nbd.c

   1 /*
   2  *  Copyright (C) 2005  Anthony Liguori <anthony@codemonkey.ws>
   3  *
   4  *  Network Block Device
   5  *
   6  *  This program is free software; you can redistribute it and/or modify
   7  *  it under the terms of the GNU General Public License as published by
   8  *  the Free Software Foundation; under version 2 of the License.
   9  *
  10  *  This program is distributed in the hope that it will be useful,
  11  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  12  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13  *  GNU General Public License for more details.
  14  *
  15  *  You should have received a copy of the GNU General Public License
  16  *  along with this program; if not, see <http://www.gnu.org/licenses/>.
  17  */
  18
  19 #include "block/nbd.h"
  20 #include "sysemu/block-backend.h"
  21
  22 #include "block/coroutine.h"
  23
  24 #include <errno.h>
  25 #include <string.h>
  26 #ifndef _WIN32
  27 #include <sys/ioctl.h>
  28 #endif
  29 #if defined(__sun__) || defined(__HAIKU__)
  30 #include <sys/ioccom.h>
  31 #endif
  32 #include <ctype.h>
  33 #include <inttypes.h>
  34
  35 #ifdef __linux__
  36 #include <linux/fs.h>
  37 #endif
  38
  39 #include "qemu/sockets.h"
  40 #include "qemu/queue.h"
  41 #include "qemu/main-loop.h"
  42
  43 //#define DEBUG_NBD
  44
  45 #ifdef DEBUG_NBD
  46 #define TRACE(msg, ...) do { \
  47     LOG(msg, ## __VA_ARGS__); \
  48 } while(0)
  49 #else
  50 #define TRACE(msg, ...) \
  51     do { } while (0)
  52 #endif
  53
  54 #define LOG(msg, ...) do { \
  55     fprintf(stderr, "%s:%s():L%d: " msg "\n", \
  56             __FILE__, __FUNCTION__, __LINE__, ## __VA_ARGS__); \
  57 } while(0)
  58
  59 /* This is all part of the "official" NBD API.
  60  *
  61  * The most up-to-date documentation is available at:
  62  * https://github.com/yoe/nbd/blob/master/doc/proto.txt
  63  */
  64
  65 #define NBD_REQUEST_SIZE        (4 + 4 + 8 + 8 + 4)
  66 #define NBD_REPLY_SIZE          (4 + 4 + 8)
  67 #define NBD_REQUEST_MAGIC       0x25609513
  68 #define NBD_REPLY_MAGIC         0x67446698
  69 #define NBD_OPTS_MAGIC          0x49484156454F5054LL
  70 #define NBD_CLIENT_MAGIC        0x0000420281861253LL
  71 #define NBD_REP_MAGIC           0x3e889045565a9LL
  72
  73 #define NBD_SET_SOCK            _IO(0xab, 0)
  74 #define NBD_SET_BLKSIZE         _IO(0xab, 1)
  75 #define NBD_SET_SIZE            _IO(0xab, 2)
  76 #define NBD_DO_IT               _IO(0xab, 3)
  77 #define NBD_CLEAR_SOCK          _IO(0xab, 4)
  78 #define NBD_CLEAR_QUE           _IO(0xab, 5)
  79 #define NBD_PRINT_DEBUG         _IO(0xab, 6)
  80 #define NBD_SET_SIZE_BLOCKS     _IO(0xab, 7)
  81 #define NBD_DISCONNECT          _IO(0xab, 8)
  82 #define NBD_SET_TIMEOUT         _IO(0xab, 9)
  83 #define NBD_SET_FLAGS           _IO(0xab, 10)
  84
  85 #define NBD_OPT_EXPORT_NAME     (1)
  86 #define NBD_OPT_ABORT           (2)
  87 #define NBD_OPT_LIST            (3)
  88
  89 /* Definitions for opaque data types */
  90
  91 typedef struct NBDRequest NBDRequest;
  92
  93 struct NBDRequest {
  94     QSIMPLEQ_ENTRY(NBDRequest) entry;
  95     NBDClient *client;
  96     uint8_t *data;
  97 };
  98
  99 struct NBDExport {
 100     int refcount;
 101     void (*close)(NBDExport *exp);
 102
 103     BlockBackend *blk;
 104     char *name;
 105     off_t dev_offset;
 106     off_t size;
 107     uint32_t nbdflags;
 108     QTAILQ_HEAD(, NBDClient) clients;
 109     QTAILQ_ENTRY(NBDExport) next;
 110
 111     AioContext *ctx;
 112 };
 113
 114 static QTAILQ_HEAD(, NBDExport) exports = QTAILQ_HEAD_INITIALIZER(exports);
 115
 116 struct NBDClient {
 117     int refcount;
 118     void (*close)(NBDClient *client);
 119
 120     NBDExport *exp;
 121     int sock;
 122
 123     Coroutine *recv_coroutine;
 124
 125     CoMutex send_lock;
 126     Coroutine *send_coroutine;
 127
 128     bool can_read;
 129
 130     QTAILQ_ENTRY(NBDClient) next;
 131     int nb_requests;
 132     bool closing;
 133 };
 134
 135 /* That's all folks */
 136
 137 static void nbd_set_handlers(NBDClient *client);
 138 static void nbd_unset_handlers(NBDClient *client);
 139 static void nbd_update_can_read(NBDClient *client);
 140
 141 ssize_t nbd_wr_sync(int fd, void *buffer, size_t size, bool do_read)
 142 {
 143     size_t offset = 0;
 144     int err;
 145
 146     if (qemu_in_coroutine()) {
 147         if (do_read) {
 148             return qemu_co_recv(fd, buffer, size);
 149         } else {
 150             return qemu_co_send(fd, buffer, size);
 151         }
 152     }
 153
 154     while (offset < size) {
 155         ssize_t len;
 156
 157         if (do_read) {
 158             len = qemu_recv(fd, buffer + offset, size - offset, 0);
 159         } else {
 160             len = send(fd, buffer + offset, size - offset, 0);
 161         }
 162
 163         if (len < 0) {
 164             err = socket_error();
 165
 166             /* recoverable error */
 167             if (err == EINTR || (offset > 0 && (err == EAGAIN || err == EWOULDBLOCK))) {
 168                 continue;
 169             }
 170
 171             /* unrecoverable error */
 172             return -err;
 173         }
 174
 175         /* eof */
 176         if (len == 0) {
 177             break;
 178         }
 179
 180         offset += len;
 181     }
 182
 183     return offset;
 184 }
 185
 186 static ssize_t read_sync(int fd, void *buffer, size_t size)
 187 {
 188     /* Sockets are kept in blocking mode in the negotiation phase.  After
 189      * that, a non-readable socket simply means that another thread stole
 190      * our request/reply.  Synchronization is done with recv_coroutine, so
 191      * that this is coroutine-safe.
 192      */
 193     return nbd_wr_sync(fd, buffer, size, true);
 194 }
 195
 196 static ssize_t write_sync(int fd, void *buffer, size_t size)
 197 {
 198     int ret;
 199     do {
 200         /* For writes, we do expect the socket to be writable.  */
 201         ret = nbd_wr_sync(fd, buffer, size, false);
 202     } while (ret == -EAGAIN);
 203     return ret;
 204 }
 205
 206 /* Basic flow for negotiation
 207
 208    Server         Client
 209    Negotiate
 210
 211    or
 212
 213    Server         Client
 214    Negotiate #1
 215                   Option
 216    Negotiate #2
 217
 218    ----
 219
 220    followed by
 221
 222    Server         Client
 223                   Request
 224    Response
 225                   Request
 226    Response
 227                   ...
 228    ...
 229                   Request (type == 2)
 230
 231 */
 232
 233 static int nbd_send_rep(int csock, uint32_t type, uint32_t opt)
 234 {
 235     uint64_t magic;
 236     uint32_t len;
 237
 238     magic = cpu_to_be64(NBD_REP_MAGIC);
 239     if (write_sync(csock, &magic, sizeof(magic)) != sizeof(magic)) {
 240         LOG("write failed (rep magic)");
 241         return -EINVAL;
 242     }
 243     opt = cpu_to_be32(opt);
 244     if (write_sync(csock, &opt, sizeof(opt)) != sizeof(opt)) {
 245         LOG("write failed (rep opt)");
 246         return -EINVAL;
 247     }
 248     type = cpu_to_be32(type);
 249     if (write_sync(csock, &type, sizeof(type)) != sizeof(type)) {
 250         LOG("write failed (rep type)");
 251         return -EINVAL;
 252     }
 253     len = cpu_to_be32(0);
 254     if (write_sync(csock, &len, sizeof(len)) != sizeof(len)) {
 255         LOG("write failed (rep data length)");
 256         return -EINVAL;
 257     }
 258     return 0;
 259 }
 260
 261 static int nbd_send_rep_list(int csock, NBDExport *exp)
 262 {
 263     uint64_t magic, name_len;
 264     uint32_t opt, type, len;
 265
 266     name_len = strlen(exp->name);
 267     magic = cpu_to_be64(NBD_REP_MAGIC);
 268     if (write_sync(csock, &magic, sizeof(magic)) != sizeof(magic)) {
 269         LOG("write failed (magic)");
 270         return -EINVAL;
 271      }
 272     opt = cpu_to_be32(NBD_OPT_LIST);
 273     if (write_sync(csock, &opt, sizeof(opt)) != sizeof(opt)) {
 274         LOG("write failed (opt)");
 275         return -EINVAL;
 276     }
 277     type = cpu_to_be32(NBD_REP_SERVER);
 278     if (write_sync(csock, &type, sizeof(type)) != sizeof(type)) {
 279         LOG("write failed (reply type)");
 280         return -EINVAL;
 281     }
 282     len = cpu_to_be32(name_len + sizeof(len));
 283     if (write_sync(csock, &len, sizeof(len)) != sizeof(len)) {
 284         LOG("write failed (length)");
 285         return -EINVAL;
 286     }
 287     len = cpu_to_be32(name_len);
 288     if (write_sync(csock, &len, sizeof(len)) != sizeof(len)) {
 289         LOG("write failed (length)");
 290         return -EINVAL;
 291     }
 292     if (write_sync(csock, exp->name, name_len) != name_len) {
 293         LOG("write failed (buffer)");
 294         return -EINVAL;
 295     }
 296     return 0;
 297 }
 298
 299 static int nbd_handle_list(NBDClient *client, uint32_t length)
 300 {
 301     int csock;
 302     NBDExport *exp;
 303
 304     csock = client->sock;
 305     if (length) {
 306         return nbd_send_rep(csock, NBD_REP_ERR_INVALID, NBD_OPT_LIST);
 307     }
 308
 309     /* For each export, send a NBD_REP_SERVER reply. */
 310     QTAILQ_FOREACH(exp, &exports, next) {
 311         if (nbd_send_rep_list(csock, exp)) {
 312             return -EINVAL;
 313         }
 314     }
 315     /* Finish with a NBD_REP_ACK. */
 316     return nbd_send_rep(csock, NBD_REP_ACK, NBD_OPT_LIST);
 317 }
 318
 319 static int nbd_handle_export_name(NBDClient *client, uint32_t length)
 320 {
 321     int rc = -EINVAL, csock = client->sock;
 322     char name[256];
 323
 324     /* Client sends:
 325         [20 ..  xx]   export name (length bytes)
 326      */
 327     TRACE("Checking length");
 328     if (length > 255) {
 329         LOG("Bad length received");
 330         goto fail;
 331     }
 332     if (read_sync(csock, name, length) != length) {
 333         LOG("read failed");
 334         goto fail;
 335     }
 336     name[length] = '\0';
 337
 338     client->exp = nbd_export_find(name);
 339     if (!client->exp) {
 340         LOG("export not found");
 341         goto fail;
 342     }
 343
 344     QTAILQ_INSERT_TAIL(&client->exp->clients, client, next);
 345     nbd_export_get(client->exp);
 346     rc = 0;
 347 fail:
 348     return rc;
 349 }
 350
 351 static int nbd_receive_options(NBDClient *client)
 352 {
 353     while (1) {
 354         int csock = client->sock;
 355         uint32_t tmp, length;
 356         uint64_t magic;
 357
 358         /* Client sends:
 359             [ 0 ..   3]   client flags
 360             [ 4 ..  11]   NBD_OPTS_MAGIC
 361             [12 ..  15]   NBD option
 362             [16 ..  19]   length
 363             ...           Rest of request
 364         */
 365
 366         if (read_sync(csock, &tmp, sizeof(tmp)) != sizeof(tmp)) {
 367             LOG("read failed");
 368             return -EINVAL;
 369         }
 370         TRACE("Checking client flags");
 371         tmp = be32_to_cpu(tmp);
 372         if (tmp != 0 && tmp != NBD_FLAG_C_FIXED_NEWSTYLE) {
 373             LOG("Bad client flags received");
 374             return -EINVAL;
 375         }
 376
 377         if (read_sync(csock, &magic, sizeof(magic)) != sizeof(magic)) {
 378             LOG("read failed");
 379             return -EINVAL;
 380         }
 381         TRACE("Checking opts magic");
 382         if (magic != be64_to_cpu(NBD_OPTS_MAGIC)) {
 383             LOG("Bad magic received");
 384             return -EINVAL;
 385         }
 386
 387         if (read_sync(csock, &tmp, sizeof(tmp)) != sizeof(tmp)) {
 388             LOG("read failed");
 389             return -EINVAL;
 390         }
 391
 392         if (read_sync(csock, &length, sizeof(length)) != sizeof(length)) {
 393             LOG("read failed");
 394             return -EINVAL;
 395         }
 396         length = be32_to_cpu(length);
 397
 398         TRACE("Checking option");
 399         switch (be32_to_cpu(tmp)) {
 400         case NBD_OPT_LIST:
 401             if (nbd_handle_list(client, length) < 0) {
 402                 return 1;
 403             }
 404             break;
 405
 406         case NBD_OPT_ABORT:
 407             return -EINVAL;
 408
 409         case NBD_OPT_EXPORT_NAME:
 410             return nbd_handle_export_name(client, length);
 411
 412         default:
 413             tmp = be32_to_cpu(tmp);
 414             LOG("Unsupported option 0x%x", tmp);
 415             nbd_send_rep(client->sock, NBD_REP_ERR_UNSUP, tmp);
 416             return -EINVAL;
 417         }
 418     }
 419 }
 420
 421 static int nbd_send_negotiate(NBDClient *client)
 422 {
 423     int csock = client->sock;
 424     char buf[8 + 8 + 8 + 128];
 425     int rc;
 426     const int myflags = (NBD_FLAG_HAS_FLAGS | NBD_FLAG_SEND_TRIM |
 427                          NBD_FLAG_SEND_FLUSH | NBD_FLAG_SEND_FUA);
 428
 429     /* Negotiation header without options:
 430         [ 0 ..   7]   passwd       ("NBDMAGIC")
 431         [ 8 ..  15]   magic        (NBD_CLIENT_MAGIC)
 432         [16 ..  23]   size
 433         [24 ..  25]   server flags (0)
 434         [26 ..  27]   export flags
 435         [28 .. 151]   reserved     (0)
 436
 437        Negotiation header with options, part 1:
 438         [ 0 ..   7]   passwd       ("NBDMAGIC")
 439         [ 8 ..  15]   magic        (NBD_OPTS_MAGIC)
 440         [16 ..  17]   server flags (0)
 441
 442        part 2 (after options are sent):
 443         [18 ..  25]   size
 444         [26 ..  27]   export flags
 445         [28 .. 151]   reserved     (0)
 446      */
 447
 448     qemu_set_block(csock);
 449     rc = -EINVAL;
 450
 451     TRACE("Beginning negotiation.");
 452     memset(buf, 0, sizeof(buf));
 453     memcpy(buf, "NBDMAGIC", 8);
 454     if (client->exp) {
 455         assert ((client->exp->nbdflags & ~65535) == 0);
 456         cpu_to_be64w((uint64_t*)(buf + 8), NBD_CLIENT_MAGIC);
 457         cpu_to_be64w((uint64_t*)(buf + 16), client->exp->size);
 458         cpu_to_be16w((uint16_t*)(buf + 26), client->exp->nbdflags | myflags);
 459     } else {
 460         cpu_to_be64w((uint64_t*)(buf + 8), NBD_OPTS_MAGIC);
 461         cpu_to_be16w((uint16_t *)(buf + 16), NBD_FLAG_FIXED_NEWSTYLE);
 462     }
 463
 464     if (client->exp) {
 465         if (write_sync(csock, buf, sizeof(buf)) != sizeof(buf)) {
 466             LOG("write failed");
 467             goto fail;
 468         }
 469     } else {
 470         if (write_sync(csock, buf, 18) != 18) {
 471             LOG("write failed");
 472             goto fail;
 473         }
 474         rc = nbd_receive_options(client);
 475         if (rc != 0) {
 476             LOG("option negotiation failed");
 477             goto fail;
 478         }
 479
 480         assert ((client->exp->nbdflags & ~65535) == 0);
 481         cpu_to_be64w((uint64_t*)(buf + 18), client->exp->size);
 482         cpu_to_be16w((uint16_t*)(buf + 26), client->exp->nbdflags | myflags);
 483         if (write_sync(csock, buf + 18, sizeof(buf) - 18) != sizeof(buf) - 18) {
 484             LOG("write failed");
 485             goto fail;
 486         }
 487     }
 488
 489     TRACE("Negotiation succeeded.");
 490     rc = 0;
 491 fail:
 492     qemu_set_nonblock(csock);
 493     return rc;
 494 }
 495
 496 int nbd_receive_negotiate(int csock, const char *name, uint32_t *flags,
 497                           off_t *size, size_t *blocksize)
 498 {
 499     char buf[256];
 500     uint64_t magic, s;
 501     uint16_t tmp;
 502     int rc;
 503
 504     TRACE("Receiving negotiation.");
 505
 506     rc = -EINVAL;
 507
 508     if (read_sync(csock, buf, 8) != 8) {
 509         LOG("read failed");
 510         goto fail;
 511     }
 512
 513     buf[8] = '\0';
 514     if (strlen(buf) == 0) {
 515         LOG("server connection closed");
 516         goto fail;
 517     }
 518
 519     TRACE("Magic is %c%c%c%c%c%c%c%c",
 520           qemu_isprint(buf[0]) ? buf[0] : '.',
 521           qemu_isprint(buf[1]) ? buf[1] : '.',
 522           qemu_isprint(buf[2]) ? buf[2] : '.',
 523           qemu_isprint(buf[3]) ? buf[3] : '.',
 524           qemu_isprint(buf[4]) ? buf[4] : '.',
 525           qemu_isprint(buf[5]) ? buf[5] : '.',
 526           qemu_isprint(buf[6]) ? buf[6] : '.',
 527           qemu_isprint(buf[7]) ? buf[7] : '.');
 528
 529     if (memcmp(buf, "NBDMAGIC", 8) != 0) {
 530         LOG("Invalid magic received");
 531         goto fail;
 532     }
 533
 534     if (read_sync(csock, &magic, sizeof(magic)) != sizeof(magic)) {
 535         LOG("read failed");
 536         goto fail;
 537     }
 538     magic = be64_to_cpu(magic);
 539     TRACE("Magic is 0x%" PRIx64, magic);
 540
 541     if (name) {
 542         uint32_t reserved = 0;
 543         uint32_t opt;
 544         uint32_t namesize;
 545
 546         TRACE("Checking magic (opts_magic)");
 547         if (magic != NBD_OPTS_MAGIC) {
 548             LOG("Bad magic received");
 549             goto fail;
 550         }
 551         if (read_sync(csock, &tmp, sizeof(tmp)) != sizeof(tmp)) {
 552             LOG("flags read failed");
 553             goto fail;
 554         }
 555         *flags = be16_to_cpu(tmp) << 16;
 556         /* reserved for future use */
 557         if (write_sync(csock, &reserved, sizeof(reserved)) !=
 558             sizeof(reserved)) {
 559             LOG("write failed (reserved)");
 560             goto fail;
 561         }
 562         /* write the export name */
 563         magic = cpu_to_be64(magic);
 564         if (write_sync(csock, &magic, sizeof(magic)) != sizeof(magic)) {
 565             LOG("write failed (magic)");
 566             goto fail;
 567         }
 568         opt = cpu_to_be32(NBD_OPT_EXPORT_NAME);
 569         if (write_sync(csock, &opt, sizeof(opt)) != sizeof(opt)) {
 570             LOG("write failed (opt)");
 571             goto fail;
 572         }
 573         namesize = cpu_to_be32(strlen(name));
 574         if (write_sync(csock, &namesize, sizeof(namesize)) !=
 575             sizeof(namesize)) {
 576             LOG("write failed (namesize)");
 577             goto fail;
 578         }
 579         if (write_sync(csock, (char*)name, strlen(name)) != strlen(name)) {
 580             LOG("write failed (name)");
 581             goto fail;
 582         }
 583     } else {
 584         TRACE("Checking magic (cli_magic)");
 585
 586         if (magic != NBD_CLIENT_MAGIC) {
 587             LOG("Bad magic received");
 588             goto fail;
 589         }
 590     }
 591
 592     if (read_sync(csock, &s, sizeof(s)) != sizeof(s)) {
 593         LOG("read failed");
 594         goto fail;
 595     }
 596     *size = be64_to_cpu(s);
 597     *blocksize = 1024;
 598     TRACE("Size is %" PRIu64, *size);
 599
 600     if (!name) {
 601         if (read_sync(csock, flags, sizeof(*flags)) != sizeof(*flags)) {
 602             LOG("read failed (flags)");
 603             goto fail;
 604         }
 605         *flags = be32_to_cpup(flags);
 606     } else {
 607         if (read_sync(csock, &tmp, sizeof(tmp)) != sizeof(tmp)) {
 608             LOG("read failed (tmp)");
 609             goto fail;
 610         }
 611         *flags |= be32_to_cpu(tmp);
 612     }
 613     if (read_sync(csock, &buf, 124) != 124) {
 614         LOG("read failed (buf)");
 615         goto fail;
 616     }
 617     rc = 0;
 618
 619 fail:
 620     return rc;
 621 }
 622
 623 #ifdef __linux__
 624 int nbd_init(int fd, int csock, uint32_t flags, off_t size, size_t blocksize)
 625 {
 626     TRACE("Setting NBD socket");
 627
 628     if (ioctl(fd, NBD_SET_SOCK, csock) < 0) {
 629         int serrno = errno;
 630         LOG("Failed to set NBD socket");
 631         return -serrno;
 632     }
 633
 634     TRACE("Setting block size to %lu", (unsigned long)blocksize);
 635
 636     if (ioctl(fd, NBD_SET_BLKSIZE, blocksize) < 0) {
 637         int serrno = errno;
 638         LOG("Failed setting NBD block size");
 639         return -serrno;
 640     }
 641
 642         TRACE("Setting size to %zd block(s)", (size_t)(size / blocksize));
 643
 644     if (ioctl(fd, NBD_SET_SIZE_BLOCKS, size / blocksize) < 0) {
 645         int serrno = errno;
 646         LOG("Failed setting size (in blocks)");
 647         return -serrno;
 648     }
 649
 650     if (ioctl(fd, NBD_SET_FLAGS, flags) < 0) {
 651         if (errno == ENOTTY) {
 652             int read_only = (flags & NBD_FLAG_READ_ONLY) != 0;
 653             TRACE("Setting readonly attribute");
 654
 655             if (ioctl(fd, BLKROSET, (unsigned long) &read_only) < 0) {
 656                 int serrno = errno;
 657                 LOG("Failed setting read-only attribute");
 658                 return -serrno;
 659             }
 660         } else {
 661             int serrno = errno;
 662             LOG("Failed setting flags");
 663             return -serrno;
 664         }
 665     }
 666
 667     TRACE("Negotiation ended");
 668
 669     return 0;
 670 }
 671
 672 int nbd_disconnect(int fd)
 673 {
 674     ioctl(fd, NBD_CLEAR_QUE);
 675     ioctl(fd, NBD_DISCONNECT);
 676     ioctl(fd, NBD_CLEAR_SOCK);
 677     return 0;
 678 }
 679
 680 int nbd_client(int fd)
 681 {
 682     int ret;
 683     int serrno;
 684
 685     TRACE("Doing NBD loop");
 686
 687     ret = ioctl(fd, NBD_DO_IT);
 688     if (ret < 0 && errno == EPIPE) {
 689         /* NBD_DO_IT normally returns EPIPE when someone has disconnected
 690          * the socket via NBD_DISCONNECT.  We do not want to return 1 in
 691          * that case.
 692          */
 693         ret = 0;
 694     }
 695     serrno = errno;
 696
 697     TRACE("NBD loop returned %d: %s", ret, strerror(serrno));
 698
 699     TRACE("Clearing NBD queue");
 700     ioctl(fd, NBD_CLEAR_QUE);
 701
 702     TRACE("Clearing NBD socket");
 703     ioctl(fd, NBD_CLEAR_SOCK);
 704
 705     errno = serrno;
 706     return ret;
 707 }
 708 #else
 709 int nbd_init(int fd, int csock, uint32_t flags, off_t size, size_t blocksize)
 710 {
 711     return -ENOTSUP;
 712 }
 713
 714 int nbd_disconnect(int fd)
 715 {
 716     return -ENOTSUP;
 717 }
 718
 719 int nbd_client(int fd)
 720 {
 721     return -ENOTSUP;
 722 }
 723 #endif
 724
 725 ssize_t nbd_send_request(int csock, struct nbd_request *request)
 726 {
 727     uint8_t buf[NBD_REQUEST_SIZE];
 728     ssize_t ret;
 729
 730     cpu_to_be32w((uint32_t*)buf, NBD_REQUEST_MAGIC);
 731     cpu_to_be32w((uint32_t*)(buf + 4), request->type);
 732     cpu_to_be64w((uint64_t*)(buf + 8), request->handle);
 733     cpu_to_be64w((uint64_t*)(buf + 16), request->from);
 734     cpu_to_be32w((uint32_t*)(buf + 24), request->len);
 735
 736     TRACE("Sending request to client: "
 737           "{ .from = %" PRIu64", .len = %u, .handle = %" PRIu64", .type=%i}",
 738           request->from, request->len, request->handle, request->type);
 739
 740     ret = write_sync(csock, buf, sizeof(buf));
 741     if (ret < 0) {
 742         return ret;
 743     }
 744
 745     if (ret != sizeof(buf)) {
 746         LOG("writing to socket failed");
 747         return -EINVAL;
 748     }
 749     return 0;
 750 }
 751
 752 static ssize_t nbd_receive_request(int csock, struct nbd_request *request)
 753 {
 754     uint8_t buf[NBD_REQUEST_SIZE];
 755     uint32_t magic;
 756     ssize_t ret;
 757
 758     ret = read_sync(csock, buf, sizeof(buf));
 759     if (ret < 0) {
 760         return ret;
 761     }
 762
 763     if (ret != sizeof(buf)) {
 764         LOG("read failed");
 765         return -EINVAL;
 766     }
 767
 768     /* Request
 769        [ 0 ..  3]   magic   (NBD_REQUEST_MAGIC)
 770        [ 4 ..  7]   type    (0 == READ, 1 == WRITE)
 771        [ 8 .. 15]   handle
 772        [16 .. 23]   from
 773        [24 .. 27]   len
 774      */
 775
 776     magic = be32_to_cpup((uint32_t*)buf);
 777     request->type  = be32_to_cpup((uint32_t*)(buf + 4));
 778     request->handle = be64_to_cpup((uint64_t*)(buf + 8));
 779     request->from  = be64_to_cpup((uint64_t*)(buf + 16));
 780     request->len   = be32_to_cpup((uint32_t*)(buf + 24));
 781
 782     TRACE("Got request: "
 783           "{ magic = 0x%x, .type = %d, from = %" PRIu64" , len = %u }",
 784           magic, request->type, request->from, request->len);
 785
 786     if (magic != NBD_REQUEST_MAGIC) {
 787         LOG("invalid magic (got 0x%x)", magic);
 788         return -EINVAL;
 789     }
 790     return 0;
 791 }
 792
 793 ssize_t nbd_receive_reply(int csock, struct nbd_reply *reply)
 794 {
 795     uint8_t buf[NBD_REPLY_SIZE];
 796     uint32_t magic;
 797     ssize_t ret;
 798
 799     ret = read_sync(csock, buf, sizeof(buf));
 800     if (ret < 0) {
 801         return ret;
 802     }
 803
 804     if (ret != sizeof(buf)) {
 805         LOG("read failed");
 806         return -EINVAL;
 807     }
 808
 809     /* Reply
 810        [ 0 ..  3]    magic   (NBD_REPLY_MAGIC)
 811        [ 4 ..  7]    error   (0 == no error)
 812        [ 7 .. 15]    handle
 813      */
 814
 815     magic = be32_to_cpup((uint32_t*)buf);
 816     reply->error  = be32_to_cpup((uint32_t*)(buf + 4));
 817     reply->handle = be64_to_cpup((uint64_t*)(buf + 8));
 818
 819     TRACE("Got reply: "
 820           "{ magic = 0x%x, .error = %d, handle = %" PRIu64" }",
 821           magic, reply->error, reply->handle);
 822
 823     if (magic != NBD_REPLY_MAGIC) {
 824         LOG("invalid magic (got 0x%x)", magic);
 825         return -EINVAL;
 826     }
 827     return 0;
 828 }
 829
 830 static ssize_t nbd_send_reply(int csock, struct nbd_reply *reply)
 831 {
 832     uint8_t buf[NBD_REPLY_SIZE];
 833     ssize_t ret;
 834
 835     /* Reply
 836        [ 0 ..  3]    magic   (NBD_REPLY_MAGIC)
 837        [ 4 ..  7]    error   (0 == no error)
 838        [ 7 .. 15]    handle
 839      */
 840     cpu_to_be32w((uint32_t*)buf, NBD_REPLY_MAGIC);
 841     cpu_to_be32w((uint32_t*)(buf + 4), reply->error);
 842     cpu_to_be64w((uint64_t*)(buf + 8), reply->handle);
 843
 844     TRACE("Sending response to client");
 845
 846     ret = write_sync(csock, buf, sizeof(buf));
 847     if (ret < 0) {
 848         return ret;
 849     }
 850
 851     if (ret != sizeof(buf)) {
 852         LOG("writing to socket failed");
 853         return -EINVAL;
 854     }
 855     return 0;
 856 }
 857
 858 #define MAX_NBD_REQUESTS 16
 859
 860 void nbd_client_get(NBDClient *client)
 861 {
 862     client->refcount++;
 863 }
 864
 865 void nbd_client_put(NBDClient *client)
 866 {
 867     if (--client->refcount == 0) {
 868         /* The last reference should be dropped by client->close,
 869          * which is called by nbd_client_close.
 870          */
 871         assert(client->closing);
 872
 873         nbd_unset_handlers(client);
 874         close(client->sock);
 875         client->sock = -1;
 876         if (client->exp) {
 877             QTAILQ_REMOVE(&client->exp->clients, client, next);
 878             nbd_export_put(client->exp);
 879         }
 880         g_free(client);
 881     }
 882 }
 883
 884 void nbd_client_close(NBDClient *client)
 885 {
 886     if (client->closing) {
 887         return;
 888     }
 889
 890     client->closing = true;
 891
 892     /* Force requests to finish.  They will drop their own references,
 893      * then we'll close the socket and free the NBDClient.
 894      */
 895     shutdown(client->sock, 2);
 896
 897     /* Also tell the client, so that they release their reference.  */
 898     if (client->close) {
 899         client->close(client);
 900     }
 901 }
 902
 903 static NBDRequest *nbd_request_get(NBDClient *client)
 904 {
 905     NBDRequest *req;
 906
 907     assert(client->nb_requests <= MAX_NBD_REQUESTS - 1);
 908     client->nb_requests++;
 909     nbd_update_can_read(client);
 910
 911     req = g_slice_new0(NBDRequest);
 912     nbd_client_get(client);
 913     req->client = client;
 914     return req;
 915 }
 916
 917 static void nbd_request_put(NBDRequest *req)
 918 {
 919     NBDClient *client = req->client;
 920
 921     if (req->data) {
 922         qemu_vfree(req->data);
 923     }
 924     g_slice_free(NBDRequest, req);
 925
 926     client->nb_requests--;
 927     nbd_update_can_read(client);
 928     nbd_client_put(client);
 929 }
 930
 931 static void blk_aio_attached(AioContext *ctx, void *opaque)
 932 {
 933     NBDExport *exp = opaque;
 934     NBDClient *client;
 935
 936     TRACE("Export %s: Attaching clients to AIO context %p\n", exp->name, ctx);
 937
 938     exp->ctx = ctx;
 939
 940     QTAILQ_FOREACH(client, &exp->clients, next) {
 941         nbd_set_handlers(client);
 942     }
 943 }
 944
 945 static void blk_aio_detach(void *opaque)
 946 {
 947     NBDExport *exp = opaque;
 948     NBDClient *client;
 949
 950     TRACE("Export %s: Detaching clients from AIO context %p\n", exp->name, exp->ctx);
 951
 952     QTAILQ_FOREACH(client, &exp->clients, next) {
 953         nbd_unset_handlers(client);
 954     }
 955
 956     exp->ctx = NULL;
 957 }
 958
 959 NBDExport *nbd_export_new(BlockBackend *blk, off_t dev_offset, off_t size,
 960                           uint32_t nbdflags, void (*close)(NBDExport *))
 961 {
 962     NBDExport *exp = g_malloc0(sizeof(NBDExport));
 963     exp->refcount = 1;
 964     QTAILQ_INIT(&exp->clients);
 965     exp->blk = blk;
 966     exp->dev_offset = dev_offset;
 967     exp->nbdflags = nbdflags;
 968     exp->size = size == -1 ? blk_getlength(blk) : size;
 969     exp->close = close;
 970     exp->ctx = blk_get_aio_context(blk);
 971     blk_ref(blk);
 972     blk_add_aio_context_notifier(blk, blk_aio_attached, blk_aio_detach, exp);
 973     /*
 974      * NBD exports are used for non-shared storage migration.  Make sure
 975      * that BDRV_O_INCOMING is cleared and the image is ready for write
 976      * access since the export could be available before migration handover.
 977      */
 978     blk_invalidate_cache(blk, NULL);
 979     return exp;
 980 }
 981
 982 NBDExport *nbd_export_find(const char *name)
 983 {
 984     NBDExport *exp;
 985     QTAILQ_FOREACH(exp, &exports, next) {
 986         if (strcmp(name, exp->name) == 0) {
 987             return exp;
 988         }
 989     }
 990
 991     return NULL;
 992 }
 993
 994 void nbd_export_set_name(NBDExport *exp, const char *name)
 995 {
 996     if (exp->name == name) {
 997         return;
 998     }
 999
1000     nbd_export_get(exp);
1001     if (exp->name != NULL) {
1002         g_free(exp->name);
1003         exp->name = NULL;
1004         QTAILQ_REMOVE(&exports, exp, next);
1005         nbd_export_put(exp);
1006     }
1007     if (name != NULL) {
1008         nbd_export_get(exp);
1009         exp->name = g_strdup(name);
1010         QTAILQ_INSERT_TAIL(&exports, exp, next);
1011     }
1012     nbd_export_put(exp);
1013 }
1014
1015 void nbd_export_close(NBDExport *exp)
1016 {
1017     NBDClient *client, *next;
1018
1019     nbd_export_get(exp);
1020     QTAILQ_FOREACH_SAFE(client, &exp->clients, next, next) {
1021         nbd_client_close(client);
1022     }
1023     nbd_export_set_name(exp, NULL);
1024     nbd_export_put(exp);
1025     if (exp->blk) {
1026         blk_remove_aio_context_notifier(exp->blk, blk_aio_attached,
1027                                         blk_aio_detach, exp);
1028         blk_unref(exp->blk);
1029         exp->blk = NULL;
1030     }
1031 }
1032
1033 void nbd_export_get(NBDExport *exp)
1034 {
1035     assert(exp->refcount > 0);
1036     exp->refcount++;
1037 }
1038
1039 void nbd_export_put(NBDExport *exp)
1040 {
1041     assert(exp->refcount > 0);
1042     if (exp->refcount == 1) {
1043         nbd_export_close(exp);
1044     }
1045
1046     if (--exp->refcount == 0) {
1047         assert(exp->name == NULL);
1048
1049         if (exp->close) {
1050             exp->close(exp);
1051         }
1052
1053         g_free(exp);
1054     }
1055 }
1056
1057 BlockBackend *nbd_export_get_blockdev(NBDExport *exp)
1058 {
1059     return exp->blk;
1060 }
1061
1062 void nbd_export_close_all(void)
1063 {
1064     NBDExport *exp, *next;
1065
1066     QTAILQ_FOREACH_SAFE(exp, &exports, next, next) {
1067         nbd_export_close(exp);
1068     }
1069 }
1070
1071 static ssize_t nbd_co_send_reply(NBDRequest *req, struct nbd_reply *reply,
1072                                  int len)
1073 {
1074     NBDClient *client = req->client;
1075     int csock = client->sock;
1076     ssize_t rc, ret;
1077
1078     qemu_co_mutex_lock(&client->send_lock);
1079     client->send_coroutine = qemu_coroutine_self();
1080     nbd_set_handlers(client);
1081
1082     if (!len) {
1083         rc = nbd_send_reply(csock, reply);
1084     } else {
1085         socket_set_cork(csock, 1);
1086         rc = nbd_send_reply(csock, reply);
1087         if (rc >= 0) {
1088             ret = qemu_co_send(csock, req->data, len);
1089             if (ret != len) {
1090                 rc = -EIO;
1091             }
1092         }
1093         socket_set_cork(csock, 0);
1094     }
1095
1096     client->send_coroutine = NULL;
1097     nbd_set_handlers(client);
1098     qemu_co_mutex_unlock(&client->send_lock);
1099     return rc;
1100 }
1101
1102 static ssize_t nbd_co_receive_request(NBDRequest *req, struct nbd_request *request)
1103 {
1104     NBDClient *client = req->client;
1105     int csock = client->sock;
1106     uint32_t command;
1107     ssize_t rc;
1108
1109     client->recv_coroutine = qemu_coroutine_self();
1110     nbd_update_can_read(client);
1111
1112     rc = nbd_receive_request(csock, request);
1113     if (rc < 0) {
1114         if (rc != -EAGAIN) {
1115             rc = -EIO;
1116         }
1117         goto out;
1118     }
1119
1120     if (request->len > NBD_MAX_BUFFER_SIZE) {
1121         LOG("len (%u) is larger than max len (%u)",
1122             request->len, NBD_MAX_BUFFER_SIZE);
1123         rc = -EINVAL;
1124         goto out;
1125     }
1126
1127     if ((request->from + request->len) < request->from) {
1128         LOG("integer overflow detected! "
1129             "you're probably being attacked");
1130         rc = -EINVAL;
1131         goto out;
1132     }
1133
1134     TRACE("Decoding type");
1135
1136     command = request->type & NBD_CMD_MASK_COMMAND;
1137     if (command == NBD_CMD_READ || command == NBD_CMD_WRITE) {
1138         req->data = blk_blockalign(client->exp->blk, request->len);
1139     }
1140     if (command == NBD_CMD_WRITE) {
1141         TRACE("Reading %u byte(s)", request->len);
1142
1143         if (qemu_co_recv(csock, req->data, request->len) != request->len) {
1144             LOG("reading from socket failed");
1145             rc = -EIO;
1146             goto out;
1147         }
1148     }
1149     rc = 0;
1150
1151 out:
1152     client->recv_coroutine = NULL;
1153     nbd_update_can_read(client);
1154
1155     return rc;
1156 }
1157
1158 static void nbd_trip(void *opaque)
1159 {
1160     NBDClient *client = opaque;
1161     NBDExport *exp = client->exp;
1162     NBDRequest *req;
1163     struct nbd_request request;
1164     struct nbd_reply reply;
1165     ssize_t ret;
1166     uint32_t command;
1167
1168     TRACE("Reading request.");
1169     if (client->closing) {
1170         return;
1171     }
1172
1173     req = nbd_request_get(client);
1174     ret = nbd_co_receive_request(req, &request);
1175     if (ret == -EAGAIN) {
1176         goto done;
1177     }
1178     if (ret == -EIO) {
1179         goto out;
1180     }
1181
1182     reply.handle = request.handle;
1183     reply.error = 0;
1184
1185     if (ret < 0) {
1186         reply.error = -ret;
1187         goto error_reply;
1188     }
1189     command = request.type & NBD_CMD_MASK_COMMAND;
1190     if (command != NBD_CMD_DISC && (request.from + request.len) > exp->size) {
1191             LOG("From: %" PRIu64 ", Len: %u, Size: %" PRIu64
1192             ", Offset: %" PRIu64 "\n",
1193                     request.from, request.len,
1194                     (uint64_t)exp->size, (uint64_t)exp->dev_offset);
1195         LOG("requested operation past EOF--bad client?");
1196         goto invalid_request;
1197     }
1198
1199     switch (command) {
1200     case NBD_CMD_READ:
1201         TRACE("Request type is READ");
1202
1203         if (request.type & NBD_CMD_FLAG_FUA) {
1204             ret = blk_co_flush(exp->blk);
1205             if (ret < 0) {
1206                 LOG("flush failed");
1207                 reply.error = -ret;
1208                 goto error_reply;
1209             }
1210         }
1211
1212         ret = blk_read(exp->blk,
1213                        (request.from + exp->dev_offset) / BDRV_SECTOR_SIZE,
1214                        req->data, request.len / BDRV_SECTOR_SIZE);
1215         if (ret < 0) {
1216             LOG("reading from file failed");
1217             reply.error = -ret;
1218             goto error_reply;
1219         }
1220
1221         TRACE("Read %u byte(s)", request.len);
1222         if (nbd_co_send_reply(req, &reply, request.len) < 0)
1223             goto out;
1224         break;
1225     case NBD_CMD_WRITE:
1226         TRACE("Request type is WRITE");
1227
1228         if (exp->nbdflags & NBD_FLAG_READ_ONLY) {
1229             TRACE("Server is read-only, return error");
1230             reply.error = EROFS;
1231             goto error_reply;
1232         }
1233
1234         TRACE("Writing to device");
1235
1236         ret = blk_write(exp->blk,
1237                         (request.from + exp->dev_offset) / BDRV_SECTOR_SIZE,
1238                         req->data, request.len / BDRV_SECTOR_SIZE);
1239         if (ret < 0) {
1240             LOG("writing to file failed");
1241             reply.error = -ret;
1242             goto error_reply;
1243         }
1244
1245         if (request.type & NBD_CMD_FLAG_FUA) {
1246             ret = blk_co_flush(exp->blk);
1247             if (ret < 0) {
1248                 LOG("flush failed");
1249                 reply.error = -ret;
1250                 goto error_reply;
1251             }
1252         }
1253
1254         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1255             goto out;
1256         }
1257         break;
1258     case NBD_CMD_DISC:
1259         TRACE("Request type is DISCONNECT");
1260         errno = 0;
1261         goto out;
1262     case NBD_CMD_FLUSH:
1263         TRACE("Request type is FLUSH");
1264
1265         ret = blk_co_flush(exp->blk);
1266         if (ret < 0) {
1267             LOG("flush failed");
1268             reply.error = -ret;
1269         }
1270         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1271             goto out;
1272         }
1273         break;
1274     case NBD_CMD_TRIM:
1275         TRACE("Request type is TRIM");
1276         ret = blk_co_discard(exp->blk, (request.from + exp->dev_offset)
1277                                        / BDRV_SECTOR_SIZE,
1278                              request.len / BDRV_SECTOR_SIZE);
1279         if (ret < 0) {
1280             LOG("discard failed");
1281             reply.error = -ret;
1282         }
1283         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1284             goto out;
1285         }
1286         break;
1287     default:
1288         LOG("invalid request type (%u) received", request.type);
1289     invalid_request:
1290         reply.error = -EINVAL;
1291     error_reply:
1292         if (nbd_co_send_reply(req, &reply, 0) < 0) {
1293             goto out;
1294         }
1295         break;
1296     }
1297
1298     TRACE("Request/Reply complete");
1299
1300 done:
1301     nbd_request_put(req);
1302     return;
1303
1304 out:
1305     nbd_request_put(req);
1306     nbd_client_close(client);
1307 }
1308
1309 static void nbd_read(void *opaque)
1310 {
1311     NBDClient *client = opaque;
1312
1313     if (client->recv_coroutine) {
1314         qemu_coroutine_enter(client->recv_coroutine, NULL);
1315     } else {
1316         qemu_coroutine_enter(qemu_coroutine_create(nbd_trip), client);
1317     }
1318 }
1319
1320 static void nbd_restart_write(void *opaque)
1321 {
1322     NBDClient *client = opaque;
1323
1324     qemu_coroutine_enter(client->send_coroutine, NULL);
1325 }
1326
1327 static void nbd_set_handlers(NBDClient *client)
1328 {
1329     if (client->exp && client->exp->ctx) {
1330         aio_set_fd_handler(client->exp->ctx, client->sock,
1331                            client->can_read ? nbd_read : NULL,
1332                            client->send_coroutine ? nbd_restart_write : NULL,
1333                            client);
1334     }
1335 }
1336
1337 static void nbd_unset_handlers(NBDClient *client)
1338 {
1339     if (client->exp && client->exp->ctx) {
1340         aio_set_fd_handler(client->exp->ctx, client->sock, NULL, NULL, NULL);
1341     }
1342 }
1343
1344 static void nbd_update_can_read(NBDClient *client)
1345 {
1346     bool can_read = client->recv_coroutine ||
1347                     client->nb_requests < MAX_NBD_REQUESTS;
1348
1349     if (can_read != client->can_read) {
1350         client->can_read = can_read;
1351         nbd_set_handlers(client);
1352
1353         /* There is no need to invoke aio_notify(), since aio_set_fd_handler()
1354          * in nbd_set_handlers() will have taken care of that */
1355     }
1356 }
1357
1358 NBDClient *nbd_client_new(NBDExport *exp, int csock,
1359                           void (*close)(NBDClient *))
1360 {
1361     NBDClient *client;
1362     client = g_malloc0(sizeof(NBDClient));
1363     client->refcount = 1;
1364     client->exp = exp;
1365     client->sock = csock;
1366     client->can_read = true;
1367     if (nbd_send_negotiate(client)) {
1368         g_free(client);
1369         return NULL;
1370     }
1371     client->close = close;
1372     qemu_co_mutex_init(&client->send_lock);
1373     nbd_set_handlers(client);
1374
1375     if (exp) {
1376         QTAILQ_INSERT_TAIL(&exp->clients, client, next);
1377         nbd_export_get(exp);
1378     }
1379     return client;
1380 }