block/iscsi.c

   1 /*
   2  * QEMU Block driver for iSCSI images
   3  *
   4  * Copyright (c) 2010-2011 Ronnie Sahlberg <ronniesahlberg@gmail.com>
   5  * Copyright (c) 2012-2014 Peter Lieven <pl@kamp.de>
   6  *
   7  * Permission is hereby granted, free of charge, to any person obtaining a copy
   8  * of this software and associated documentation files (the "Software"), to deal
   9  * in the Software without restriction, including without limitation the rights
  10  * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
  11  * copies of the Software, and to permit persons to whom the Software is
  12  * furnished to do so, subject to the following conditions:
  13  *
  14  * The above copyright notice and this permission notice shall be included in
  15  * all copies or substantial portions of the Software.
  16  *
  17  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  18  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  19  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
  20  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  21  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  22  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
  23  * THE SOFTWARE.
  24  */
  25
  26 #include "config-host.h"
  27
  28 #include <poll.h>
  29 #include <arpa/inet.h>
  30 #include "qemu-common.h"
  31 #include "qemu/config-file.h"
  32 #include "qemu/error-report.h"
  33 #include "qemu/bitops.h"
  34 #include "qemu/bitmap.h"
  35 #include "block/block_int.h"
  36 #include "trace.h"
  37 #include "block/scsi.h"
  38 #include "qemu/iov.h"
  39 #include "sysemu/sysemu.h"
  40 #include "qmp-commands.h"
  41
  42 #include <iscsi/iscsi.h>
  43 #include <iscsi/scsi-lowlevel.h>
  44
  45 #ifdef __linux__
  46 #include <scsi/sg.h>
  47 #include <block/scsi.h>
  48 #endif
  49
  50 typedef struct IscsiLun {
  51     struct iscsi_context *iscsi;
  52     int lun;
  53     enum scsi_inquiry_peripheral_device_type type;
  54     int block_size;
  55     uint64_t num_blocks;
  56     int events;
  57     QEMUTimer *nop_timer;
  58     uint8_t lbpme;
  59     uint8_t lbprz;
  60     uint8_t has_write_same;
  61     struct scsi_inquiry_logical_block_provisioning lbp;
  62     struct scsi_inquiry_block_limits bl;
  63     unsigned char *zeroblock;
  64     unsigned long *allocationmap;
  65     int cluster_sectors;
  66 } IscsiLun;
  67
  68 typedef struct IscsiTask {
  69     int status;
  70     int complete;
  71     int retries;
  72     int do_retry;
  73     struct scsi_task *task;
  74     Coroutine *co;
  75     QEMUBH *bh;
  76 } IscsiTask;
  77
  78 typedef struct IscsiAIOCB {
  79     BlockDriverAIOCB common;
  80     QEMUIOVector *qiov;
  81     QEMUBH *bh;
  82     IscsiLun *iscsilun;
  83     struct scsi_task *task;
  84     uint8_t *buf;
  85     int status;
  86     int canceled;
  87     int retries;
  88     int64_t sector_num;
  89     int nb_sectors;
  90 #ifdef __linux__
  91     sg_io_hdr_t *ioh;
  92 #endif
  93 } IscsiAIOCB;
  94
  95 #define NOP_INTERVAL 5000
  96 #define MAX_NOP_FAILURES 3
  97 #define ISCSI_CMD_RETRIES 5
  98
  99 /* this threshhold is a trade-off knob to choose between
 100  * the potential additional overhead of an extra GET_LBA_STATUS request
 101  * vs. unnecessarily reading a lot of zero sectors over the wire.
 102  * If a read request is greater or equal than ISCSI_CHECKALLOC_THRES
 103  * sectors we check the allocation status of the area covered by the
 104  * request first if the allocationmap indicates that the area might be
 105  * unallocated. */
 106 #define ISCSI_CHECKALLOC_THRES 64
 107
 108 static void
 109 iscsi_bh_cb(void *p)
 110 {
 111     IscsiAIOCB *acb = p;
 112
 113     qemu_bh_delete(acb->bh);
 114
 115     g_free(acb->buf);
 116     acb->buf = NULL;
 117
 118     if (acb->canceled == 0) {
 119         acb->common.cb(acb->common.opaque, acb->status);
 120     }
 121
 122     if (acb->task != NULL) {
 123         scsi_free_scsi_task(acb->task);
 124         acb->task = NULL;
 125     }
 126
 127     qemu_aio_release(acb);
 128 }
 129
 130 static void
 131 iscsi_schedule_bh(IscsiAIOCB *acb)
 132 {
 133     if (acb->bh) {
 134         return;
 135     }
 136     acb->bh = qemu_bh_new(iscsi_bh_cb, acb);
 137     qemu_bh_schedule(acb->bh);
 138 }
 139
 140 static void iscsi_co_generic_bh_cb(void *opaque)
 141 {
 142     struct IscsiTask *iTask = opaque;
 143     qemu_bh_delete(iTask->bh);
 144     qemu_coroutine_enter(iTask->co, NULL);
 145 }
 146
 147 static void
 148 iscsi_co_generic_cb(struct iscsi_context *iscsi, int status,
 149                         void *command_data, void *opaque)
 150 {
 151     struct IscsiTask *iTask = opaque;
 152     struct scsi_task *task = command_data;
 153
 154     iTask->complete = 1;
 155     iTask->status = status;
 156     iTask->do_retry = 0;
 157     iTask->task = task;
 158
 159     if (iTask->retries-- > 0 && status == SCSI_STATUS_CHECK_CONDITION
 160         && task->sense.key == SCSI_SENSE_UNIT_ATTENTION) {
 161         error_report("iSCSI CheckCondition: %s", iscsi_get_error(iscsi));
 162         iTask->do_retry = 1;
 163         goto out;
 164     }
 165
 166     if (status != SCSI_STATUS_GOOD) {
 167         error_report("iSCSI Failure: %s", iscsi_get_error(iscsi));
 168     }
 169
 170 out:
 171     if (iTask->co) {
 172         iTask->bh = qemu_bh_new(iscsi_co_generic_bh_cb, iTask);
 173         qemu_bh_schedule(iTask->bh);
 174     }
 175 }
 176
 177 static void iscsi_co_init_iscsitask(IscsiLun *iscsilun, struct IscsiTask *iTask)
 178 {
 179     *iTask = (struct IscsiTask) {
 180         .co         = qemu_coroutine_self(),
 181         .retries    = ISCSI_CMD_RETRIES,
 182     };
 183 }
 184
 185 static void
 186 iscsi_abort_task_cb(struct iscsi_context *iscsi, int status, void *command_data,
 187                     void *private_data)
 188 {
 189     IscsiAIOCB *acb = private_data;
 190
 191     acb->status = -ECANCELED;
 192     iscsi_schedule_bh(acb);
 193 }
 194
 195 static void
 196 iscsi_aio_cancel(BlockDriverAIOCB *blockacb)
 197 {
 198     IscsiAIOCB *acb = (IscsiAIOCB *)blockacb;
 199     IscsiLun *iscsilun = acb->iscsilun;
 200
 201     if (acb->status != -EINPROGRESS) {
 202         return;
 203     }
 204
 205     acb->canceled = 1;
 206
 207     /* send a task mgmt call to the target to cancel the task on the target */
 208     iscsi_task_mgmt_abort_task_async(iscsilun->iscsi, acb->task,
 209                                      iscsi_abort_task_cb, acb);
 210
 211     while (acb->status == -EINPROGRESS) {
 212         qemu_aio_wait();
 213     }
 214 }
 215
 216 static const AIOCBInfo iscsi_aiocb_info = {
 217     .aiocb_size         = sizeof(IscsiAIOCB),
 218     .cancel             = iscsi_aio_cancel,
 219 };
 220
 221
 222 static void iscsi_process_read(void *arg);
 223 static void iscsi_process_write(void *arg);
 224
 225 static void
 226 iscsi_set_events(IscsiLun *iscsilun)
 227 {
 228     struct iscsi_context *iscsi = iscsilun->iscsi;
 229     int ev;
 230
 231     /* We always register a read handler.  */
 232     ev = POLLIN;
 233     ev |= iscsi_which_events(iscsi);
 234     if (ev != iscsilun->events) {
 235         qemu_aio_set_fd_handler(iscsi_get_fd(iscsi),
 236                       iscsi_process_read,
 237                       (ev & POLLOUT) ? iscsi_process_write : NULL,
 238                       iscsilun);
 239
 240     }
 241
 242     iscsilun->events = ev;
 243 }
 244
 245 static void
 246 iscsi_process_read(void *arg)
 247 {
 248     IscsiLun *iscsilun = arg;
 249     struct iscsi_context *iscsi = iscsilun->iscsi;
 250
 251     iscsi_service(iscsi, POLLIN);
 252     iscsi_set_events(iscsilun);
 253 }
 254
 255 static void
 256 iscsi_process_write(void *arg)
 257 {
 258     IscsiLun *iscsilun = arg;
 259     struct iscsi_context *iscsi = iscsilun->iscsi;
 260
 261     iscsi_service(iscsi, POLLOUT);
 262     iscsi_set_events(iscsilun);
 263 }
 264
 265 static int64_t sector_lun2qemu(int64_t sector, IscsiLun *iscsilun)
 266 {
 267     return sector * iscsilun->block_size / BDRV_SECTOR_SIZE;
 268 }
 269
 270 static int64_t sector_qemu2lun(int64_t sector, IscsiLun *iscsilun)
 271 {
 272     return sector * BDRV_SECTOR_SIZE / iscsilun->block_size;
 273 }
 274
 275 static bool is_request_lun_aligned(int64_t sector_num, int nb_sectors,
 276                                       IscsiLun *iscsilun)
 277 {
 278     if ((sector_num * BDRV_SECTOR_SIZE) % iscsilun->block_size ||
 279         (nb_sectors * BDRV_SECTOR_SIZE) % iscsilun->block_size) {
 280             error_report("iSCSI misaligned request: "
 281                          "iscsilun->block_size %u, sector_num %" PRIi64
 282                          ", nb_sectors %d",
 283                          iscsilun->block_size, sector_num, nb_sectors);
 284             return 0;
 285     }
 286     return 1;
 287 }
 288
 289 static void iscsi_allocationmap_set(IscsiLun *iscsilun, int64_t sector_num,
 290                                     int nb_sectors)
 291 {
 292     if (iscsilun->allocationmap == NULL) {
 293         return;
 294     }
 295     bitmap_set(iscsilun->allocationmap,
 296                sector_num / iscsilun->cluster_sectors,
 297                DIV_ROUND_UP(nb_sectors, iscsilun->cluster_sectors));
 298 }
 299
 300 static void iscsi_allocationmap_clear(IscsiLun *iscsilun, int64_t sector_num,
 301                                       int nb_sectors)
 302 {
 303     int64_t cluster_num, nb_clusters;
 304     if (iscsilun->allocationmap == NULL) {
 305         return;
 306     }
 307     cluster_num = DIV_ROUND_UP(sector_num, iscsilun->cluster_sectors);
 308     nb_clusters = (sector_num + nb_sectors) / iscsilun->cluster_sectors
 309                   - cluster_num;
 310     if (nb_clusters > 0) {
 311         bitmap_clear(iscsilun->allocationmap, cluster_num, nb_clusters);
 312     }
 313 }
 314
 315 static int coroutine_fn iscsi_co_writev(BlockDriverState *bs,
 316                                         int64_t sector_num, int nb_sectors,
 317                                         QEMUIOVector *iov)
 318 {
 319     IscsiLun *iscsilun = bs->opaque;
 320     struct IscsiTask iTask;
 321     uint64_t lba;
 322     uint32_t num_sectors;
 323     uint8_t *data = NULL;
 324     uint8_t *buf = NULL;
 325
 326     if (!is_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
 327         return -EINVAL;
 328     }
 329
 330     lba = sector_qemu2lun(sector_num, iscsilun);
 331     num_sectors = sector_qemu2lun(nb_sectors, iscsilun);
 332 #if !defined(LIBISCSI_FEATURE_IOVECTOR)
 333     /* if the iovec only contains one buffer we can pass it directly */
 334     if (iov->niov == 1) {
 335         data = iov->iov[0].iov_base;
 336     } else {
 337         size_t size = MIN(nb_sectors * BDRV_SECTOR_SIZE, iov->size);
 338         buf = g_malloc(size);
 339         qemu_iovec_to_buf(iov, 0, buf, size);
 340         data = buf;
 341     }
 342 #endif
 343     iscsi_co_init_iscsitask(iscsilun, &iTask);
 344 retry:
 345     iTask.task = iscsi_write16_task(iscsilun->iscsi, iscsilun->lun, lba,
 346                                     data, num_sectors * iscsilun->block_size,
 347                                     iscsilun->block_size, 0, 0, 0, 0, 0,
 348                                     iscsi_co_generic_cb, &iTask);
 349     if (iTask.task == NULL) {
 350         g_free(buf);
 351         return -ENOMEM;
 352     }
 353 #if defined(LIBISCSI_FEATURE_IOVECTOR)
 354     scsi_task_set_iov_out(iTask.task, (struct scsi_iovec *) iov->iov,
 355                           iov->niov);
 356 #endif
 357     while (!iTask.complete) {
 358         iscsi_set_events(iscsilun);
 359         qemu_coroutine_yield();
 360     }
 361
 362     if (iTask.task != NULL) {
 363         scsi_free_scsi_task(iTask.task);
 364         iTask.task = NULL;
 365     }
 366
 367     if (iTask.do_retry) {
 368         iTask.complete = 0;
 369         goto retry;
 370     }
 371
 372     g_free(buf);
 373
 374     if (iTask.status != SCSI_STATUS_GOOD) {
 375         return -EIO;
 376     }
 377
 378     iscsi_allocationmap_set(iscsilun, sector_num, nb_sectors);
 379
 380     return 0;
 381 }
 382
 383
 384 #if defined(LIBISCSI_FEATURE_IOVECTOR)
 385 static bool iscsi_allocationmap_is_allocated(IscsiLun *iscsilun,
 386                                              int64_t sector_num, int nb_sectors)
 387 {
 388     unsigned long size;
 389     if (iscsilun->allocationmap == NULL) {
 390         return true;
 391     }
 392     size = DIV_ROUND_UP(sector_num + nb_sectors, iscsilun->cluster_sectors);
 393     return !(find_next_bit(iscsilun->allocationmap, size,
 394                            sector_num / iscsilun->cluster_sectors) == size);
 395 }
 396
 397 static int64_t coroutine_fn iscsi_co_get_block_status(BlockDriverState *bs,
 398                                                   int64_t sector_num,
 399                                                   int nb_sectors, int *pnum)
 400 {
 401     IscsiLun *iscsilun = bs->opaque;
 402     struct scsi_get_lba_status *lbas = NULL;
 403     struct scsi_lba_status_descriptor *lbasd = NULL;
 404     struct IscsiTask iTask;
 405     int64_t ret;
 406
 407     iscsi_co_init_iscsitask(iscsilun, &iTask);
 408
 409     if (!is_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
 410         ret = -EINVAL;
 411         goto out;
 412     }
 413
 414     /* default to all sectors allocated */
 415     ret = BDRV_BLOCK_DATA;
 416     ret |= (sector_num << BDRV_SECTOR_BITS) | BDRV_BLOCK_OFFSET_VALID;
 417     *pnum = nb_sectors;
 418
 419     /* LUN does not support logical block provisioning */
 420     if (iscsilun->lbpme == 0) {
 421         goto out;
 422     }
 423
 424 retry:
 425     if (iscsi_get_lba_status_task(iscsilun->iscsi, iscsilun->lun,
 426                                   sector_qemu2lun(sector_num, iscsilun),
 427                                   8 + 16, iscsi_co_generic_cb,
 428                                   &iTask) == NULL) {
 429         ret = -ENOMEM;
 430         goto out;
 431     }
 432
 433     while (!iTask.complete) {
 434         iscsi_set_events(iscsilun);
 435         qemu_coroutine_yield();
 436     }
 437
 438     if (iTask.do_retry) {
 439         if (iTask.task != NULL) {
 440             scsi_free_scsi_task(iTask.task);
 441             iTask.task = NULL;
 442         }
 443         iTask.complete = 0;
 444         goto retry;
 445     }
 446
 447     if (iTask.status != SCSI_STATUS_GOOD) {
 448         /* in case the get_lba_status_callout fails (i.e.
 449          * because the device is busy or the cmd is not
 450          * supported) we pretend all blocks are allocated
 451          * for backwards compatibility */
 452         goto out;
 453     }
 454
 455     lbas = scsi_datain_unmarshall(iTask.task);
 456     if (lbas == NULL) {
 457         ret = -EIO;
 458         goto out;
 459     }
 460
 461     lbasd = &lbas->descriptors[0];
 462
 463     if (sector_qemu2lun(sector_num, iscsilun) != lbasd->lba) {
 464         ret = -EIO;
 465         goto out;
 466     }
 467
 468     *pnum = sector_lun2qemu(lbasd->num_blocks, iscsilun);
 469
 470     if (lbasd->provisioning == SCSI_PROVISIONING_TYPE_DEALLOCATED ||
 471         lbasd->provisioning == SCSI_PROVISIONING_TYPE_ANCHORED) {
 472         ret &= ~BDRV_BLOCK_DATA;
 473         if (iscsilun->lbprz) {
 474             ret |= BDRV_BLOCK_ZERO;
 475         }
 476     }
 477
 478     if (ret & BDRV_BLOCK_ZERO) {
 479         iscsi_allocationmap_clear(iscsilun, sector_num, *pnum);
 480     } else {
 481         iscsi_allocationmap_set(iscsilun, sector_num, *pnum);
 482     }
 483
 484     if (*pnum > nb_sectors) {
 485         *pnum = nb_sectors;
 486     }
 487 out:
 488     if (iTask.task != NULL) {
 489         scsi_free_scsi_task(iTask.task);
 490     }
 491     return ret;
 492 }
 493
 494 #endif /* LIBISCSI_FEATURE_IOVECTOR */
 495
 496
 497 static int coroutine_fn iscsi_co_readv(BlockDriverState *bs,
 498                                        int64_t sector_num, int nb_sectors,
 499                                        QEMUIOVector *iov)
 500 {
 501     IscsiLun *iscsilun = bs->opaque;
 502     struct IscsiTask iTask;
 503     uint64_t lba;
 504     uint32_t num_sectors;
 505 #if !defined(LIBISCSI_FEATURE_IOVECTOR)
 506     int i;
 507 #endif
 508
 509     if (!is_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
 510         return -EINVAL;
 511     }
 512
 513 #if defined(LIBISCSI_FEATURE_IOVECTOR)
 514     if (iscsilun->lbprz && nb_sectors >= ISCSI_CHECKALLOC_THRES &&
 515         !iscsi_allocationmap_is_allocated(iscsilun, sector_num, nb_sectors)) {
 516         int64_t ret;
 517         int pnum;
 518         ret = iscsi_co_get_block_status(bs, sector_num, INT_MAX, &pnum);
 519         if (ret < 0) {
 520             return ret;
 521         }
 522         if (ret & BDRV_BLOCK_ZERO && pnum >= nb_sectors) {
 523             qemu_iovec_memset(iov, 0, 0x00, iov->size);
 524             return 0;
 525         }
 526     }
 527 #endif
 528
 529     lba = sector_qemu2lun(sector_num, iscsilun);
 530     num_sectors = sector_qemu2lun(nb_sectors, iscsilun);
 531
 532     iscsi_co_init_iscsitask(iscsilun, &iTask);
 533 retry:
 534     switch (iscsilun->type) {
 535     case TYPE_DISK:
 536         iTask.task = iscsi_read16_task(iscsilun->iscsi, iscsilun->lun, lba,
 537                                        num_sectors * iscsilun->block_size,
 538                                        iscsilun->block_size, 0, 0, 0, 0, 0,
 539                                        iscsi_co_generic_cb, &iTask);
 540         break;
 541     default:
 542         iTask.task = iscsi_read10_task(iscsilun->iscsi, iscsilun->lun, lba,
 543                                        num_sectors * iscsilun->block_size,
 544                                        iscsilun->block_size,
 545 #if !defined(CONFIG_LIBISCSI_1_4) /* API change from 1.4.0 to 1.5.0 */
 546                                        0, 0, 0, 0, 0,
 547 #endif
 548                                        iscsi_co_generic_cb, &iTask);
 549         break;
 550     }
 551     if (iTask.task == NULL) {
 552         return -ENOMEM;
 553     }
 554 #if defined(LIBISCSI_FEATURE_IOVECTOR)
 555     scsi_task_set_iov_in(iTask.task, (struct scsi_iovec *) iov->iov, iov->niov);
 556 #else
 557     for (i = 0; i < iov->niov; i++) {
 558         scsi_task_add_data_in_buffer(iTask.task,
 559                                      iov->iov[i].iov_len,
 560                                      iov->iov[i].iov_base);
 561     }
 562 #endif
 563
 564     while (!iTask.complete) {
 565         iscsi_set_events(iscsilun);
 566         qemu_coroutine_yield();
 567     }
 568
 569     if (iTask.task != NULL) {
 570         scsi_free_scsi_task(iTask.task);
 571         iTask.task = NULL;
 572     }
 573
 574     if (iTask.do_retry) {
 575         iTask.complete = 0;
 576         goto retry;
 577     }
 578
 579     if (iTask.status != SCSI_STATUS_GOOD) {
 580         return -EIO;
 581     }
 582
 583     return 0;
 584 }
 585
 586 static int coroutine_fn iscsi_co_flush(BlockDriverState *bs)
 587 {
 588     IscsiLun *iscsilun = bs->opaque;
 589     struct IscsiTask iTask;
 590
 591     if (bs->sg) {
 592         return 0;
 593     }
 594
 595     iscsi_co_init_iscsitask(iscsilun, &iTask);
 596
 597 retry:
 598     if (iscsi_synchronizecache10_task(iscsilun->iscsi, iscsilun->lun, 0, 0, 0,
 599                                       0, iscsi_co_generic_cb, &iTask) == NULL) {
 600         return -ENOMEM;
 601     }
 602
 603     while (!iTask.complete) {
 604         iscsi_set_events(iscsilun);
 605         qemu_coroutine_yield();
 606     }
 607
 608     if (iTask.task != NULL) {
 609         scsi_free_scsi_task(iTask.task);
 610         iTask.task = NULL;
 611     }
 612
 613     if (iTask.do_retry) {
 614         iTask.complete = 0;
 615         goto retry;
 616     }
 617
 618     if (iTask.status != SCSI_STATUS_GOOD) {
 619         return -EIO;
 620     }
 621
 622     return 0;
 623 }
 624
 625 #ifdef __linux__
 626 static void
 627 iscsi_aio_ioctl_cb(struct iscsi_context *iscsi, int status,
 628                      void *command_data, void *opaque)
 629 {
 630     IscsiAIOCB *acb = opaque;
 631
 632     g_free(acb->buf);
 633     acb->buf = NULL;
 634
 635     if (acb->canceled != 0) {
 636         return;
 637     }
 638
 639     acb->status = 0;
 640     if (status < 0) {
 641         error_report("Failed to ioctl(SG_IO) to iSCSI lun. %s",
 642                      iscsi_get_error(iscsi));
 643         acb->status = -EIO;
 644     }
 645
 646     acb->ioh->driver_status = 0;
 647     acb->ioh->host_status   = 0;
 648     acb->ioh->resid         = 0;
 649
 650 #define SG_ERR_DRIVER_SENSE    0x08
 651
 652     if (status == SCSI_STATUS_CHECK_CONDITION && acb->task->datain.size >= 2) {
 653         int ss;
 654
 655         acb->ioh->driver_status |= SG_ERR_DRIVER_SENSE;
 656
 657         acb->ioh->sb_len_wr = acb->task->datain.size - 2;
 658         ss = (acb->ioh->mx_sb_len >= acb->ioh->sb_len_wr) ?
 659              acb->ioh->mx_sb_len : acb->ioh->sb_len_wr;
 660         memcpy(acb->ioh->sbp, &acb->task->datain.data[2], ss);
 661     }
 662
 663     iscsi_schedule_bh(acb);
 664 }
 665
 666 static BlockDriverAIOCB *iscsi_aio_ioctl(BlockDriverState *bs,
 667         unsigned long int req, void *buf,
 668         BlockDriverCompletionFunc *cb, void *opaque)
 669 {
 670     IscsiLun *iscsilun = bs->opaque;
 671     struct iscsi_context *iscsi = iscsilun->iscsi;
 672     struct iscsi_data data;
 673     IscsiAIOCB *acb;
 674
 675     assert(req == SG_IO);
 676
 677     acb = qemu_aio_get(&iscsi_aiocb_info, bs, cb, opaque);
 678
 679     acb->iscsilun = iscsilun;
 680     acb->canceled    = 0;
 681     acb->bh          = NULL;
 682     acb->status      = -EINPROGRESS;
 683     acb->buf         = NULL;
 684     acb->ioh         = buf;
 685
 686     acb->task = malloc(sizeof(struct scsi_task));
 687     if (acb->task == NULL) {
 688         error_report("iSCSI: Failed to allocate task for scsi command. %s",
 689                      iscsi_get_error(iscsi));
 690         qemu_aio_release(acb);
 691         return NULL;
 692     }
 693     memset(acb->task, 0, sizeof(struct scsi_task));
 694
 695     switch (acb->ioh->dxfer_direction) {
 696     case SG_DXFER_TO_DEV:
 697         acb->task->xfer_dir = SCSI_XFER_WRITE;
 698         break;
 699     case SG_DXFER_FROM_DEV:
 700         acb->task->xfer_dir = SCSI_XFER_READ;
 701         break;
 702     default:
 703         acb->task->xfer_dir = SCSI_XFER_NONE;
 704         break;
 705     }
 706
 707     acb->task->cdb_size = acb->ioh->cmd_len;
 708     memcpy(&acb->task->cdb[0], acb->ioh->cmdp, acb->ioh->cmd_len);
 709     acb->task->expxferlen = acb->ioh->dxfer_len;
 710
 711     data.size = 0;
 712     if (acb->task->xfer_dir == SCSI_XFER_WRITE) {
 713         if (acb->ioh->iovec_count == 0) {
 714             data.data = acb->ioh->dxferp;
 715             data.size = acb->ioh->dxfer_len;
 716         } else {
 717 #if defined(LIBISCSI_FEATURE_IOVECTOR)
 718             scsi_task_set_iov_out(acb->task,
 719                                  (struct scsi_iovec *) acb->ioh->dxferp,
 720                                  acb->ioh->iovec_count);
 721 #else
 722             struct iovec *iov = (struct iovec *)acb->ioh->dxferp;
 723
 724             acb->buf = g_malloc(acb->ioh->dxfer_len);
 725             data.data = acb->buf;
 726             data.size = iov_to_buf(iov, acb->ioh->iovec_count, 0,
 727                                    acb->buf, acb->ioh->dxfer_len);
 728 #endif
 729         }
 730     }
 731
 732     if (iscsi_scsi_command_async(iscsi, iscsilun->lun, acb->task,
 733                                  iscsi_aio_ioctl_cb,
 734                                  (data.size > 0) ? &data : NULL,
 735                                  acb) != 0) {
 736         scsi_free_scsi_task(acb->task);
 737         qemu_aio_release(acb);
 738         return NULL;
 739     }
 740
 741     /* tell libiscsi to read straight into the buffer we got from ioctl */
 742     if (acb->task->xfer_dir == SCSI_XFER_READ) {
 743         if (acb->ioh->iovec_count == 0) {
 744             scsi_task_add_data_in_buffer(acb->task,
 745                                          acb->ioh->dxfer_len,
 746                                          acb->ioh->dxferp);
 747         } else {
 748 #if defined(LIBISCSI_FEATURE_IOVECTOR)
 749             scsi_task_set_iov_in(acb->task,
 750                                  (struct scsi_iovec *) acb->ioh->dxferp,
 751                                  acb->ioh->iovec_count);
 752 #else
 753             int i;
 754             for (i = 0; i < acb->ioh->iovec_count; i++) {
 755                 struct iovec *iov = (struct iovec *)acb->ioh->dxferp;
 756
 757                 scsi_task_add_data_in_buffer(acb->task,
 758                     iov[i].iov_len,
 759                     iov[i].iov_base);
 760             }
 761 #endif
 762         }
 763     }
 764
 765     iscsi_set_events(iscsilun);
 766
 767     return &acb->common;
 768 }
 769
 770
 771 static void ioctl_cb(void *opaque, int status)
 772 {
 773     int *p_status = opaque;
 774     *p_status = status;
 775 }
 776
 777 static int iscsi_ioctl(BlockDriverState *bs, unsigned long int req, void *buf)
 778 {
 779     IscsiLun *iscsilun = bs->opaque;
 780     int status;
 781
 782     switch (req) {
 783     case SG_GET_VERSION_NUM:
 784         *(int *)buf = 30000;
 785         break;
 786     case SG_GET_SCSI_ID:
 787         ((struct sg_scsi_id *)buf)->scsi_type = iscsilun->type;
 788         break;
 789     case SG_IO:
 790         status = -EINPROGRESS;
 791         iscsi_aio_ioctl(bs, req, buf, ioctl_cb, &status);
 792
 793         while (status == -EINPROGRESS) {
 794             qemu_aio_wait();
 795         }
 796
 797         return 0;
 798     default:
 799         return -1;
 800     }
 801     return 0;
 802 }
 803 #endif
 804
 805 static int64_t
 806 iscsi_getlength(BlockDriverState *bs)
 807 {
 808     IscsiLun *iscsilun = bs->opaque;
 809     int64_t len;
 810
 811     len  = iscsilun->num_blocks;
 812     len *= iscsilun->block_size;
 813
 814     return len;
 815 }
 816
 817 static int
 818 coroutine_fn iscsi_co_discard(BlockDriverState *bs, int64_t sector_num,
 819                                    int nb_sectors)
 820 {
 821     IscsiLun *iscsilun = bs->opaque;
 822     struct IscsiTask iTask;
 823     struct unmap_list list;
 824
 825     if (!is_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
 826         return -EINVAL;
 827     }
 828
 829     if (!iscsilun->lbp.lbpu) {
 830         /* UNMAP is not supported by the target */
 831         return 0;
 832     }
 833
 834     list.lba = sector_qemu2lun(sector_num, iscsilun);
 835     list.num = sector_qemu2lun(nb_sectors, iscsilun);
 836
 837     iscsi_co_init_iscsitask(iscsilun, &iTask);
 838 retry:
 839     if (iscsi_unmap_task(iscsilun->iscsi, iscsilun->lun, 0, 0, &list, 1,
 840                      iscsi_co_generic_cb, &iTask) == NULL) {
 841         return -ENOMEM;
 842     }
 843
 844     while (!iTask.complete) {
 845         iscsi_set_events(iscsilun);
 846         qemu_coroutine_yield();
 847     }
 848
 849     if (iTask.task != NULL) {
 850         scsi_free_scsi_task(iTask.task);
 851         iTask.task = NULL;
 852     }
 853
 854     if (iTask.do_retry) {
 855         iTask.complete = 0;
 856         goto retry;
 857     }
 858
 859     if (iTask.status == SCSI_STATUS_CHECK_CONDITION) {
 860         /* the target might fail with a check condition if it
 861            is not happy with the alignment of the UNMAP request
 862            we silently fail in this case */
 863         return 0;
 864     }
 865
 866     if (iTask.status != SCSI_STATUS_GOOD) {
 867         return -EIO;
 868     }
 869
 870     iscsi_allocationmap_clear(iscsilun, sector_num, nb_sectors);
 871
 872     return 0;
 873 }
 874
 875 #if defined(SCSI_SENSE_ASCQ_CAPACITY_DATA_HAS_CHANGED)
 876
 877 static int
 878 coroutine_fn iscsi_co_write_zeroes(BlockDriverState *bs, int64_t sector_num,
 879                                    int nb_sectors, BdrvRequestFlags flags)
 880 {
 881     IscsiLun *iscsilun = bs->opaque;
 882     struct IscsiTask iTask;
 883     uint64_t lba;
 884     uint32_t nb_blocks;
 885
 886     if (!is_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
 887         return -EINVAL;
 888     }
 889
 890     if ((flags & BDRV_REQ_MAY_UNMAP) && !iscsilun->lbp.lbpws) {
 891         /* WRITE SAME with UNMAP is not supported by the target,
 892          * fall back and try WRITE SAME without UNMAP */
 893         flags &= ~BDRV_REQ_MAY_UNMAP;
 894     }
 895
 896     if (!(flags & BDRV_REQ_MAY_UNMAP) && !iscsilun->has_write_same) {
 897         /* WRITE SAME without UNMAP is not supported by the target */
 898         return -ENOTSUP;
 899     }
 900
 901     lba = sector_qemu2lun(sector_num, iscsilun);
 902     nb_blocks = sector_qemu2lun(nb_sectors, iscsilun);
 903
 904     if (iscsilun->zeroblock == NULL) {
 905         iscsilun->zeroblock = g_malloc0(iscsilun->block_size);
 906     }
 907
 908     iscsi_co_init_iscsitask(iscsilun, &iTask);
 909 retry:
 910     if (iscsi_writesame16_task(iscsilun->iscsi, iscsilun->lun, lba,
 911                                iscsilun->zeroblock, iscsilun->block_size,
 912                                nb_blocks, 0, !!(flags & BDRV_REQ_MAY_UNMAP),
 913                                0, 0, iscsi_co_generic_cb, &iTask) == NULL) {
 914         return -ENOMEM;
 915     }
 916
 917     while (!iTask.complete) {
 918         iscsi_set_events(iscsilun);
 919         qemu_coroutine_yield();
 920     }
 921
 922     if (iTask.status == SCSI_STATUS_CHECK_CONDITION &&
 923         iTask.task->sense.key == SCSI_SENSE_ILLEGAL_REQUEST &&
 924         (iTask.task->sense.ascq == SCSI_SENSE_ASCQ_INVALID_OPERATION_CODE ||
 925          iTask.task->sense.ascq == SCSI_SENSE_ASCQ_INVALID_FIELD_IN_CDB)) {
 926         /* WRITE SAME is not supported by the target */
 927         iscsilun->has_write_same = false;
 928         scsi_free_scsi_task(iTask.task);
 929         return -ENOTSUP;
 930     }
 931
 932     if (iTask.task != NULL) {
 933         scsi_free_scsi_task(iTask.task);
 934         iTask.task = NULL;
 935     }
 936
 937     if (iTask.do_retry) {
 938         iTask.complete = 0;
 939         goto retry;
 940     }
 941
 942     if (iTask.status != SCSI_STATUS_GOOD) {
 943         return -EIO;
 944     }
 945
 946     if (flags & BDRV_REQ_MAY_UNMAP) {
 947         iscsi_allocationmap_clear(iscsilun, sector_num, nb_sectors);
 948     } else {
 949         iscsi_allocationmap_set(iscsilun, sector_num, nb_sectors);
 950     }
 951
 952     return 0;
 953 }
 954
 955 #endif /* SCSI_SENSE_ASCQ_CAPACITY_DATA_HAS_CHANGED */
 956
 957 static void parse_chap(struct iscsi_context *iscsi, const char *target,
 958                        Error **errp)
 959 {
 960     QemuOptsList *list;
 961     QemuOpts *opts;
 962     const char *user = NULL;
 963     const char *password = NULL;
 964
 965     list = qemu_find_opts("iscsi");
 966     if (!list) {
 967         return;
 968     }
 969
 970     opts = qemu_opts_find(list, target);
 971     if (opts == NULL) {
 972         opts = QTAILQ_FIRST(&list->head);
 973         if (!opts) {
 974             return;
 975         }
 976     }
 977
 978     user = qemu_opt_get(opts, "user");
 979     if (!user) {
 980         return;
 981     }
 982
 983     password = qemu_opt_get(opts, "password");
 984     if (!password) {
 985         error_setg(errp, "CHAP username specified but no password was given");
 986         return;
 987     }
 988
 989     if (iscsi_set_initiator_username_pwd(iscsi, user, password)) {
 990         error_setg(errp, "Failed to set initiator username and password");
 991     }
 992 }
 993
 994 static void parse_header_digest(struct iscsi_context *iscsi, const char *target,
 995                                 Error **errp)
 996 {
 997     QemuOptsList *list;
 998     QemuOpts *opts;
 999     const char *digest = NULL;
1000
1001     list = qemu_find_opts("iscsi");
1002     if (!list) {
1003         return;
1004     }
1005
1006     opts = qemu_opts_find(list, target);
1007     if (opts == NULL) {
1008         opts = QTAILQ_FIRST(&list->head);
1009         if (!opts) {
1010             return;
1011         }
1012     }
1013
1014     digest = qemu_opt_get(opts, "header-digest");
1015     if (!digest) {
1016         return;
1017     }
1018
1019     if (!strcmp(digest, "CRC32C")) {
1020         iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_CRC32C);
1021     } else if (!strcmp(digest, "NONE")) {
1022         iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_NONE);
1023     } else if (!strcmp(digest, "CRC32C-NONE")) {
1024         iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_CRC32C_NONE);
1025     } else if (!strcmp(digest, "NONE-CRC32C")) {
1026         iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_NONE_CRC32C);
1027     } else {
1028         error_setg(errp, "Invalid header-digest setting : %s", digest);
1029     }
1030 }
1031
1032 static char *parse_initiator_name(const char *target)
1033 {
1034     QemuOptsList *list;
1035     QemuOpts *opts;
1036     const char *name;
1037     char *iscsi_name;
1038     UuidInfo *uuid_info;
1039
1040     list = qemu_find_opts("iscsi");
1041     if (list) {
1042         opts = qemu_opts_find(list, target);
1043         if (!opts) {
1044             opts = QTAILQ_FIRST(&list->head);
1045         }
1046         if (opts) {
1047             name = qemu_opt_get(opts, "initiator-name");
1048             if (name) {
1049                 return g_strdup(name);
1050             }
1051         }
1052     }
1053
1054     uuid_info = qmp_query_uuid(NULL);
1055     if (strcmp(uuid_info->UUID, UUID_NONE) == 0) {
1056         name = qemu_get_vm_name();
1057     } else {
1058         name = uuid_info->UUID;
1059     }
1060     iscsi_name = g_strdup_printf("iqn.2008-11.org.linux-kvm%s%s",
1061                                  name ? ":" : "", name ? name : "");
1062     qapi_free_UuidInfo(uuid_info);
1063     return iscsi_name;
1064 }
1065
1066 #if defined(LIBISCSI_FEATURE_NOP_COUNTER)
1067 static void iscsi_nop_timed_event(void *opaque)
1068 {
1069     IscsiLun *iscsilun = opaque;
1070
1071     if (iscsi_get_nops_in_flight(iscsilun->iscsi) > MAX_NOP_FAILURES) {
1072         error_report("iSCSI: NOP timeout. Reconnecting...");
1073         iscsi_reconnect(iscsilun->iscsi);
1074     }
1075
1076     if (iscsi_nop_out_async(iscsilun->iscsi, NULL, NULL, 0, NULL) != 0) {
1077         error_report("iSCSI: failed to sent NOP-Out. Disabling NOP messages.");
1078         return;
1079     }
1080
1081     timer_mod(iscsilun->nop_timer, qemu_clock_get_ms(QEMU_CLOCK_REALTIME) + NOP_INTERVAL);
1082     iscsi_set_events(iscsilun);
1083 }
1084 #endif
1085
1086 static void iscsi_readcapacity_sync(IscsiLun *iscsilun, Error **errp)
1087 {
1088     struct scsi_task *task = NULL;
1089     struct scsi_readcapacity10 *rc10 = NULL;
1090     struct scsi_readcapacity16 *rc16 = NULL;
1091     int retries = ISCSI_CMD_RETRIES;
1092
1093     do {
1094         if (task != NULL) {
1095             scsi_free_scsi_task(task);
1096             task = NULL;
1097         }
1098
1099         switch (iscsilun->type) {
1100         case TYPE_DISK:
1101             task = iscsi_readcapacity16_sync(iscsilun->iscsi, iscsilun->lun);
1102             if (task != NULL && task->status == SCSI_STATUS_GOOD) {
1103                 rc16 = scsi_datain_unmarshall(task);
1104                 if (rc16 == NULL) {
1105                     error_setg(errp, "iSCSI: Failed to unmarshall readcapacity16 data.");
1106                 } else {
1107                     iscsilun->block_size = rc16->block_length;
1108                     iscsilun->num_blocks = rc16->returned_lba + 1;
1109                     iscsilun->lbpme = rc16->lbpme;
1110                     iscsilun->lbprz = rc16->lbprz;
1111                 }
1112             }
1113             break;
1114         case TYPE_ROM:
1115             task = iscsi_readcapacity10_sync(iscsilun->iscsi, iscsilun->lun, 0, 0);
1116             if (task != NULL && task->status == SCSI_STATUS_GOOD) {
1117                 rc10 = scsi_datain_unmarshall(task);
1118                 if (rc10 == NULL) {
1119                     error_setg(errp, "iSCSI: Failed to unmarshall readcapacity10 data.");
1120                 } else {
1121                     iscsilun->block_size = rc10->block_size;
1122                     if (rc10->lba == 0) {
1123                         /* blank disk loaded */
1124                         iscsilun->num_blocks = 0;
1125                     } else {
1126                         iscsilun->num_blocks = rc10->lba + 1;
1127                     }
1128                 }
1129             }
1130             break;
1131         default:
1132             return;
1133         }
1134     } while (task != NULL && task->status == SCSI_STATUS_CHECK_CONDITION
1135              && task->sense.key == SCSI_SENSE_UNIT_ATTENTION
1136              && retries-- > 0);
1137
1138     if (task == NULL || task->status != SCSI_STATUS_GOOD) {
1139         error_setg(errp, "iSCSI: failed to send readcapacity10 command.");
1140     }
1141     if (task) {
1142         scsi_free_scsi_task(task);
1143     }
1144 }
1145
1146 /* TODO Convert to fine grained options */
1147 static QemuOptsList runtime_opts = {
1148     .name = "iscsi",
1149     .head = QTAILQ_HEAD_INITIALIZER(runtime_opts.head),
1150     .desc = {
1151         {
1152             .name = "filename",
1153             .type = QEMU_OPT_STRING,
1154             .help = "URL to the iscsi image",
1155         },
1156         { /* end of list */ }
1157     },
1158 };
1159
1160 static struct scsi_task *iscsi_do_inquiry(struct iscsi_context *iscsi, int lun,
1161                                           int evpd, int pc, void **inq, Error **errp)
1162 {
1163     int full_size;
1164     struct scsi_task *task = NULL;
1165     task = iscsi_inquiry_sync(iscsi, lun, evpd, pc, 64);
1166     if (task == NULL || task->status != SCSI_STATUS_GOOD) {
1167         goto fail;
1168     }
1169     full_size = scsi_datain_getfullsize(task);
1170     if (full_size > task->datain.size) {
1171         scsi_free_scsi_task(task);
1172
1173         /* we need more data for the full list */
1174         task = iscsi_inquiry_sync(iscsi, lun, evpd, pc, full_size);
1175         if (task == NULL || task->status != SCSI_STATUS_GOOD) {
1176             goto fail;
1177         }
1178     }
1179
1180     *inq = scsi_datain_unmarshall(task);
1181     if (*inq == NULL) {
1182         error_setg(errp, "iSCSI: failed to unmarshall inquiry datain blob");
1183         goto fail_with_err;
1184     }
1185
1186     return task;
1187
1188 fail:
1189     error_setg(errp, "iSCSI: Inquiry command failed : %s",
1190                iscsi_get_error(iscsi));
1191 fail_with_err:
1192     if (task != NULL) {
1193         scsi_free_scsi_task(task);
1194     }
1195     return NULL;
1196 }
1197
1198 /*
1199  * We support iscsi url's on the form
1200  * iscsi://[<username>%<password>@]<host>[:<port>]/<targetname>/<lun>
1201  *
1202  * Note: flags are currently not used by iscsi_open.  If this function
1203  * is changed such that flags are used, please examine iscsi_reopen_prepare()
1204  * to see if needs to be changed as well.
1205  */
1206 static int iscsi_open(BlockDriverState *bs, QDict *options, int flags,
1207                       Error **errp)
1208 {
1209     IscsiLun *iscsilun = bs->opaque;
1210     struct iscsi_context *iscsi = NULL;
1211     struct iscsi_url *iscsi_url = NULL;
1212     struct scsi_task *task = NULL;
1213     struct scsi_inquiry_standard *inq = NULL;
1214     struct scsi_inquiry_supported_pages *inq_vpd;
1215     char *initiator_name = NULL;
1216     QemuOpts *opts;
1217     Error *local_err = NULL;
1218     const char *filename;
1219     int i, ret;
1220
1221     if ((BDRV_SECTOR_SIZE % 512) != 0) {
1222         error_setg(errp, "iSCSI: Invalid BDRV_SECTOR_SIZE. "
1223                    "BDRV_SECTOR_SIZE(%lld) is not a multiple "
1224                    "of 512", BDRV_SECTOR_SIZE);
1225         return -EINVAL;
1226     }
1227
1228     opts = qemu_opts_create(&runtime_opts, NULL, 0, &error_abort);
1229     qemu_opts_absorb_qdict(opts, options, &local_err);
1230     if (local_err) {
1231         error_propagate(errp, local_err);
1232         ret = -EINVAL;
1233         goto out;
1234     }
1235
1236     filename = qemu_opt_get(opts, "filename");
1237
1238     iscsi_url = iscsi_parse_full_url(iscsi, filename);
1239     if (iscsi_url == NULL) {
1240         error_setg(errp, "Failed to parse URL : %s", filename);
1241         ret = -EINVAL;
1242         goto out;
1243     }
1244
1245     memset(iscsilun, 0, sizeof(IscsiLun));
1246
1247     initiator_name = parse_initiator_name(iscsi_url->target);
1248
1249     iscsi = iscsi_create_context(initiator_name);
1250     if (iscsi == NULL) {
1251         error_setg(errp, "iSCSI: Failed to create iSCSI context.");
1252         ret = -ENOMEM;
1253         goto out;
1254     }
1255
1256     if (iscsi_set_targetname(iscsi, iscsi_url->target)) {
1257         error_setg(errp, "iSCSI: Failed to set target name.");
1258         ret = -EINVAL;
1259         goto out;
1260     }
1261
1262     if (iscsi_url->user != NULL) {
1263         ret = iscsi_set_initiator_username_pwd(iscsi, iscsi_url->user,
1264                                               iscsi_url->passwd);
1265         if (ret != 0) {
1266             error_setg(errp, "Failed to set initiator username and password");
1267             ret = -EINVAL;
1268             goto out;
1269         }
1270     }
1271
1272     /* check if we got CHAP username/password via the options */
1273     parse_chap(iscsi, iscsi_url->target, &local_err);
1274     if (local_err != NULL) {
1275         error_propagate(errp, local_err);
1276         ret = -EINVAL;
1277         goto out;
1278     }
1279
1280     if (iscsi_set_session_type(iscsi, ISCSI_SESSION_NORMAL) != 0) {
1281         error_setg(errp, "iSCSI: Failed to set session type to normal.");
1282         ret = -EINVAL;
1283         goto out;
1284     }
1285
1286     iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_NONE_CRC32C);
1287
1288     /* check if we got HEADER_DIGEST via the options */
1289     parse_header_digest(iscsi, iscsi_url->target, &local_err);
1290     if (local_err != NULL) {
1291         error_propagate(errp, local_err);
1292         ret = -EINVAL;
1293         goto out;
1294     }
1295
1296     if (iscsi_full_connect_sync(iscsi, iscsi_url->portal, iscsi_url->lun) != 0) {
1297         error_setg(errp, "iSCSI: Failed to connect to LUN : %s",
1298             iscsi_get_error(iscsi));
1299         ret = -EINVAL;
1300         goto out;
1301     }
1302
1303     iscsilun->iscsi = iscsi;
1304     iscsilun->lun   = iscsi_url->lun;
1305     iscsilun->has_write_same = true;
1306
1307     task = iscsi_do_inquiry(iscsilun->iscsi, iscsilun->lun, 0, 0,
1308                             (void **) &inq, errp);
1309     if (task == NULL) {
1310         ret = -EINVAL;
1311         goto out;
1312     }
1313     iscsilun->type = inq->periperal_device_type;
1314     scsi_free_scsi_task(task);
1315     task = NULL;
1316
1317     iscsi_readcapacity_sync(iscsilun, &local_err);
1318     if (local_err != NULL) {
1319         error_propagate(errp, local_err);
1320         ret = -EINVAL;
1321         goto out;
1322     }
1323     bs->total_sectors = sector_lun2qemu(iscsilun->num_blocks, iscsilun);
1324     bs->request_alignment = iscsilun->block_size;
1325
1326     /* We don't have any emulation for devices other than disks and CD-ROMs, so
1327      * this must be sg ioctl compatible. We force it to be sg, otherwise qemu
1328      * will try to read from the device to guess the image format.
1329      */
1330     if (iscsilun->type != TYPE_DISK && iscsilun->type != TYPE_ROM) {
1331         bs->sg = 1;
1332     }
1333
1334     task = iscsi_do_inquiry(iscsilun->iscsi, iscsilun->lun, 1,
1335                             SCSI_INQUIRY_PAGECODE_SUPPORTED_VPD_PAGES,
1336                             (void **) &inq_vpd, errp);
1337     if (task == NULL) {
1338         ret = -EINVAL;
1339         goto out;
1340     }
1341     for (i = 0; i < inq_vpd->num_pages; i++) {
1342         struct scsi_task *inq_task;
1343         struct scsi_inquiry_logical_block_provisioning *inq_lbp;
1344         struct scsi_inquiry_block_limits *inq_bl;
1345         switch (inq_vpd->pages[i]) {
1346         case SCSI_INQUIRY_PAGECODE_LOGICAL_BLOCK_PROVISIONING:
1347             inq_task = iscsi_do_inquiry(iscsilun->iscsi, iscsilun->lun, 1,
1348                                         SCSI_INQUIRY_PAGECODE_LOGICAL_BLOCK_PROVISIONING,
1349                                         (void **) &inq_lbp, errp);
1350             if (inq_task == NULL) {
1351                 ret = -EINVAL;
1352                 goto out;
1353             }
1354             memcpy(&iscsilun->lbp, inq_lbp,
1355                    sizeof(struct scsi_inquiry_logical_block_provisioning));
1356             scsi_free_scsi_task(inq_task);
1357             break;
1358         case SCSI_INQUIRY_PAGECODE_BLOCK_LIMITS:
1359             inq_task = iscsi_do_inquiry(iscsilun->iscsi, iscsilun->lun, 1,
1360                                     SCSI_INQUIRY_PAGECODE_BLOCK_LIMITS,
1361                                     (void **) &inq_bl, errp);
1362             if (inq_task == NULL) {
1363                 ret = -EINVAL;
1364                 goto out;
1365             }
1366             memcpy(&iscsilun->bl, inq_bl,
1367                    sizeof(struct scsi_inquiry_block_limits));
1368             scsi_free_scsi_task(inq_task);
1369             break;
1370         default:
1371             break;
1372         }
1373     }
1374     scsi_free_scsi_task(task);
1375     task = NULL;
1376
1377 #if defined(LIBISCSI_FEATURE_NOP_COUNTER)
1378     /* Set up a timer for sending out iSCSI NOPs */
1379     iscsilun->nop_timer = timer_new_ms(QEMU_CLOCK_REALTIME, iscsi_nop_timed_event, iscsilun);
1380     timer_mod(iscsilun->nop_timer, qemu_clock_get_ms(QEMU_CLOCK_REALTIME) + NOP_INTERVAL);
1381 #endif
1382
1383     /* Guess the internal cluster (page) size of the iscsi target by the means
1384      * of opt_unmap_gran. Transfer the unmap granularity only if it has a
1385      * reasonable size */
1386     if (iscsilun->bl.opt_unmap_gran * iscsilun->block_size >= 4 * 1024 &&
1387         iscsilun->bl.opt_unmap_gran * iscsilun->block_size <= 16 * 1024 * 1024) {
1388         iscsilun->cluster_sectors = (iscsilun->bl.opt_unmap_gran *
1389                                      iscsilun->block_size) >> BDRV_SECTOR_BITS;
1390 #if defined(LIBISCSI_FEATURE_IOVECTOR)
1391         if (iscsilun->lbprz && !(bs->open_flags & BDRV_O_NOCACHE)) {
1392             iscsilun->allocationmap =
1393                 bitmap_new(DIV_ROUND_UP(bs->total_sectors,
1394                                         iscsilun->cluster_sectors));
1395         }
1396 #endif
1397     }
1398
1399 out:
1400     qemu_opts_del(opts);
1401     if (initiator_name != NULL) {
1402         g_free(initiator_name);
1403     }
1404     if (iscsi_url != NULL) {
1405         iscsi_destroy_url(iscsi_url);
1406     }
1407     if (task != NULL) {
1408         scsi_free_scsi_task(task);
1409     }
1410
1411     if (ret) {
1412         if (iscsi != NULL) {
1413             iscsi_destroy_context(iscsi);
1414         }
1415         memset(iscsilun, 0, sizeof(IscsiLun));
1416     }
1417     return ret;
1418 }
1419
1420 static void iscsi_close(BlockDriverState *bs)
1421 {
1422     IscsiLun *iscsilun = bs->opaque;
1423     struct iscsi_context *iscsi = iscsilun->iscsi;
1424
1425     if (iscsilun->nop_timer) {
1426         timer_del(iscsilun->nop_timer);
1427         timer_free(iscsilun->nop_timer);
1428     }
1429     qemu_aio_set_fd_handler(iscsi_get_fd(iscsi), NULL, NULL, NULL);
1430     iscsi_destroy_context(iscsi);
1431     g_free(iscsilun->zeroblock);
1432     g_free(iscsilun->allocationmap);
1433     memset(iscsilun, 0, sizeof(IscsiLun));
1434 }
1435
1436 static int iscsi_refresh_limits(BlockDriverState *bs)
1437 {
1438     IscsiLun *iscsilun = bs->opaque;
1439
1440     /* We don't actually refresh here, but just return data queried in
1441      * iscsi_open(): iscsi targets don't change their limits. */
1442     if (iscsilun->lbp.lbpu) {
1443         if (iscsilun->bl.max_unmap < 0xffffffff) {
1444             bs->bl.max_discard = sector_lun2qemu(iscsilun->bl.max_unmap,
1445                                                  iscsilun);
1446         }
1447         bs->bl.discard_alignment = sector_lun2qemu(iscsilun->bl.opt_unmap_gran,
1448                                                    iscsilun);
1449     }
1450
1451     if (iscsilun->bl.max_ws_len < 0xffffffff) {
1452         bs->bl.max_write_zeroes = sector_lun2qemu(iscsilun->bl.max_ws_len,
1453                                                   iscsilun);
1454     }
1455     if (iscsilun->lbp.lbpws) {
1456         bs->bl.write_zeroes_alignment = sector_lun2qemu(iscsilun->bl.opt_unmap_gran,
1457                                                         iscsilun);
1458     }
1459     bs->bl.opt_transfer_length = sector_lun2qemu(iscsilun->bl.opt_xfer_len,
1460                                                  iscsilun);
1461     return 0;
1462 }
1463
1464 /* Since iscsi_open() ignores bdrv_flags, there is nothing to do here in
1465  * prepare.  Note that this will not re-establish a connection with an iSCSI
1466  * target - it is effectively a NOP.  */
1467 static int iscsi_reopen_prepare(BDRVReopenState *state,
1468                                 BlockReopenQueue *queue, Error **errp)
1469 {
1470     /* NOP */
1471     return 0;
1472 }
1473
1474 static int iscsi_truncate(BlockDriverState *bs, int64_t offset)
1475 {
1476     IscsiLun *iscsilun = bs->opaque;
1477     Error *local_err = NULL;
1478
1479     if (iscsilun->type != TYPE_DISK) {
1480         return -ENOTSUP;
1481     }
1482
1483     iscsi_readcapacity_sync(iscsilun, &local_err);
1484     if (local_err != NULL) {
1485         error_free(local_err);
1486         return -EIO;
1487     }
1488
1489     if (offset > iscsi_getlength(bs)) {
1490         return -EINVAL;
1491     }
1492
1493     if (iscsilun->allocationmap != NULL) {
1494         g_free(iscsilun->allocationmap);
1495         iscsilun->allocationmap =
1496             bitmap_new(DIV_ROUND_UP(bs->total_sectors,
1497                                     iscsilun->cluster_sectors));
1498     }
1499
1500     return 0;
1501 }
1502
1503 static int iscsi_create(const char *filename, QEMUOptionParameter *options,
1504                         Error **errp)
1505 {
1506     int ret = 0;
1507     int64_t total_size = 0;
1508     BlockDriverState *bs;
1509     IscsiLun *iscsilun = NULL;
1510     QDict *bs_options;
1511
1512     bs = bdrv_new("", &error_abort);
1513
1514     /* Read out options */
1515     while (options && options->name) {
1516         if (!strcmp(options->name, "size")) {
1517             total_size = options->value.n / BDRV_SECTOR_SIZE;
1518         }
1519         options++;
1520     }
1521
1522     bs->opaque = g_malloc0(sizeof(struct IscsiLun));
1523     iscsilun = bs->opaque;
1524
1525     bs_options = qdict_new();
1526     qdict_put(bs_options, "filename", qstring_from_str(filename));
1527     ret = iscsi_open(bs, bs_options, 0, NULL);
1528     QDECREF(bs_options);
1529
1530     if (ret != 0) {
1531         goto out;
1532     }
1533     if (iscsilun->nop_timer) {
1534         timer_del(iscsilun->nop_timer);
1535         timer_free(iscsilun->nop_timer);
1536     }
1537     if (iscsilun->type != TYPE_DISK) {
1538         ret = -ENODEV;
1539         goto out;
1540     }
1541     if (bs->total_sectors < total_size) {
1542         ret = -ENOSPC;
1543         goto out;
1544     }
1545
1546     ret = 0;
1547 out:
1548     if (iscsilun->iscsi != NULL) {
1549         iscsi_destroy_context(iscsilun->iscsi);
1550     }
1551     g_free(bs->opaque);
1552     bs->opaque = NULL;
1553     bdrv_unref(bs);
1554     return ret;
1555 }
1556
1557 static int iscsi_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
1558 {
1559     IscsiLun *iscsilun = bs->opaque;
1560     bdi->unallocated_blocks_are_zero = !!iscsilun->lbprz;
1561     bdi->can_write_zeroes_with_unmap = iscsilun->lbprz && iscsilun->lbp.lbpws;
1562     bdi->cluster_size = iscsilun->cluster_sectors * BDRV_SECTOR_SIZE;
1563     return 0;
1564 }
1565
1566 static QEMUOptionParameter iscsi_create_options[] = {
1567     {
1568         .name = BLOCK_OPT_SIZE,
1569         .type = OPT_SIZE,
1570         .help = "Virtual disk size"
1571     },
1572     { NULL }
1573 };
1574
1575 static BlockDriver bdrv_iscsi = {
1576     .format_name     = "iscsi",
1577     .protocol_name   = "iscsi",
1578
1579     .instance_size   = sizeof(IscsiLun),
1580     .bdrv_needs_filename = true,
1581     .bdrv_file_open  = iscsi_open,
1582     .bdrv_close      = iscsi_close,
1583     .bdrv_create     = iscsi_create,
1584     .create_options  = iscsi_create_options,
1585     .bdrv_reopen_prepare  = iscsi_reopen_prepare,
1586
1587     .bdrv_getlength  = iscsi_getlength,
1588     .bdrv_get_info   = iscsi_get_info,
1589     .bdrv_truncate   = iscsi_truncate,
1590     .bdrv_refresh_limits = iscsi_refresh_limits,
1591
1592 #if defined(LIBISCSI_FEATURE_IOVECTOR)
1593     .bdrv_co_get_block_status = iscsi_co_get_block_status,
1594 #endif
1595     .bdrv_co_discard      = iscsi_co_discard,
1596 #if defined(SCSI_SENSE_ASCQ_CAPACITY_DATA_HAS_CHANGED)
1597     .bdrv_co_write_zeroes = iscsi_co_write_zeroes,
1598 #endif
1599     .bdrv_co_readv         = iscsi_co_readv,
1600     .bdrv_co_writev        = iscsi_co_writev,
1601     .bdrv_co_flush_to_disk = iscsi_co_flush,
1602
1603 #ifdef __linux__
1604     .bdrv_ioctl       = iscsi_ioctl,
1605     .bdrv_aio_ioctl   = iscsi_aio_ioctl,
1606 #endif
1607 };
1608
1609 static QemuOptsList qemu_iscsi_opts = {
1610     .name = "iscsi",
1611     .head = QTAILQ_HEAD_INITIALIZER(qemu_iscsi_opts.head),
1612     .desc = {
1613         {
1614             .name = "user",
1615             .type = QEMU_OPT_STRING,
1616             .help = "username for CHAP authentication to target",
1617         },{
1618             .name = "password",
1619             .type = QEMU_OPT_STRING,
1620             .help = "password for CHAP authentication to target",
1621         },{
1622             .name = "header-digest",
1623             .type = QEMU_OPT_STRING,
1624             .help = "HeaderDigest setting. "
1625                     "{CRC32C|CRC32C-NONE|NONE-CRC32C|NONE}",
1626         },{
1627             .name = "initiator-name",
1628             .type = QEMU_OPT_STRING,
1629             .help = "Initiator iqn name to use when connecting",
1630         },
1631         { /* end of list */ }
1632     },
1633 };
1634
1635 static void iscsi_block_init(void)
1636 {
1637     bdrv_register(&bdrv_iscsi);
1638     qemu_add_opts(&qemu_iscsi_opts);
1639 }
1640
1641 block_init(iscsi_block_init);