }
1629 
1630 static int nbd_client_co_pwrite_zeroes(BlockDriverState *bs, int64_t offset,
1631                                        int bytes, BdrvRequestFlags flags)
1632 {
1633     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
1634     NBDRequest request = {
1635         .type = NBD_CMD_WRITE_ZEROES,
1636         .from = offset,
1637         .len = bytes,
1638     };
1639 
1640     assert(!(s->info.flags & NBD_FLAG_READ_ONLY));
1641     if (!(s->info.flags & NBD_FLAG_SEND_WRITE_ZEROES)) {
1642         return -ENOTSUP;
1643     }
1644 
1645     if (flags & BDRV_REQ_FUA) {
1646         assert(s->info.flags & NBD_FLAG_SEND_FUA);
1647         request.flags |= NBD_CMD_FLAG_FUA;
1648     }
1649     if (!(flags & BDRV_REQ_MAY_UNMAP)) {
1650         request.flags |= NBD_CMD_FLAG_NO_HOLE;
1651     }
1652     if (flags & BDRV_REQ_NO_FALLBACK) {
1653         assert(s->info.flags & NBD_FLAG_SEND_FAST_ZERO);
1654         request.flags |= NBD_CMD_FLAG_FAST_ZERO;
1655     }
1656 
1657     if (!bytes) {
1658         return 0;
1659     }
1660     return nbd_co_request(bs, &request, NULL);
1661 }
1662 
1663 static int nbd_client_co_flush(BlockDriverState *bs)
1664 {
1665     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
1666     NBDRequest request = { .type = NBD_CMD_FLUSH };
1667 
1668     if (!(s->info.flags & NBD_FLAG_SEND_FLUSH)) {
1669         return 0;
1670     }
1671 
1672     request.from = 0;
1673     request.len = 0;
1674 
1675     return nbd_co_request(bs, &request, NULL);
1676 }
1677 
1678 static int nbd_client_co_pdiscard(BlockDriverState *bs, int64_t offset,
1679                                   int bytes)
1680 {
1681     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
1682     NBDRequest request = {
1683         .type = NBD_CMD_TRIM,
1684         .from = offset,
1685         .len = bytes,
1686     };
1687 
1688     assert(!(s->info.flags & NBD_FLAG_READ_ONLY));
1689     if (!(s->info.flags & NBD_FLAG_SEND_TRIM) || !bytes) {
1690         return 0;
1691     }
1692 
1693     return nbd_co_request(bs, &request, NULL);
1694 }
1695 
1696 static int coroutine_fn nbd_client_co_block_status(
1697         BlockDriverState *bs, bool want_zero, int64_t offset, int64_t bytes,
1698         int64_t *pnum, int64_t *map, BlockDriverState **file)
1699 {
1700     int ret, request_ret;
1701     NBDExtent extent = { 0 };
1702     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
1703     Error *local_err = NULL;
1704 
1705     NBDRequest request = {
1706         .type = NBD_CMD_BLOCK_STATUS,
1707         .from = offset,
1708         .len = MIN(QEMU_ALIGN_DOWN(INT_MAX, bs->bl.request_alignment),
1709                    MIN(bytes, s->info.size - offset)),
1710         .flags = NBD_CMD_FLAG_REQ_ONE,
1711     };
1712 
1713     if (!s->info.base_allocation) {
1714         *pnum = bytes;
1715         *map = offset;
1716         *file = bs;
1717         return BDRV_BLOCK_DATA | BDRV_BLOCK_OFFSET_VALID;
1718     }
1719 
1720     /*
1721      * Work around the fact that the block layer doesn't do
1722      * byte-accurate sizing yet - if the status request exceeds the
1723      * server's advertised size because the block layer rounded size
1724      * up, we truncated the request to the server (above), or are
1725      * called on just the hole.
1726      */
1727     if (offset >= s->info.size) {
1728         *pnum = bytes;
1729         assert(bytes < BDRV_SECTOR_SIZE);
1730         /* Intentionally don't report offset_valid for the hole */
1731         return BDRV_BLOCK_ZERO;
1732     }
1733 
1734     if (s->info.min_block) {
1735         assert(QEMU_IS_ALIGNED(request.len, s->info.min_block));
1736     }
1737     do {
1738         ret = nbd_co_send_request(bs, &request, NULL);
1739         if (ret < 0) {
1740             continue;
1741         }
1742 
1743         ret = nbd_co_receive_blockstatus_reply(s, request.handle, bytes,
1744                                                &extent, &request_ret,
1745                                                &local_err);
1746         if (local_err) {
1747             trace_nbd_co_request_fail(request.from, request.len, request.handle,
1748                                       request.flags, request.type,
1749                                       nbd_cmd_lookup(request.type),
1750                                       ret, error_get_pretty(local_err));
1751             error_free(local_err);
1752             local_err = NULL;
1753         }
1754     } while (ret < 0 && nbd_client_connecting_wait(s));
1755 
1756     if (ret < 0 || request_ret < 0) {
1757         return ret ? ret : request_ret;
1758     }
1759 
1760     assert(extent.length);
1761     *pnum = extent.length;
1762     *map = offset;
1763     *file = bs;
1764     return (extent.flags & NBD_STATE_HOLE ? 0 : BDRV_BLOCK_DATA) |
1765         (extent.flags & NBD_STATE_ZERO ? BDRV_BLOCK_ZERO : 0) |
1766         BDRV_BLOCK_OFFSET_VALID;
1767 }
1768 
1769 static int nbd_client_reopen_prepare(BDRVReopenState *state,
1770                                      BlockReopenQueue *queue, Error **errp)
1771 {
1772     BDRVNBDState *s = (BDRVNBDState *)state->bs->opaque;
1773 
1774     if ((state->flags & BDRV_O_RDWR) && (s->info.flags & NBD_FLAG_READ_ONLY)) {
1775         error_setg(errp, "Can't reopen read-only NBD mount as read/write");
1776         return -EACCES;
1777     }
1778     return 0;
1779 }
1780 
1781 static void nbd_yank(void *opaque)
1782 {
1783     BlockDriverState *bs = opaque;
1784     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
1785 
1786     qatomic_store_release(&s->state, NBD_CLIENT_QUIT);
1787     qio_channel_shutdown(QIO_CHANNEL(s->sioc), QIO_CHANNEL_SHUTDOWN_BOTH, NULL);
1788 }
1789 
1790 static void nbd_client_close(BlockDriverState *bs)
1791 {
1792     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
1793     NBDRequest request = { .type = NBD_CMD_DISC };
1794 
1795     if (s->ioc) {
1796         nbd_send_request(s->ioc, &request);
1797     }
1798 
1799     nbd_teardown_connection(bs);
1800 }
1801 
1802 static int nbd_establish_connection(BlockDriverState *bs,
1803                                     SocketAddress *saddr,
1804                                     Error **errp)
1805 {
1806     ERRP_GUARD();
1807     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
1808 
1809     s->sioc = qio_channel_socket_new();
1810     qio_channel_set_name(QIO_CHANNEL(s->sioc), "nbd-client");
1811 
1812     qio_channel_socket_connect_sync(s->sioc, saddr, errp);
1813     if (*errp) {
1814         object_unref(OBJECT(s->sioc));
1815         s->sioc = NULL;
1816         return -1;
1817     }
1818 
1819     yank_register_function(BLOCKDEV_YANK_INSTANCE(bs->node_name), nbd_yank, bs);
1820     qio_channel_set_delay(QIO_CHANNEL(s->sioc), false);
1821 
1822     return 0;
1823 }
1824 
1825 /* nbd_client_handshake takes ownership on s->sioc. On failure it's unref'ed. */
1826 static int nbd_client_handshake(BlockDriverState *bs, Error **errp)
1827 {
1828     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
1829     AioContext *aio_context = bdrv_get_aio_context(bs);
1830     int ret;
1831 
1832     trace_nbd_client_handshake(s->export);
1833     qio_channel_set_blocking(QIO_CHANNEL(s->sioc), false, NULL);
1834     qio_channel_attach_aio_context(QIO_CHANNEL(s->sioc), aio_context);
1835 
1836     s->info.request_sizes = true;
1837     s->info.structured_reply = true;
1838     s->info.base_allocation = true;
1839     s->info.x_dirty_bitmap = g_strdup(s->x_dirty_bitmap);
1840     s->info.name = g_strdup(s->export ?: "");
1841     ret = nbd_receive_negotiate(aio_context, QIO_CHANNEL(s->sioc), s->tlscreds,
1842                                 s->hostname, &s->ioc, &s->info, errp);
1843     g_free(s->info.x_dirty_bitmap);
1844     g_free(s->info.name);
1845     if (ret < 0) {
1846         yank_unregister_function(BLOCKDEV_YANK_INSTANCE(bs->node_name),
1847                                  nbd_yank, bs);
1848         object_unref(OBJECT(s->sioc));
1849         s->sioc = NULL;
1850         return ret;
1851     }
1852     if (s->x_dirty_bitmap) {
1853         if (!s->info.base_allocation) {
1854             error_setg(errp, "requested x-dirty-bitmap %s not found",
1855                        s->x_dirty_bitmap);
1856             ret = -EINVAL;
1857             goto fail;
1858         }
1859         if (strcmp(s->x_dirty_bitmap, "qemu:allocation-depth") == 0) {
1860             s->alloc_depth = true;
1861         }
1862     }
1863     if (s->info.flags & NBD_FLAG_READ_ONLY) {
1864         ret = bdrv_apply_auto_read_only(bs, "NBD export is read-only", errp);
1865         if (ret < 0) {
1866             goto fail;
1867         }
1868     }
1869     if (s->info.flags & NBD_FLAG_SEND_FUA) {
1870         bs->supported_write_flags = BDRV_REQ_FUA;
1871         bs->supported_zero_flags |= BDRV_REQ_FUA;
1872     }
1873     if (s->info.flags & NBD_FLAG_SEND_WRITE_ZEROES) {
1874         bs->supported_zero_flags |= BDRV_REQ_MAY_UNMAP;
1875         if (s->info.flags & NBD_FLAG_SEND_FAST_ZERO) {
1876             bs->supported_zero_flags |= BDRV_REQ_NO_FALLBACK;
1877         }
1878     }
1879 
1880     if (!s->ioc) {
1881         s->ioc = QIO_CHANNEL(s->sioc);
1882         object_ref(OBJECT(s->ioc));
1883     }
1884 
1885     trace_nbd_client_handshake_success(s->export);
1886 
1887     return 0;
1888 
1889  fail:
1890     /*
1891      * We have connected, but must fail for other reasons.
1892      * Send NBD_CMD_DISC as a courtesy to the server.
1893      */
1894     {
1895         NBDRequest request = { .type = NBD_CMD_DISC };
1896 
1897         nbd_send_request(s->ioc ?: QIO_CHANNEL(s->sioc), &request);
1898 
1899         yank_unregister_function(BLOCKDEV_YANK_INSTANCE(bs->node_name),
1900                                  nbd_yank, bs);
1901         object_unref(OBJECT(s->sioc));
1902         s->sioc = NULL;
1903 
1904         return ret;
1905     }
1906 }
1907 
1908 /*
1909  * Parse nbd_open options
1910  */
1911 
1912 static int nbd_parse_uri(const char *filename, QDict *options)
1913 {
1914     URI *uri;
1915     const char *p;
1916     QueryParams *qp = NULL;
1917     int ret = 0;
1918     bool is_unix;
1919 
1920     uri = uri_parse(filename);
1921     if (!uri) {
1922         return -EINVAL;
1923     }
1924 
1925     /* transport */
1926     if (!g_strcmp0(uri->scheme, "nbd")) {
1927         is_unix = false;
1928     } else if (!g_strcmp0(uri->scheme, "nbd+tcp")) {
1929         is_unix = false;
1930     } else if (!g_strcmp0(uri->scheme, "nbd+unix")) {
1931         is_unix = true;
1932     } else {
1933         ret = -EINVAL;
1934         goto out;
1935     }
1936 
1937     p = uri->path ? uri->path : "";
1938     if (p[0] == '/') {
1939         p++;
1940     }
1941     if (p[0]) {
1942         qdict_put_str(options, "export", p);
1943     }
1944 
1945     qp = query_params_parse(uri->query);
1946     if (qp->n > 1 || (is_unix && !qp->n) || (!is_unix && qp->n)) {
1947         ret = -EINVAL;
1948         goto out;
1949     }
1950 
1951     if (is_unix) {
1952         /* nbd+unix:///export?socket=path */
1953         if (uri->server || uri->port || strcmp(qp->p[0].name, "socket")) {
1954             ret = -EINVAL;
1955             goto out;
1956         }
1957         qdict_put_str(options, "server.type", "unix");
1958         qdict_put_str(options, "server.path", qp->p[0].value);
1959     } else {
1960         QString *host;
1961         char *port_str;
1962 
1963         /* nbd[+tcp]://host[:port]/export */
1964         if (!uri->server) {
1965             ret = -EINVAL;
1966             goto out;
1967         }
1968 
1969         /* strip braces from literal IPv6 address */
1970         if (uri->server[0] == '[') {
1971             host = qstring_from_substr(uri->server, 1,
1972                                        strlen(uri->server) - 1);
1973         } else {
1974             host = qstring_from_str(uri->server);
1975         }
1976 
1977         qdict_put_str(options, "server.type", "inet");
1978         qdict_put(options, "server.host", host);
1979 
1980         port_str = g_strdup_printf("%d", uri->port ?: NBD_DEFAULT_PORT);
1981         qdict_put_str(options, "server.port", port_str);
1982         g_free(port_str);
1983     }
1984 
1985 out:
1986     if (qp) {
1987         query_params_free(qp);
1988     }
1989     uri_free(uri);
1990     return ret;
1991 }
1992 
1993 static bool nbd_has_filename_options_conflict(QDict *options, Error **errp)
1994 {
1995     const QDictEntry *e;
1996 
1997     for (e = qdict_first(options); e; e = qdict_next(options, e)) {
1998         if (!strcmp(e->key, "host") ||
1999             !strcmp(e->key, "port") ||
2000             !strcmp(e->key, "path") ||
2001             !strcmp(e->key, "export") ||
2002             strstart(e->key, "server.", NULL))
2003         {
2004             error_setg(errp, "Option '%s' cannot be used with a file name",
2005                        e->key);
2006             return true;
2007         }
2008     }
2009 
2010     return false;
2011 }
2012 
2013 static void nbd_parse_filename(const char *filename, QDict *options,
2014                                Error **errp)
2015 {
2016     g_autofree char *file = NULL;
2017     char *export_name;
2018     const char *host_spec;
2019     const char *unixpath;
2020 
2021     if (nbd_has_filename_options_conflict(options, errp)) {
2022         return;
2023     }
2024 
2025     if (strstr(filename, "://")) {
2026         int ret = nbd_parse_uri(filename, options);
2027         if (ret < 0) {
2028             error_setg(errp, "No valid URL specified");
2029         }
2030         return;
2031     }
2032 
2033     file = g_strdup(filename);
2034 
2035     export_name = strstr(file, EN_OPTSTR);
2036     if (export_name) {
2037         if (export_name[strlen(EN_OPTSTR)] == 0) {
2038             return;
2039         }
2040         export_name[0] = 0; /* truncate 'file' */
2041         export_name += strlen(EN_OPTSTR);
2042 
2043         qdict_put_str(options, "export", export_name);
2044     }
2045 
2046     /* extract the host_spec - fail if it's not nbd:... */
2047     if (!strstart(file, "nbd:", &host_spec)) {
2048         error_setg(errp, "File name string for NBD must start with 'nbd:'");
2049         return;
2050     }
2051 
2052     if (!*host_spec) {
2053         return;
2054     }
2055 
2056     /* are we a UNIX or TCP socket? */
2057     if (strstart(host_spec, "unix:", &unixpath)) {
2058         qdict_put_str(options, "server.type", "unix");
2059         qdict_put_str(options, "server.path", unixpath);
2060     } else {
2061         InetSocketAddress *addr = g_new(InetSocketAddress, 1);
2062 
2063         if (inet_parse(addr, host_spec, errp)) {
2064             goto out_inet;
2065         }
2066 
2067         qdict_put_str(options, "server.type", "inet");
2068         qdict_put_str(options, "server.host", addr->host);
2069         qdict_put_str(options, "server.port", addr->port);
2070     out_inet:
2071         qapi_free_InetSocketAddress(addr);
2072     }
2073 }
2074 
2075 static bool nbd_process_legacy_socket_options(QDict *output_options,
2076                                               QemuOpts *legacy_opts,
2077                                               Error **errp)
2078 {
2079     const char *path = qemu_opt_get(legacy_opts, "path");
2080     const char *host = qemu_opt_get(legacy_opts, "host");
2081     const char *port = qemu_opt_get(legacy_opts, "port");
2082     const QDictEntry *e;
2083 
2084     if (!path && !host && !port) {
2085         return true;
2086     }
2087 
2088     for (e = qdict_first(output_options); e; e = qdict_next(output_options, e))
2089     {
2090         if (strstart(e->key, "server.", NULL)) {
2091             error_setg(errp, "Cannot use 'server' and path/host/port at the "
2092                        "same time");
2093             return false;
2094         }
2095     }
2096 
2097     if (path && host) {
2098         error_setg(errp, "path and host may not be used at the same time");
2099         return false;
2100     } else if (path) {
2101         if (port) {
2102             error_setg(errp, "port may not be used without host");
2103             return false;
2104         }
2105 
2106         qdict_put_str(output_options, "server.type", "unix");
2107         qdict_put_str(output_options, "server.path", path);
2108     } else if (host) {
2109         qdict_put_str(output_options, "server.type", "inet");
2110         qdict_put_str(output_options, "server.host", host);
2111         qdict_put_str(output_options, "server.port",
2112                       port ?: stringify(NBD_DEFAULT_PORT));
2113     }
2114 
2115     return true;
2116 }
2117 
2118 static SocketAddress *nbd_config(BDRVNBDState *s, QDict *options,
2119                                  Error **errp)
2120 {
2121     SocketAddress *saddr = NULL;
2122     QDict *addr = NULL;
2123     Visitor *iv = NULL;
2124 
2125     qdict_extract_subqdict(options, &addr, "server.");
2126     if (!qdict_size(addr)) {
2127         error_setg(errp, "NBD server address missing");
2128         goto done;
2129     }
2130 
2131     iv = qobject_input_visitor_new_flat_confused(addr, errp);
2132     if (!iv) {
2133         goto done;
2134     }
2135 
2136     if (!visit_type_SocketAddress(iv, NULL, &saddr, errp)) {
2137         goto done;
2138     }
2139 
2140 done:
2141     qobject_unref(addr);
2142     visit_free(iv);
2143     return saddr;
2144 }
2145 
2146 static QCryptoTLSCreds *nbd_get_tls_creds(const char *id, Error **errp)
2147 {
2148     Object *obj;
2149     QCryptoTLSCreds *creds;
2150 
2151     obj = object_resolve_path_component(
2152         object_get_objects_root(), id);
2153     if (!obj) {
2154         error_setg(errp, "No TLS credentials with id '%s'",
2155                    id);
2156         return NULL;
2157     }
2158     creds = (QCryptoTLSCreds *)
2159         object_dynamic_cast(obj, TYPE_QCRYPTO_TLS_CREDS);
2160     if (!creds) {
2161         error_setg(errp, "Object with id '%s' is not TLS credentials",
2162                    id);
2163         return NULL;
2164     }
2165 
2166     if (creds->endpoint != QCRYPTO_TLS_CREDS_ENDPOINT_CLIENT) {
2167         error_setg(errp,
2168                    "Expecting TLS credentials with a client endpoint");
2169         return NULL;
2170     }
2171     object_ref(obj);
2172     return creds;
2173 }
2174 
2175 
2176 static QemuOptsList nbd_runtime_opts = {
2177     .name = "nbd",
2178     .head = QTAILQ_HEAD_INITIALIZER(nbd_runtime_opts.head),
2179     .desc = {
2180         {
2181             .name = "host",
2182             .type = QEMU_OPT_STRING,
2183             .help = "TCP host to connect to",
2184         },
2185         {
2186             .name = "port",
2187             .type = QEMU_OPT_STRING,
2188             .help = "TCP port to connect to",
2189         },
2190         {
2191             .name = "path",
2192             .type = QEMU_OPT_STRING,
2193             .help = "Unix socket path to connect to",
2194         },
2195         {
2196             .name = "export",
2197             .type = QEMU_OPT_STRING,
2198             .help = "Name of the NBD export to open",
2199         },
2200         {
2201             .name = "tls-creds",
2202             .type = QEMU_OPT_STRING,
2203             .help = "ID of the TLS credentials to use",
2204         },
2205         {
2206             .name = "x-dirty-bitmap",
2207             .type = QEMU_OPT_STRING,
2208             .help = "experimental: expose named dirty bitmap in place of "
2209                     "block status",
2210         },
2211         {
2212             .name = "reconnect-delay",
2213             .type = QEMU_OPT_NUMBER,
2214             .help = "On an unexpected disconnect, the nbd client tries to "
2215                     "connect again until succeeding or encountering a serious "
2216                     "error.  During the first @reconnect-delay seconds, all "
2217                     "requests are paused and will be rerun on a successful "
2218                     "reconnect. After that time, any delayed requests and all "
2219                     "future requests before a successful reconnect will "
2220                     "immediately fail. Default 0",
2221         },
2222         { /* end of list */ }
2223     },
2224 };
2225 
2226 static int nbd_process_options(BlockDriverState *bs, QDict *options,
2227                                Error **errp)
2228 {
2229     BDRVNBDState *s = bs->opaque;
2230     QemuOpts *opts;
2231     int ret = -EINVAL;
2232 
2233     opts = qemu_opts_create(&nbd_runtime_opts, NULL, 0, &error_abort);
2234     if (!qemu_opts_absorb_qdict(opts, options, errp)) {
2235         goto error;
2236     }
2237 
2238     /* Translate @host, @port, and @path to a SocketAddress */
2239     if (!nbd_process_legacy_socket_options(options, opts, errp)) {
2240         goto error;
2241     }
2242 
2243     /* Pop the config into our state object. Exit if invalid. */
2244     s->saddr = nbd_config(s, options, errp);
2245     if (!s->saddr) {
2246         goto error;
2247     }
2248 
2249     s->export = g_strdup(qemu_opt_get(opts, "export"));
2250     if (s->export && strlen(s->export) > NBD_MAX_STRING_SIZE) {
2251         error_setg(errp, "export name too long to send to server");
2252         goto error;
2253     }
2254 
2255     s->tlscredsid = g_strdup(qemu_opt_get(opts, "tls-creds"));
2256     if (s->tlscredsid) {
2257         s->tlscreds = nbd_get_tls_creds(s->tlscredsid, errp);
2258         if (!s->tlscreds) {
2259             goto error;
2260         }
2261 
2262         /* TODO SOCKET_ADDRESS_KIND_FD where fd has AF_INET or AF_INET6 */
2263         if (s->saddr->type != SOCKET_ADDRESS_TYPE_INET) {
2264             error_setg(errp, "TLS only supported over IP sockets");
2265             goto error;
2266         }
2267         s->hostname = s->saddr->u.inet.host;
2268     }
2269 
2270     s->x_dirty_bitmap = g_strdup(qemu_opt_get(opts, "x-dirty-bitmap"));
2271     if (s->x_dirty_bitmap && strlen(s->x_dirty_bitmap) > NBD_MAX_STRING_SIZE) {
2272         error_setg(errp, "x-dirty-bitmap query too long to send to server");
2273         goto error;
2274     }
2275 
2276     s->reconnect_delay = qemu_opt_get_number(opts, "reconnect-delay", 0);
2277 
2278     ret = 0;
2279 
2280  error:
2281     if (ret < 0) {
2282         nbd_clear_bdrvstate(s);
2283     }
2284     qemu_opts_del(opts);
2285     return ret;
2286 }
2287 
2288 static int nbd_open(BlockDriverState *bs, QDict *options, int flags,
2289                     Error **errp)
2290 {
2291     int ret;
2292     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
2293 
2294     ret = nbd_process_options(bs, options, errp);
2295     if (ret < 0) {
2296         return ret;
2297     }
2298 
2299     s->bs = bs;
2300     qemu_co_mutex_init(&s->send_mutex);
2301     qemu_co_queue_init(&s->free_sema);
2302 
2303     if (!yank_register_instance(BLOCKDEV_YANK_INSTANCE(bs->node_name), errp)) {
2304         return -EEXIST;
2305     }
2306 
2307     /*
2308      * establish TCP connection, return error if it fails
2309      * TODO: Configurable retry-until-timeout behaviour.
2310      */
2311     if (nbd_establish_connection(bs, s->saddr, errp) < 0) {
2312         yank_unregister_instance(BLOCKDEV_YANK_INSTANCE(bs->node_name));
2313         return -ECONNREFUSED;
2314     }
2315 
2316     ret = nbd_client_handshake(bs, errp);
2317     if (ret < 0) {
2318         yank_unregister_instance(BLOCKDEV_YANK_INSTANCE(bs->node_name));
2319         nbd_clear_bdrvstate(s);
2320         return ret;
2321     }
2322     /* successfully connected */
2323     s->state = NBD_CLIENT_CONNECTED;
2324 
2325     nbd_init_connect_thread(s);
2326 
2327     s->connection_co = qemu_coroutine_create(nbd_connection_entry, s);
2328     bdrv_inc_in_flight(bs);
2329     aio_co_schedule(bdrv_get_aio_context(bs), s->connection_co);
2330 
2331     return 0;
2332 }
2333 
2334 static int nbd_co_flush(BlockDriverState *bs)
2335 {
2336     return nbd_client_co_flush(bs);
2337 }
2338 
2339 static void nbd_refresh_limits(BlockDriverState *bs, Error **errp)
2340 {
2341     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
2342     uint32_t min = s->info.min_block;
2343     uint32_t max = MIN_NON_ZERO(NBD_MAX_BUFFER_SIZE, s->info.max_block);
2344 
2345     /*
2346      * If the server did not advertise an alignment:
2347      * - a size that is not sector-aligned implies that an alignment
2348      *   of 1 can be used to access those tail bytes
2349      * - advertisement of block status requires an alignment of 1, so
2350      *   that we don't violate block layer constraints that block
2351      *   status is always aligned (as we can't control whether the
2352      *   server will report sub-sector extents, such as a hole at EOF
2353      *   on an unaligned POSIX file)
2354      * - otherwise, assume the server is so old that we are safer avoiding
2355      *   sub-sector requests
2356      */
2357     if (!min) {
2358         min = (!QEMU_IS_ALIGNED(s->info.size, BDRV_SECTOR_SIZE) ||
2359                s->info.base_allocation) ? 1 : BDRV_SECTOR_SIZE;
2360     }
2361 
2362     bs->bl.request_alignment = min;
2363     bs->bl.max_pdiscard = QEMU_ALIGN_DOWN(INT_MAX, min);
2364     bs->bl.max_pwrite_zeroes = max;
2365     bs->bl.max_transfer = max;
2366 
2367     if (s->info.opt_block &&
2368         s->info.opt_block > bs->bl.opt_transfer) {
2369         bs->bl.opt_transfer = s->info.opt_block;
2370     }
2371 }
2372 
2373 static void nbd_close(BlockDriverState *bs)
2374 {
2375     BDRVNBDState *s = bs->opaque;
2376 
2377     nbd_client_close(bs);
2378     yank_unregister_instance(BLOCKDEV_YANK_INSTANCE(bs->node_name));
2379     nbd_clear_bdrvstate(s);
2380 }
2381 
2382 /*
2383  * NBD cannot truncate, but if the caller asks to truncate to the same size, or
2384  * to a smaller size with exact=false, there is no reason to fail the
2385  * operation.
2386  *
2387  * Preallocation mode is ignored since it does not seems useful to fail when
2388  * we never change anything.
2389  */
2390 static int coroutine_fn nbd_co_truncate(BlockDriverState *bs, int64_t offset,
2391                                         bool exact, PreallocMode prealloc,
2392                                         BdrvRequestFlags flags, Error **errp)
2393 {
2394     BDRVNBDState *s = bs->opaque;
2395 
2396     if (offset != s->info.size && exact) {
2397         error_setg(errp, "Cannot resize NBD nodes");
2398         return -ENOTSUP;
2399     }
2400 
2401     if (offset > s->info.size) {
2402         error_setg(errp, "Cannot grow NBD nodes");
2403         return -EINVAL;
2404     }
2405 
2406     return 0;
2407 }
2408 
2409 static int64_t nbd_getlength(BlockDriverState *bs)
2410 {
2411     BDRVNBDState *s = bs->opaque;
2412 
2413     return s->info.size;
2414 }
2415 
2416 static void nbd_refresh_filename(BlockDriverState *bs)
2417 {
2418     BDRVNBDState *s = bs->opaque;
2419     const char *host = NULL, *port = NULL, *path = NULL;
2420     size_t len = 0;
2421 
2422     if (s->saddr->type == SOCKET_ADDRESS_TYPE_INET) {
2423         const InetSocketAddress *inet = &s->saddr->u.inet;
2424         if (!inet->has_ipv4 && !inet->has_ipv6 && !inet->has_to) {
2425             host = inet->host;
2426             port = inet->port;
2427         }
2428     } else if (s->saddr->type == SOCKET_ADDRESS_TYPE_UNIX) {
2429         path = s->saddr->u.q_unix.path;
2430     } /* else can't represent as pseudo-filename */
2431 
2432     if (path && s->export) {
2433         len = snprintf(bs->exact_filename, sizeof(bs->exact_filename),
2434                        "nbd+unix:///%s?socket=%s", s->export, path);
2435     } else if (path && !s->export) {
2436         len = snprintf(bs->exact_filename, sizeof(bs->exact_filename),
2437                        "nbd+unix://?socket=%s", path);
2438     } else if (host && s->export) {
2439         len = snprintf(bs->exact_filename, sizeof(bs->exact_filename),
2440                        "nbd://%s:%s/%s", host, port, s->export);
2441     } else if (host && !s->export) {
2442         len = snprintf(bs->exact_filename, sizeof(bs->exact_filename),
2443                        "nbd://%s:%s", host, port);
2444     }
2445     if (len >= sizeof(bs->exact_filename)) {
2446         /* Name is too long to represent exactly, so leave it empty. */
2447         bs->exact_filename[0] = '\0';
2448     }
2449 }
2450 
2451 static char *nbd_dirname(BlockDriverState *bs, Error **errp)
2452 {
2453     /* The generic bdrv_dirname() implementation is able to work out some
2454      * directory name for NBD nodes, but that would be wrong. So far there is no
2455      * specification for how "export paths" would work, so NBD does not have
2456      * directory names. */
2457     error_setg(errp, "Cannot generate a base directory for NBD nodes");
2458     return NULL;
2459 }
2460 
2461 static const char *const nbd_strong_runtime_opts[] = {
2462     "path",
2463     "host",
2464     "port",
2465     "export",
2466     "tls-creds",
2467     "server.",
2468 
2469     NULL
2470 };
2471 
2472 static void nbd_cancel_in_flight(BlockDriverState *bs)
2473 {
2474     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
2475 
2476     reconnect_delay_timer_del(s);
2477 
2478     if (s->state == NBD_CLIENT_CONNECTING_WAIT) {
2479         s->state = NBD_CLIENT_CONNECTING_NOWAIT;
2480         qemu_co_queue_restart_all(&s->free_sema);
2481     }
2482 }
2483 
2484 static BlockDriver bdrv_nbd = {
2485     .format_name                = "nbd",
2486     .protocol_name              = "nbd",
2487     .instance_size              = sizeof(BDRVNBDState),
2488     .bdrv_parse_filename        = nbd_parse_filename,
2489     .bdrv_co_create_opts        = bdrv_co_create_opts_simple,
2490     .create_opts                = &bdrv_create_opts_simple,
2491     .bdrv_file_open             = nbd_open,
2492     .bdrv_reopen_prepare        = nbd_client_reopen_prepare,
2493     .bdrv_co_preadv             = nbd_client_co_preadv,
2494     .bdrv_co_pwritev            = nbd_client_co_pwritev,
2495     .bdrv_co_pwrite_zeroes      = nbd_client_co_pwrite_zeroes,
2496     .bdrv_close                 = nbd_close,
2497     .bdrv_co_flush_to_os        = nbd_co_flush,
2498     .bdrv_co_pdiscard           = nbd_client_co_pdiscard,
2499     .bdrv_refresh_limits        = nbd_refresh_limits,
2500     .bdrv_co_truncate           = nbd_co_truncate,
2501     .bdrv_getlength             = nbd_getlength,
2502     .bdrv_detach_aio_context    = nbd_client_detach_aio_context,
2503     .bdrv_attach_aio_context    = nbd_client_attach_aio_context,
2504     .bdrv_co_drain_begin        = nbd_client_co_drain_begin,
2505     .bdrv_co_drain_end          = nbd_client_co_drain_end,
2506     .bdrv_refresh_filename      = nbd_refresh_filename,
2507     .bdrv_co_block_status       = nbd_client_co_block_status,
2508     .bdrv_dirname               = nbd_dirname,
2509     .strong_runtime_opts        = nbd_strong_runtime_opts,
2510     .bdrv_cancel_in_flight      = nbd_cancel_in_flight,
2511 };
2512 
2513 static BlockDriver bdrv_nbd_tcp = {
2514     .format_name                = "nbd",
2515     .protocol_name              = "nbd+tcp",
2516     .instance_size              = sizeof(BDRVNBDState),
2517     .bdrv_parse_filename        = nbd_parse_filename,
2518     .bdrv_co_create_opts        = bdrv_co_create_opts_simple,
2519     .create_opts                = &bdrv_create_opts_simple,
2520     .bdrv_file_open             = nbd_open,
2521     .bdrv_reopen_prepare        = nbd_client_reopen_prepare,
2522     .bdrv_co_preadv             = nbd_client_co_preadv,
2523     .bdrv_co_pwritev            = nbd_client_co_pwritev,
2524     .bdrv_co_pwrite_zeroes      = nbd_client_co_pwrite_zeroes,
2525     .bdrv_close                 = nbd_close,
2526     .bdrv_co_flush_to_os        = nbd_co_flush,
2527     .bdrv_co_pdiscard           = nbd_client_co_pdiscard,
2528     .bdrv_refresh_limits        = nbd_refresh_limits,
2529     .bdrv_co_truncate           = nbd_co_truncate,
2530     .bdrv_getlength             = nbd_getlength,
2531     .bdrv_detach_aio_context    = nbd_client_detach_aio_context,
2532     .bdrv_attach_aio_context    = nbd_client_attach_aio_context,
2533     .bdrv_co_drain_begin        = nbd_client_co_drain_begin,
2534     .bdrv_co_drain_end          = nbd_client_co_drain_end,
2535     .bdrv_refresh_filename      = nbd_refresh_filename,
2536     .bdrv_co_block_status       = nbd_client_co_block_status,
2537     .bdrv_dirname               = nbd_dirname,
2538     .strong_runtime_opts        = nbd_strong_runtime_opts,
2539     .bdrv_cancel_in_flight      = nbd_cancel_in_flight,
2540 };
2541 
2542 static BlockDriver bdrv_nbd_unix = {
2543     .format_name                = "nbd",
2544     .protocol_name              = "nbd+unix",
2545     .instance_size              = sizeof(BDRVNBDState),
2546     .bdrv_parse_filename        = nbd_parse_filename,
2547     .bdrv_co_create_opts        = bdrv_co_create_opts_simple,
2548     .create_opts                = &bdrv_create_opts_simple,
2549     .bdrv_file_open             = nbd_open,
2550     .bdrv_reopen_prepare        = nbd_client_reopen_prepare,
2551     .bdrv_co_preadv             = nbd_client_co_preadv,
2552     .bdrv_co_pwritev            = nbd_client_co_pwritev,
2553     .bdrv_co_pwrite_zeroes      = nbd_client_co_pwrite_zeroes,
2554     .bdrv_close                 = nbd_close,
2555     .bdrv_co_flush_to_os        = nbd_co_flush,
2556     .bdrv_co_pdiscard           = nbd_client_co_pdiscard,
2557     .bdrv_refresh_limits        = nbd_refresh_limits,
2558     .bdrv_co_truncate           = nbd_co_truncate,
2559     .bdrv_getlength             = nbd_getlength,
2560     .bdrv_detach_aio_context    = nbd_client_detach_aio_context,
2561     .bdrv_attach_aio_context    = nbd_client_attach_aio_context,
2562     .bdrv_co_drain_begin        = nbd_client_co_drain_begin,
2563     .bdrv_co_drain_end          = nbd_client_co_drain_end,
2564     .bdrv_refresh_filename      = nbd_refresh_filename,
2565     .bdrv_co_block_status       = nbd_client_co_block_status,
2566     .bdrv_dirname               = nbd_dirname,
2567     .strong_runtime_opts        = nbd_strong_runtime_opts,
2568     .bdrv_cancel_in_flight      = nbd_cancel_in_flight,
2569 };
2570 
2571 static void bdrv_nbd_init(void)
2572 {
2573     bdrv_register(&bdrv_nbd);
2574     bdrv_register(&bdrv_nbd_tcp);
2575     bdrv_register(&bdrv_nbd_unix);
2576 }
2577 
2578 block_init(bdrv_nbd_init);
2579