xref: /openbmc/qemu/block/nbd.c (revision 278f064e)
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     QemuCoSleep reconnect_sleep;
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     qemu_co_sleep_wake(&s->reconnect_sleep);
293 
294     nbd_co_establish_connection_cancel(bs, false);
295 
296     reconnect_delay_timer_del(s);
297 
298     if (qatomic_load_acquire(&s->state) == NBD_CLIENT_CONNECTING_WAIT) {
299         s->state = NBD_CLIENT_CONNECTING_NOWAIT;
300         qemu_co_queue_restart_all(&s->free_sema);
301     }
302 }
303 
304 static void coroutine_fn nbd_client_co_drain_end(BlockDriverState *bs)
305 {
306     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
307 
308     s->drained = false;
309     if (s->wait_drained_end) {
310         s->wait_drained_end = false;
311         aio_co_wake(s->connection_co);
312     }
313 }
314 
315 
316 static void nbd_teardown_connection(BlockDriverState *bs)
317 {
318     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
319 
320     if (s->ioc) {
321         /* finish any pending coroutines */
322         qio_channel_shutdown(s->ioc, QIO_CHANNEL_SHUTDOWN_BOTH, NULL);
323     } else if (s->sioc) {
324         /* abort negotiation */
325         qio_channel_shutdown(QIO_CHANNEL(s->sioc), QIO_CHANNEL_SHUTDOWN_BOTH,
326                              NULL);
327     }
328 
329     s->state = NBD_CLIENT_QUIT;
330     if (s->connection_co) {
331         qemu_co_sleep_wake(&s->reconnect_sleep);
332         nbd_co_establish_connection_cancel(bs, true);
333     }
334     if (qemu_in_coroutine()) {
335         s->teardown_co = qemu_coroutine_self();
336         /* connection_co resumes us when it terminates */
337         qemu_coroutine_yield();
338         s->teardown_co = NULL;
339     } else {
340         BDRV_POLL_WHILE(bs, s->connection_co);
341     }
342     assert(!s->connection_co);
343 }
344 
345 static bool nbd_client_connecting(BDRVNBDState *s)
346 {
347     NBDClientState state = qatomic_load_acquire(&s->state);
348     return state == NBD_CLIENT_CONNECTING_WAIT ||
349         state == NBD_CLIENT_CONNECTING_NOWAIT;
350 }
351 
352 static bool nbd_client_connecting_wait(BDRVNBDState *s)
353 {
354     return qatomic_load_acquire(&s->state) == NBD_CLIENT_CONNECTING_WAIT;
355 }
356 
357 static void connect_bh(void *opaque)
358 {
359     BDRVNBDState *state = opaque;
360 
361     assert(state->wait_connect);
362     state->wait_connect = false;
363     aio_co_wake(state->connection_co);
364 }
365 
366 static void nbd_init_connect_thread(BDRVNBDState *s)
367 {
368     s->connect_thread = g_new(NBDConnectThread, 1);
369 
370     *s->connect_thread = (NBDConnectThread) {
371         .saddr = QAPI_CLONE(SocketAddress, s->saddr),
372         .state = CONNECT_THREAD_NONE,
373         .bh_func = connect_bh,
374         .bh_opaque = s,
375     };
376 
377     qemu_mutex_init(&s->connect_thread->mutex);
378 }
379 
380 static void nbd_free_connect_thread(NBDConnectThread *thr)
381 {
382     if (thr->sioc) {
383         qio_channel_close(QIO_CHANNEL(thr->sioc), NULL);
384     }
385     error_free(thr->err);
386     qapi_free_SocketAddress(thr->saddr);
387     g_free(thr);
388 }
389 
390 static void *connect_thread_func(void *opaque)
391 {
392     NBDConnectThread *thr = opaque;
393     int ret;
394     bool do_free = false;
395 
396     thr->sioc = qio_channel_socket_new();
397 
398     error_free(thr->err);
399     thr->err = NULL;
400     ret = qio_channel_socket_connect_sync(thr->sioc, thr->saddr, &thr->err);
401     if (ret < 0) {
402         object_unref(OBJECT(thr->sioc));
403         thr->sioc = NULL;
404     }
405 
406     qemu_mutex_lock(&thr->mutex);
407 
408     switch (thr->state) {
409     case CONNECT_THREAD_RUNNING:
410         thr->state = ret < 0 ? CONNECT_THREAD_FAIL : CONNECT_THREAD_SUCCESS;
411         if (thr->bh_ctx) {
412             aio_bh_schedule_oneshot(thr->bh_ctx, thr->bh_func, thr->bh_opaque);
413 
414             /* play safe, don't reuse bh_ctx on further connection attempts */
415             thr->bh_ctx = NULL;
416         }
417         break;
418     case CONNECT_THREAD_RUNNING_DETACHED:
419         do_free = true;
420         break;
421     default:
422         abort();
423     }
424 
425     qemu_mutex_unlock(&thr->mutex);
426 
427     if (do_free) {
428         nbd_free_connect_thread(thr);
429     }
430 
431     return NULL;
432 }
433 
434 static int coroutine_fn
435 nbd_co_establish_connection(BlockDriverState *bs, Error **errp)
436 {
437     int ret;
438     QemuThread thread;
439     BDRVNBDState *s = bs->opaque;
440     NBDConnectThread *thr = s->connect_thread;
441 
442     if (!thr) {
443         /* detached */
444         return -1;
445     }
446 
447     qemu_mutex_lock(&thr->mutex);
448 
449     switch (thr->state) {
450     case CONNECT_THREAD_FAIL:
451     case CONNECT_THREAD_NONE:
452         error_free(thr->err);
453         thr->err = NULL;
454         thr->state = CONNECT_THREAD_RUNNING;
455         qemu_thread_create(&thread, "nbd-connect",
456                            connect_thread_func, thr, QEMU_THREAD_DETACHED);
457         break;
458     case CONNECT_THREAD_SUCCESS:
459         /* Previous attempt finally succeeded in background */
460         thr->state = CONNECT_THREAD_NONE;
461         s->sioc = thr->sioc;
462         thr->sioc = NULL;
463         yank_register_function(BLOCKDEV_YANK_INSTANCE(bs->node_name),
464                                nbd_yank, bs);
465         qemu_mutex_unlock(&thr->mutex);
466         return 0;
467     case CONNECT_THREAD_RUNNING:
468         /* Already running, will wait */
469         break;
470     default:
471         abort();
472     }
473 
474     thr->bh_ctx = qemu_get_current_aio_context();
475 
476     qemu_mutex_unlock(&thr->mutex);
477 
478 
479     /*
480      * We are going to wait for connect-thread finish, but
481      * nbd_client_co_drain_begin() can interrupt.
482      *
483      * Note that wait_connect variable is not visible for connect-thread. It
484      * doesn't need mutex protection, it used only inside home aio context of
485      * bs.
486      */
487     s->wait_connect = true;
488     qemu_coroutine_yield();
489 
490     if (!s->connect_thread) {
491         /* detached */
492         return -1;
493     }
494     assert(thr == s->connect_thread);
495 
496     qemu_mutex_lock(&thr->mutex);
497 
498     switch (thr->state) {
499     case CONNECT_THREAD_SUCCESS:
500     case CONNECT_THREAD_FAIL:
501         thr->state = CONNECT_THREAD_NONE;
502         error_propagate(errp, thr->err);
503         thr->err = NULL;
504         s->sioc = thr->sioc;
505         thr->sioc = NULL;
506         if (s->sioc) {
507             yank_register_function(BLOCKDEV_YANK_INSTANCE(bs->node_name),
508                                    nbd_yank, bs);
509         }
510         ret = (s->sioc ? 0 : -1);
511         break;
512     case CONNECT_THREAD_RUNNING:
513     case CONNECT_THREAD_RUNNING_DETACHED:
514         /*
515          * Obviously, drained section wants to start. Report the attempt as
516          * failed. Still connect thread is executing in background, and its
517          * result may be used for next connection attempt.
518          */
519         ret = -1;
520         error_setg(errp, "Connection attempt cancelled by other operation");
521         break;
522 
523     case CONNECT_THREAD_NONE:
524         /*
525          * Impossible. We've seen this thread running. So it should be
526          * running or at least give some results.
527          */
528         abort();
529 
530     default:
531         abort();
532     }
533 
534     qemu_mutex_unlock(&thr->mutex);
535 
536     return ret;
537 }
538 
539 /*
540  * nbd_co_establish_connection_cancel
541  * Cancel nbd_co_establish_connection asynchronously: it will finish soon, to
542  * allow drained section to begin.
543  *
544  * If detach is true, also cleanup the state (or if thread is running, move it
545  * to CONNECT_THREAD_RUNNING_DETACHED state). s->connect_thread becomes NULL if
546  * detach is true.
547  */
548 static void nbd_co_establish_connection_cancel(BlockDriverState *bs,
549                                                bool detach)
550 {
551     BDRVNBDState *s = bs->opaque;
552     NBDConnectThread *thr = s->connect_thread;
553     bool wake = false;
554     bool do_free = false;
555 
556     qemu_mutex_lock(&thr->mutex);
557 
558     if (thr->state == CONNECT_THREAD_RUNNING) {
559         /* We can cancel only in running state, when bh is not yet scheduled */
560         thr->bh_ctx = NULL;
561         if (s->wait_connect) {
562             s->wait_connect = false;
563             wake = true;
564         }
565         if (detach) {
566             thr->state = CONNECT_THREAD_RUNNING_DETACHED;
567             s->connect_thread = NULL;
568         }
569     } else if (detach) {
570         do_free = true;
571     }
572 
573     qemu_mutex_unlock(&thr->mutex);
574 
575     if (do_free) {
576         nbd_free_connect_thread(thr);
577         s->connect_thread = NULL;
578     }
579 
580     if (wake) {
581         aio_co_wake(s->connection_co);
582     }
583 }
584 
585 static coroutine_fn void nbd_reconnect_attempt(BDRVNBDState *s)
586 {
587     int ret;
588     Error *local_err = NULL;
589 
590     if (!nbd_client_connecting(s)) {
591         return;
592     }
593 
594     /* Wait for completion of all in-flight requests */
595 
596     qemu_co_mutex_lock(&s->send_mutex);
597 
598     while (s->in_flight > 0) {
599         qemu_co_mutex_unlock(&s->send_mutex);
600         nbd_recv_coroutines_wake_all(s);
601         s->wait_in_flight = true;
602         qemu_coroutine_yield();
603         s->wait_in_flight = false;
604         qemu_co_mutex_lock(&s->send_mutex);
605     }
606 
607     qemu_co_mutex_unlock(&s->send_mutex);
608 
609     if (!nbd_client_connecting(s)) {
610         return;
611     }
612 
613     /*
614      * Now we are sure that nobody is accessing the channel, and no one will
615      * try until we set the state to CONNECTED.
616      */
617 
618     /* Finalize previous connection if any */
619     if (s->ioc) {
620         qio_channel_detach_aio_context(QIO_CHANNEL(s->ioc));
621         yank_unregister_function(BLOCKDEV_YANK_INSTANCE(s->bs->node_name),
622                                  nbd_yank, s->bs);
623         object_unref(OBJECT(s->sioc));
624         s->sioc = NULL;
625         object_unref(OBJECT(s->ioc));
626         s->ioc = NULL;
627     }
628 
629     if (nbd_co_establish_connection(s->bs, &local_err) < 0) {
630         ret = -ECONNREFUSED;
631         goto out;
632     }
633 
634     bdrv_dec_in_flight(s->bs);
635 
636     ret = nbd_client_handshake(s->bs, &local_err);
637 
638     if (s->drained) {
639         s->wait_drained_end = true;
640         while (s->drained) {
641             /*
642              * We may be entered once from nbd_client_attach_aio_context_bh
643              * and then from nbd_client_co_drain_end. So here is a loop.
644              */
645             qemu_coroutine_yield();
646         }
647     }
648     bdrv_inc_in_flight(s->bs);
649 
650 out:
651     s->connect_status = ret;
652     error_free(s->connect_err);
653     s->connect_err = NULL;
654     error_propagate(&s->connect_err, local_err);
655 
656     if (ret >= 0) {
657         /* successfully connected */
658         s->state = NBD_CLIENT_CONNECTED;
659         qemu_co_queue_restart_all(&s->free_sema);
660     }
661 }
662 
663 static coroutine_fn void nbd_co_reconnect_loop(BDRVNBDState *s)
664 {
665     uint64_t timeout = 1 * NANOSECONDS_PER_SECOND;
666     uint64_t max_timeout = 16 * NANOSECONDS_PER_SECOND;
667 
668     if (qatomic_load_acquire(&s->state) == NBD_CLIENT_CONNECTING_WAIT) {
669         reconnect_delay_timer_init(s, qemu_clock_get_ns(QEMU_CLOCK_REALTIME) +
670                                    s->reconnect_delay * NANOSECONDS_PER_SECOND);
671     }
672 
673     nbd_reconnect_attempt(s);
674 
675     while (nbd_client_connecting(s)) {
676         if (s->drained) {
677             bdrv_dec_in_flight(s->bs);
678             s->wait_drained_end = true;
679             while (s->drained) {
680                 /*
681                  * We may be entered once from nbd_client_attach_aio_context_bh
682                  * and then from nbd_client_co_drain_end. So here is a loop.
683                  */
684                 qemu_coroutine_yield();
685             }
686             bdrv_inc_in_flight(s->bs);
687         } else {
688             qemu_co_sleep_ns_wakeable(&s->reconnect_sleep,
689                                       QEMU_CLOCK_REALTIME, timeout);
690             if (s->drained) {
691                 continue;
692             }
693             if (timeout < max_timeout) {
694                 timeout *= 2;
695             }
696         }
697 
698         nbd_reconnect_attempt(s);
699     }
700 
701     reconnect_delay_timer_del(s);
702 }
703 
704 static coroutine_fn void nbd_connection_entry(void *opaque)
705 {
706     BDRVNBDState *s = opaque;
707     uint64_t i;
708     int ret = 0;
709     Error *local_err = NULL;
710 
711     while (qatomic_load_acquire(&s->state) != NBD_CLIENT_QUIT) {
712         /*
713          * The NBD client can only really be considered idle when it has
714          * yielded from qio_channel_readv_all_eof(), waiting for data. This is
715          * the point where the additional scheduled coroutine entry happens
716          * after nbd_client_attach_aio_context().
717          *
718          * Therefore we keep an additional in_flight reference all the time and
719          * only drop it temporarily here.
720          */
721 
722         if (nbd_client_connecting(s)) {
723             nbd_co_reconnect_loop(s);
724         }
725 
726         if (qatomic_load_acquire(&s->state) != NBD_CLIENT_CONNECTED) {
727             continue;
728         }
729 
730         assert(s->reply.handle == 0);
731         ret = nbd_receive_reply(s->bs, s->ioc, &s->reply, &local_err);
732 
733         if (local_err) {
734             trace_nbd_read_reply_entry_fail(ret, error_get_pretty(local_err));
735             error_free(local_err);
736             local_err = NULL;
737         }
738         if (ret <= 0) {
739             nbd_channel_error(s, ret ? ret : -EIO);
740             continue;
741         }
742 
743         /*
744          * There's no need for a mutex on the receive side, because the
745          * handler acts as a synchronization point and ensures that only
746          * one coroutine is called until the reply finishes.
747          */
748         i = HANDLE_TO_INDEX(s, s->reply.handle);
749         if (i >= MAX_NBD_REQUESTS ||
750             !s->requests[i].coroutine ||
751             !s->requests[i].receiving ||
752             (nbd_reply_is_structured(&s->reply) && !s->info.structured_reply))
753         {
754             nbd_channel_error(s, -EINVAL);
755             continue;
756         }
757 
758         /*
759          * We're woken up again by the request itself.  Note that there
760          * is no race between yielding and reentering connection_co.  This
761          * is because:
762          *
763          * - if the request runs on the same AioContext, it is only
764          *   entered after we yield
765          *
766          * - if the request runs on a different AioContext, reentering
767          *   connection_co happens through a bottom half, which can only
768          *   run after we yield.
769          */
770         aio_co_wake(s->requests[i].coroutine);
771         qemu_coroutine_yield();
772     }
773 
774     qemu_co_queue_restart_all(&s->free_sema);
775     nbd_recv_coroutines_wake_all(s);
776     bdrv_dec_in_flight(s->bs);
777 
778     s->connection_co = NULL;
779     if (s->ioc) {
780         qio_channel_detach_aio_context(QIO_CHANNEL(s->ioc));
781         yank_unregister_function(BLOCKDEV_YANK_INSTANCE(s->bs->node_name),
782                                  nbd_yank, s->bs);
783         object_unref(OBJECT(s->sioc));
784         s->sioc = NULL;
785         object_unref(OBJECT(s->ioc));
786         s->ioc = NULL;
787     }
788 
789     if (s->teardown_co) {
790         aio_co_wake(s->teardown_co);
791     }
792     aio_wait_kick();
793 }
794 
795 static int nbd_co_send_request(BlockDriverState *bs,
796                                NBDRequest *request,
797                                QEMUIOVector *qiov)
798 {
799     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
800     int rc, i = -1;
801 
802     qemu_co_mutex_lock(&s->send_mutex);
803     while (s->in_flight == MAX_NBD_REQUESTS || nbd_client_connecting_wait(s)) {
804         qemu_co_queue_wait(&s->free_sema, &s->send_mutex);
805     }
806 
807     if (qatomic_load_acquire(&s->state) != NBD_CLIENT_CONNECTED) {
808         rc = -EIO;
809         goto err;
810     }
811 
812     s->in_flight++;
813 
814     for (i = 0; i < MAX_NBD_REQUESTS; i++) {
815         if (s->requests[i].coroutine == NULL) {
816             break;
817         }
818     }
819 
820     g_assert(qemu_in_coroutine());
821     assert(i < MAX_NBD_REQUESTS);
822 
823     s->requests[i].coroutine = qemu_coroutine_self();
824     s->requests[i].offset = request->from;
825     s->requests[i].receiving = false;
826 
827     request->handle = INDEX_TO_HANDLE(s, i);
828 
829     assert(s->ioc);
830 
831     if (qiov) {
832         qio_channel_set_cork(s->ioc, true);
833         rc = nbd_send_request(s->ioc, request);
834         if (qatomic_load_acquire(&s->state) == NBD_CLIENT_CONNECTED &&
835             rc >= 0) {
836             if (qio_channel_writev_all(s->ioc, qiov->iov, qiov->niov,
837                                        NULL) < 0) {
838                 rc = -EIO;
839             }
840         } else if (rc >= 0) {
841             rc = -EIO;
842         }
843         qio_channel_set_cork(s->ioc, false);
844     } else {
845         rc = nbd_send_request(s->ioc, request);
846     }
847 
848 err:
849     if (rc < 0) {
850         nbd_channel_error(s, rc);
851         if (i != -1) {
852             s->requests[i].coroutine = NULL;
853             s->in_flight--;
854         }
855         if (s->in_flight == 0 && s->wait_in_flight) {
856             aio_co_wake(s->connection_co);
857         } else {
858             qemu_co_queue_next(&s->free_sema);
859         }
860     }
861     qemu_co_mutex_unlock(&s->send_mutex);
862     return rc;
863 }
864 
865 static inline uint16_t payload_advance16(uint8_t **payload)
866 {
867     *payload += 2;
868     return lduw_be_p(*payload - 2);
869 }
870 
871 static inline uint32_t payload_advance32(uint8_t **payload)
872 {
873     *payload += 4;
874     return ldl_be_p(*payload - 4);
875 }
876 
877 static inline uint64_t payload_advance64(uint8_t **payload)
878 {
879     *payload += 8;
880     return ldq_be_p(*payload - 8);
881 }
882 
883 static int nbd_parse_offset_hole_payload(BDRVNBDState *s,
884                                          NBDStructuredReplyChunk *chunk,
885                                          uint8_t *payload, uint64_t orig_offset,
886                                          QEMUIOVector *qiov, Error **errp)
887 {
888     uint64_t offset;
889     uint32_t hole_size;
890 
891     if (chunk->length != sizeof(offset) + sizeof(hole_size)) {
892         error_setg(errp, "Protocol error: invalid payload for "
893                          "NBD_REPLY_TYPE_OFFSET_HOLE");
894         return -EINVAL;
895     }
896 
897     offset = payload_advance64(&payload);
898     hole_size = payload_advance32(&payload);
899 
900     if (!hole_size || offset < orig_offset || hole_size > qiov->size ||
901         offset > orig_offset + qiov->size - hole_size) {
902         error_setg(errp, "Protocol error: server sent chunk exceeding requested"
903                          " region");
904         return -EINVAL;
905     }
906     if (s->info.min_block &&
907         !QEMU_IS_ALIGNED(hole_size, s->info.min_block)) {
908         trace_nbd_structured_read_compliance("hole");
909     }
910 
911     qemu_iovec_memset(qiov, offset - orig_offset, 0, hole_size);
912 
913     return 0;
914 }
915 
916 /*
917  * nbd_parse_blockstatus_payload
918  * Based on our request, we expect only one extent in reply, for the
919  * base:allocation context.
920  */
921 static int nbd_parse_blockstatus_payload(BDRVNBDState *s,
922                                          NBDStructuredReplyChunk *chunk,
923                                          uint8_t *payload, uint64_t orig_length,
924                                          NBDExtent *extent, Error **errp)
925 {
926     uint32_t context_id;
927 
928     /* The server succeeded, so it must have sent [at least] one extent */
929     if (chunk->length < sizeof(context_id) + sizeof(*extent)) {
930         error_setg(errp, "Protocol error: invalid payload for "
931                          "NBD_REPLY_TYPE_BLOCK_STATUS");
932         return -EINVAL;
933     }
934 
935     context_id = payload_advance32(&payload);
936     if (s->info.context_id != context_id) {
937         error_setg(errp, "Protocol error: unexpected context id %d for "
938                          "NBD_REPLY_TYPE_BLOCK_STATUS, when negotiated context "
939                          "id is %d", context_id,
940                          s->info.context_id);
941         return -EINVAL;
942     }
943 
944     extent->length = payload_advance32(&payload);
945     extent->flags = payload_advance32(&payload);
946 
947     if (extent->length == 0) {
948         error_setg(errp, "Protocol error: server sent status chunk with "
949                    "zero length");
950         return -EINVAL;
951     }
952 
953     /*
954      * A server sending unaligned block status is in violation of the
955      * protocol, but as qemu-nbd 3.1 is such a server (at least for
956      * POSIX files that are not a multiple of 512 bytes, since qemu
957      * rounds files up to 512-byte multiples but lseek(SEEK_HOLE)
958      * still sees an implicit hole beyond the real EOF), it's nicer to
959      * work around the misbehaving server. If the request included
960      * more than the final unaligned block, truncate it back to an
961      * aligned result; if the request was only the final block, round
962      * up to the full block and change the status to fully-allocated
963      * (always a safe status, even if it loses information).
964      */
965     if (s->info.min_block && !QEMU_IS_ALIGNED(extent->length,
966                                                    s->info.min_block)) {
967         trace_nbd_parse_blockstatus_compliance("extent length is unaligned");
968         if (extent->length > s->info.min_block) {
969             extent->length = QEMU_ALIGN_DOWN(extent->length,
970                                              s->info.min_block);
971         } else {
972             extent->length = s->info.min_block;
973             extent->flags = 0;
974         }
975     }
976 
977     /*
978      * We used NBD_CMD_FLAG_REQ_ONE, so the server should not have
979      * sent us any more than one extent, nor should it have included
980      * status beyond our request in that extent. However, it's easy
981      * enough to ignore the server's noncompliance without killing the
982      * connection; just ignore trailing extents, and clamp things to
983      * the length of our request.
984      */
985     if (chunk->length > sizeof(context_id) + sizeof(*extent)) {
986         trace_nbd_parse_blockstatus_compliance("more than one extent");
987     }
988     if (extent->length > orig_length) {
989         extent->length = orig_length;
990         trace_nbd_parse_blockstatus_compliance("extent length too large");
991     }
992 
993     /*
994      * HACK: if we are using x-dirty-bitmaps to access
995      * qemu:allocation-depth, treat all depths > 2 the same as 2,
996      * since nbd_client_co_block_status is only expecting the low two
997      * bits to be set.
998      */
999     if (s->alloc_depth && extent->flags > 2) {
1000         extent->flags = 2;
1001     }
1002 
1003     return 0;
1004 }
1005 
1006 /*
1007  * nbd_parse_error_payload
1008  * on success @errp contains message describing nbd error reply
1009  */
1010 static int nbd_parse_error_payload(NBDStructuredReplyChunk *chunk,
1011                                    uint8_t *payload, int *request_ret,
1012                                    Error **errp)
1013 {
1014     uint32_t error;
1015     uint16_t message_size;
1016 
1017     assert(chunk->type & (1 << 15));
1018 
1019     if (chunk->length < sizeof(error) + sizeof(message_size)) {
1020         error_setg(errp,
1021                    "Protocol error: invalid payload for structured error");
1022         return -EINVAL;
1023     }
1024 
1025     error = nbd_errno_to_system_errno(payload_advance32(&payload));
1026     if (error == 0) {
1027         error_setg(errp, "Protocol error: server sent structured error chunk "
1028                          "with error = 0");
1029         return -EINVAL;
1030     }
1031 
1032     *request_ret = -error;
1033     message_size = payload_advance16(&payload);
1034 
1035     if (message_size > chunk->length - sizeof(error) - sizeof(message_size)) {
1036         error_setg(errp, "Protocol error: server sent structured error chunk "
1037                          "with incorrect message size");
1038         return -EINVAL;
1039     }
1040 
1041     /* TODO: Add a trace point to mention the server complaint */
1042 
1043     /* TODO handle ERROR_OFFSET */
1044 
1045     return 0;
1046 }
1047 
1048 static int nbd_co_receive_offset_data_payload(BDRVNBDState *s,
1049                                               uint64_t orig_offset,
1050                                               QEMUIOVector *qiov, Error **errp)
1051 {
1052     QEMUIOVector sub_qiov;
1053     uint64_t offset;
1054     size_t data_size;
1055     int ret;
1056     NBDStructuredReplyChunk *chunk = &s->reply.structured;
1057 
1058     assert(nbd_reply_is_structured(&s->reply));
1059 
1060     /* The NBD spec requires at least one byte of payload */
1061     if (chunk->length <= sizeof(offset)) {
1062         error_setg(errp, "Protocol error: invalid payload for "
1063                          "NBD_REPLY_TYPE_OFFSET_DATA");
1064         return -EINVAL;
1065     }
1066 
1067     if (nbd_read64(s->ioc, &offset, "OFFSET_DATA offset", errp) < 0) {
1068         return -EIO;
1069     }
1070 
1071     data_size = chunk->length - sizeof(offset);
1072     assert(data_size);
1073     if (offset < orig_offset || data_size > qiov->size ||
1074         offset > orig_offset + qiov->size - data_size) {
1075         error_setg(errp, "Protocol error: server sent chunk exceeding requested"
1076                          " region");
1077         return -EINVAL;
1078     }
1079     if (s->info.min_block && !QEMU_IS_ALIGNED(data_size, s->info.min_block)) {
1080         trace_nbd_structured_read_compliance("data");
1081     }
1082 
1083     qemu_iovec_init(&sub_qiov, qiov->niov);
1084     qemu_iovec_concat(&sub_qiov, qiov, offset - orig_offset, data_size);
1085     ret = qio_channel_readv_all(s->ioc, sub_qiov.iov, sub_qiov.niov, errp);
1086     qemu_iovec_destroy(&sub_qiov);
1087 
1088     return ret < 0 ? -EIO : 0;
1089 }
1090 
1091 #define NBD_MAX_MALLOC_PAYLOAD 1000
1092 static coroutine_fn int nbd_co_receive_structured_payload(
1093         BDRVNBDState *s, void **payload, Error **errp)
1094 {
1095     int ret;
1096     uint32_t len;
1097 
1098     assert(nbd_reply_is_structured(&s->reply));
1099 
1100     len = s->reply.structured.length;
1101 
1102     if (len == 0) {
1103         return 0;
1104     }
1105 
1106     if (payload == NULL) {
1107         error_setg(errp, "Unexpected structured payload");
1108         return -EINVAL;
1109     }
1110 
1111     if (len > NBD_MAX_MALLOC_PAYLOAD) {
1112         error_setg(errp, "Payload too large");
1113         return -EINVAL;
1114     }
1115 
1116     *payload = g_new(char, len);
1117     ret = nbd_read(s->ioc, *payload, len, "structured payload", errp);
1118     if (ret < 0) {
1119         g_free(*payload);
1120         *payload = NULL;
1121         return ret;
1122     }
1123 
1124     return 0;
1125 }
1126 
1127 /*
1128  * nbd_co_do_receive_one_chunk
1129  * for simple reply:
1130  *   set request_ret to received reply error
1131  *   if qiov is not NULL: read payload to @qiov
1132  * for structured reply chunk:
1133  *   if error chunk: read payload, set @request_ret, do not set @payload
1134  *   else if offset_data chunk: read payload data to @qiov, do not set @payload
1135  *   else: read payload to @payload
1136  *
1137  * If function fails, @errp contains corresponding error message, and the
1138  * connection with the server is suspect.  If it returns 0, then the
1139  * transaction succeeded (although @request_ret may be a negative errno
1140  * corresponding to the server's error reply), and errp is unchanged.
1141  */
1142 static coroutine_fn int nbd_co_do_receive_one_chunk(
1143         BDRVNBDState *s, uint64_t handle, bool only_structured,
1144         int *request_ret, QEMUIOVector *qiov, void **payload, Error **errp)
1145 {
1146     int ret;
1147     int i = HANDLE_TO_INDEX(s, handle);
1148     void *local_payload = NULL;
1149     NBDStructuredReplyChunk *chunk;
1150 
1151     if (payload) {
1152         *payload = NULL;
1153     }
1154     *request_ret = 0;
1155 
1156     /* Wait until we're woken up by nbd_connection_entry.  */
1157     s->requests[i].receiving = true;
1158     qemu_coroutine_yield();
1159     s->requests[i].receiving = false;
1160     if (qatomic_load_acquire(&s->state) != NBD_CLIENT_CONNECTED) {
1161         error_setg(errp, "Connection closed");
1162         return -EIO;
1163     }
1164     assert(s->ioc);
1165 
1166     assert(s->reply.handle == handle);
1167 
1168     if (nbd_reply_is_simple(&s->reply)) {
1169         if (only_structured) {
1170             error_setg(errp, "Protocol error: simple reply when structured "
1171                              "reply chunk was expected");
1172             return -EINVAL;
1173         }
1174 
1175         *request_ret = -nbd_errno_to_system_errno(s->reply.simple.error);
1176         if (*request_ret < 0 || !qiov) {
1177             return 0;
1178         }
1179 
1180         return qio_channel_readv_all(s->ioc, qiov->iov, qiov->niov,
1181                                      errp) < 0 ? -EIO : 0;
1182     }
1183 
1184     /* handle structured reply chunk */
1185     assert(s->info.structured_reply);
1186     chunk = &s->reply.structured;
1187 
1188     if (chunk->type == NBD_REPLY_TYPE_NONE) {
1189         if (!(chunk->flags & NBD_REPLY_FLAG_DONE)) {
1190             error_setg(errp, "Protocol error: NBD_REPLY_TYPE_NONE chunk without"
1191                        " NBD_REPLY_FLAG_DONE flag set");
1192             return -EINVAL;
1193         }
1194         if (chunk->length) {
1195             error_setg(errp, "Protocol error: NBD_REPLY_TYPE_NONE chunk with"
1196                        " nonzero length");
1197             return -EINVAL;
1198         }
1199         return 0;
1200     }
1201 
1202     if (chunk->type == NBD_REPLY_TYPE_OFFSET_DATA) {
1203         if (!qiov) {
1204             error_setg(errp, "Unexpected NBD_REPLY_TYPE_OFFSET_DATA chunk");
1205             return -EINVAL;
1206         }
1207 
1208         return nbd_co_receive_offset_data_payload(s, s->requests[i].offset,
1209                                                   qiov, errp);
1210     }
1211 
1212     if (nbd_reply_type_is_error(chunk->type)) {
1213         payload = &local_payload;
1214     }
1215 
1216     ret = nbd_co_receive_structured_payload(s, payload, errp);
1217     if (ret < 0) {
1218         return ret;
1219     }
1220 
1221     if (nbd_reply_type_is_error(chunk->type)) {
1222         ret = nbd_parse_error_payload(chunk, local_payload, request_ret, errp);
1223         g_free(local_payload);
1224         return ret;
1225     }
1226 
1227     return 0;
1228 }
1229 
1230 /*
1231  * nbd_co_receive_one_chunk
1232  * Read reply, wake up connection_co and set s->quit if needed.
1233  * Return value is a fatal error code or normal nbd reply error code
1234  */
1235 static coroutine_fn int nbd_co_receive_one_chunk(
1236         BDRVNBDState *s, uint64_t handle, bool only_structured,
1237         int *request_ret, QEMUIOVector *qiov, NBDReply *reply, void **payload,
1238         Error **errp)
1239 {
1240     int ret = nbd_co_do_receive_one_chunk(s, handle, only_structured,
1241                                           request_ret, qiov, payload, errp);
1242 
1243     if (ret < 0) {
1244         memset(reply, 0, sizeof(*reply));
1245         nbd_channel_error(s, ret);
1246     } else {
1247         /* For assert at loop start in nbd_connection_entry */
1248         *reply = s->reply;
1249     }
1250     s->reply.handle = 0;
1251 
1252     if (s->connection_co && !s->wait_in_flight) {
1253         /*
1254          * We must check s->wait_in_flight, because we may entered by
1255          * nbd_recv_coroutines_wake_all(), in this case we should not
1256          * wake connection_co here, it will woken by last request.
1257          */
1258         aio_co_wake(s->connection_co);
1259     }
1260 
1261     return ret;
1262 }
1263 
1264 typedef struct NBDReplyChunkIter {
1265     int ret;
1266     int request_ret;
1267     Error *err;
1268     bool done, only_structured;
1269 } NBDReplyChunkIter;
1270 
1271 static void nbd_iter_channel_error(NBDReplyChunkIter *iter,
1272                                    int ret, Error **local_err)
1273 {
1274     assert(local_err && *local_err);
1275     assert(ret < 0);
1276 
1277     if (!iter->ret) {
1278         iter->ret = ret;
1279         error_propagate(&iter->err, *local_err);
1280     } else {
1281         error_free(*local_err);
1282     }
1283 
1284     *local_err = NULL;
1285 }
1286 
1287 static void nbd_iter_request_error(NBDReplyChunkIter *iter, int ret)
1288 {
1289     assert(ret < 0);
1290 
1291     if (!iter->request_ret) {
1292         iter->request_ret = ret;
1293     }
1294 }
1295 
1296 /*
1297  * NBD_FOREACH_REPLY_CHUNK
1298  * The pointer stored in @payload requires g_free() to free it.
1299  */
1300 #define NBD_FOREACH_REPLY_CHUNK(s, iter, handle, structured, \
1301                                 qiov, reply, payload) \
1302     for (iter = (NBDReplyChunkIter) { .only_structured = structured }; \
1303          nbd_reply_chunk_iter_receive(s, &iter, handle, qiov, reply, payload);)
1304 
1305 /*
1306  * nbd_reply_chunk_iter_receive
1307  * The pointer stored in @payload requires g_free() to free it.
1308  */
1309 static bool nbd_reply_chunk_iter_receive(BDRVNBDState *s,
1310                                          NBDReplyChunkIter *iter,
1311                                          uint64_t handle,
1312                                          QEMUIOVector *qiov, NBDReply *reply,
1313                                          void **payload)
1314 {
1315     int ret, request_ret;
1316     NBDReply local_reply;
1317     NBDStructuredReplyChunk *chunk;
1318     Error *local_err = NULL;
1319     if (qatomic_load_acquire(&s->state) != NBD_CLIENT_CONNECTED) {
1320         error_setg(&local_err, "Connection closed");
1321         nbd_iter_channel_error(iter, -EIO, &local_err);
1322         goto break_loop;
1323     }
1324 
1325     if (iter->done) {
1326         /* Previous iteration was last. */
1327         goto break_loop;
1328     }
1329 
1330     if (reply == NULL) {
1331         reply = &local_reply;
1332     }
1333 
1334     ret = nbd_co_receive_one_chunk(s, handle, iter->only_structured,
1335                                    &request_ret, qiov, reply, payload,
1336                                    &local_err);
1337     if (ret < 0) {
1338         nbd_iter_channel_error(iter, ret, &local_err);
1339     } else if (request_ret < 0) {
1340         nbd_iter_request_error(iter, request_ret);
1341     }
1342 
1343     /* Do not execute the body of NBD_FOREACH_REPLY_CHUNK for simple reply. */
1344     if (nbd_reply_is_simple(reply) ||
1345         qatomic_load_acquire(&s->state) != NBD_CLIENT_CONNECTED) {
1346         goto break_loop;
1347     }
1348 
1349     chunk = &reply->structured;
1350     iter->only_structured = true;
1351 
1352     if (chunk->type == NBD_REPLY_TYPE_NONE) {
1353         /* NBD_REPLY_FLAG_DONE is already checked in nbd_co_receive_one_chunk */
1354         assert(chunk->flags & NBD_REPLY_FLAG_DONE);
1355         goto break_loop;
1356     }
1357 
1358     if (chunk->flags & NBD_REPLY_FLAG_DONE) {
1359         /* This iteration is last. */
1360         iter->done = true;
1361     }
1362 
1363     /* Execute the loop body */
1364     return true;
1365 
1366 break_loop:
1367     s->requests[HANDLE_TO_INDEX(s, handle)].coroutine = NULL;
1368 
1369     qemu_co_mutex_lock(&s->send_mutex);
1370     s->in_flight--;
1371     if (s->in_flight == 0 && s->wait_in_flight) {
1372         aio_co_wake(s->connection_co);
1373     } else {
1374         qemu_co_queue_next(&s->free_sema);
1375     }
1376     qemu_co_mutex_unlock(&s->send_mutex);
1377 
1378     return false;
1379 }
1380 
1381 static int nbd_co_receive_return_code(BDRVNBDState *s, uint64_t handle,
1382                                       int *request_ret, Error **errp)
1383 {
1384     NBDReplyChunkIter iter;
1385 
1386     NBD_FOREACH_REPLY_CHUNK(s, iter, handle, false, NULL, NULL, NULL) {
1387         /* nbd_reply_chunk_iter_receive does all the work */
1388     }
1389 
1390     error_propagate(errp, iter.err);
1391     *request_ret = iter.request_ret;
1392     return iter.ret;
1393 }
1394 
1395 static int nbd_co_receive_cmdread_reply(BDRVNBDState *s, uint64_t handle,
1396                                         uint64_t offset, QEMUIOVector *qiov,
1397                                         int *request_ret, Error **errp)
1398 {
1399     NBDReplyChunkIter iter;
1400     NBDReply reply;
1401     void *payload = NULL;
1402     Error *local_err = NULL;
1403 
1404     NBD_FOREACH_REPLY_CHUNK(s, iter, handle, s->info.structured_reply,
1405                             qiov, &reply, &payload)
1406     {
1407         int ret;
1408         NBDStructuredReplyChunk *chunk = &reply.structured;
1409 
1410         assert(nbd_reply_is_structured(&reply));
1411 
1412         switch (chunk->type) {
1413         case NBD_REPLY_TYPE_OFFSET_DATA:
1414             /*
1415              * special cased in nbd_co_receive_one_chunk, data is already
1416              * in qiov
1417              */
1418             break;
1419         case NBD_REPLY_TYPE_OFFSET_HOLE:
1420             ret = nbd_parse_offset_hole_payload(s, &reply.structured, payload,
1421                                                 offset, qiov, &local_err);
1422             if (ret < 0) {
1423                 nbd_channel_error(s, ret);
1424                 nbd_iter_channel_error(&iter, ret, &local_err);
1425             }
1426             break;
1427         default:
1428             if (!nbd_reply_type_is_error(chunk->type)) {
1429                 /* not allowed reply type */
1430                 nbd_channel_error(s, -EINVAL);
1431                 error_setg(&local_err,
1432                            "Unexpected reply type: %d (%s) for CMD_READ",
1433                            chunk->type, nbd_reply_type_lookup(chunk->type));
1434                 nbd_iter_channel_error(&iter, -EINVAL, &local_err);
1435             }
1436         }
1437 
1438         g_free(payload);
1439         payload = NULL;
1440     }
1441 
1442     error_propagate(errp, iter.err);
1443     *request_ret = iter.request_ret;
1444     return iter.ret;
1445 }
1446 
1447 static int nbd_co_receive_blockstatus_reply(BDRVNBDState *s,
1448                                             uint64_t handle, uint64_t length,
1449                                             NBDExtent *extent,
1450                                             int *request_ret, Error **errp)
1451 {
1452     NBDReplyChunkIter iter;
1453     NBDReply reply;
1454     void *payload = NULL;
1455     Error *local_err = NULL;
1456     bool received = false;
1457 
1458     assert(!extent->length);
1459     NBD_FOREACH_REPLY_CHUNK(s, iter, handle, false, NULL, &reply, &payload) {
1460         int ret;
1461         NBDStructuredReplyChunk *chunk = &reply.structured;
1462 
1463         assert(nbd_reply_is_structured(&reply));
1464 
1465         switch (chunk->type) {
1466         case NBD_REPLY_TYPE_BLOCK_STATUS:
1467             if (received) {
1468                 nbd_channel_error(s, -EINVAL);
1469                 error_setg(&local_err, "Several BLOCK_STATUS chunks in reply");
1470                 nbd_iter_channel_error(&iter, -EINVAL, &local_err);
1471             }
1472             received = true;
1473 
1474             ret = nbd_parse_blockstatus_payload(s, &reply.structured,
1475                                                 payload, length, extent,
1476                                                 &local_err);
1477             if (ret < 0) {
1478                 nbd_channel_error(s, ret);
1479                 nbd_iter_channel_error(&iter, ret, &local_err);
1480             }
1481             break;
1482         default:
1483             if (!nbd_reply_type_is_error(chunk->type)) {
1484                 nbd_channel_error(s, -EINVAL);
1485                 error_setg(&local_err,
1486                            "Unexpected reply type: %d (%s) "
1487                            "for CMD_BLOCK_STATUS",
1488                            chunk->type, nbd_reply_type_lookup(chunk->type));
1489                 nbd_iter_channel_error(&iter, -EINVAL, &local_err);
1490             }
1491         }
1492 
1493         g_free(payload);
1494         payload = NULL;
1495     }
1496 
1497     if (!extent->length && !iter.request_ret) {
1498         error_setg(&local_err, "Server did not reply with any status extents");
1499         nbd_iter_channel_error(&iter, -EIO, &local_err);
1500     }
1501 
1502     error_propagate(errp, iter.err);
1503     *request_ret = iter.request_ret;
1504     return iter.ret;
1505 }
1506 
1507 static int nbd_co_request(BlockDriverState *bs, NBDRequest *request,
1508                           QEMUIOVector *write_qiov)
1509 {
1510     int ret, request_ret;
1511     Error *local_err = NULL;
1512     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
1513 
1514     assert(request->type != NBD_CMD_READ);
1515     if (write_qiov) {
1516         assert(request->type == NBD_CMD_WRITE);
1517         assert(request->len == iov_size(write_qiov->iov, write_qiov->niov));
1518     } else {
1519         assert(request->type != NBD_CMD_WRITE);
1520     }
1521 
1522     do {
1523         ret = nbd_co_send_request(bs, request, write_qiov);
1524         if (ret < 0) {
1525             continue;
1526         }
1527 
1528         ret = nbd_co_receive_return_code(s, request->handle,
1529                                          &request_ret, &local_err);
1530         if (local_err) {
1531             trace_nbd_co_request_fail(request->from, request->len,
1532                                       request->handle, request->flags,
1533                                       request->type,
1534                                       nbd_cmd_lookup(request->type),
1535                                       ret, error_get_pretty(local_err));
1536             error_free(local_err);
1537             local_err = NULL;
1538         }
1539     } while (ret < 0 && nbd_client_connecting_wait(s));
1540 
1541     return ret ? ret : request_ret;
1542 }
1543 
1544 static int nbd_client_co_preadv(BlockDriverState *bs, uint64_t offset,
1545                                 uint64_t bytes, QEMUIOVector *qiov, int flags)
1546 {
1547     int ret, request_ret;
1548     Error *local_err = NULL;
1549     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
1550     NBDRequest request = {
1551         .type = NBD_CMD_READ,
1552         .from = offset,
1553         .len = bytes,
1554     };
1555 
1556     assert(bytes <= NBD_MAX_BUFFER_SIZE);
1557     assert(!flags);
1558 
1559     if (!bytes) {
1560         return 0;
1561     }
1562     /*
1563      * Work around the fact that the block layer doesn't do
1564      * byte-accurate sizing yet - if the read exceeds the server's
1565      * advertised size because the block layer rounded size up, then
1566      * truncate the request to the server and tail-pad with zero.
1567      */
1568     if (offset >= s->info.size) {
1569         assert(bytes < BDRV_SECTOR_SIZE);
1570         qemu_iovec_memset(qiov, 0, 0, bytes);
1571         return 0;
1572     }
1573     if (offset + bytes > s->info.size) {
1574         uint64_t slop = offset + bytes - s->info.size;
1575 
1576         assert(slop < BDRV_SECTOR_SIZE);
1577         qemu_iovec_memset(qiov, bytes - slop, 0, slop);
1578         request.len -= slop;
1579     }
1580 
1581     do {
1582         ret = nbd_co_send_request(bs, &request, NULL);
1583         if (ret < 0) {
1584             continue;
1585         }
1586 
1587         ret = nbd_co_receive_cmdread_reply(s, request.handle, offset, qiov,
1588                                            &request_ret, &local_err);
1589         if (local_err) {
1590             trace_nbd_co_request_fail(request.from, request.len, request.handle,
1591                                       request.flags, request.type,
1592                                       nbd_cmd_lookup(request.type),
1593                                       ret, error_get_pretty(local_err));
1594             error_free(local_err);
1595             local_err = NULL;
1596         }
1597     } while (ret < 0 && nbd_client_connecting_wait(s));
1598 
1599     return ret ? ret : request_ret;
1600 }
1601 
1602 static int nbd_client_co_pwritev(BlockDriverState *bs, uint64_t offset,
1603                                  uint64_t bytes, QEMUIOVector *qiov, int flags)
1604 {
1605     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
1606     NBDRequest request = {
1607         .type = NBD_CMD_WRITE,
1608         .from = offset,
1609         .len = bytes,
1610     };
1611 
1612     assert(!(s->info.flags & NBD_FLAG_READ_ONLY));
1613     if (flags & BDRV_REQ_FUA) {
1614         assert(s->info.flags & NBD_FLAG_SEND_FUA);
1615         request.flags |= NBD_CMD_FLAG_FUA;
1616     }
1617 
1618     assert(bytes <= NBD_MAX_BUFFER_SIZE);
1619 
1620     if (!bytes) {
1621         return 0;
1622     }
1623     return nbd_co_request(bs, &request, qiov);
1624 }
1625 
1626 static int nbd_client_co_pwrite_zeroes(BlockDriverState *bs, int64_t offset,
1627                                        int bytes, BdrvRequestFlags flags)
1628 {
1629     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
1630     NBDRequest request = {
1631         .type = NBD_CMD_WRITE_ZEROES,
1632         .from = offset,
1633         .len = bytes,
1634     };
1635 
1636     assert(!(s->info.flags & NBD_FLAG_READ_ONLY));
1637     if (!(s->info.flags & NBD_FLAG_SEND_WRITE_ZEROES)) {
1638         return -ENOTSUP;
1639     }
1640 
1641     if (flags & BDRV_REQ_FUA) {
1642         assert(s->info.flags & NBD_FLAG_SEND_FUA);
1643         request.flags |= NBD_CMD_FLAG_FUA;
1644     }
1645     if (!(flags & BDRV_REQ_MAY_UNMAP)) {
1646         request.flags |= NBD_CMD_FLAG_NO_HOLE;
1647     }
1648     if (flags & BDRV_REQ_NO_FALLBACK) {
1649         assert(s->info.flags & NBD_FLAG_SEND_FAST_ZERO);
1650         request.flags |= NBD_CMD_FLAG_FAST_ZERO;
1651     }
1652 
1653     if (!bytes) {
1654         return 0;
1655     }
1656     return nbd_co_request(bs, &request, NULL);
1657 }
1658 
1659 static int nbd_client_co_flush(BlockDriverState *bs)
1660 {
1661     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
1662     NBDRequest request = { .type = NBD_CMD_FLUSH };
1663 
1664     if (!(s->info.flags & NBD_FLAG_SEND_FLUSH)) {
1665         return 0;
1666     }
1667 
1668     request.from = 0;
1669     request.len = 0;
1670 
1671     return nbd_co_request(bs, &request, NULL);
1672 }
1673 
1674 static int nbd_client_co_pdiscard(BlockDriverState *bs, int64_t offset,
1675                                   int bytes)
1676 {
1677     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
1678     NBDRequest request = {
1679         .type = NBD_CMD_TRIM,
1680         .from = offset,
1681         .len = bytes,
1682     };
1683 
1684     assert(!(s->info.flags & NBD_FLAG_READ_ONLY));
1685     if (!(s->info.flags & NBD_FLAG_SEND_TRIM) || !bytes) {
1686         return 0;
1687     }
1688 
1689     return nbd_co_request(bs, &request, NULL);
1690 }
1691 
1692 static int coroutine_fn nbd_client_co_block_status(
1693         BlockDriverState *bs, bool want_zero, int64_t offset, int64_t bytes,
1694         int64_t *pnum, int64_t *map, BlockDriverState **file)
1695 {
1696     int ret, request_ret;
1697     NBDExtent extent = { 0 };
1698     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
1699     Error *local_err = NULL;
1700 
1701     NBDRequest request = {
1702         .type = NBD_CMD_BLOCK_STATUS,
1703         .from = offset,
1704         .len = MIN(QEMU_ALIGN_DOWN(INT_MAX, bs->bl.request_alignment),
1705                    MIN(bytes, s->info.size - offset)),
1706         .flags = NBD_CMD_FLAG_REQ_ONE,
1707     };
1708 
1709     if (!s->info.base_allocation) {
1710         *pnum = bytes;
1711         *map = offset;
1712         *file = bs;
1713         return BDRV_BLOCK_DATA | BDRV_BLOCK_OFFSET_VALID;
1714     }
1715 
1716     /*
1717      * Work around the fact that the block layer doesn't do
1718      * byte-accurate sizing yet - if the status request exceeds the
1719      * server's advertised size because the block layer rounded size
1720      * up, we truncated the request to the server (above), or are
1721      * called on just the hole.
1722      */
1723     if (offset >= s->info.size) {
1724         *pnum = bytes;
1725         assert(bytes < BDRV_SECTOR_SIZE);
1726         /* Intentionally don't report offset_valid for the hole */
1727         return BDRV_BLOCK_ZERO;
1728     }
1729 
1730     if (s->info.min_block) {
1731         assert(QEMU_IS_ALIGNED(request.len, s->info.min_block));
1732     }
1733     do {
1734         ret = nbd_co_send_request(bs, &request, NULL);
1735         if (ret < 0) {
1736             continue;
1737         }
1738 
1739         ret = nbd_co_receive_blockstatus_reply(s, request.handle, bytes,
1740                                                &extent, &request_ret,
1741                                                &local_err);
1742         if (local_err) {
1743             trace_nbd_co_request_fail(request.from, request.len, request.handle,
1744                                       request.flags, request.type,
1745                                       nbd_cmd_lookup(request.type),
1746                                       ret, error_get_pretty(local_err));
1747             error_free(local_err);
1748             local_err = NULL;
1749         }
1750     } while (ret < 0 && nbd_client_connecting_wait(s));
1751 
1752     if (ret < 0 || request_ret < 0) {
1753         return ret ? ret : request_ret;
1754     }
1755 
1756     assert(extent.length);
1757     *pnum = extent.length;
1758     *map = offset;
1759     *file = bs;
1760     return (extent.flags & NBD_STATE_HOLE ? 0 : BDRV_BLOCK_DATA) |
1761         (extent.flags & NBD_STATE_ZERO ? BDRV_BLOCK_ZERO : 0) |
1762         BDRV_BLOCK_OFFSET_VALID;
1763 }
1764 
1765 static int nbd_client_reopen_prepare(BDRVReopenState *state,
1766                                      BlockReopenQueue *queue, Error **errp)
1767 {
1768     BDRVNBDState *s = (BDRVNBDState *)state->bs->opaque;
1769 
1770     if ((state->flags & BDRV_O_RDWR) && (s->info.flags & NBD_FLAG_READ_ONLY)) {
1771         error_setg(errp, "Can't reopen read-only NBD mount as read/write");
1772         return -EACCES;
1773     }
1774     return 0;
1775 }
1776 
1777 static void nbd_yank(void *opaque)
1778 {
1779     BlockDriverState *bs = opaque;
1780     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
1781 
1782     qatomic_store_release(&s->state, NBD_CLIENT_QUIT);
1783     qio_channel_shutdown(QIO_CHANNEL(s->sioc), QIO_CHANNEL_SHUTDOWN_BOTH, NULL);
1784 }
1785 
1786 static void nbd_client_close(BlockDriverState *bs)
1787 {
1788     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
1789     NBDRequest request = { .type = NBD_CMD_DISC };
1790 
1791     if (s->ioc) {
1792         nbd_send_request(s->ioc, &request);
1793     }
1794 
1795     nbd_teardown_connection(bs);
1796 }
1797 
1798 static int nbd_establish_connection(BlockDriverState *bs,
1799                                     SocketAddress *saddr,
1800                                     Error **errp)
1801 {
1802     ERRP_GUARD();
1803     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
1804 
1805     s->sioc = qio_channel_socket_new();
1806     qio_channel_set_name(QIO_CHANNEL(s->sioc), "nbd-client");
1807 
1808     qio_channel_socket_connect_sync(s->sioc, saddr, errp);
1809     if (*errp) {
1810         object_unref(OBJECT(s->sioc));
1811         s->sioc = NULL;
1812         return -1;
1813     }
1814 
1815     yank_register_function(BLOCKDEV_YANK_INSTANCE(bs->node_name), nbd_yank, bs);
1816     qio_channel_set_delay(QIO_CHANNEL(s->sioc), false);
1817 
1818     return 0;
1819 }
1820 
1821 /* nbd_client_handshake takes ownership on s->sioc. On failure it's unref'ed. */
1822 static int nbd_client_handshake(BlockDriverState *bs, Error **errp)
1823 {
1824     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
1825     AioContext *aio_context = bdrv_get_aio_context(bs);
1826     int ret;
1827 
1828     trace_nbd_client_handshake(s->export);
1829     qio_channel_set_blocking(QIO_CHANNEL(s->sioc), false, NULL);
1830     qio_channel_attach_aio_context(QIO_CHANNEL(s->sioc), aio_context);
1831 
1832     s->info.request_sizes = true;
1833     s->info.structured_reply = true;
1834     s->info.base_allocation = true;
1835     s->info.x_dirty_bitmap = g_strdup(s->x_dirty_bitmap);
1836     s->info.name = g_strdup(s->export ?: "");
1837     ret = nbd_receive_negotiate(aio_context, QIO_CHANNEL(s->sioc), s->tlscreds,
1838                                 s->hostname, &s->ioc, &s->info, errp);
1839     g_free(s->info.x_dirty_bitmap);
1840     g_free(s->info.name);
1841     if (ret < 0) {
1842         yank_unregister_function(BLOCKDEV_YANK_INSTANCE(bs->node_name),
1843                                  nbd_yank, bs);
1844         object_unref(OBJECT(s->sioc));
1845         s->sioc = NULL;
1846         return ret;
1847     }
1848     if (s->x_dirty_bitmap) {
1849         if (!s->info.base_allocation) {
1850             error_setg(errp, "requested x-dirty-bitmap %s not found",
1851                        s->x_dirty_bitmap);
1852             ret = -EINVAL;
1853             goto fail;
1854         }
1855         if (strcmp(s->x_dirty_bitmap, "qemu:allocation-depth") == 0) {
1856             s->alloc_depth = true;
1857         }
1858     }
1859     if (s->info.flags & NBD_FLAG_READ_ONLY) {
1860         ret = bdrv_apply_auto_read_only(bs, "NBD export is read-only", errp);
1861         if (ret < 0) {
1862             goto fail;
1863         }
1864     }
1865     if (s->info.flags & NBD_FLAG_SEND_FUA) {
1866         bs->supported_write_flags = BDRV_REQ_FUA;
1867         bs->supported_zero_flags |= BDRV_REQ_FUA;
1868     }
1869     if (s->info.flags & NBD_FLAG_SEND_WRITE_ZEROES) {
1870         bs->supported_zero_flags |= BDRV_REQ_MAY_UNMAP;
1871         if (s->info.flags & NBD_FLAG_SEND_FAST_ZERO) {
1872             bs->supported_zero_flags |= BDRV_REQ_NO_FALLBACK;
1873         }
1874     }
1875 
1876     if (!s->ioc) {
1877         s->ioc = QIO_CHANNEL(s->sioc);
1878         object_ref(OBJECT(s->ioc));
1879     }
1880 
1881     trace_nbd_client_handshake_success(s->export);
1882 
1883     return 0;
1884 
1885  fail:
1886     /*
1887      * We have connected, but must fail for other reasons.
1888      * Send NBD_CMD_DISC as a courtesy to the server.
1889      */
1890     {
1891         NBDRequest request = { .type = NBD_CMD_DISC };
1892 
1893         nbd_send_request(s->ioc ?: QIO_CHANNEL(s->sioc), &request);
1894 
1895         yank_unregister_function(BLOCKDEV_YANK_INSTANCE(bs->node_name),
1896                                  nbd_yank, bs);
1897         object_unref(OBJECT(s->sioc));
1898         s->sioc = NULL;
1899 
1900         return ret;
1901     }
1902 }
1903 
1904 /*
1905  * Parse nbd_open options
1906  */
1907 
1908 static int nbd_parse_uri(const char *filename, QDict *options)
1909 {
1910     URI *uri;
1911     const char *p;
1912     QueryParams *qp = NULL;
1913     int ret = 0;
1914     bool is_unix;
1915 
1916     uri = uri_parse(filename);
1917     if (!uri) {
1918         return -EINVAL;
1919     }
1920 
1921     /* transport */
1922     if (!g_strcmp0(uri->scheme, "nbd")) {
1923         is_unix = false;
1924     } else if (!g_strcmp0(uri->scheme, "nbd+tcp")) {
1925         is_unix = false;
1926     } else if (!g_strcmp0(uri->scheme, "nbd+unix")) {
1927         is_unix = true;
1928     } else {
1929         ret = -EINVAL;
1930         goto out;
1931     }
1932 
1933     p = uri->path ? uri->path : "";
1934     if (p[0] == '/') {
1935         p++;
1936     }
1937     if (p[0]) {
1938         qdict_put_str(options, "export", p);
1939     }
1940 
1941     qp = query_params_parse(uri->query);
1942     if (qp->n > 1 || (is_unix && !qp->n) || (!is_unix && qp->n)) {
1943         ret = -EINVAL;
1944         goto out;
1945     }
1946 
1947     if (is_unix) {
1948         /* nbd+unix:///export?socket=path */
1949         if (uri->server || uri->port || strcmp(qp->p[0].name, "socket")) {
1950             ret = -EINVAL;
1951             goto out;
1952         }
1953         qdict_put_str(options, "server.type", "unix");
1954         qdict_put_str(options, "server.path", qp->p[0].value);
1955     } else {
1956         QString *host;
1957         char *port_str;
1958 
1959         /* nbd[+tcp]://host[:port]/export */
1960         if (!uri->server) {
1961             ret = -EINVAL;
1962             goto out;
1963         }
1964 
1965         /* strip braces from literal IPv6 address */
1966         if (uri->server[0] == '[') {
1967             host = qstring_from_substr(uri->server, 1,
1968                                        strlen(uri->server) - 1);
1969         } else {
1970             host = qstring_from_str(uri->server);
1971         }
1972 
1973         qdict_put_str(options, "server.type", "inet");
1974         qdict_put(options, "server.host", host);
1975 
1976         port_str = g_strdup_printf("%d", uri->port ?: NBD_DEFAULT_PORT);
1977         qdict_put_str(options, "server.port", port_str);
1978         g_free(port_str);
1979     }
1980 
1981 out:
1982     if (qp) {
1983         query_params_free(qp);
1984     }
1985     uri_free(uri);
1986     return ret;
1987 }
1988 
1989 static bool nbd_has_filename_options_conflict(QDict *options, Error **errp)
1990 {
1991     const QDictEntry *e;
1992 
1993     for (e = qdict_first(options); e; e = qdict_next(options, e)) {
1994         if (!strcmp(e->key, "host") ||
1995             !strcmp(e->key, "port") ||
1996             !strcmp(e->key, "path") ||
1997             !strcmp(e->key, "export") ||
1998             strstart(e->key, "server.", NULL))
1999         {
2000             error_setg(errp, "Option '%s' cannot be used with a file name",
2001                        e->key);
2002             return true;
2003         }
2004     }
2005 
2006     return false;
2007 }
2008 
2009 static void nbd_parse_filename(const char *filename, QDict *options,
2010                                Error **errp)
2011 {
2012     g_autofree char *file = NULL;
2013     char *export_name;
2014     const char *host_spec;
2015     const char *unixpath;
2016 
2017     if (nbd_has_filename_options_conflict(options, errp)) {
2018         return;
2019     }
2020 
2021     if (strstr(filename, "://")) {
2022         int ret = nbd_parse_uri(filename, options);
2023         if (ret < 0) {
2024             error_setg(errp, "No valid URL specified");
2025         }
2026         return;
2027     }
2028 
2029     file = g_strdup(filename);
2030 
2031     export_name = strstr(file, EN_OPTSTR);
2032     if (export_name) {
2033         if (export_name[strlen(EN_OPTSTR)] == 0) {
2034             return;
2035         }
2036         export_name[0] = 0; /* truncate 'file' */
2037         export_name += strlen(EN_OPTSTR);
2038 
2039         qdict_put_str(options, "export", export_name);
2040     }
2041 
2042     /* extract the host_spec - fail if it's not nbd:... */
2043     if (!strstart(file, "nbd:", &host_spec)) {
2044         error_setg(errp, "File name string for NBD must start with 'nbd:'");
2045         return;
2046     }
2047 
2048     if (!*host_spec) {
2049         return;
2050     }
2051 
2052     /* are we a UNIX or TCP socket? */
2053     if (strstart(host_spec, "unix:", &unixpath)) {
2054         qdict_put_str(options, "server.type", "unix");
2055         qdict_put_str(options, "server.path", unixpath);
2056     } else {
2057         InetSocketAddress *addr = g_new(InetSocketAddress, 1);
2058 
2059         if (inet_parse(addr, host_spec, errp)) {
2060             goto out_inet;
2061         }
2062 
2063         qdict_put_str(options, "server.type", "inet");
2064         qdict_put_str(options, "server.host", addr->host);
2065         qdict_put_str(options, "server.port", addr->port);
2066     out_inet:
2067         qapi_free_InetSocketAddress(addr);
2068     }
2069 }
2070 
2071 static bool nbd_process_legacy_socket_options(QDict *output_options,
2072                                               QemuOpts *legacy_opts,
2073                                               Error **errp)
2074 {
2075     const char *path = qemu_opt_get(legacy_opts, "path");
2076     const char *host = qemu_opt_get(legacy_opts, "host");
2077     const char *port = qemu_opt_get(legacy_opts, "port");
2078     const QDictEntry *e;
2079 
2080     if (!path && !host && !port) {
2081         return true;
2082     }
2083 
2084     for (e = qdict_first(output_options); e; e = qdict_next(output_options, e))
2085     {
2086         if (strstart(e->key, "server.", NULL)) {
2087             error_setg(errp, "Cannot use 'server' and path/host/port at the "
2088                        "same time");
2089             return false;
2090         }
2091     }
2092 
2093     if (path && host) {
2094         error_setg(errp, "path and host may not be used at the same time");
2095         return false;
2096     } else if (path) {
2097         if (port) {
2098             error_setg(errp, "port may not be used without host");
2099             return false;
2100         }
2101 
2102         qdict_put_str(output_options, "server.type", "unix");
2103         qdict_put_str(output_options, "server.path", path);
2104     } else if (host) {
2105         qdict_put_str(output_options, "server.type", "inet");
2106         qdict_put_str(output_options, "server.host", host);
2107         qdict_put_str(output_options, "server.port",
2108                       port ?: stringify(NBD_DEFAULT_PORT));
2109     }
2110 
2111     return true;
2112 }
2113 
2114 static SocketAddress *nbd_config(BDRVNBDState *s, QDict *options,
2115                                  Error **errp)
2116 {
2117     SocketAddress *saddr = NULL;
2118     QDict *addr = NULL;
2119     Visitor *iv = NULL;
2120 
2121     qdict_extract_subqdict(options, &addr, "server.");
2122     if (!qdict_size(addr)) {
2123         error_setg(errp, "NBD server address missing");
2124         goto done;
2125     }
2126 
2127     iv = qobject_input_visitor_new_flat_confused(addr, errp);
2128     if (!iv) {
2129         goto done;
2130     }
2131 
2132     if (!visit_type_SocketAddress(iv, NULL, &saddr, errp)) {
2133         goto done;
2134     }
2135 
2136 done:
2137     qobject_unref(addr);
2138     visit_free(iv);
2139     return saddr;
2140 }
2141 
2142 static QCryptoTLSCreds *nbd_get_tls_creds(const char *id, Error **errp)
2143 {
2144     Object *obj;
2145     QCryptoTLSCreds *creds;
2146 
2147     obj = object_resolve_path_component(
2148         object_get_objects_root(), id);
2149     if (!obj) {
2150         error_setg(errp, "No TLS credentials with id '%s'",
2151                    id);
2152         return NULL;
2153     }
2154     creds = (QCryptoTLSCreds *)
2155         object_dynamic_cast(obj, TYPE_QCRYPTO_TLS_CREDS);
2156     if (!creds) {
2157         error_setg(errp, "Object with id '%s' is not TLS credentials",
2158                    id);
2159         return NULL;
2160     }
2161 
2162     if (creds->endpoint != QCRYPTO_TLS_CREDS_ENDPOINT_CLIENT) {
2163         error_setg(errp,
2164                    "Expecting TLS credentials with a client endpoint");
2165         return NULL;
2166     }
2167     object_ref(obj);
2168     return creds;
2169 }
2170 
2171 
2172 static QemuOptsList nbd_runtime_opts = {
2173     .name = "nbd",
2174     .head = QTAILQ_HEAD_INITIALIZER(nbd_runtime_opts.head),
2175     .desc = {
2176         {
2177             .name = "host",
2178             .type = QEMU_OPT_STRING,
2179             .help = "TCP host to connect to",
2180         },
2181         {
2182             .name = "port",
2183             .type = QEMU_OPT_STRING,
2184             .help = "TCP port to connect to",
2185         },
2186         {
2187             .name = "path",
2188             .type = QEMU_OPT_STRING,
2189             .help = "Unix socket path to connect to",
2190         },
2191         {
2192             .name = "export",
2193             .type = QEMU_OPT_STRING,
2194             .help = "Name of the NBD export to open",
2195         },
2196         {
2197             .name = "tls-creds",
2198             .type = QEMU_OPT_STRING,
2199             .help = "ID of the TLS credentials to use",
2200         },
2201         {
2202             .name = "x-dirty-bitmap",
2203             .type = QEMU_OPT_STRING,
2204             .help = "experimental: expose named dirty bitmap in place of "
2205                     "block status",
2206         },
2207         {
2208             .name = "reconnect-delay",
2209             .type = QEMU_OPT_NUMBER,
2210             .help = "On an unexpected disconnect, the nbd client tries to "
2211                     "connect again until succeeding or encountering a serious "
2212                     "error.  During the first @reconnect-delay seconds, all "
2213                     "requests are paused and will be rerun on a successful "
2214                     "reconnect. After that time, any delayed requests and all "
2215                     "future requests before a successful reconnect will "
2216                     "immediately fail. Default 0",
2217         },
2218         { /* end of list */ }
2219     },
2220 };
2221 
2222 static int nbd_process_options(BlockDriverState *bs, QDict *options,
2223                                Error **errp)
2224 {
2225     BDRVNBDState *s = bs->opaque;
2226     QemuOpts *opts;
2227     int ret = -EINVAL;
2228 
2229     opts = qemu_opts_create(&nbd_runtime_opts, NULL, 0, &error_abort);
2230     if (!qemu_opts_absorb_qdict(opts, options, errp)) {
2231         goto error;
2232     }
2233 
2234     /* Translate @host, @port, and @path to a SocketAddress */
2235     if (!nbd_process_legacy_socket_options(options, opts, errp)) {
2236         goto error;
2237     }
2238 
2239     /* Pop the config into our state object. Exit if invalid. */
2240     s->saddr = nbd_config(s, options, errp);
2241     if (!s->saddr) {
2242         goto error;
2243     }
2244 
2245     s->export = g_strdup(qemu_opt_get(opts, "export"));
2246     if (s->export && strlen(s->export) > NBD_MAX_STRING_SIZE) {
2247         error_setg(errp, "export name too long to send to server");
2248         goto error;
2249     }
2250 
2251     s->tlscredsid = g_strdup(qemu_opt_get(opts, "tls-creds"));
2252     if (s->tlscredsid) {
2253         s->tlscreds = nbd_get_tls_creds(s->tlscredsid, errp);
2254         if (!s->tlscreds) {
2255             goto error;
2256         }
2257 
2258         /* TODO SOCKET_ADDRESS_KIND_FD where fd has AF_INET or AF_INET6 */
2259         if (s->saddr->type != SOCKET_ADDRESS_TYPE_INET) {
2260             error_setg(errp, "TLS only supported over IP sockets");
2261             goto error;
2262         }
2263         s->hostname = s->saddr->u.inet.host;
2264     }
2265 
2266     s->x_dirty_bitmap = g_strdup(qemu_opt_get(opts, "x-dirty-bitmap"));
2267     if (s->x_dirty_bitmap && strlen(s->x_dirty_bitmap) > NBD_MAX_STRING_SIZE) {
2268         error_setg(errp, "x-dirty-bitmap query too long to send to server");
2269         goto error;
2270     }
2271 
2272     s->reconnect_delay = qemu_opt_get_number(opts, "reconnect-delay", 0);
2273 
2274     ret = 0;
2275 
2276  error:
2277     if (ret < 0) {
2278         nbd_clear_bdrvstate(s);
2279     }
2280     qemu_opts_del(opts);
2281     return ret;
2282 }
2283 
2284 static int nbd_open(BlockDriverState *bs, QDict *options, int flags,
2285                     Error **errp)
2286 {
2287     int ret;
2288     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
2289 
2290     ret = nbd_process_options(bs, options, errp);
2291     if (ret < 0) {
2292         return ret;
2293     }
2294 
2295     s->bs = bs;
2296     qemu_co_mutex_init(&s->send_mutex);
2297     qemu_co_queue_init(&s->free_sema);
2298 
2299     if (!yank_register_instance(BLOCKDEV_YANK_INSTANCE(bs->node_name), errp)) {
2300         return -EEXIST;
2301     }
2302 
2303     /*
2304      * establish TCP connection, return error if it fails
2305      * TODO: Configurable retry-until-timeout behaviour.
2306      */
2307     if (nbd_establish_connection(bs, s->saddr, errp) < 0) {
2308         yank_unregister_instance(BLOCKDEV_YANK_INSTANCE(bs->node_name));
2309         return -ECONNREFUSED;
2310     }
2311 
2312     ret = nbd_client_handshake(bs, errp);
2313     if (ret < 0) {
2314         yank_unregister_instance(BLOCKDEV_YANK_INSTANCE(bs->node_name));
2315         nbd_clear_bdrvstate(s);
2316         return ret;
2317     }
2318     /* successfully connected */
2319     s->state = NBD_CLIENT_CONNECTED;
2320 
2321     nbd_init_connect_thread(s);
2322 
2323     s->connection_co = qemu_coroutine_create(nbd_connection_entry, s);
2324     bdrv_inc_in_flight(bs);
2325     aio_co_schedule(bdrv_get_aio_context(bs), s->connection_co);
2326 
2327     return 0;
2328 }
2329 
2330 static int nbd_co_flush(BlockDriverState *bs)
2331 {
2332     return nbd_client_co_flush(bs);
2333 }
2334 
2335 static void nbd_refresh_limits(BlockDriverState *bs, Error **errp)
2336 {
2337     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
2338     uint32_t min = s->info.min_block;
2339     uint32_t max = MIN_NON_ZERO(NBD_MAX_BUFFER_SIZE, s->info.max_block);
2340 
2341     /*
2342      * If the server did not advertise an alignment:
2343      * - a size that is not sector-aligned implies that an alignment
2344      *   of 1 can be used to access those tail bytes
2345      * - advertisement of block status requires an alignment of 1, so
2346      *   that we don't violate block layer constraints that block
2347      *   status is always aligned (as we can't control whether the
2348      *   server will report sub-sector extents, such as a hole at EOF
2349      *   on an unaligned POSIX file)
2350      * - otherwise, assume the server is so old that we are safer avoiding
2351      *   sub-sector requests
2352      */
2353     if (!min) {
2354         min = (!QEMU_IS_ALIGNED(s->info.size, BDRV_SECTOR_SIZE) ||
2355                s->info.base_allocation) ? 1 : BDRV_SECTOR_SIZE;
2356     }
2357 
2358     bs->bl.request_alignment = min;
2359     bs->bl.max_pdiscard = QEMU_ALIGN_DOWN(INT_MAX, min);
2360     bs->bl.max_pwrite_zeroes = max;
2361     bs->bl.max_transfer = max;
2362 
2363     if (s->info.opt_block &&
2364         s->info.opt_block > bs->bl.opt_transfer) {
2365         bs->bl.opt_transfer = s->info.opt_block;
2366     }
2367 }
2368 
2369 static void nbd_close(BlockDriverState *bs)
2370 {
2371     BDRVNBDState *s = bs->opaque;
2372 
2373     nbd_client_close(bs);
2374     yank_unregister_instance(BLOCKDEV_YANK_INSTANCE(bs->node_name));
2375     nbd_clear_bdrvstate(s);
2376 }
2377 
2378 /*
2379  * NBD cannot truncate, but if the caller asks to truncate to the same size, or
2380  * to a smaller size with exact=false, there is no reason to fail the
2381  * operation.
2382  *
2383  * Preallocation mode is ignored since it does not seems useful to fail when
2384  * we never change anything.
2385  */
2386 static int coroutine_fn nbd_co_truncate(BlockDriverState *bs, int64_t offset,
2387                                         bool exact, PreallocMode prealloc,
2388                                         BdrvRequestFlags flags, Error **errp)
2389 {
2390     BDRVNBDState *s = bs->opaque;
2391 
2392     if (offset != s->info.size && exact) {
2393         error_setg(errp, "Cannot resize NBD nodes");
2394         return -ENOTSUP;
2395     }
2396 
2397     if (offset > s->info.size) {
2398         error_setg(errp, "Cannot grow NBD nodes");
2399         return -EINVAL;
2400     }
2401 
2402     return 0;
2403 }
2404 
2405 static int64_t nbd_getlength(BlockDriverState *bs)
2406 {
2407     BDRVNBDState *s = bs->opaque;
2408 
2409     return s->info.size;
2410 }
2411 
2412 static void nbd_refresh_filename(BlockDriverState *bs)
2413 {
2414     BDRVNBDState *s = bs->opaque;
2415     const char *host = NULL, *port = NULL, *path = NULL;
2416     size_t len = 0;
2417 
2418     if (s->saddr->type == SOCKET_ADDRESS_TYPE_INET) {
2419         const InetSocketAddress *inet = &s->saddr->u.inet;
2420         if (!inet->has_ipv4 && !inet->has_ipv6 && !inet->has_to) {
2421             host = inet->host;
2422             port = inet->port;
2423         }
2424     } else if (s->saddr->type == SOCKET_ADDRESS_TYPE_UNIX) {
2425         path = s->saddr->u.q_unix.path;
2426     } /* else can't represent as pseudo-filename */
2427 
2428     if (path && s->export) {
2429         len = snprintf(bs->exact_filename, sizeof(bs->exact_filename),
2430                        "nbd+unix:///%s?socket=%s", s->export, path);
2431     } else if (path && !s->export) {
2432         len = snprintf(bs->exact_filename, sizeof(bs->exact_filename),
2433                        "nbd+unix://?socket=%s", path);
2434     } else if (host && s->export) {
2435         len = snprintf(bs->exact_filename, sizeof(bs->exact_filename),
2436                        "nbd://%s:%s/%s", host, port, s->export);
2437     } else if (host && !s->export) {
2438         len = snprintf(bs->exact_filename, sizeof(bs->exact_filename),
2439                        "nbd://%s:%s", host, port);
2440     }
2441     if (len >= sizeof(bs->exact_filename)) {
2442         /* Name is too long to represent exactly, so leave it empty. */
2443         bs->exact_filename[0] = '\0';
2444     }
2445 }
2446 
2447 static char *nbd_dirname(BlockDriverState *bs, Error **errp)
2448 {
2449     /* The generic bdrv_dirname() implementation is able to work out some
2450      * directory name for NBD nodes, but that would be wrong. So far there is no
2451      * specification for how "export paths" would work, so NBD does not have
2452      * directory names. */
2453     error_setg(errp, "Cannot generate a base directory for NBD nodes");
2454     return NULL;
2455 }
2456 
2457 static const char *const nbd_strong_runtime_opts[] = {
2458     "path",
2459     "host",
2460     "port",
2461     "export",
2462     "tls-creds",
2463     "server.",
2464 
2465     NULL
2466 };
2467 
2468 static void nbd_cancel_in_flight(BlockDriverState *bs)
2469 {
2470     BDRVNBDState *s = (BDRVNBDState *)bs->opaque;
2471 
2472     reconnect_delay_timer_del(s);
2473 
2474     if (s->state == NBD_CLIENT_CONNECTING_WAIT) {
2475         s->state = NBD_CLIENT_CONNECTING_NOWAIT;
2476         qemu_co_queue_restart_all(&s->free_sema);
2477     }
2478 }
2479 
2480 static BlockDriver bdrv_nbd = {
2481     .format_name                = "nbd",
2482     .protocol_name              = "nbd",
2483     .instance_size              = sizeof(BDRVNBDState),
2484     .bdrv_parse_filename        = nbd_parse_filename,
2485     .bdrv_co_create_opts        = bdrv_co_create_opts_simple,
2486     .create_opts                = &bdrv_create_opts_simple,
2487     .bdrv_file_open             = nbd_open,
2488     .bdrv_reopen_prepare        = nbd_client_reopen_prepare,
2489     .bdrv_co_preadv             = nbd_client_co_preadv,
2490     .bdrv_co_pwritev            = nbd_client_co_pwritev,
2491     .bdrv_co_pwrite_zeroes      = nbd_client_co_pwrite_zeroes,
2492     .bdrv_close                 = nbd_close,
2493     .bdrv_co_flush_to_os        = nbd_co_flush,
2494     .bdrv_co_pdiscard           = nbd_client_co_pdiscard,
2495     .bdrv_refresh_limits        = nbd_refresh_limits,
2496     .bdrv_co_truncate           = nbd_co_truncate,
2497     .bdrv_getlength             = nbd_getlength,
2498     .bdrv_detach_aio_context    = nbd_client_detach_aio_context,
2499     .bdrv_attach_aio_context    = nbd_client_attach_aio_context,
2500     .bdrv_co_drain_begin        = nbd_client_co_drain_begin,
2501     .bdrv_co_drain_end          = nbd_client_co_drain_end,
2502     .bdrv_refresh_filename      = nbd_refresh_filename,
2503     .bdrv_co_block_status       = nbd_client_co_block_status,
2504     .bdrv_dirname               = nbd_dirname,
2505     .strong_runtime_opts        = nbd_strong_runtime_opts,
2506     .bdrv_cancel_in_flight      = nbd_cancel_in_flight,
2507 };
2508 
2509 static BlockDriver bdrv_nbd_tcp = {
2510     .format_name                = "nbd",
2511     .protocol_name              = "nbd+tcp",
2512     .instance_size              = sizeof(BDRVNBDState),
2513     .bdrv_parse_filename        = nbd_parse_filename,
2514     .bdrv_co_create_opts        = bdrv_co_create_opts_simple,
2515     .create_opts                = &bdrv_create_opts_simple,
2516     .bdrv_file_open             = nbd_open,
2517     .bdrv_reopen_prepare        = nbd_client_reopen_prepare,
2518     .bdrv_co_preadv             = nbd_client_co_preadv,
2519     .bdrv_co_pwritev            = nbd_client_co_pwritev,
2520     .bdrv_co_pwrite_zeroes      = nbd_client_co_pwrite_zeroes,
2521     .bdrv_close                 = nbd_close,
2522     .bdrv_co_flush_to_os        = nbd_co_flush,
2523     .bdrv_co_pdiscard           = nbd_client_co_pdiscard,
2524     .bdrv_refresh_limits        = nbd_refresh_limits,
2525     .bdrv_co_truncate           = nbd_co_truncate,
2526     .bdrv_getlength             = nbd_getlength,
2527     .bdrv_detach_aio_context    = nbd_client_detach_aio_context,
2528     .bdrv_attach_aio_context    = nbd_client_attach_aio_context,
2529     .bdrv_co_drain_begin        = nbd_client_co_drain_begin,
2530     .bdrv_co_drain_end          = nbd_client_co_drain_end,
2531     .bdrv_refresh_filename      = nbd_refresh_filename,
2532     .bdrv_co_block_status       = nbd_client_co_block_status,
2533     .bdrv_dirname               = nbd_dirname,
2534     .strong_runtime_opts        = nbd_strong_runtime_opts,
2535     .bdrv_cancel_in_flight      = nbd_cancel_in_flight,
2536 };
2537 
2538 static BlockDriver bdrv_nbd_unix = {
2539     .format_name                = "nbd",
2540     .protocol_name              = "nbd+unix",
2541     .instance_size              = sizeof(BDRVNBDState),
2542     .bdrv_parse_filename        = nbd_parse_filename,
2543     .bdrv_co_create_opts        = bdrv_co_create_opts_simple,
2544     .create_opts                = &bdrv_create_opts_simple,
2545     .bdrv_file_open             = nbd_open,
2546     .bdrv_reopen_prepare        = nbd_client_reopen_prepare,
2547     .bdrv_co_preadv             = nbd_client_co_preadv,
2548     .bdrv_co_pwritev            = nbd_client_co_pwritev,
2549     .bdrv_co_pwrite_zeroes      = nbd_client_co_pwrite_zeroes,
2550     .bdrv_close                 = nbd_close,
2551     .bdrv_co_flush_to_os        = nbd_co_flush,
2552     .bdrv_co_pdiscard           = nbd_client_co_pdiscard,
2553     .bdrv_refresh_limits        = nbd_refresh_limits,
2554     .bdrv_co_truncate           = nbd_co_truncate,
2555     .bdrv_getlength             = nbd_getlength,
2556     .bdrv_detach_aio_context    = nbd_client_detach_aio_context,
2557     .bdrv_attach_aio_context    = nbd_client_attach_aio_context,
2558     .bdrv_co_drain_begin        = nbd_client_co_drain_begin,
2559     .bdrv_co_drain_end          = nbd_client_co_drain_end,
2560     .bdrv_refresh_filename      = nbd_refresh_filename,
2561     .bdrv_co_block_status       = nbd_client_co_block_status,
2562     .bdrv_dirname               = nbd_dirname,
2563     .strong_runtime_opts        = nbd_strong_runtime_opts,
2564     .bdrv_cancel_in_flight      = nbd_cancel_in_flight,
2565 };
2566 
2567 static void bdrv_nbd_init(void)
2568 {
2569     bdrv_register(&bdrv_nbd);
2570     bdrv_register(&bdrv_nbd_tcp);
2571     bdrv_register(&bdrv_nbd_unix);
2572 }
2573 
2574 block_init(bdrv_nbd_init);
2575