block/nbd-client.c

   1 /*
   2  * QEMU Block driver for  NBD
   3  *
   4  * Copyright (C) 2016 Red Hat, Inc.
   5  * Copyright (C) 2008 Bull S.A.S.
   6  *     Author: Laurent Vivier <[email protected]>
   7  *
   8  * Some parts:
   9  *    Copyright (C) 2007 Anthony Liguori <[email protected]>
  10  *
  11  * Permission is hereby granted, free of charge, to any person obtaining a copy
  12  * of this software and associated documentation files (the "Software"), to deal
  13  * in the Software without restriction, including without limitation the rights
  14  * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
  15  * copies of the Software, and to permit persons to whom the Software is
  16  * furnished to do so, subject to the following conditions:
  17  *
  18  * The above copyright notice and this permission notice shall be included in
  19  * all copies or substantial portions of the Software.
  20  *
  21  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  22  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  23  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
  24  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  25  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  26  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
  27  * THE SOFTWARE.
  28  */
  29
  30 #include "qemu/osdep.h"
  31 #include "nbd-client.h"
  32
  33 #define HANDLE_TO_INDEX(bs, handle) ((handle) ^ ((uint64_t)(intptr_t)bs))
  34 #define INDEX_TO_HANDLE(bs, index)  ((index)  ^ ((uint64_t)(intptr_t)bs))
  35
  36 static void nbd_recv_coroutines_enter_all(NBDClientSession *s)
  37 {
  38     int i;
  39
  40     for (i = 0; i < MAX_NBD_REQUESTS; i++) {
  41         if (s->recv_coroutine[i]) {
  42             qemu_coroutine_enter(s->recv_coroutine[i]);
  43         }
  44     }
  45 }
  46
  47 static void nbd_teardown_connection(BlockDriverState *bs)
  48 {
  49     NBDClientSession *client = nbd_get_client_session(bs);
  50
  51     if (!client->ioc) { /* Already closed */
  52         return;
  53     }
  54
  55     /* finish any pending coroutines */
  56     qio_channel_shutdown(client->ioc,
  57                          QIO_CHANNEL_SHUTDOWN_BOTH,
  58                          NULL);
  59     nbd_recv_coroutines_enter_all(client);
  60
  61     nbd_client_detach_aio_context(bs);
  62     object_unref(OBJECT(client->sioc));
  63     client->sioc = NULL;
  64     object_unref(OBJECT(client->ioc));
  65     client->ioc = NULL;
  66 }
  67
  68 static void nbd_reply_ready(void *opaque)
  69 {
  70     BlockDriverState *bs = opaque;
  71     NBDClientSession *s = nbd_get_client_session(bs);
  72     uint64_t i;
  73     int ret;
  74
  75     if (!s->ioc) { /* Already closed */
  76         return;
  77     }
  78
  79     if (s->reply.handle == 0) {
  80         /* No reply already in flight.  Fetch a header.  It is possible
  81          * that another thread has done the same thing in parallel, so
  82          * the socket is not readable anymore.
  83          */
  84         ret = nbd_receive_reply(s->ioc, &s->reply);
  85         if (ret == -EAGAIN) {
  86             return;
  87         }
  88         if (ret < 0) {
  89             s->reply.handle = 0;
  90             goto fail;
  91         }
  92     }
  93
  94     /* There's no need for a mutex on the receive side, because the
  95      * handler acts as a synchronization point and ensures that only
  96      * one coroutine is called until the reply finishes.  */
  97     i = HANDLE_TO_INDEX(s, s->reply.handle);
  98     if (i >= MAX_NBD_REQUESTS) {
  99         goto fail;
 100     }
 101
 102     if (s->recv_coroutine[i]) {
 103         qemu_coroutine_enter(s->recv_coroutine[i]);
 104         return;
 105     }
 106
 107 fail:
 108     nbd_teardown_connection(bs);
 109 }
 110
 111 static void nbd_restart_write(void *opaque)
 112 {
 113     BlockDriverState *bs = opaque;
 114
 115     qemu_coroutine_enter(nbd_get_client_session(bs)->send_coroutine);
 116 }
 117
 118 static int nbd_co_send_request(BlockDriverState *bs,
 119                                NBDRequest *request,
 120                                QEMUIOVector *qiov)
 121 {
 122     NBDClientSession *s = nbd_get_client_session(bs);
 123     AioContext *aio_context;
 124     int rc, ret, i;
 125
 126     qemu_co_mutex_lock(&s->send_mutex);
 127
 128     for (i = 0; i < MAX_NBD_REQUESTS; i++) {
 129         if (s->recv_coroutine[i] == NULL) {
 130             s->recv_coroutine[i] = qemu_coroutine_self();
 131             break;
 132         }
 133     }
 134
 135     g_assert(qemu_in_coroutine());
 136     assert(i < MAX_NBD_REQUESTS);
 137     request->handle = INDEX_TO_HANDLE(s, i);
 138
 139     if (!s->ioc) {
 140         qemu_co_mutex_unlock(&s->send_mutex);
 141         return -EPIPE;
 142     }
 143
 144     s->send_coroutine = qemu_coroutine_self();
 145     aio_context = bdrv_get_aio_context(bs);
 146
 147     aio_set_fd_handler(aio_context, s->sioc->fd, false,
 148                        nbd_reply_ready, nbd_restart_write, bs);
 149     if (qiov) {
 150         qio_channel_set_cork(s->ioc, true);
 151         rc = nbd_send_request(s->ioc, request);
 152         if (rc >= 0) {
 153             ret = nbd_wr_syncv(s->ioc, qiov->iov, qiov->niov, request->len,
 154                                false);
 155             if (ret != request->len) {
 156                 rc = -EIO;
 157             }
 158         }
 159         qio_channel_set_cork(s->ioc, false);
 160     } else {
 161         rc = nbd_send_request(s->ioc, request);
 162     }
 163     aio_set_fd_handler(aio_context, s->sioc->fd, false,
 164                        nbd_reply_ready, NULL, bs);
 165     s->send_coroutine = NULL;
 166     qemu_co_mutex_unlock(&s->send_mutex);
 167     return rc;
 168 }
 169
 170 static void nbd_co_receive_reply(NBDClientSession *s,
 171                                  NBDRequest *request,
 172                                  NBDReply *reply,
 173                                  QEMUIOVector *qiov)
 174 {
 175     int ret;
 176
 177     /* Wait until we're woken up by the read handler.  TODO: perhaps
 178      * peek at the next reply and avoid yielding if it's ours?  */
 179     qemu_coroutine_yield();
 180     *reply = s->reply;
 181     if (reply->handle != request->handle ||
 182         !s->ioc) {
 183         reply->error = EIO;
 184     } else {
 185         if (qiov && reply->error == 0) {
 186             ret = nbd_wr_syncv(s->ioc, qiov->iov, qiov->niov, request->len,
 187                                true);
 188             if (ret != request->len) {
 189                 reply->error = EIO;
 190             }
 191         }
 192
 193         /* Tell the read handler to read another header.  */
 194         s->reply.handle = 0;
 195     }
 196 }
 197
 198 static void nbd_coroutine_start(NBDClientSession *s,
 199                                 NBDRequest *request)
 200 {
 201     /* Poor man semaphore.  The free_sema is locked when no other request
 202      * can be accepted, and unlocked after receiving one reply.  */
 203     if (s->in_flight == MAX_NBD_REQUESTS) {
 204         qemu_co_queue_wait(&s->free_sema);
 205         assert(s->in_flight < MAX_NBD_REQUESTS);
 206     }
 207     s->in_flight++;
 208
 209     /* s->recv_coroutine[i] is set as soon as we get the send_lock.  */
 210 }
 211
 212 static void nbd_coroutine_end(NBDClientSession *s,
 213                               NBDRequest *request)
 214 {
 215     int i = HANDLE_TO_INDEX(s, request->handle);
 216     s->recv_coroutine[i] = NULL;
 217     if (s->in_flight-- == MAX_NBD_REQUESTS) {
 218         qemu_co_queue_next(&s->free_sema);
 219     }
 220 }
 221
 222 int nbd_client_co_preadv(BlockDriverState *bs, uint64_t offset,
 223                          uint64_t bytes, QEMUIOVector *qiov, int flags)
 224 {
 225     NBDClientSession *client = nbd_get_client_session(bs);
 226     NBDRequest request = {
 227         .type = NBD_CMD_READ,
 228         .from = offset,
 229         .len = bytes,
 230     };
 231     NBDReply reply;
 232     ssize_t ret;
 233
 234     assert(bytes <= NBD_MAX_BUFFER_SIZE);
 235     assert(!flags);
 236
 237     nbd_coroutine_start(client, &request);
 238     ret = nbd_co_send_request(bs, &request, NULL);
 239     if (ret < 0) {
 240         reply.error = -ret;
 241     } else {
 242         nbd_co_receive_reply(client, &request, &reply, qiov);
 243     }
 244     nbd_coroutine_end(client, &request);
 245     return -reply.error;
 246 }
 247
 248 int nbd_client_co_pwritev(BlockDriverState *bs, uint64_t offset,
 249                           uint64_t bytes, QEMUIOVector *qiov, int flags)
 250 {
 251     NBDClientSession *client = nbd_get_client_session(bs);
 252     NBDRequest request = {
 253         .type = NBD_CMD_WRITE,
 254         .from = offset,
 255         .len = bytes,
 256     };
 257     NBDReply reply;
 258     ssize_t ret;
 259
 260     if (flags & BDRV_REQ_FUA) {
 261         assert(client->nbdflags & NBD_FLAG_SEND_FUA);
 262         request.flags |= NBD_CMD_FLAG_FUA;
 263     }
 264
 265     assert(bytes <= NBD_MAX_BUFFER_SIZE);
 266
 267     nbd_coroutine_start(client, &request);
 268     ret = nbd_co_send_request(bs, &request, qiov);
 269     if (ret < 0) {
 270         reply.error = -ret;
 271     } else {
 272         nbd_co_receive_reply(client, &request, &reply, NULL);
 273     }
 274     nbd_coroutine_end(client, &request);
 275     return -reply.error;
 276 }
 277
 278 int nbd_client_co_pwrite_zeroes(BlockDriverState *bs, int64_t offset,
 279                                 int count, BdrvRequestFlags flags)
 280 {
 281     ssize_t ret;
 282     NBDClientSession *client = nbd_get_client_session(bs);
 283     NBDRequest request = {
 284         .type = NBD_CMD_WRITE_ZEROES,
 285         .from = offset,
 286         .len = count,
 287     };
 288     NBDReply reply;
 289
 290     if (!(client->nbdflags & NBD_FLAG_SEND_WRITE_ZEROES)) {
 291         return -ENOTSUP;
 292     }
 293
 294     if (flags & BDRV_REQ_FUA) {
 295         assert(client->nbdflags & NBD_FLAG_SEND_FUA);
 296         request.flags |= NBD_CMD_FLAG_FUA;
 297     }
 298     if (!(flags & BDRV_REQ_MAY_UNMAP)) {
 299         request.flags |= NBD_CMD_FLAG_NO_HOLE;
 300     }
 301
 302     nbd_coroutine_start(client, &request);
 303     ret = nbd_co_send_request(bs, &request, NULL);
 304     if (ret < 0) {
 305         reply.error = -ret;
 306     } else {
 307         nbd_co_receive_reply(client, &request, &reply, NULL);
 308     }
 309     nbd_coroutine_end(client, &request);
 310     return -reply.error;
 311 }
 312
 313 int nbd_client_co_flush(BlockDriverState *bs)
 314 {
 315     NBDClientSession *client = nbd_get_client_session(bs);
 316     NBDRequest request = { .type = NBD_CMD_FLUSH };
 317     NBDReply reply;
 318     ssize_t ret;
 319
 320     if (!(client->nbdflags & NBD_FLAG_SEND_FLUSH)) {
 321         return 0;
 322     }
 323
 324     request.from = 0;
 325     request.len = 0;
 326
 327     nbd_coroutine_start(client, &request);
 328     ret = nbd_co_send_request(bs, &request, NULL);
 329     if (ret < 0) {
 330         reply.error = -ret;
 331     } else {
 332         nbd_co_receive_reply(client, &request, &reply, NULL);
 333     }
 334     nbd_coroutine_end(client, &request);
 335     return -reply.error;
 336 }
 337
 338 int nbd_client_co_pdiscard(BlockDriverState *bs, int64_t offset, int count)
 339 {
 340     NBDClientSession *client = nbd_get_client_session(bs);
 341     NBDRequest request = {
 342         .type = NBD_CMD_TRIM,
 343         .from = offset,
 344         .len = count,
 345     };
 346     NBDReply reply;
 347     ssize_t ret;
 348
 349     if (!(client->nbdflags & NBD_FLAG_SEND_TRIM)) {
 350         return 0;
 351     }
 352
 353     nbd_coroutine_start(client, &request);
 354     ret = nbd_co_send_request(bs, &request, NULL);
 355     if (ret < 0) {
 356         reply.error = -ret;
 357     } else {
 358         nbd_co_receive_reply(client, &request, &reply, NULL);
 359     }
 360     nbd_coroutine_end(client, &request);
 361     return -reply.error;
 362
 363 }
 364
 365 void nbd_client_detach_aio_context(BlockDriverState *bs)
 366 {
 367     aio_set_fd_handler(bdrv_get_aio_context(bs),
 368                        nbd_get_client_session(bs)->sioc->fd,
 369                        false, NULL, NULL, NULL);
 370 }
 371
 372 void nbd_client_attach_aio_context(BlockDriverState *bs,
 373                                    AioContext *new_context)
 374 {
 375     aio_set_fd_handler(new_context, nbd_get_client_session(bs)->sioc->fd,
 376                        false, nbd_reply_ready, NULL, bs);
 377 }
 378
 379 void nbd_client_close(BlockDriverState *bs)
 380 {
 381     NBDClientSession *client = nbd_get_client_session(bs);
 382     NBDRequest request = { .type = NBD_CMD_DISC };
 383
 384     if (client->ioc == NULL) {
 385         return;
 386     }
 387
 388     nbd_send_request(client->ioc, &request);
 389
 390     nbd_teardown_connection(bs);
 391 }
 392
 393 int nbd_client_init(BlockDriverState *bs,
 394                     QIOChannelSocket *sioc,
 395                     const char *export,
 396                     QCryptoTLSCreds *tlscreds,
 397                     const char *hostname,
 398                     Error **errp)
 399 {
 400     NBDClientSession *client = nbd_get_client_session(bs);
 401     int ret;
 402
 403     /* NBD handshake */
 404     logout("session init %s\n", export);
 405     qio_channel_set_blocking(QIO_CHANNEL(sioc), true, NULL);
 406
 407     ret = nbd_receive_negotiate(QIO_CHANNEL(sioc), export,
 408                                 &client->nbdflags,
 409                                 tlscreds, hostname,
 410                                 &client->ioc,
 411                                 &client->size, errp);
 412     if (ret < 0) {
 413         logout("Failed to negotiate with the NBD server\n");
 414         return ret;
 415     }
 416     if (client->nbdflags & NBD_FLAG_SEND_FUA) {
 417         bs->supported_write_flags = BDRV_REQ_FUA;
 418     }
 419
 420     qemu_co_mutex_init(&client->send_mutex);
 421     qemu_co_queue_init(&client->free_sema);
 422     client->sioc = sioc;
 423     object_ref(OBJECT(client->sioc));
 424
 425     if (!client->ioc) {
 426         client->ioc = QIO_CHANNEL(sioc);
 427         object_ref(OBJECT(client->ioc));
 428     }
 429
 430     /* Now that we're connected, set the socket to be non-blocking and
 431      * kick the reply mechanism.  */
 432     qio_channel_set_blocking(QIO_CHANNEL(sioc), false, NULL);
 433
 434     nbd_client_attach_aio_context(bs, bdrv_get_aio_context(bs));
 435
 436     logout("Established connection with NBD server\n");
 437     return 0;
 438 }