xref: /openbmc/qemu/block/nbd.c (revision e9e51b71)
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     qemu_mutex_lock(&thr->mutex);
447 
448     switch (thr->state) {
449     case CONNECT_THREAD_FAIL:
450     case CONNECT_THREAD_NONE:
451         error_free(thr->err);
452         thr->err = NULL;
453         thr->state = CONNECT_THREAD_RUNNING;
454         qemu_thread_create(&thread, "nbd-connect",
455                            connect_thread_func, thr, QEMU_THREAD_DETACHED);
456         break;
457     case CONNECT_THREAD_SUCCESS:
458         /* Previous attempt finally succeeded in background */
459         thr->state = CONNECT_THREAD_NONE;
460         s->sioc = thr->sioc;
461         thr->sioc = NULL;
462         yank_register_function(BLOCKDEV_YANK_INSTANCE(bs->node_name),
463                                nbd_yank, bs);
464         qemu_mutex_unlock(&thr->mutex);
465         return 0;
466     case CONNECT_THREAD_RUNNING:
467         /* Already running, will wait */
468         break;
469     default:
470         abort();
471     }
472 
473     thr->bh_ctx = qemu_get_current_aio_context();
474 
475     qemu_mutex_unlock(&thr->mutex);
476 
477 
478     /*
479      * We are going to wait for connect-thread finish, but
480      * nbd_client_co_drain_begin() can interrupt.
481      *
482      * Note that wait_connect variable is not visible for connect-thread. It
483      * doesn't need mutex protection, it used only inside home aio context of
484      * bs.
485      */
486     s->wait_connect = true;
487     qemu_coroutine_yield();
488 
489     qemu_mutex_lock(&thr->mutex);
490 
491     switch (thr->state) {
492     case CONNECT_THREAD_SUCCESS:
493     case CONNECT_THREAD_FAIL:
494         thr->state = CONNECT_THREAD_NONE;
495         error_propagate(errp, thr->err);
496         thr->err = NULL;
497         s->sioc = thr->sioc;
498         thr->sioc = NULL;
499         if (s->sioc) {
500             yank_register_function(BLOCKDEV_YANK_INSTANCE(bs->node_name),
501                                    nbd_yank, bs);
502         }
503         ret = (s->sioc ? 0 : -1);
504         break;
505     case CONNECT_THREAD_RUNNING:
506     case CONNECT_THREAD_RUNNING_DETACHED:
507         /*
508          * Obviously, drained section wants to start. Report the attempt as
509          * failed. Still connect thread is executing in background, and its
510          * result may be used for next connection attempt.
511          */
512         ret = -1;
513         error_setg(errp, "Connection attempt cancelled by other operation");
514         break;
515 
516     case CONNECT_THREAD_NONE:
517         /*
518          * Impossible. We've seen this thread running. So it should be
519          * running or at least give some results.
520          */
521         abort();
522 
523     default:
524         abort();
525     }
526 
527     qemu_mutex_unlock(&thr->mutex);
528 
529     return ret;
530 }
531 
532 /*
533  * nbd_co_establish_connection_cancel
534  * Cancel nbd_co_establish_connection asynchronously: it will finish soon, to
535  * allow drained section to begin.
536  *
537  * If detach is true, also cleanup the state (or if thread is running, move it
538  * to CONNECT_THREAD_RUNNING_DETACHED state). s->connect_thread becomes NULL if
539  * detach is true.
540  */
541 static void nbd_co_establish_connection_cancel(BlockDriverState *bs,
542                                                bool detach)
543 {
544     BDRVNBDState *s = bs->opaque;
545     NBDConnectThread *thr = s->connect_thread;
546     bool wake = false;
547     bool do_free = false;
548 
549     qemu_mutex_lock(&thr->mutex);
550 
551     if (thr->state == CONNECT_THREAD_RUNNING) {
552         /* We can cancel only in running state, when bh is not yet scheduled */
553         thr->bh_ctx = NULL;
554         if (s->wait_connect) {
555             s->wait_connect = false;
556             wake = true;
557         }
558         if (detach) {
559             thr->state = CONNECT_THREAD_RUNNING_DETACHED;
560             s->connect_thread = NULL;
561         }
562     } else if (detach) {
563         do_free = true;
564     }
565 
566     qemu_mutex_unlock(&thr->mutex);
567 
568     if (do_free) {
569         nbd_free_connect_thread(thr);
570         s->connect_thread = NULL;
571     }
572 
573     if (wake) {
574         aio_co_wake(s->connection_co);
575     }
576 }
577 
578 static coroutine_fn void nbd_reconnect_attempt(BDRVNBDState *s)
579 {
580     int ret;
581     Error *local_err = NULL;
582 
583     if (!nbd_client_connecting(s)) {
584         return;
585     }
586 
587     /* Wait for completion of all in-flight requests */
588 
589     qemu_co_mutex_lock(&s->send_mutex);
590 
591     while (s->in_flight > 0) {
592         qemu_co_mutex_unlock(&s->send_mutex);
593         nbd_recv_coroutines_wake_all(s);
594         s->wait_in_flight = true;
595         qemu_coroutine_yield();
596         s->wait_in_flight = false;
597         qemu_co_mutex_lock(&s->send_mutex);
598     }
599 
600     qemu_co_mutex_unlock(&s->send_mutex);
601 
602     if (!nbd_client_connecting(s)) {
603         return;
604     }
605 
606     /*
607      * Now we are sure that nobody is accessing the channel, and no one will
608      * try until we set the state to CONNECTED.
609      */
610 
611     /* Finalize previous connection if any */
612     if (s->ioc) {
613         qio_channel_detach_aio_context(QIO_CHANNEL(s->ioc));
614         yank_unregister_function(BLOCKDEV_YANK_INSTANCE(s->bs->node_name),
615                                  nbd_yank, s->bs);
616         object_unref(OBJECT(s->sioc));
617         s->sioc = NULL;
618         object_unref(OBJECT(s->ioc));
619         s->ioc = NULL;
620     }
621 
622     if (nbd_co_establish_connection(s->bs, &local_err) < 0) {
623         ret = -ECONNREFUSED;
624         goto out;
625     }
626 
627     bdrv_dec_in_flight(s->bs);
628 
629     ret = nbd_client_handshake(s->bs, &local_err);
630 
631     if (s->drained) {
632         s->wait_drained_end = true;
633         while (s->drained) {
634             /*
635              * We may be entered once from nbd_client_attach_aio_context_bh
636              * and then from nbd_client_co_drain_end. So here is a loop.
637              */
638             qemu_coroutine_yield();
639         }
640     }
641     bdrv_inc_in_flight(s->bs);
642 
643 out:
644     s->connect_status = ret;
645     error_free(s->connect_err);
646     s->connect_err = NULL;
647     error_propagate(&s->connect_err, local_err);
648 
649     if (ret >= 0) {
650         /* successfully connected */
651         s->state = NBD_CLIENT_CONNECTED;
652         qemu_co_queue_restart_all(&s->free_sema);
653     }
654 }
655 
656 static coroutine_fn void nbd_co_reconnect_loop(BDRVNBDState *s)
657 {
658     uint64_t timeout = 1 * NANOSECONDS_PER_SECOND;
659     uint64_t max_timeout = 16 * NANOSECONDS_PER_SECOND;
660 
661     if (qatomic_load_acquire(&s->state) == NBD_CLIENT_CONNECTING_WAIT) {
662         reconnect_delay_timer_init(s, qemu_clock_get_ns(QEMU_CLOCK_REALTIME) +
663                                    s->reconnect_delay * NANOSECONDS_PER_SECOND);
664     }
665 
666     nbd_reconnect_attempt(s);
667 
668     while (nbd_client_connecting(s)) {
669         if (s->drained) {
670             bdrv_dec_in_flight(s->bs);
671             s->wait_drained_end = true;
672             while (s->drained) {
673                 /*
674                  * We may be entered once from nbd_client_attach_aio_context_bh
675                  * and then from nbd_client_co_drain_end. So here is a loop.
676                  */
677                 qemu_coroutine_yield();
678             }
679             bdrv_inc_in_flight(s->bs);
680         } else {
681             qemu_co_sleep_ns_wakeable(QEMU_CLOCK_REALTIME, timeout,
682                                       &s->connection_co_sleep_ns_state);
683             if (s->drained) {
684                 continue;
685             }
686             if (timeout < max_timeout) {
687                 timeout *= 2;
688             }
689         }
690 
691         nbd_reconnect_attempt(s);
692     }
693 
694     reconnect_delay_timer_del(s);
695 }
696 
697 static coroutine_fn void nbd_connection_entry(void *opaque)
698 {
699     BDRVNBDState *s = opaque;
700     uint64_t i;
701     int ret = 0;
702     Error *local_err = NULL;
703 
704     while (qatomic_load_acquire(&s->state) != NBD_CLIENT_QUIT) {
705         /*
706          * The NBD client can only really be considered idle when it has
707          * yielded from qio_channel_readv_all_eof(), waiting for data. This is
708          * the point where the additional scheduled coroutine entry happens
709          * after nbd_client_attach_aio_context().
710          *
711          * Therefore we keep an additional in_flight reference all the time and
712          * only drop it temporarily here.
713          */
714 
715         if (nbd_client_connecting(s)) {
716             nbd_co_reconnect_loop(s);
717         }
718 
719         if (qatomic_load_acquire(&s->state) != NBD_CLIENT_CONNECTED) {
720             continue;
721         }
722 
723         assert(s->reply.handle == 0);
724         ret = nbd_receive_reply(s->bs, s->ioc, &s->reply, &local_err);
725 
726         if (local_err) {
727             trace_nbd_read_reply_entry_fail(ret, error_get_pretty(local_err));
728             error_free(local_err);
729             local_err = NULL;
730         }
731         if (ret <= 0) {
732             nbd_channel_error(s, ret ? ret : -EIO);
733             continue;
734         }
735 
736         /*
737          * There's no need for a mutex on the receive side, because the
738          * handler acts as a synchronization point and ensures that only
739          * one coroutine is called until the reply finishes.
740          */
741         i = HANDLE_TO_INDEX(s, s->reply.handle);
742         if (i >= MAX_NBD_REQUESTS ||
743             !s->requests[i].coroutine ||
744             !s->requests[i].receiving ||
745             (nbd_reply_is_structured(&s->reply) && !s->info.structured_reply))
746         {
747             nbd_channel_error(s, -EINVAL);
748             continue;
749         }
750 
751         /*
752          * We're woken up again by the request itself.  Note that there
753          * is no race between yielding and reentering connection_co.  This
754          * is because:
755          *
756          * - if the request runs on the same AioContext, it is only
757          *   entered after we yield
758          *
759          * - if the request runs on a different AioContext, reentering
760          *   connection_co happens through a bottom half, which can only
761          *   run after we yield.
762          */
763         aio_co_wake(s->requests[i].coroutine);
764         qemu_coroutine_yield();
765     }
766 
767     qemu_co_queue_restart_all(&s->free_sema);
768     nbd_recv_coroutines_wake_all(s);
769     bdrv_dec_in_flight(s->bs);
770 
771     s->connection_co = NULL;
772     if (s->ioc) {
773         qio_channel_detach_aio_context(QIO_CHANNEL(s->ioc));
774         yank_unregister_function(BLOCKDEV_YANK_INSTANCE(s->bs->node_name),
775                                  nbd_yank, s->bs);
776         object_unref(OBJECT(s->sioc));
777         s->sioc = NULL;
778         object_unref(OBJECT(s->ioc));
779         s->ioc = NULL;
780     }
781 
782     if (s->teardown_co) {
783         aio_co_wake(s->teardown_co);
784     }
785     aio_wait_kick();
786 }
787 
788 static int nbd_co_send_request(BlockDriverState *bs,
789                                NBDRequest *request,
790                                QEMUIOVector *qiov)
791 {
792     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
793     int rc, i = -1;
794 
795     qemu_co_mutex_lock(&s->send_mutex);
796     while (s->in_flight == MAX_NBD_REQUESTS || nbd_client_connecting_wait(s)) {
797         qemu_co_queue_wait(&s->free_sema, &s->send_mutex);
798     }
799 
800     if (qatomic_load_acquire(&s->state) != NBD_CLIENT_CONNECTED) {
801         rc = -EIO;
802         goto err;
803     }
804 
805     s->in_flight++;
806 
807     for (i = 0; i < MAX_NBD_REQUESTS; i++) {
808         if (s->requests[i].coroutine == NULL) {
809             break;
810         }
811     }
812 
813     g_assert(qemu_in_coroutine());
814     assert(i < MAX_NBD_REQUESTS);
815 
816     s->requests[i].coroutine = qemu_coroutine_self();
817     s->requests[i].offset = request->from;
818     s->requests[i].receiving = false;
819 
820     request->handle = INDEX_TO_HANDLE(s, i);
821 
822     assert(s->ioc);
823 
824     if (qiov) {
825         qio_channel_set_cork(s->ioc, true);
826         rc = nbd_send_request(s->ioc, request);
827         if (qatomic_load_acquire(&s->state) == NBD_CLIENT_CONNECTED &&
828             rc >= 0) {
829             if (qio_channel_writev_all(s->ioc, qiov->iov, qiov->niov,
830                                        NULL) < 0) {
831                 rc = -EIO;
832             }
833         } else if (rc >= 0) {
834             rc = -EIO;
835         }
836         qio_channel_set_cork(s->ioc, false);
837     } else {
838         rc = nbd_send_request(s->ioc, request);
839     }
840 
841 err:
842     if (rc < 0) {
843         nbd_channel_error(s, rc);
844         if (i != -1) {
845             s->requests[i].coroutine = NULL;
846             s->in_flight--;
847         }
848         if (s->in_flight == 0 && s->wait_in_flight) {
849             aio_co_wake(s->connection_co);
850         } else {
851             qemu_co_queue_next(&s->free_sema);
852         }
853     }
854     qemu_co_mutex_unlock(&s->send_mutex);
855     return rc;
856 }
857 
858 static inline uint16_t payload_advance16(uint8_t **payload)
859 {
860     *payload += 2;
861     return lduw_be_p(*payload - 2);
862 }
863 
864 static inline uint32_t payload_advance32(uint8_t **payload)
865 {
866     *payload += 4;
867     return ldl_be_p(*payload - 4);
868 }
869 
870 static inline uint64_t payload_advance64(uint8_t **payload)
871 {
872     *payload += 8;
873     return ldq_be_p(*payload - 8);
874 }
875 
876 static int nbd_parse_offset_hole_payload(BDRVNBDState *s,
877                                          NBDStructuredReplyChunk *chunk,
878                                          uint8_t *payload, uint64_t orig_offset,
879                                          QEMUIOVector *qiov, Error **errp)
880 {
881     uint64_t offset;
882     uint32_t hole_size;
883 
884     if (chunk->length != sizeof(offset) + sizeof(hole_size)) {
885         error_setg(errp, "Protocol error: invalid payload for "
886                          "NBD_REPLY_TYPE_OFFSET_HOLE");
887         return -EINVAL;
888     }
889 
890     offset = payload_advance64(&payload);
891     hole_size = payload_advance32(&payload);
892 
893     if (!hole_size || offset < orig_offset || hole_size > qiov->size ||
894         offset > orig_offset + qiov->size - hole_size) {
895         error_setg(errp, "Protocol error: server sent chunk exceeding requested"
896                          " region");
897         return -EINVAL;
898     }
899     if (s->info.min_block &&
900         !QEMU_IS_ALIGNED(hole_size, s->info.min_block)) {
901         trace_nbd_structured_read_compliance("hole");
902     }
903 
904     qemu_iovec_memset(qiov, offset - orig_offset, 0, hole_size);
905 
906     return 0;
907 }
908 
909 /*
910  * nbd_parse_blockstatus_payload
911  * Based on our request, we expect only one extent in reply, for the
912  * base:allocation context.
913  */
914 static int nbd_parse_blockstatus_payload(BDRVNBDState *s,
915                                          NBDStructuredReplyChunk *chunk,
916                                          uint8_t *payload, uint64_t orig_length,
917                                          NBDExtent *extent, Error **errp)
918 {
919     uint32_t context_id;
920 
921     /* The server succeeded, so it must have sent [at least] one extent */
922     if (chunk->length < sizeof(context_id) + sizeof(*extent)) {
923         error_setg(errp, "Protocol error: invalid payload for "
924                          "NBD_REPLY_TYPE_BLOCK_STATUS");
925         return -EINVAL;
926     }
927 
928     context_id = payload_advance32(&payload);
929     if (s->info.context_id != context_id) {
930         error_setg(errp, "Protocol error: unexpected context id %d for "
931                          "NBD_REPLY_TYPE_BLOCK_STATUS, when negotiated context "
932                          "id is %d", context_id,
933                          s->info.context_id);
934         return -EINVAL;
935     }
936 
937     extent->length = payload_advance32(&payload);
938     extent->flags = payload_advance32(&payload);
939 
940     if (extent->length == 0) {
941         error_setg(errp, "Protocol error: server sent status chunk with "
942                    "zero length");
943         return -EINVAL;
944     }
945 
946     /*
947      * A server sending unaligned block status is in violation of the
948      * protocol, but as qemu-nbd 3.1 is such a server (at least for
949      * POSIX files that are not a multiple of 512 bytes, since qemu
950      * rounds files up to 512-byte multiples but lseek(SEEK_HOLE)
951      * still sees an implicit hole beyond the real EOF), it's nicer to
952      * work around the misbehaving server. If the request included
953      * more than the final unaligned block, truncate it back to an
954      * aligned result; if the request was only the final block, round
955      * up to the full block and change the status to fully-allocated
956      * (always a safe status, even if it loses information).
957      */
958     if (s->info.min_block && !QEMU_IS_ALIGNED(extent->length,
959                                                    s->info.min_block)) {
960         trace_nbd_parse_blockstatus_compliance("extent length is unaligned");
961         if (extent->length > s->info.min_block) {
962             extent->length = QEMU_ALIGN_DOWN(extent->length,
963                                              s->info.min_block);
964         } else {
965             extent->length = s->info.min_block;
966             extent->flags = 0;
967         }
968     }
969 
970     /*
971      * We used NBD_CMD_FLAG_REQ_ONE, so the server should not have
972      * sent us any more than one extent, nor should it have included
973      * status beyond our request in that extent. However, it's easy
974      * enough to ignore the server's noncompliance without killing the
975      * connection; just ignore trailing extents, and clamp things to
976      * the length of our request.
977      */
978     if (chunk->length > sizeof(context_id) + sizeof(*extent)) {
979         trace_nbd_parse_blockstatus_compliance("more than one extent");
980     }
981     if (extent->length > orig_length) {
982         extent->length = orig_length;
983         trace_nbd_parse_blockstatus_compliance("extent length too large");
984     }
985 
986     /*
987      * HACK: if we are using x-dirty-bitmaps to access
988      * qemu:allocation-depth, treat all depths > 2 the same as 2,
989      * since nbd_client_co_block_status is only expecting the low two
990      * bits to be set.
991      */
992     if (s->alloc_depth && extent->flags > 2) {
993         extent->flags = 2;
994     }
995 
996     return 0;
997 }
998 
999 /*
1000  * nbd_parse_error_payload
1001  * on success @errp contains message describing nbd error reply
1002  */
1003 static int nbd_parse_error_payload(NBDStructuredReplyChunk *chunk,
1004                                    uint8_t *payload, int *request_ret,
1005                                    Error **errp)
1006 {
1007     uint32_t error;
1008     uint16_t message_size;
1009 
1010     assert(chunk->type & (1 << 15));
1011 
1012     if (chunk->length < sizeof(error) + sizeof(message_size)) {
1013         error_setg(errp,
1014                    "Protocol error: invalid payload for structured error");
1015         return -EINVAL;
1016     }
1017 
1018     error = nbd_errno_to_system_errno(payload_advance32(&payload));
1019     if (error == 0) {
1020         error_setg(errp, "Protocol error: server sent structured error chunk "
1021                          "with error = 0");
1022         return -EINVAL;
1023     }
1024 
1025     *request_ret = -error;
1026     message_size = payload_advance16(&payload);
1027 
1028     if (message_size > chunk->length - sizeof(error) - sizeof(message_size)) {
1029         error_setg(errp, "Protocol error: server sent structured error chunk "
1030                          "with incorrect message size");
1031         return -EINVAL;
1032     }
1033 
1034     /* TODO: Add a trace point to mention the server complaint */
1035 
1036     /* TODO handle ERROR_OFFSET */
1037 
1038     return 0;
1039 }
1040 
1041 static int nbd_co_receive_offset_data_payload(BDRVNBDState *s,
1042                                               uint64_t orig_offset,
1043                                               QEMUIOVector *qiov, Error **errp)
1044 {
1045     QEMUIOVector sub_qiov;
1046     uint64_t offset;
1047     size_t data_size;
1048     int ret;
1049     NBDStructuredReplyChunk *chunk = &s->reply.structured;
1050 
1051     assert(nbd_reply_is_structured(&s->reply));
1052 
1053     /* The NBD spec requires at least one byte of payload */
1054     if (chunk->length <= sizeof(offset)) {
1055         error_setg(errp, "Protocol error: invalid payload for "
1056                          "NBD_REPLY_TYPE_OFFSET_DATA");
1057         return -EINVAL;
1058     }
1059 
1060     if (nbd_read64(s->ioc, &offset, "OFFSET_DATA offset", errp) < 0) {
1061         return -EIO;
1062     }
1063 
1064     data_size = chunk->length - sizeof(offset);
1065     assert(data_size);
1066     if (offset < orig_offset || data_size > qiov->size ||
1067         offset > orig_offset + qiov->size - data_size) {
1068         error_setg(errp, "Protocol error: server sent chunk exceeding requested"
1069                          " region");
1070         return -EINVAL;
1071     }
1072     if (s->info.min_block && !QEMU_IS_ALIGNED(data_size, s->info.min_block)) {
1073         trace_nbd_structured_read_compliance("data");
1074     }
1075 
1076     qemu_iovec_init(&sub_qiov, qiov->niov);
1077     qemu_iovec_concat(&sub_qiov, qiov, offset - orig_offset, data_size);
1078     ret = qio_channel_readv_all(s->ioc, sub_qiov.iov, sub_qiov.niov, errp);
1079     qemu_iovec_destroy(&sub_qiov);
1080 
1081     return ret < 0 ? -EIO : 0;
1082 }
1083 
1084 #define NBD_MAX_MALLOC_PAYLOAD 1000
1085 static coroutine_fn int nbd_co_receive_structured_payload(
1086         BDRVNBDState *s, void **payload, Error **errp)
1087 {
1088     int ret;
1089     uint32_t len;
1090 
1091     assert(nbd_reply_is_structured(&s->reply));
1092 
1093     len = s->reply.structured.length;
1094 
1095     if (len == 0) {
1096         return 0;
1097     }
1098 
1099     if (payload == NULL) {
1100         error_setg(errp, "Unexpected structured payload");
1101         return -EINVAL;
1102     }
1103 
1104     if (len > NBD_MAX_MALLOC_PAYLOAD) {
1105         error_setg(errp, "Payload too large");
1106         return -EINVAL;
1107     }
1108 
1109     *payload = g_new(char, len);
1110     ret = nbd_read(s->ioc, *payload, len, "structured payload", errp);
1111     if (ret < 0) {
1112         g_free(*payload);
1113         *payload = NULL;
1114         return ret;
1115     }
1116 
1117     return 0;
1118 }
1119 
1120 /*
1121  * nbd_co_do_receive_one_chunk
1122  * for simple reply:
1123  *   set request_ret to received reply error
1124  *   if qiov is not NULL: read payload to @qiov
1125  * for structured reply chunk:
1126  *   if error chunk: read payload, set @request_ret, do not set @payload
1127  *   else if offset_data chunk: read payload data to @qiov, do not set @payload
1128  *   else: read payload to @payload
1129  *
1130  * If function fails, @errp contains corresponding error message, and the
1131  * connection with the server is suspect.  If it returns 0, then the
1132  * transaction succeeded (although @request_ret may be a negative errno
1133  * corresponding to the server's error reply), and errp is unchanged.
1134  */
1135 static coroutine_fn int nbd_co_do_receive_one_chunk(
1136         BDRVNBDState *s, uint64_t handle, bool only_structured,
1137         int *request_ret, QEMUIOVector *qiov, void **payload, Error **errp)
1138 {
1139     int ret;
1140     int i = HANDLE_TO_INDEX(s, handle);
1141     void *local_payload = NULL;
1142     NBDStructuredReplyChunk *chunk;
1143 
1144     if (payload) {
1145         *payload = NULL;
1146     }
1147     *request_ret = 0;
1148 
1149     /* Wait until we're woken up by nbd_connection_entry.  */
1150     s->requests[i].receiving = true;
1151     qemu_coroutine_yield();
1152     s->requests[i].receiving = false;
1153     if (qatomic_load_acquire(&s->state) != NBD_CLIENT_CONNECTED) {
1154         error_setg(errp, "Connection closed");
1155         return -EIO;
1156     }
1157     assert(s->ioc);
1158 
1159     assert(s->reply.handle == handle);
1160 
1161     if (nbd_reply_is_simple(&s->reply)) {
1162         if (only_structured) {
1163             error_setg(errp, "Protocol error: simple reply when structured "
1164                              "reply chunk was expected");
1165             return -EINVAL;
1166         }
1167 
1168         *request_ret = -nbd_errno_to_system_errno(s->reply.simple.error);
1169         if (*request_ret < 0 || !qiov) {
1170             return 0;
1171         }
1172 
1173         return qio_channel_readv_all(s->ioc, qiov->iov, qiov->niov,
1174                                      errp) < 0 ? -EIO : 0;
1175     }
1176 
1177     /* handle structured reply chunk */
1178     assert(s->info.structured_reply);
1179     chunk = &s->reply.structured;
1180 
1181     if (chunk->type == NBD_REPLY_TYPE_NONE) {
1182         if (!(chunk->flags & NBD_REPLY_FLAG_DONE)) {
1183             error_setg(errp, "Protocol error: NBD_REPLY_TYPE_NONE chunk without"
1184                        " NBD_REPLY_FLAG_DONE flag set");
1185             return -EINVAL;
1186         }
1187         if (chunk->length) {
1188             error_setg(errp, "Protocol error: NBD_REPLY_TYPE_NONE chunk with"
1189                        " nonzero length");
1190             return -EINVAL;
1191         }
1192         return 0;
1193     }
1194 
1195     if (chunk->type == NBD_REPLY_TYPE_OFFSET_DATA) {
1196         if (!qiov) {
1197             error_setg(errp, "Unexpected NBD_REPLY_TYPE_OFFSET_DATA chunk");
1198             return -EINVAL;
1199         }
1200 
1201         return nbd_co_receive_offset_data_payload(s, s->requests[i].offset,
1202                                                   qiov, errp);
1203     }
1204 
1205     if (nbd_reply_type_is_error(chunk->type)) {
1206         payload = &local_payload;
1207     }
1208 
1209     ret = nbd_co_receive_structured_payload(s, payload, errp);
1210     if (ret < 0) {
1211         return ret;
1212     }
1213 
1214     if (nbd_reply_type_is_error(chunk->type)) {
1215         ret = nbd_parse_error_payload(chunk, local_payload, request_ret, errp);
1216         g_free(local_payload);
1217         return ret;
1218     }
1219 
1220     return 0;
1221 }
1222 
1223 /*
1224  * nbd_co_receive_one_chunk
1225  * Read reply, wake up connection_co and set s->quit if needed.
1226  * Return value is a fatal error code or normal nbd reply error code
1227  */
1228 static coroutine_fn int nbd_co_receive_one_chunk(
1229         BDRVNBDState *s, uint64_t handle, bool only_structured,
1230         int *request_ret, QEMUIOVector *qiov, NBDReply *reply, void **payload,
1231         Error **errp)
1232 {
1233     int ret = nbd_co_do_receive_one_chunk(s, handle, only_structured,
1234                                           request_ret, qiov, payload, errp);
1235 
1236     if (ret < 0) {
1237         memset(reply, 0, sizeof(*reply));
1238         nbd_channel_error(s, ret);
1239     } else {
1240         /* For assert at loop start in nbd_connection_entry */
1241         *reply = s->reply;
1242     }
1243     s->reply.handle = 0;
1244 
1245     if (s->connection_co && !s->wait_in_flight) {
1246         /*
1247          * We must check s->wait_in_flight, because we may entered by
1248          * nbd_recv_coroutines_wake_all(), in this case we should not
1249          * wake connection_co here, it will woken by last request.
1250          */
1251         aio_co_wake(s->connection_co);
1252     }
1253 
1254     return ret;
1255 }
1256 
1257 typedef struct NBDReplyChunkIter {
1258     int ret;
1259     int request_ret;
1260     Error *err;
1261     bool done, only_structured;
1262 } NBDReplyChunkIter;
1263 
1264 static void nbd_iter_channel_error(NBDReplyChunkIter *iter,
1265                                    int ret, Error **local_err)
1266 {
1267     assert(local_err && *local_err);
1268     assert(ret < 0);
1269 
1270     if (!iter->ret) {
1271         iter->ret = ret;
1272         error_propagate(&iter->err, *local_err);
1273     } else {
1274         error_free(*local_err);
1275     }
1276 
1277     *local_err = NULL;
1278 }
1279 
1280 static void nbd_iter_request_error(NBDReplyChunkIter *iter, int ret)
1281 {
1282     assert(ret < 0);
1283 
1284     if (!iter->request_ret) {
1285         iter->request_ret = ret;
1286     }
1287 }
1288 
1289 /*
1290  * NBD_FOREACH_REPLY_CHUNK
1291  * The pointer stored in @payload requires g_free() to free it.
1292  */
1293 #define NBD_FOREACH_REPLY_CHUNK(s, iter, handle, structured, \
1294                                 qiov, reply, payload) \
1295     for (iter = (NBDReplyChunkIter) { .only_structured = structured }; \
1296          nbd_reply_chunk_iter_receive(s, &iter, handle, qiov, reply, payload);)
1297 
1298 /*
1299  * nbd_reply_chunk_iter_receive
1300  * The pointer stored in @payload requires g_free() to free it.
1301  */
1302 static bool nbd_reply_chunk_iter_receive(BDRVNBDState *s,
1303                                          NBDReplyChunkIter *iter,
1304                                          uint64_t handle,
1305                                          QEMUIOVector *qiov, NBDReply *reply,
1306                                          void **payload)
1307 {
1308     int ret, request_ret;
1309     NBDReply local_reply;
1310     NBDStructuredReplyChunk *chunk;
1311     Error *local_err = NULL;
1312     if (qatomic_load_acquire(&s->state) != NBD_CLIENT_CONNECTED) {
1313         error_setg(&local_err, "Connection closed");
1314         nbd_iter_channel_error(iter, -EIO, &local_err);
1315         goto break_loop;
1316     }
1317 
1318     if (iter->done) {
1319         /* Previous iteration was last. */
1320         goto break_loop;
1321     }
1322 
1323     if (reply == NULL) {
1324         reply = &local_reply;
1325     }
1326 
1327     ret = nbd_co_receive_one_chunk(s, handle, iter->only_structured,
1328                                    &request_ret, qiov, reply, payload,
1329                                    &local_err);
1330     if (ret < 0) {
1331         nbd_iter_channel_error(iter, ret, &local_err);
1332     } else if (request_ret < 0) {
1333         nbd_iter_request_error(iter, request_ret);
1334     }
1335 
1336     /* Do not execute the body of NBD_FOREACH_REPLY_CHUNK for simple reply. */
1337     if (nbd_reply_is_simple(reply) ||
1338         qatomic_load_acquire(&s->state) != NBD_CLIENT_CONNECTED) {
1339         goto break_loop;
1340     }
1341 
1342     chunk = &reply->structured;
1343     iter->only_structured = true;
1344 
1345     if (chunk->type == NBD_REPLY_TYPE_NONE) {
1346         /* NBD_REPLY_FLAG_DONE is already checked in nbd_co_receive_one_chunk */
1347         assert(chunk->flags & NBD_REPLY_FLAG_DONE);
1348         goto break_loop;
1349     }
1350 
1351     if (chunk->flags & NBD_REPLY_FLAG_DONE) {
1352         /* This iteration is last. */
1353         iter->done = true;
1354     }
1355 
1356     /* Execute the loop body */
1357     return true;
1358 
1359 break_loop:
1360     s->requests[HANDLE_TO_INDEX(s, handle)].coroutine = NULL;
1361 
1362     qemu_co_mutex_lock(&s->send_mutex);
1363     s->in_flight--;
1364     if (s->in_flight == 0 && s->wait_in_flight) {
1365         aio_co_wake(s->connection_co);
1366     } else {
1367         qemu_co_queue_next(&s->free_sema);
1368     }
1369     qemu_co_mutex_unlock(&s->send_mutex);
1370 
1371     return false;
1372 }
1373 
1374 static int nbd_co_receive_return_code(BDRVNBDState *s, uint64_t handle,
1375                                       int *request_ret, Error **errp)
1376 {
1377     NBDReplyChunkIter iter;
1378 
1379     NBD_FOREACH_REPLY_CHUNK(s, iter, handle, false, NULL, NULL, NULL) {
1380         /* nbd_reply_chunk_iter_receive does all the work */
1381     }
1382 
1383     error_propagate(errp, iter.err);
1384     *request_ret = iter.request_ret;
1385     return iter.ret;
1386 }
1387 
1388 static int nbd_co_receive_cmdread_reply(BDRVNBDState *s, uint64_t handle,
1389                                         uint64_t offset, QEMUIOVector *qiov,
1390                                         int *request_ret, Error **errp)
1391 {
1392     NBDReplyChunkIter iter;
1393     NBDReply reply;
1394     void *payload = NULL;
1395     Error *local_err = NULL;
1396 
1397     NBD_FOREACH_REPLY_CHUNK(s, iter, handle, s->info.structured_reply,
1398                             qiov, &reply, &payload)
1399     {
1400         int ret;
1401         NBDStructuredReplyChunk *chunk = &reply.structured;
1402 
1403         assert(nbd_reply_is_structured(&reply));
1404 
1405         switch (chunk->type) {
1406         case NBD_REPLY_TYPE_OFFSET_DATA:
1407             /*
1408              * special cased in nbd_co_receive_one_chunk, data is already
1409              * in qiov
1410              */
1411             break;
1412         case NBD_REPLY_TYPE_OFFSET_HOLE:
1413             ret = nbd_parse_offset_hole_payload(s, &reply.structured, payload,
1414                                                 offset, qiov, &local_err);
1415             if (ret < 0) {
1416                 nbd_channel_error(s, ret);
1417                 nbd_iter_channel_error(&iter, ret, &local_err);
1418             }
1419             break;
1420         default:
1421             if (!nbd_reply_type_is_error(chunk->type)) {
1422                 /* not allowed reply type */
1423                 nbd_channel_error(s, -EINVAL);
1424                 error_setg(&local_err,
1425                            "Unexpected reply type: %d (%s) for CMD_READ",
1426                            chunk->type, nbd_reply_type_lookup(chunk->type));
1427                 nbd_iter_channel_error(&iter, -EINVAL, &local_err);
1428             }
1429         }
1430 
1431         g_free(payload);
1432         payload = NULL;
1433     }
1434 
1435     error_propagate(errp, iter.err);
1436     *request_ret = iter.request_ret;
1437     return iter.ret;
1438 }
1439 
1440 static int nbd_co_receive_blockstatus_reply(BDRVNBDState *s,
1441                                             uint64_t handle, uint64_t length,
1442                                             NBDExtent *extent,
1443                                             int *request_ret, Error **errp)
1444 {
1445     NBDReplyChunkIter iter;
1446     NBDReply reply;
1447     void *payload = NULL;
1448     Error *local_err = NULL;
1449     bool received = false;
1450 
1451     assert(!extent->length);
1452     NBD_FOREACH_REPLY_CHUNK(s, iter, handle, false, NULL, &reply, &payload) {
1453         int ret;
1454         NBDStructuredReplyChunk *chunk = &reply.structured;
1455 
1456         assert(nbd_reply_is_structured(&reply));
1457 
1458         switch (chunk->type) {
1459         case NBD_REPLY_TYPE_BLOCK_STATUS:
1460             if (received) {
1461                 nbd_channel_error(s, -EINVAL);
1462                 error_setg(&local_err, "Several BLOCK_STATUS chunks in reply");
1463                 nbd_iter_channel_error(&iter, -EINVAL, &local_err);
1464             }
1465             received = true;
1466 
1467             ret = nbd_parse_blockstatus_payload(s, &reply.structured,
1468                                                 payload, length, extent,
1469                                                 &local_err);
1470             if (ret < 0) {
1471                 nbd_channel_error(s, ret);
1472                 nbd_iter_channel_error(&iter, ret, &local_err);
1473             }
1474             break;
1475         default:
1476             if (!nbd_reply_type_is_error(chunk->type)) {
1477                 nbd_channel_error(s, -EINVAL);
1478                 error_setg(&local_err,
1479                            "Unexpected reply type: %d (%s) "
1480                            "for CMD_BLOCK_STATUS",
1481                            chunk->type, nbd_reply_type_lookup(chunk->type));
1482                 nbd_iter_channel_error(&iter, -EINVAL, &local_err);
1483             }
1484         }
1485 
1486         g_free(payload);
1487         payload = NULL;
1488     }
1489 
1490     if (!extent->length && !iter.request_ret) {
1491         error_setg(&local_err, "Server did not reply with any status extents");
1492         nbd_iter_channel_error(&iter, -EIO, &local_err);
1493     }
1494 
1495     error_propagate(errp, iter.err);
1496     *request_ret = iter.request_ret;
1497     return iter.ret;
1498 }
1499 
1500 static int nbd_co_request(BlockDriverState *bs, NBDRequest *request,
1501                           QEMUIOVector *write_qiov)
1502 {
1503     int ret, request_ret;
1504     Error *local_err = NULL;
1505     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
1506 
1507     assert(request->type != NBD_CMD_READ);
1508     if (write_qiov) {
1509         assert(request->type == NBD_CMD_WRITE);
1510         assert(request->len == iov_size(write_qiov->iov, write_qiov->niov));
1511     } else {
1512         assert(request->type != NBD_CMD_WRITE);
1513     }
1514 
1515     do {
1516         ret = nbd_co_send_request(bs, request, write_qiov);
1517         if (ret < 0) {
1518             continue;
1519         }
1520 
1521         ret = nbd_co_receive_return_code(s, request->handle,
1522                                          &request_ret, &local_err);
1523         if (local_err) {
1524             trace_nbd_co_request_fail(request->from, request->len,
1525                                       request->handle, request->flags,
1526                                       request->type,
1527                                       nbd_cmd_lookup(request->type),
1528                                       ret, error_get_pretty(local_err));
1529             error_free(local_err);
1530             local_err = NULL;
1531         }
1532     } while (ret < 0 && nbd_client_connecting_wait(s));
1533 
1534     return ret ? ret : request_ret;
1535 }
1536 
1537 static int nbd_client_co_preadv(BlockDriverState *bs, uint64_t offset,
1538                                 uint64_t bytes, QEMUIOVector *qiov, int flags)
1539 {
1540     int ret, request_ret;
1541     Error *local_err = NULL;
1542     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
1543     NBDRequest request = {
1544         .type = NBD_CMD_READ,
1545         .from = offset,
1546         .len = bytes,
1547     };
1548 
1549     assert(bytes <= NBD_MAX_BUFFER_SIZE);
1550     assert(!flags);
1551 
1552     if (!bytes) {
1553         return 0;
1554     }
1555     /*
1556      * Work around the fact that the block layer doesn't do
1557      * byte-accurate sizing yet - if the read exceeds the server's
1558      * advertised size because the block layer rounded size up, then
1559      * truncate the request to the server and tail-pad with zero.
1560      */
1561     if (offset >= s->info.size) {
1562         assert(bytes < BDRV_SECTOR_SIZE);
1563         qemu_iovec_memset(qiov, 0, 0, bytes);
1564         return 0;
1565     }
1566     if (offset + bytes > s->info.size) {
1567         uint64_t slop = offset + bytes - s->info.size;
1568 
1569         assert(slop < BDRV_SECTOR_SIZE);
1570         qemu_iovec_memset(qiov, bytes - slop, 0, slop);
1571         request.len -= slop;
1572     }
1573 
1574     do {
1575         ret = nbd_co_send_request(bs, &request, NULL);
1576         if (ret < 0) {
1577             continue;
1578         }
1579 
1580         ret = nbd_co_receive_cmdread_reply(s, request.handle, offset, qiov,
1581                                            &request_ret, &local_err);
1582         if (local_err) {
1583             trace_nbd_co_request_fail(request.from, request.len, request.handle,
1584                                       request.flags, request.type,
1585                                       nbd_cmd_lookup(request.type),
1586                                       ret, error_get_pretty(local_err));
1587             error_free(local_err);
1588             local_err = NULL;
1589         }
1590     } while (ret < 0 && nbd_client_connecting_wait(s));
1591 
1592     return ret ? ret : request_ret;
1593 }
1594 
1595 static int nbd_client_co_pwritev(BlockDriverState *bs, uint64_t offset,
1596                                  uint64_t bytes, QEMUIOVector *qiov, int flags)
1597 {
1598     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
1599     NBDRequest request = {
1600         .type = NBD_CMD_WRITE,
1601         .from = offset,
1602         .len = bytes,
1603     };
1604 
1605     assert(!(s->info.flags & NBD_FLAG_READ_ONLY));
1606     if (flags & BDRV_REQ_FUA) {
1607         assert(s->info.flags & NBD_FLAG_SEND_FUA);
1608         request.flags |= NBD_CMD_FLAG_FUA;
1609     }
1610 
1611     assert(bytes <= NBD_MAX_BUFFER_SIZE);
1612 
1613     if (!bytes) {
1614         return 0;
1615     }
1616     return nbd_co_request(bs, &request, qiov);
1617 }
1618 
1619 static int nbd_client_co_pwrite_zeroes(BlockDriverState *bs, int64_t offset,
1620                                        int bytes, BdrvRequestFlags flags)
1621 {
1622     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
1623     NBDRequest request = {
1624         .type = NBD_CMD_WRITE_ZEROES,
1625         .from = offset,
1626         .len = bytes,
1627     };
1628 
1629     assert(!(s->info.flags & NBD_FLAG_READ_ONLY));
1630     if (!(s->info.flags & NBD_FLAG_SEND_WRITE_ZEROES)) {
1631         return -ENOTSUP;
1632     }
1633 
1634     if (flags & BDRV_REQ_FUA) {
1635         assert(s->info.flags & NBD_FLAG_SEND_FUA);
1636         request.flags |= NBD_CMD_FLAG_FUA;
1637     }
1638     if (!(flags & BDRV_REQ_MAY_UNMAP)) {
1639         request.flags |= NBD_CMD_FLAG_NO_HOLE;
1640     }
1641     if (flags & BDRV_REQ_NO_FALLBACK) {
1642         assert(s->info.flags & NBD_FLAG_SEND_FAST_ZERO);
1643         request.flags |= NBD_CMD_FLAG_FAST_ZERO;
1644     }
1645 
1646     if (!bytes) {
1647         return 0;
1648     }
1649     return nbd_co_request(bs, &request, NULL);
1650 }
1651 
1652 static int nbd_client_co_flush(BlockDriverState *bs)
1653 {
1654     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
1655     NBDRequest request = { .type = NBD_CMD_FLUSH };
1656 
1657     if (!(s->info.flags & NBD_FLAG_SEND_FLUSH)) {
1658         return 0;
1659     }
1660 
1661     request.from = 0;
1662     request.len = 0;
1663 
1664     return nbd_co_request(bs, &request, NULL);
1665 }
1666 
1667 static int nbd_client_co_pdiscard(BlockDriverState *bs, int64_t offset,
1668                                   int bytes)
1669 {
1670     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
1671     NBDRequest request = {
1672         .type = NBD_CMD_TRIM,
1673         .from = offset,
1674         .len = bytes,
1675     };
1676 
1677     assert(!(s->info.flags & NBD_FLAG_READ_ONLY));
1678     if (!(s->info.flags & NBD_FLAG_SEND_TRIM) || !bytes) {
1679         return 0;
1680     }
1681 
1682     return nbd_co_request(bs, &request, NULL);
1683 }
1684 
1685 static int coroutine_fn nbd_client_co_block_status(
1686         BlockDriverState *bs, bool want_zero, int64_t offset, int64_t bytes,
1687         int64_t *pnum, int64_t *map, BlockDriverState **file)
1688 {
1689     int ret, request_ret;
1690     NBDExtent extent = { 0 };
1691     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
1692     Error *local_err = NULL;
1693 
1694     NBDRequest request = {
1695         .type = NBD_CMD_BLOCK_STATUS,
1696         .from = offset,
1697         .len = MIN(QEMU_ALIGN_DOWN(INT_MAX, bs->bl.request_alignment),
1698                    MIN(bytes, s->info.size - offset)),
1699         .flags = NBD_CMD_FLAG_REQ_ONE,
1700     };
1701 
1702     if (!s->info.base_allocation) {
1703         *pnum = bytes;
1704         *map = offset;
1705         *file = bs;
1706         return BDRV_BLOCK_DATA | BDRV_BLOCK_OFFSET_VALID;
1707     }
1708 
1709     /*
1710      * Work around the fact that the block layer doesn't do
1711      * byte-accurate sizing yet - if the status request exceeds the
1712      * server's advertised size because the block layer rounded size
1713      * up, we truncated the request to the server (above), or are
1714      * called on just the hole.
1715      */
1716     if (offset >= s->info.size) {
1717         *pnum = bytes;
1718         assert(bytes < BDRV_SECTOR_SIZE);
1719         /* Intentionally don't report offset_valid for the hole */
1720         return BDRV_BLOCK_ZERO;
1721     }
1722 
1723     if (s->info.min_block) {
1724         assert(QEMU_IS_ALIGNED(request.len, s->info.min_block));
1725     }
1726     do {
1727         ret = nbd_co_send_request(bs, &request, NULL);
1728         if (ret < 0) {
1729             continue;
1730         }
1731 
1732         ret = nbd_co_receive_blockstatus_reply(s, request.handle, bytes,
1733                                                &extent, &request_ret,
1734                                                &local_err);
1735         if (local_err) {
1736             trace_nbd_co_request_fail(request.from, request.len, request.handle,
1737                                       request.flags, request.type,
1738                                       nbd_cmd_lookup(request.type),
1739                                       ret, error_get_pretty(local_err));
1740             error_free(local_err);
1741             local_err = NULL;
1742         }
1743     } while (ret < 0 && nbd_client_connecting_wait(s));
1744 
1745     if (ret < 0 || request_ret < 0) {
1746         return ret ? ret : request_ret;
1747     }
1748 
1749     assert(extent.length);
1750     *pnum = extent.length;
1751     *map = offset;
1752     *file = bs;
1753     return (extent.flags & NBD_STATE_HOLE ? 0 : BDRV_BLOCK_DATA) |
1754         (extent.flags & NBD_STATE_ZERO ? BDRV_BLOCK_ZERO : 0) |
1755         BDRV_BLOCK_OFFSET_VALID;
1756 }
1757 
1758 static int nbd_client_reopen_prepare(BDRVReopenState *state,
1759                                      BlockReopenQueue *queue, Error **errp)
1760 {
1761     BDRVNBDState *s = (BDRVNBDState *)state->bs->opaque;
1762 
1763     if ((state->flags & BDRV_O_RDWR) && (s->info.flags & NBD_FLAG_READ_ONLY)) {
1764         error_setg(errp, "Can't reopen read-only NBD mount as read/write");
1765         return -EACCES;
1766     }
1767     return 0;
1768 }
1769 
1770 static void nbd_yank(void *opaque)
1771 {
1772     BlockDriverState *bs = opaque;
1773     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
1774 
1775     qatomic_store_release(&s->state, NBD_CLIENT_QUIT);
1776     qio_channel_shutdown(QIO_CHANNEL(s->sioc), QIO_CHANNEL_SHUTDOWN_BOTH, NULL);
1777 }
1778 
1779 static void nbd_client_close(BlockDriverState *bs)
1780 {
1781     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
1782     NBDRequest request = { .type = NBD_CMD_DISC };
1783 
1784     if (s->ioc) {
1785         nbd_send_request(s->ioc, &request);
1786     }
1787 
1788     nbd_teardown_connection(bs);
1789 }
1790 
1791 static int nbd_establish_connection(BlockDriverState *bs,
1792                                     SocketAddress *saddr,
1793                                     Error **errp)
1794 {
1795     ERRP_GUARD();
1796     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
1797 
1798     s->sioc = qio_channel_socket_new();
1799     qio_channel_set_name(QIO_CHANNEL(s->sioc), "nbd-client");
1800 
1801     qio_channel_socket_connect_sync(s->sioc, saddr, errp);
1802     if (*errp) {
1803         object_unref(OBJECT(s->sioc));
1804         s->sioc = NULL;
1805         return -1;
1806     }
1807 
1808     yank_register_function(BLOCKDEV_YANK_INSTANCE(bs->node_name), nbd_yank, bs);
1809     qio_channel_set_delay(QIO_CHANNEL(s->sioc), false);
1810 
1811     return 0;
1812 }
1813 
1814 /* nbd_client_handshake takes ownership on s->sioc. On failure it's unref'ed. */
1815 static int nbd_client_handshake(BlockDriverState *bs, Error **errp)
1816 {
1817     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
1818     AioContext *aio_context = bdrv_get_aio_context(bs);
1819     int ret;
1820 
1821     trace_nbd_client_handshake(s->export);
1822     qio_channel_set_blocking(QIO_CHANNEL(s->sioc), false, NULL);
1823     qio_channel_attach_aio_context(QIO_CHANNEL(s->sioc), aio_context);
1824 
1825     s->info.request_sizes = true;
1826     s->info.structured_reply = true;
1827     s->info.base_allocation = true;
1828     s->info.x_dirty_bitmap = g_strdup(s->x_dirty_bitmap);
1829     s->info.name = g_strdup(s->export ?: "");
1830     ret = nbd_receive_negotiate(aio_context, QIO_CHANNEL(s->sioc), s->tlscreds,
1831                                 s->hostname, &s->ioc, &s->info, errp);
1832     g_free(s->info.x_dirty_bitmap);
1833     g_free(s->info.name);
1834     if (ret < 0) {
1835         yank_unregister_function(BLOCKDEV_YANK_INSTANCE(bs->node_name),
1836                                  nbd_yank, bs);
1837         object_unref(OBJECT(s->sioc));
1838         s->sioc = NULL;
1839         return ret;
1840     }
1841     if (s->x_dirty_bitmap) {
1842         if (!s->info.base_allocation) {
1843             error_setg(errp, "requested x-dirty-bitmap %s not found",
1844                        s->x_dirty_bitmap);
1845             ret = -EINVAL;
1846             goto fail;
1847         }
1848         if (strcmp(s->x_dirty_bitmap, "qemu:allocation-depth") == 0) {
1849             s->alloc_depth = true;
1850         }
1851     }
1852     if (s->info.flags & NBD_FLAG_READ_ONLY) {
1853         ret = bdrv_apply_auto_read_only(bs, "NBD export is read-only", errp);
1854         if (ret < 0) {
1855             goto fail;
1856         }
1857     }
1858     if (s->info.flags & NBD_FLAG_SEND_FUA) {
1859         bs->supported_write_flags = BDRV_REQ_FUA;
1860         bs->supported_zero_flags |= BDRV_REQ_FUA;
1861     }
1862     if (s->info.flags & NBD_FLAG_SEND_WRITE_ZEROES) {
1863         bs->supported_zero_flags |= BDRV_REQ_MAY_UNMAP;
1864         if (s->info.flags & NBD_FLAG_SEND_FAST_ZERO) {
1865             bs->supported_zero_flags |= BDRV_REQ_NO_FALLBACK;
1866         }
1867     }
1868 
1869     if (!s->ioc) {
1870         s->ioc = QIO_CHANNEL(s->sioc);
1871         object_ref(OBJECT(s->ioc));
1872     }
1873 
1874     trace_nbd_client_handshake_success(s->export);
1875 
1876     return 0;
1877 
1878  fail:
1879     /*
1880      * We have connected, but must fail for other reasons.
1881      * Send NBD_CMD_DISC as a courtesy to the server.
1882      */
1883     {
1884         NBDRequest request = { .type = NBD_CMD_DISC };
1885 
1886         nbd_send_request(s->ioc ?: QIO_CHANNEL(s->sioc), &request);
1887 
1888         yank_unregister_function(BLOCKDEV_YANK_INSTANCE(bs->node_name),
1889                                  nbd_yank, bs);
1890         object_unref(OBJECT(s->sioc));
1891         s->sioc = NULL;
1892 
1893         return ret;
1894     }
1895 }
1896 
1897 /*
1898  * Parse nbd_open options
1899  */
1900 
1901 static int nbd_parse_uri(const char *filename, QDict *options)
1902 {
1903     URI *uri;
1904     const char *p;
1905     QueryParams *qp = NULL;
1906     int ret = 0;
1907     bool is_unix;
1908 
1909     uri = uri_parse(filename);
1910     if (!uri) {
1911         return -EINVAL;
1912     }
1913 
1914     /* transport */
1915     if (!g_strcmp0(uri->scheme, "nbd")) {
1916         is_unix = false;
1917     } else if (!g_strcmp0(uri->scheme, "nbd+tcp")) {
1918         is_unix = false;
1919     } else if (!g_strcmp0(uri->scheme, "nbd+unix")) {
1920         is_unix = true;
1921     } else {
1922         ret = -EINVAL;
1923         goto out;
1924     }
1925 
1926     p = uri->path ? uri->path : "";
1927     if (p[0] == '/') {
1928         p++;
1929     }
1930     if (p[0]) {
1931         qdict_put_str(options, "export", p);
1932     }
1933 
1934     qp = query_params_parse(uri->query);
1935     if (qp->n > 1 || (is_unix && !qp->n) || (!is_unix && qp->n)) {
1936         ret = -EINVAL;
1937         goto out;
1938     }
1939 
1940     if (is_unix) {
1941         /* nbd+unix:///export?socket=path */
1942         if (uri->server || uri->port || strcmp(qp->p[0].name, "socket")) {
1943             ret = -EINVAL;
1944             goto out;
1945         }
1946         qdict_put_str(options, "server.type", "unix");
1947         qdict_put_str(options, "server.path", qp->p[0].value);
1948     } else {
1949         QString *host;
1950         char *port_str;
1951 
1952         /* nbd[+tcp]://host[:port]/export */
1953         if (!uri->server) {
1954             ret = -EINVAL;
1955             goto out;
1956         }
1957 
1958         /* strip braces from literal IPv6 address */
1959         if (uri->server[0] == '[') {
1960             host = qstring_from_substr(uri->server, 1,
1961                                        strlen(uri->server) - 1);
1962         } else {
1963             host = qstring_from_str(uri->server);
1964         }
1965 
1966         qdict_put_str(options, "server.type", "inet");
1967         qdict_put(options, "server.host", host);
1968 
1969         port_str = g_strdup_printf("%d", uri->port ?: NBD_DEFAULT_PORT);
1970         qdict_put_str(options, "server.port", port_str);
1971         g_free(port_str);
1972     }
1973 
1974 out:
1975     if (qp) {
1976         query_params_free(qp);
1977     }
1978     uri_free(uri);
1979     return ret;
1980 }
1981 
1982 static bool nbd_has_filename_options_conflict(QDict *options, Error **errp)
1983 {
1984     const QDictEntry *e;
1985 
1986     for (e = qdict_first(options); e; e = qdict_next(options, e)) {
1987         if (!strcmp(e->key, "host") ||
1988             !strcmp(e->key, "port") ||
1989             !strcmp(e->key, "path") ||
1990             !strcmp(e->key, "export") ||
1991             strstart(e->key, "server.", NULL))
1992         {
1993             error_setg(errp, "Option '%s' cannot be used with a file name",
1994                        e->key);
1995             return true;
1996         }
1997     }
1998 
1999     return false;
2000 }
2001 
2002 static void nbd_parse_filename(const char *filename, QDict *options,
2003                                Error **errp)
2004 {
2005     g_autofree char *file = NULL;
2006     char *export_name;
2007     const char *host_spec;
2008     const char *unixpath;
2009 
2010     if (nbd_has_filename_options_conflict(options, errp)) {
2011         return;
2012     }
2013 
2014     if (strstr(filename, "://")) {
2015         int ret = nbd_parse_uri(filename, options);
2016         if (ret < 0) {
2017             error_setg(errp, "No valid URL specified");
2018         }
2019         return;
2020     }
2021 
2022     file = g_strdup(filename);
2023 
2024     export_name = strstr(file, EN_OPTSTR);
2025     if (export_name) {
2026         if (export_name[strlen(EN_OPTSTR)] == 0) {
2027             return;
2028         }
2029         export_name[0] = 0; /* truncate 'file' */
2030         export_name += strlen(EN_OPTSTR);
2031 
2032         qdict_put_str(options, "export", export_name);
2033     }
2034 
2035     /* extract the host_spec - fail if it's not nbd:... */
2036     if (!strstart(file, "nbd:", &host_spec)) {
2037         error_setg(errp, "File name string for NBD must start with 'nbd:'");
2038         return;
2039     }
2040 
2041     if (!*host_spec) {
2042         return;
2043     }
2044 
2045     /* are we a UNIX or TCP socket? */
2046     if (strstart(host_spec, "unix:", &unixpath)) {
2047         qdict_put_str(options, "server.type", "unix");
2048         qdict_put_str(options, "server.path", unixpath);
2049     } else {
2050         InetSocketAddress *addr = g_new(InetSocketAddress, 1);
2051 
2052         if (inet_parse(addr, host_spec, errp)) {
2053             goto out_inet;
2054         }
2055 
2056         qdict_put_str(options, "server.type", "inet");
2057         qdict_put_str(options, "server.host", addr->host);
2058         qdict_put_str(options, "server.port", addr->port);
2059     out_inet:
2060         qapi_free_InetSocketAddress(addr);
2061     }
2062 }
2063 
2064 static bool nbd_process_legacy_socket_options(QDict *output_options,
2065                                               QemuOpts *legacy_opts,
2066                                               Error **errp)
2067 {
2068     const char *path = qemu_opt_get(legacy_opts, "path");
2069     const char *host = qemu_opt_get(legacy_opts, "host");
2070     const char *port = qemu_opt_get(legacy_opts, "port");
2071     const QDictEntry *e;
2072 
2073     if (!path && !host && !port) {
2074         return true;
2075     }
2076 
2077     for (e = qdict_first(output_options); e; e = qdict_next(output_options, e))
2078     {
2079         if (strstart(e->key, "server.", NULL)) {
2080             error_setg(errp, "Cannot use 'server' and path/host/port at the "
2081                        "same time");
2082             return false;
2083         }
2084     }
2085 
2086     if (path && host) {
2087         error_setg(errp, "path and host may not be used at the same time");
2088         return false;
2089     } else if (path) {
2090         if (port) {
2091             error_setg(errp, "port may not be used without host");
2092             return false;
2093         }
2094 
2095         qdict_put_str(output_options, "server.type", "unix");
2096         qdict_put_str(output_options, "server.path", path);
2097     } else if (host) {
2098         qdict_put_str(output_options, "server.type", "inet");
2099         qdict_put_str(output_options, "server.host", host);
2100         qdict_put_str(output_options, "server.port",
2101                       port ?: stringify(NBD_DEFAULT_PORT));
2102     }
2103 
2104     return true;
2105 }
2106 
2107 static SocketAddress *nbd_config(BDRVNBDState *s, QDict *options,
2108                                  Error **errp)
2109 {
2110     SocketAddress *saddr = NULL;
2111     QDict *addr = NULL;
2112     Visitor *iv = NULL;
2113 
2114     qdict_extract_subqdict(options, &addr, "server.");
2115     if (!qdict_size(addr)) {
2116         error_setg(errp, "NBD server address missing");
2117         goto done;
2118     }
2119 
2120     iv = qobject_input_visitor_new_flat_confused(addr, errp);
2121     if (!iv) {
2122         goto done;
2123     }
2124 
2125     if (!visit_type_SocketAddress(iv, NULL, &saddr, errp)) {
2126         goto done;
2127     }
2128 
2129 done:
2130     qobject_unref(addr);
2131     visit_free(iv);
2132     return saddr;
2133 }
2134 
2135 static QCryptoTLSCreds *nbd_get_tls_creds(const char *id, Error **errp)
2136 {
2137     Object *obj;
2138     QCryptoTLSCreds *creds;
2139 
2140     obj = object_resolve_path_component(
2141         object_get_objects_root(), id);
2142     if (!obj) {
2143         error_setg(errp, "No TLS credentials with id '%s'",
2144                    id);
2145         return NULL;
2146     }
2147     creds = (QCryptoTLSCreds *)
2148         object_dynamic_cast(obj, TYPE_QCRYPTO_TLS_CREDS);
2149     if (!creds) {
2150         error_setg(errp, "Object with id '%s' is not TLS credentials",
2151                    id);
2152         return NULL;
2153     }
2154 
2155     if (creds->endpoint != QCRYPTO_TLS_CREDS_ENDPOINT_CLIENT) {
2156         error_setg(errp,
2157                    "Expecting TLS credentials with a client endpoint");
2158         return NULL;
2159     }
2160     object_ref(obj);
2161     return creds;
2162 }
2163 
2164 
2165 static QemuOptsList nbd_runtime_opts = {
2166     .name = "nbd",
2167     .head = QTAILQ_HEAD_INITIALIZER(nbd_runtime_opts.head),
2168     .desc = {
2169         {
2170             .name = "host",
2171             .type = QEMU_OPT_STRING,
2172             .help = "TCP host to connect to",
2173         },
2174         {
2175             .name = "port",
2176             .type = QEMU_OPT_STRING,
2177             .help = "TCP port to connect to",
2178         },
2179         {
2180             .name = "path",
2181             .type = QEMU_OPT_STRING,
2182             .help = "Unix socket path to connect to",
2183         },
2184         {
2185             .name = "export",
2186             .type = QEMU_OPT_STRING,
2187             .help = "Name of the NBD export to open",
2188         },
2189         {
2190             .name = "tls-creds",
2191             .type = QEMU_OPT_STRING,
2192             .help = "ID of the TLS credentials to use",
2193         },
2194         {
2195             .name = "x-dirty-bitmap",
2196             .type = QEMU_OPT_STRING,
2197             .help = "experimental: expose named dirty bitmap in place of "
2198                     "block status",
2199         },
2200         {
2201             .name = "reconnect-delay",
2202             .type = QEMU_OPT_NUMBER,
2203             .help = "On an unexpected disconnect, the nbd client tries to "
2204                     "connect again until succeeding or encountering a serious "
2205                     "error.  During the first @reconnect-delay seconds, all "
2206                     "requests are paused and will be rerun on a successful "
2207                     "reconnect. After that time, any delayed requests and all "
2208                     "future requests before a successful reconnect will "
2209                     "immediately fail. Default 0",
2210         },
2211         { /* end of list */ }
2212     },
2213 };
2214 
2215 static int nbd_process_options(BlockDriverState *bs, QDict *options,
2216                                Error **errp)
2217 {
2218     BDRVNBDState *s = bs->opaque;
2219     QemuOpts *opts;
2220     int ret = -EINVAL;
2221 
2222     opts = qemu_opts_create(&nbd_runtime_opts, NULL, 0, &error_abort);
2223     if (!qemu_opts_absorb_qdict(opts, options, errp)) {
2224         goto error;
2225     }
2226 
2227     /* Translate @host, @port, and @path to a SocketAddress */
2228     if (!nbd_process_legacy_socket_options(options, opts, errp)) {
2229         goto error;
2230     }
2231 
2232     /* Pop the config into our state object. Exit if invalid. */
2233     s->saddr = nbd_config(s, options, errp);
2234     if (!s->saddr) {
2235         goto error;
2236     }
2237 
2238     s->export = g_strdup(qemu_opt_get(opts, "export"));
2239     if (s->export && strlen(s->export) > NBD_MAX_STRING_SIZE) {
2240         error_setg(errp, "export name too long to send to server");
2241         goto error;
2242     }
2243 
2244     s->tlscredsid = g_strdup(qemu_opt_get(opts, "tls-creds"));
2245     if (s->tlscredsid) {
2246         s->tlscreds = nbd_get_tls_creds(s->tlscredsid, errp);
2247         if (!s->tlscreds) {
2248             goto error;
2249         }
2250 
2251         /* TODO SOCKET_ADDRESS_KIND_FD where fd has AF_INET or AF_INET6 */
2252         if (s->saddr->type != SOCKET_ADDRESS_TYPE_INET) {
2253             error_setg(errp, "TLS only supported over IP sockets");
2254             goto error;
2255         }
2256         s->hostname = s->saddr->u.inet.host;
2257     }
2258 
2259     s->x_dirty_bitmap = g_strdup(qemu_opt_get(opts, "x-dirty-bitmap"));
2260     if (s->x_dirty_bitmap && strlen(s->x_dirty_bitmap) > NBD_MAX_STRING_SIZE) {
2261         error_setg(errp, "x-dirty-bitmap query too long to send to server");
2262         goto error;
2263     }
2264 
2265     s->reconnect_delay = qemu_opt_get_number(opts, "reconnect-delay", 0);
2266 
2267     ret = 0;
2268 
2269  error:
2270     if (ret < 0) {
2271         nbd_clear_bdrvstate(s);
2272     }
2273     qemu_opts_del(opts);
2274     return ret;
2275 }
2276 
2277 static int nbd_open(BlockDriverState *bs, QDict *options, int flags,
2278                     Error **errp)
2279 {
2280     int ret;
2281     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
2282 
2283     ret = nbd_process_options(bs, options, errp);
2284     if (ret < 0) {
2285         return ret;
2286     }
2287 
2288     s->bs = bs;
2289     qemu_co_mutex_init(&s->send_mutex);
2290     qemu_co_queue_init(&s->free_sema);
2291 
2292     if (!yank_register_instance(BLOCKDEV_YANK_INSTANCE(bs->node_name), errp)) {
2293         return -EEXIST;
2294     }
2295 
2296     /*
2297      * establish TCP connection, return error if it fails
2298      * TODO: Configurable retry-until-timeout behaviour.
2299      */
2300     if (nbd_establish_connection(bs, s->saddr, errp) < 0) {
2301         yank_unregister_instance(BLOCKDEV_YANK_INSTANCE(bs->node_name));
2302         return -ECONNREFUSED;
2303     }
2304 
2305     ret = nbd_client_handshake(bs, errp);
2306     if (ret < 0) {
2307         yank_unregister_instance(BLOCKDEV_YANK_INSTANCE(bs->node_name));
2308         nbd_clear_bdrvstate(s);
2309         return ret;
2310     }
2311     /* successfully connected */
2312     s->state = NBD_CLIENT_CONNECTED;
2313 
2314     nbd_init_connect_thread(s);
2315 
2316     s->connection_co = qemu_coroutine_create(nbd_connection_entry, s);
2317     bdrv_inc_in_flight(bs);
2318     aio_co_schedule(bdrv_get_aio_context(bs), s->connection_co);
2319 
2320     return 0;
2321 }
2322 
2323 static int nbd_co_flush(BlockDriverState *bs)
2324 {
2325     return nbd_client_co_flush(bs);
2326 }
2327 
2328 static void nbd_refresh_limits(BlockDriverState *bs, Error **errp)
2329 {
2330     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
2331     uint32_t min = s->info.min_block;
2332     uint32_t max = MIN_NON_ZERO(NBD_MAX_BUFFER_SIZE, s->info.max_block);
2333 
2334     /*
2335      * If the server did not advertise an alignment:
2336      * - a size that is not sector-aligned implies that an alignment
2337      *   of 1 can be used to access those tail bytes
2338      * - advertisement of block status requires an alignment of 1, so
2339      *   that we don't violate block layer constraints that block
2340      *   status is always aligned (as we can't control whether the
2341      *   server will report sub-sector extents, such as a hole at EOF
2342      *   on an unaligned POSIX file)
2343      * - otherwise, assume the server is so old that we are safer avoiding
2344      *   sub-sector requests
2345      */
2346     if (!min) {
2347         min = (!QEMU_IS_ALIGNED(s->info.size, BDRV_SECTOR_SIZE) ||
2348                s->info.base_allocation) ? 1 : BDRV_SECTOR_SIZE;
2349     }
2350 
2351     bs->bl.request_alignment = min;
2352     bs->bl.max_pdiscard = QEMU_ALIGN_DOWN(INT_MAX, min);
2353     bs->bl.max_pwrite_zeroes = max;
2354     bs->bl.max_transfer = max;
2355 
2356     if (s->info.opt_block &&
2357         s->info.opt_block > bs->bl.opt_transfer) {
2358         bs->bl.opt_transfer = s->info.opt_block;
2359     }
2360 }
2361 
2362 static void nbd_close(BlockDriverState *bs)
2363 {
2364     BDRVNBDState *s = bs->opaque;
2365 
2366     nbd_client_close(bs);
2367     yank_unregister_instance(BLOCKDEV_YANK_INSTANCE(bs->node_name));
2368     nbd_clear_bdrvstate(s);
2369 }
2370 
2371 /*
2372  * NBD cannot truncate, but if the caller asks to truncate to the same size, or
2373  * to a smaller size with exact=false, there is no reason to fail the
2374  * operation.
2375  *
2376  * Preallocation mode is ignored since it does not seems useful to fail when
2377  * we never change anything.
2378  */
2379 static int coroutine_fn nbd_co_truncate(BlockDriverState *bs, int64_t offset,
2380                                         bool exact, PreallocMode prealloc,
2381                                         BdrvRequestFlags flags, Error **errp)
2382 {
2383     BDRVNBDState *s = bs->opaque;
2384 
2385     if (offset != s->info.size && exact) {
2386         error_setg(errp, "Cannot resize NBD nodes");
2387         return -ENOTSUP;
2388     }
2389 
2390     if (offset > s->info.size) {
2391         error_setg(errp, "Cannot grow NBD nodes");
2392         return -EINVAL;
2393     }
2394 
2395     return 0;
2396 }
2397 
2398 static int64_t nbd_getlength(BlockDriverState *bs)
2399 {
2400     BDRVNBDState *s = bs->opaque;
2401 
2402     return s->info.size;
2403 }
2404 
2405 static void nbd_refresh_filename(BlockDriverState *bs)
2406 {
2407     BDRVNBDState *s = bs->opaque;
2408     const char *host = NULL, *port = NULL, *path = NULL;
2409     size_t len = 0;
2410 
2411     if (s->saddr->type == SOCKET_ADDRESS_TYPE_INET) {
2412         const InetSocketAddress *inet = &s->saddr->u.inet;
2413         if (!inet->has_ipv4 && !inet->has_ipv6 && !inet->has_to) {
2414             host = inet->host;
2415             port = inet->port;
2416         }
2417     } else if (s->saddr->type == SOCKET_ADDRESS_TYPE_UNIX) {
2418         path = s->saddr->u.q_unix.path;
2419     } /* else can't represent as pseudo-filename */
2420 
2421     if (path && s->export) {
2422         len = snprintf(bs->exact_filename, sizeof(bs->exact_filename),
2423                        "nbd+unix:///%s?socket=%s", s->export, path);
2424     } else if (path && !s->export) {
2425         len = snprintf(bs->exact_filename, sizeof(bs->exact_filename),
2426                        "nbd+unix://?socket=%s", path);
2427     } else if (host && s->export) {
2428         len = snprintf(bs->exact_filename, sizeof(bs->exact_filename),
2429                        "nbd://%s:%s/%s", host, port, s->export);
2430     } else if (host && !s->export) {
2431         len = snprintf(bs->exact_filename, sizeof(bs->exact_filename),
2432                        "nbd://%s:%s", host, port);
2433     }
2434     if (len >= sizeof(bs->exact_filename)) {
2435         /* Name is too long to represent exactly, so leave it empty. */
2436         bs->exact_filename[0] = '\0';
2437     }
2438 }
2439 
2440 static char *nbd_dirname(BlockDriverState *bs, Error **errp)
2441 {
2442     /* The generic bdrv_dirname() implementation is able to work out some
2443      * directory name for NBD nodes, but that would be wrong. So far there is no
2444      * specification for how "export paths" would work, so NBD does not have
2445      * directory names. */
2446     error_setg(errp, "Cannot generate a base directory for NBD nodes");
2447     return NULL;
2448 }
2449 
2450 static const char *const nbd_strong_runtime_opts[] = {
2451     "path",
2452     "host",
2453     "port",
2454     "export",
2455     "tls-creds",
2456     "server.",
2457 
2458     NULL
2459 };
2460 
2461 static BlockDriver bdrv_nbd = {
2462     .format_name                = "nbd",
2463     .protocol_name              = "nbd",
2464     .instance_size              = sizeof(BDRVNBDState),
2465     .bdrv_parse_filename        = nbd_parse_filename,
2466     .bdrv_co_create_opts        = bdrv_co_create_opts_simple,
2467     .create_opts                = &bdrv_create_opts_simple,
2468     .bdrv_file_open             = nbd_open,
2469     .bdrv_reopen_prepare        = nbd_client_reopen_prepare,
2470     .bdrv_co_preadv             = nbd_client_co_preadv,
2471     .bdrv_co_pwritev            = nbd_client_co_pwritev,
2472     .bdrv_co_pwrite_zeroes      = nbd_client_co_pwrite_zeroes,
2473     .bdrv_close                 = nbd_close,
2474     .bdrv_co_flush_to_os        = nbd_co_flush,
2475     .bdrv_co_pdiscard           = nbd_client_co_pdiscard,
2476     .bdrv_refresh_limits        = nbd_refresh_limits,
2477     .bdrv_co_truncate           = nbd_co_truncate,
2478     .bdrv_getlength             = nbd_getlength,
2479     .bdrv_detach_aio_context    = nbd_client_detach_aio_context,
2480     .bdrv_attach_aio_context    = nbd_client_attach_aio_context,
2481     .bdrv_co_drain_begin        = nbd_client_co_drain_begin,
2482     .bdrv_co_drain_end          = nbd_client_co_drain_end,
2483     .bdrv_refresh_filename      = nbd_refresh_filename,
2484     .bdrv_co_block_status       = nbd_client_co_block_status,
2485     .bdrv_dirname               = nbd_dirname,
2486     .strong_runtime_opts        = nbd_strong_runtime_opts,
2487 };
2488 
2489 static BlockDriver bdrv_nbd_tcp = {
2490     .format_name                = "nbd",
2491     .protocol_name              = "nbd+tcp",
2492     .instance_size              = sizeof(BDRVNBDState),
2493     .bdrv_parse_filename        = nbd_parse_filename,
2494     .bdrv_co_create_opts        = bdrv_co_create_opts_simple,
2495     .create_opts                = &bdrv_create_opts_simple,
2496     .bdrv_file_open             = nbd_open,
2497     .bdrv_reopen_prepare        = nbd_client_reopen_prepare,
2498     .bdrv_co_preadv             = nbd_client_co_preadv,
2499     .bdrv_co_pwritev            = nbd_client_co_pwritev,
2500     .bdrv_co_pwrite_zeroes      = nbd_client_co_pwrite_zeroes,
2501     .bdrv_close                 = nbd_close,
2502     .bdrv_co_flush_to_os        = nbd_co_flush,
2503     .bdrv_co_pdiscard           = nbd_client_co_pdiscard,
2504     .bdrv_refresh_limits        = nbd_refresh_limits,
2505     .bdrv_co_truncate           = nbd_co_truncate,
2506     .bdrv_getlength             = nbd_getlength,
2507     .bdrv_detach_aio_context    = nbd_client_detach_aio_context,
2508     .bdrv_attach_aio_context    = nbd_client_attach_aio_context,
2509     .bdrv_co_drain_begin        = nbd_client_co_drain_begin,
2510     .bdrv_co_drain_end          = nbd_client_co_drain_end,
2511     .bdrv_refresh_filename      = nbd_refresh_filename,
2512     .bdrv_co_block_status       = nbd_client_co_block_status,
2513     .bdrv_dirname               = nbd_dirname,
2514     .strong_runtime_opts        = nbd_strong_runtime_opts,
2515 };
2516 
2517 static BlockDriver bdrv_nbd_unix = {
2518     .format_name                = "nbd",
2519     .protocol_name              = "nbd+unix",
2520     .instance_size              = sizeof(BDRVNBDState),
2521     .bdrv_parse_filename        = nbd_parse_filename,
2522     .bdrv_co_create_opts        = bdrv_co_create_opts_simple,
2523     .create_opts                = &bdrv_create_opts_simple,
2524     .bdrv_file_open             = nbd_open,
2525     .bdrv_reopen_prepare        = nbd_client_reopen_prepare,
2526     .bdrv_co_preadv             = nbd_client_co_preadv,
2527     .bdrv_co_pwritev            = nbd_client_co_pwritev,
2528     .bdrv_co_pwrite_zeroes      = nbd_client_co_pwrite_zeroes,
2529     .bdrv_close                 = nbd_close,
2530     .bdrv_co_flush_to_os        = nbd_co_flush,
2531     .bdrv_co_pdiscard           = nbd_client_co_pdiscard,
2532     .bdrv_refresh_limits        = nbd_refresh_limits,
2533     .bdrv_co_truncate           = nbd_co_truncate,
2534     .bdrv_getlength             = nbd_getlength,
2535     .bdrv_detach_aio_context    = nbd_client_detach_aio_context,
2536     .bdrv_attach_aio_context    = nbd_client_attach_aio_context,
2537     .bdrv_co_drain_begin        = nbd_client_co_drain_begin,
2538     .bdrv_co_drain_end          = nbd_client_co_drain_end,
2539     .bdrv_refresh_filename      = nbd_refresh_filename,
2540     .bdrv_co_block_status       = nbd_client_co_block_status,
2541     .bdrv_dirname               = nbd_dirname,
2542     .strong_runtime_opts        = nbd_strong_runtime_opts,
2543 };
2544 
2545 static void bdrv_nbd_init(void)
2546 {
2547     bdrv_register(&bdrv_nbd);
2548     bdrv_register(&bdrv_nbd_tcp);
2549     bdrv_register(&bdrv_nbd_unix);
2550 }
2551 
2552 block_init(bdrv_nbd_init);
2553