block/nbd-client.c

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