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 <math.h>
  30 #include <arpa/inet.h>
  31 #include "qemu-common.h"
  32 #include "qemu/config-file.h"
  33 #include "qemu/error-report.h"
  34 #include "qemu/bitops.h"
  35 #include "qemu/bitmap.h"
  36 #include "block/block_int.h"
  37 #include "trace.h"
  38 #include "block/scsi.h"
  39 #include "qemu/iov.h"
  40 #include "sysemu/sysemu.h"
  41 #include "qmp-commands.h"
  42
  43 #include <iscsi/iscsi.h>
  44 #include <iscsi/scsi-lowlevel.h>
  45
  46 #ifdef __linux__
  47 #include <scsi/sg.h>
  48 #include <block/scsi.h>
  49 #endif
  50
  51 typedef struct IscsiLun {
  52     struct iscsi_context *iscsi;
  53     AioContext *aio_context;
  54     int lun;
  55     enum scsi_inquiry_peripheral_device_type type;
  56     int block_size;
  57     uint64_t num_blocks;
  58     int events;
  59     QEMUTimer *nop_timer;
  60     uint8_t lbpme;
  61     uint8_t lbprz;
  62     uint8_t has_write_same;
  63     struct scsi_inquiry_logical_block_provisioning lbp;
  64     struct scsi_inquiry_block_limits bl;
  65     unsigned char *zeroblock;
  66     unsigned long *allocationmap;
  67     int cluster_sectors;
  68     bool use_16_for_rw;
  69 } IscsiLun;
  70
  71 typedef struct IscsiTask {
  72     int status;
  73     int complete;
  74     int retries;
  75     int do_retry;
  76     struct scsi_task *task;
  77     Coroutine *co;
  78     QEMUBH *bh;
  79     IscsiLun *iscsilun;
  80     QEMUTimer retry_timer;
  81 } IscsiTask;
  82
  83 typedef struct IscsiAIOCB {
  84     BlockDriverAIOCB common;
  85     QEMUIOVector *qiov;
  86     QEMUBH *bh;
  87     IscsiLun *iscsilun;
  88     struct scsi_task *task;
  89     uint8_t *buf;
  90     int status;
  91     int canceled;
  92     int64_t sector_num;
  93     int nb_sectors;
  94 #ifdef __linux__
  95     sg_io_hdr_t *ioh;
  96 #endif
  97 } IscsiAIOCB;
  98
  99 #define NOP_INTERVAL 5000
 100 #define MAX_NOP_FAILURES 3
 101 #define ISCSI_CMD_RETRIES ARRAY_SIZE(iscsi_retry_times)
 102 static const unsigned iscsi_retry_times[] = {8, 32, 128, 512, 2048};
 103
 104 /* this threshhold is a trade-off knob to choose between
 105  * the potential additional overhead of an extra GET_LBA_STATUS request
 106  * vs. unnecessarily reading a lot of zero sectors over the wire.
 107  * If a read request is greater or equal than ISCSI_CHECKALLOC_THRES
 108  * sectors we check the allocation status of the area covered by the
 109  * request first if the allocationmap indicates that the area might be
 110  * unallocated. */
 111 #define ISCSI_CHECKALLOC_THRES 64
 112
 113 static void
 114 iscsi_bh_cb(void *p)
 115 {
 116     IscsiAIOCB *acb = p;
 117
 118     qemu_bh_delete(acb->bh);
 119
 120     g_free(acb->buf);
 121     acb->buf = NULL;
 122
 123     if (acb->canceled == 0) {
 124         acb->common.cb(acb->common.opaque, acb->status);
 125     }
 126
 127     if (acb->task != NULL) {
 128         scsi_free_scsi_task(acb->task);
 129         acb->task = NULL;
 130     }
 131
 132     qemu_aio_release(acb);
 133 }
 134
 135 static void
 136 iscsi_schedule_bh(IscsiAIOCB *acb)
 137 {
 138     if (acb->bh) {
 139         return;
 140     }
 141     acb->bh = aio_bh_new(acb->iscsilun->aio_context, iscsi_bh_cb, acb);
 142     qemu_bh_schedule(acb->bh);
 143 }
 144
 145 static void iscsi_co_generic_bh_cb(void *opaque)
 146 {
 147     struct IscsiTask *iTask = opaque;
 148     iTask->complete = 1;
 149     qemu_bh_delete(iTask->bh);
 150     qemu_coroutine_enter(iTask->co, NULL);
 151 }
 152
 153 static void iscsi_retry_timer_expired(void *opaque)
 154 {
 155     struct IscsiTask *iTask = opaque;
 156     iTask->complete = 1;
 157     if (iTask->co) {
 158         qemu_coroutine_enter(iTask->co, NULL);
 159     }
 160 }
 161
 162 static inline unsigned exp_random(double mean)
 163 {
 164     return -mean * log((double)rand() / RAND_MAX);
 165 }
 166
 167 static void
 168 iscsi_co_generic_cb(struct iscsi_context *iscsi, int status,
 169                         void *command_data, void *opaque)
 170 {
 171     struct IscsiTask *iTask = opaque;
 172     struct scsi_task *task = command_data;
 173
 174     iTask->status = status;
 175     iTask->do_retry = 0;
 176     iTask->task = task;
 177
 178     if (status != SCSI_STATUS_GOOD) {
 179         if (iTask->retries++ < ISCSI_CMD_RETRIES) {
 180             if (status == SCSI_STATUS_CHECK_CONDITION
 181                 && task->sense.key == SCSI_SENSE_UNIT_ATTENTION) {
 182                 error_report("iSCSI CheckCondition: %s",
 183                              iscsi_get_error(iscsi));
 184                 iTask->do_retry = 1;
 185                 goto out;
 186             }
 187             if (status == SCSI_STATUS_BUSY) {
 188                 unsigned retry_time =
 189                     exp_random(iscsi_retry_times[iTask->retries - 1]);
 190                 error_report("iSCSI Busy (retry #%u in %u ms): %s",
 191                              iTask->retries, retry_time,
 192                              iscsi_get_error(iscsi));
 193                 aio_timer_init(iTask->iscsilun->aio_context,
 194                                &iTask->retry_timer, QEMU_CLOCK_REALTIME,
 195                                SCALE_MS, iscsi_retry_timer_expired, iTask);
 196                 timer_mod(&iTask->retry_timer,
 197                           qemu_clock_get_ms(QEMU_CLOCK_REALTIME) + retry_time);
 198                 iTask->do_retry = 1;
 199                 return;
 200             }
 201         }
 202         error_report("iSCSI Failure: %s", iscsi_get_error(iscsi));
 203     }
 204
 205 out:
 206     if (iTask->co) {
 207         iTask->bh = aio_bh_new(iTask->iscsilun->aio_context,
 208                                iscsi_co_generic_bh_cb, iTask);
 209         qemu_bh_schedule(iTask->bh);
 210     } else {
 211         iTask->complete = 1;
 212     }
 213 }
 214
 215 static void iscsi_co_init_iscsitask(IscsiLun *iscsilun, struct IscsiTask *iTask)
 216 {
 217     *iTask = (struct IscsiTask) {
 218         .co         = qemu_coroutine_self(),
 219         .retries    = ISCSI_CMD_RETRIES,
 220         .iscsilun   = iscsilun,
 221     };
 222 }
 223
 224 static void
 225 iscsi_abort_task_cb(struct iscsi_context *iscsi, int status, void *command_data,
 226                     void *private_data)
 227 {
 228     IscsiAIOCB *acb = private_data;
 229
 230     acb->status = -ECANCELED;
 231     iscsi_schedule_bh(acb);
 232 }
 233
 234 static void
 235 iscsi_aio_cancel(BlockDriverAIOCB *blockacb)
 236 {
 237     IscsiAIOCB *acb = (IscsiAIOCB *)blockacb;
 238     IscsiLun *iscsilun = acb->iscsilun;
 239
 240     if (acb->status != -EINPROGRESS) {
 241         return;
 242     }
 243
 244     acb->canceled = 1;
 245
 246     /* send a task mgmt call to the target to cancel the task on the target */
 247     iscsi_task_mgmt_abort_task_async(iscsilun->iscsi, acb->task,
 248                                      iscsi_abort_task_cb, acb);
 249
 250     while (acb->status == -EINPROGRESS) {
 251         aio_poll(iscsilun->aio_context, true);
 252     }
 253 }
 254
 255 static const AIOCBInfo iscsi_aiocb_info = {
 256     .aiocb_size         = sizeof(IscsiAIOCB),
 257     .cancel             = iscsi_aio_cancel,
 258 };
 259
 260
 261 static void iscsi_process_read(void *arg);
 262 static void iscsi_process_write(void *arg);
 263
 264 static void
 265 iscsi_set_events(IscsiLun *iscsilun)
 266 {
 267     struct iscsi_context *iscsi = iscsilun->iscsi;
 268     int ev;
 269
 270     /* We always register a read handler.  */
 271     ev = POLLIN;
 272     ev |= iscsi_which_events(iscsi);
 273     if (ev != iscsilun->events) {
 274         aio_set_fd_handler(iscsilun->aio_context,
 275                            iscsi_get_fd(iscsi),
 276                            iscsi_process_read,
 277                            (ev & POLLOUT) ? iscsi_process_write : NULL,
 278                            iscsilun);
 279
 280     }
 281
 282     iscsilun->events = ev;
 283 }
 284
 285 static void
 286 iscsi_process_read(void *arg)
 287 {
 288     IscsiLun *iscsilun = arg;
 289     struct iscsi_context *iscsi = iscsilun->iscsi;
 290
 291     iscsi_service(iscsi, POLLIN);
 292     iscsi_set_events(iscsilun);
 293 }
 294
 295 static void
 296 iscsi_process_write(void *arg)
 297 {
 298     IscsiLun *iscsilun = arg;
 299     struct iscsi_context *iscsi = iscsilun->iscsi;
 300
 301     iscsi_service(iscsi, POLLOUT);
 302     iscsi_set_events(iscsilun);
 303 }
 304
 305 static int64_t sector_lun2qemu(int64_t sector, IscsiLun *iscsilun)
 306 {
 307     return sector * iscsilun->block_size / BDRV_SECTOR_SIZE;
 308 }
 309
 310 static int64_t sector_qemu2lun(int64_t sector, IscsiLun *iscsilun)
 311 {
 312     return sector * BDRV_SECTOR_SIZE / iscsilun->block_size;
 313 }
 314
 315 static bool is_request_lun_aligned(int64_t sector_num, int nb_sectors,
 316                                       IscsiLun *iscsilun)
 317 {
 318     if ((sector_num * BDRV_SECTOR_SIZE) % iscsilun->block_size ||
 319         (nb_sectors * BDRV_SECTOR_SIZE) % iscsilun->block_size) {
 320             error_report("iSCSI misaligned request: "
 321                          "iscsilun->block_size %u, sector_num %" PRIi64
 322                          ", nb_sectors %d",
 323                          iscsilun->block_size, sector_num, nb_sectors);
 324             return 0;
 325     }
 326     return 1;
 327 }
 328
 329 static void iscsi_allocationmap_set(IscsiLun *iscsilun, int64_t sector_num,
 330                                     int nb_sectors)
 331 {
 332     if (iscsilun->allocationmap == NULL) {
 333         return;
 334     }
 335     bitmap_set(iscsilun->allocationmap,
 336                sector_num / iscsilun->cluster_sectors,
 337                DIV_ROUND_UP(nb_sectors, iscsilun->cluster_sectors));
 338 }
 339
 340 static void iscsi_allocationmap_clear(IscsiLun *iscsilun, int64_t sector_num,
 341                                       int nb_sectors)
 342 {
 343     int64_t cluster_num, nb_clusters;
 344     if (iscsilun->allocationmap == NULL) {
 345         return;
 346     }
 347     cluster_num = DIV_ROUND_UP(sector_num, iscsilun->cluster_sectors);
 348     nb_clusters = (sector_num + nb_sectors) / iscsilun->cluster_sectors
 349                   - cluster_num;
 350     if (nb_clusters > 0) {
 351         bitmap_clear(iscsilun->allocationmap, cluster_num, nb_clusters);
 352     }
 353 }
 354
 355 static int coroutine_fn iscsi_co_writev(BlockDriverState *bs,
 356                                         int64_t sector_num, int nb_sectors,
 357                                         QEMUIOVector *iov)
 358 {
 359     IscsiLun *iscsilun = bs->opaque;
 360     struct IscsiTask iTask;
 361     uint64_t lba;
 362     uint32_t num_sectors;
 363     uint8_t *data = NULL;
 364     uint8_t *buf = NULL;
 365
 366     if (!is_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
 367         return -EINVAL;
 368     }
 369
 370     lba = sector_qemu2lun(sector_num, iscsilun);
 371     num_sectors = sector_qemu2lun(nb_sectors, iscsilun);
 372     iscsi_co_init_iscsitask(iscsilun, &iTask);
 373 retry:
 374     if (iscsilun->use_16_for_rw) {
 375         iTask.task = iscsi_write16_task(iscsilun->iscsi, iscsilun->lun, lba,
 376                                         data, num_sectors * iscsilun->block_size,
 377                                         iscsilun->block_size, 0, 0, 0, 0, 0,
 378                                         iscsi_co_generic_cb, &iTask);
 379     } else {
 380         iTask.task = iscsi_write10_task(iscsilun->iscsi, iscsilun->lun, lba,
 381                                         data, num_sectors * iscsilun->block_size,
 382                                         iscsilun->block_size, 0, 0, 0, 0, 0,
 383                                         iscsi_co_generic_cb, &iTask);
 384     }
 385     if (iTask.task == NULL) {
 386         g_free(buf);
 387         return -ENOMEM;
 388     }
 389     scsi_task_set_iov_out(iTask.task, (struct scsi_iovec *) iov->iov,
 390                           iov->niov);
 391     while (!iTask.complete) {
 392         iscsi_set_events(iscsilun);
 393         qemu_coroutine_yield();
 394     }
 395
 396     if (iTask.task != NULL) {
 397         scsi_free_scsi_task(iTask.task);
 398         iTask.task = NULL;
 399     }
 400
 401     if (iTask.do_retry) {
 402         iTask.complete = 0;
 403         goto retry;
 404     }
 405
 406     g_free(buf);
 407
 408     if (iTask.status != SCSI_STATUS_GOOD) {
 409         return -EIO;
 410     }
 411
 412     iscsi_allocationmap_set(iscsilun, sector_num, nb_sectors);
 413
 414     return 0;
 415 }
 416
 417
 418 static bool iscsi_allocationmap_is_allocated(IscsiLun *iscsilun,
 419                                              int64_t sector_num, int nb_sectors)
 420 {
 421     unsigned long size;
 422     if (iscsilun->allocationmap == NULL) {
 423         return true;
 424     }
 425     size = DIV_ROUND_UP(sector_num + nb_sectors, iscsilun->cluster_sectors);
 426     return !(find_next_bit(iscsilun->allocationmap, size,
 427                            sector_num / iscsilun->cluster_sectors) == size);
 428 }
 429
 430 static int64_t coroutine_fn iscsi_co_get_block_status(BlockDriverState *bs,
 431                                                   int64_t sector_num,
 432                                                   int nb_sectors, int *pnum)
 433 {
 434     IscsiLun *iscsilun = bs->opaque;
 435     struct scsi_get_lba_status *lbas = NULL;
 436     struct scsi_lba_status_descriptor *lbasd = NULL;
 437     struct IscsiTask iTask;
 438     int64_t ret;
 439
 440     iscsi_co_init_iscsitask(iscsilun, &iTask);
 441
 442     if (!is_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
 443         ret = -EINVAL;
 444         goto out;
 445     }
 446
 447     /* default to all sectors allocated */
 448     ret = BDRV_BLOCK_DATA;
 449     ret |= (sector_num << BDRV_SECTOR_BITS) | BDRV_BLOCK_OFFSET_VALID;
 450     *pnum = nb_sectors;
 451
 452     /* LUN does not support logical block provisioning */
 453     if (iscsilun->lbpme == 0) {
 454         goto out;
 455     }
 456
 457 retry:
 458     if (iscsi_get_lba_status_task(iscsilun->iscsi, iscsilun->lun,
 459                                   sector_qemu2lun(sector_num, iscsilun),
 460                                   8 + 16, iscsi_co_generic_cb,
 461                                   &iTask) == NULL) {
 462         ret = -ENOMEM;
 463         goto out;
 464     }
 465
 466     while (!iTask.complete) {
 467         iscsi_set_events(iscsilun);
 468         qemu_coroutine_yield();
 469     }
 470
 471     if (iTask.do_retry) {
 472         if (iTask.task != NULL) {
 473             scsi_free_scsi_task(iTask.task);
 474             iTask.task = NULL;
 475         }
 476         iTask.complete = 0;
 477         goto retry;
 478     }
 479
 480     if (iTask.status != SCSI_STATUS_GOOD) {
 481         /* in case the get_lba_status_callout fails (i.e.
 482          * because the device is busy or the cmd is not
 483          * supported) we pretend all blocks are allocated
 484          * for backwards compatibility */
 485         goto out;
 486     }
 487
 488     lbas = scsi_datain_unmarshall(iTask.task);
 489     if (lbas == NULL) {
 490         ret = -EIO;
 491         goto out;
 492     }
 493
 494     lbasd = &lbas->descriptors[0];
 495
 496     if (sector_qemu2lun(sector_num, iscsilun) != lbasd->lba) {
 497         ret = -EIO;
 498         goto out;
 499     }
 500
 501     *pnum = sector_lun2qemu(lbasd->num_blocks, iscsilun);
 502
 503     if (lbasd->provisioning == SCSI_PROVISIONING_TYPE_DEALLOCATED ||
 504         lbasd->provisioning == SCSI_PROVISIONING_TYPE_ANCHORED) {
 505         ret &= ~BDRV_BLOCK_DATA;
 506         if (iscsilun->lbprz) {
 507             ret |= BDRV_BLOCK_ZERO;
 508         }
 509     }
 510
 511     if (ret & BDRV_BLOCK_ZERO) {
 512         iscsi_allocationmap_clear(iscsilun, sector_num, *pnum);
 513     } else {
 514         iscsi_allocationmap_set(iscsilun, sector_num, *pnum);
 515     }
 516
 517     if (*pnum > nb_sectors) {
 518         *pnum = nb_sectors;
 519     }
 520 out:
 521     if (iTask.task != NULL) {
 522         scsi_free_scsi_task(iTask.task);
 523     }
 524     return ret;
 525 }
 526
 527 static int coroutine_fn iscsi_co_readv(BlockDriverState *bs,
 528                                        int64_t sector_num, int nb_sectors,
 529                                        QEMUIOVector *iov)
 530 {
 531     IscsiLun *iscsilun = bs->opaque;
 532     struct IscsiTask iTask;
 533     uint64_t lba;
 534     uint32_t num_sectors;
 535
 536     if (!is_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
 537         return -EINVAL;
 538     }
 539
 540     if (iscsilun->lbprz && nb_sectors >= ISCSI_CHECKALLOC_THRES &&
 541         !iscsi_allocationmap_is_allocated(iscsilun, sector_num, nb_sectors)) {
 542         int64_t ret;
 543         int pnum;
 544         ret = iscsi_co_get_block_status(bs, sector_num, INT_MAX, &pnum);
 545         if (ret < 0) {
 546             return ret;
 547         }
 548         if (ret & BDRV_BLOCK_ZERO && pnum >= nb_sectors) {
 549             qemu_iovec_memset(iov, 0, 0x00, iov->size);
 550             return 0;
 551         }
 552     }
 553
 554     lba = sector_qemu2lun(sector_num, iscsilun);
 555     num_sectors = sector_qemu2lun(nb_sectors, iscsilun);
 556
 557     iscsi_co_init_iscsitask(iscsilun, &iTask);
 558 retry:
 559     if (iscsilun->use_16_for_rw) {
 560         iTask.task = iscsi_read16_task(iscsilun->iscsi, iscsilun->lun, lba,
 561                                        num_sectors * iscsilun->block_size,
 562                                        iscsilun->block_size, 0, 0, 0, 0, 0,
 563                                        iscsi_co_generic_cb, &iTask);
 564     } else {
 565         iTask.task = iscsi_read10_task(iscsilun->iscsi, iscsilun->lun, lba,
 566                                        num_sectors * iscsilun->block_size,
 567                                        iscsilun->block_size,
 568                                        0, 0, 0, 0, 0,
 569                                        iscsi_co_generic_cb, &iTask);
 570     }
 571     if (iTask.task == NULL) {
 572         return -ENOMEM;
 573     }
 574     scsi_task_set_iov_in(iTask.task, (struct scsi_iovec *) iov->iov, iov->niov);
 575
 576     while (!iTask.complete) {
 577         iscsi_set_events(iscsilun);
 578         qemu_coroutine_yield();
 579     }
 580
 581     if (iTask.task != NULL) {
 582         scsi_free_scsi_task(iTask.task);
 583         iTask.task = NULL;
 584     }
 585
 586     if (iTask.do_retry) {
 587         iTask.complete = 0;
 588         goto retry;
 589     }
 590
 591     if (iTask.status != SCSI_STATUS_GOOD) {
 592         return -EIO;
 593     }
 594
 595     return 0;
 596 }
 597
 598 static int coroutine_fn iscsi_co_flush(BlockDriverState *bs)
 599 {
 600     IscsiLun *iscsilun = bs->opaque;
 601     struct IscsiTask iTask;
 602
 603     if (bs->sg) {
 604         return 0;
 605     }
 606
 607     iscsi_co_init_iscsitask(iscsilun, &iTask);
 608
 609 retry:
 610     if (iscsi_synchronizecache10_task(iscsilun->iscsi, iscsilun->lun, 0, 0, 0,
 611                                       0, iscsi_co_generic_cb, &iTask) == NULL) {
 612         return -ENOMEM;
 613     }
 614
 615     while (!iTask.complete) {
 616         iscsi_set_events(iscsilun);
 617         qemu_coroutine_yield();
 618     }
 619
 620     if (iTask.task != NULL) {
 621         scsi_free_scsi_task(iTask.task);
 622         iTask.task = NULL;
 623     }
 624
 625     if (iTask.do_retry) {
 626         iTask.complete = 0;
 627         goto retry;
 628     }
 629
 630     if (iTask.status != SCSI_STATUS_GOOD) {
 631         return -EIO;
 632     }
 633
 634     return 0;
 635 }
 636
 637 #ifdef __linux__
 638 static void
 639 iscsi_aio_ioctl_cb(struct iscsi_context *iscsi, int status,
 640                      void *command_data, void *opaque)
 641 {
 642     IscsiAIOCB *acb = opaque;
 643
 644     g_free(acb->buf);
 645     acb->buf = NULL;
 646
 647     if (acb->canceled != 0) {
 648         return;
 649     }
 650
 651     acb->status = 0;
 652     if (status < 0) {
 653         error_report("Failed to ioctl(SG_IO) to iSCSI lun. %s",
 654                      iscsi_get_error(iscsi));
 655         acb->status = -EIO;
 656     }
 657
 658     acb->ioh->driver_status = 0;
 659     acb->ioh->host_status   = 0;
 660     acb->ioh->resid         = 0;
 661
 662 #define SG_ERR_DRIVER_SENSE    0x08
 663
 664     if (status == SCSI_STATUS_CHECK_CONDITION && acb->task->datain.size >= 2) {
 665         int ss;
 666
 667         acb->ioh->driver_status |= SG_ERR_DRIVER_SENSE;
 668
 669         acb->ioh->sb_len_wr = acb->task->datain.size - 2;
 670         ss = (acb->ioh->mx_sb_len >= acb->ioh->sb_len_wr) ?
 671              acb->ioh->mx_sb_len : acb->ioh->sb_len_wr;
 672         memcpy(acb->ioh->sbp, &acb->task->datain.data[2], ss);
 673     }
 674
 675     iscsi_schedule_bh(acb);
 676 }
 677
 678 static BlockDriverAIOCB *iscsi_aio_ioctl(BlockDriverState *bs,
 679         unsigned long int req, void *buf,
 680         BlockDriverCompletionFunc *cb, void *opaque)
 681 {
 682     IscsiLun *iscsilun = bs->opaque;
 683     struct iscsi_context *iscsi = iscsilun->iscsi;
 684     struct iscsi_data data;
 685     IscsiAIOCB *acb;
 686
 687     assert(req == SG_IO);
 688
 689     acb = qemu_aio_get(&iscsi_aiocb_info, bs, cb, opaque);
 690
 691     acb->iscsilun = iscsilun;
 692     acb->canceled    = 0;
 693     acb->bh          = NULL;
 694     acb->status      = -EINPROGRESS;
 695     acb->buf         = NULL;
 696     acb->ioh         = buf;
 697
 698     acb->task = malloc(sizeof(struct scsi_task));
 699     if (acb->task == NULL) {
 700         error_report("iSCSI: Failed to allocate task for scsi command. %s",
 701                      iscsi_get_error(iscsi));
 702         qemu_aio_release(acb);
 703         return NULL;
 704     }
 705     memset(acb->task, 0, sizeof(struct scsi_task));
 706
 707     switch (acb->ioh->dxfer_direction) {
 708     case SG_DXFER_TO_DEV:
 709         acb->task->xfer_dir = SCSI_XFER_WRITE;
 710         break;
 711     case SG_DXFER_FROM_DEV:
 712         acb->task->xfer_dir = SCSI_XFER_READ;
 713         break;
 714     default:
 715         acb->task->xfer_dir = SCSI_XFER_NONE;
 716         break;
 717     }
 718
 719     acb->task->cdb_size = acb->ioh->cmd_len;
 720     memcpy(&acb->task->cdb[0], acb->ioh->cmdp, acb->ioh->cmd_len);
 721     acb->task->expxferlen = acb->ioh->dxfer_len;
 722
 723     data.size = 0;
 724     if (acb->task->xfer_dir == SCSI_XFER_WRITE) {
 725         if (acb->ioh->iovec_count == 0) {
 726             data.data = acb->ioh->dxferp;
 727             data.size = acb->ioh->dxfer_len;
 728         } else {
 729             scsi_task_set_iov_out(acb->task,
 730                                  (struct scsi_iovec *) acb->ioh->dxferp,
 731                                  acb->ioh->iovec_count);
 732         }
 733     }
 734
 735     if (iscsi_scsi_command_async(iscsi, iscsilun->lun, acb->task,
 736                                  iscsi_aio_ioctl_cb,
 737                                  (data.size > 0) ? &data : NULL,
 738                                  acb) != 0) {
 739         scsi_free_scsi_task(acb->task);
 740         qemu_aio_release(acb);
 741         return NULL;
 742     }
 743
 744     /* tell libiscsi to read straight into the buffer we got from ioctl */
 745     if (acb->task->xfer_dir == SCSI_XFER_READ) {
 746         if (acb->ioh->iovec_count == 0) {
 747             scsi_task_add_data_in_buffer(acb->task,
 748                                          acb->ioh->dxfer_len,
 749                                          acb->ioh->dxferp);
 750         } else {
 751             scsi_task_set_iov_in(acb->task,
 752                                  (struct scsi_iovec *) acb->ioh->dxferp,
 753                                  acb->ioh->iovec_count);
 754         }
 755     }
 756
 757     iscsi_set_events(iscsilun);
 758
 759     return &acb->common;
 760 }
 761
 762 static void ioctl_cb(void *opaque, int status)
 763 {
 764     int *p_status = opaque;
 765     *p_status = status;
 766 }
 767
 768 static int iscsi_ioctl(BlockDriverState *bs, unsigned long int req, void *buf)
 769 {
 770     IscsiLun *iscsilun = bs->opaque;
 771     int status;
 772
 773     switch (req) {
 774     case SG_GET_VERSION_NUM:
 775         *(int *)buf = 30000;
 776         break;
 777     case SG_GET_SCSI_ID:
 778         ((struct sg_scsi_id *)buf)->scsi_type = iscsilun->type;
 779         break;
 780     case SG_IO:
 781         status = -EINPROGRESS;
 782         iscsi_aio_ioctl(bs, req, buf, ioctl_cb, &status);
 783
 784         while (status == -EINPROGRESS) {
 785             aio_poll(iscsilun->aio_context, true);
 786         }
 787
 788         return 0;
 789     default:
 790         return -1;
 791     }
 792     return 0;
 793 }
 794 #endif
 795
 796 static int64_t
 797 iscsi_getlength(BlockDriverState *bs)
 798 {
 799     IscsiLun *iscsilun = bs->opaque;
 800     int64_t len;
 801
 802     len  = iscsilun->num_blocks;
 803     len *= iscsilun->block_size;
 804
 805     return len;
 806 }
 807
 808 static int
 809 coroutine_fn iscsi_co_discard(BlockDriverState *bs, int64_t sector_num,
 810                                    int nb_sectors)
 811 {
 812     IscsiLun *iscsilun = bs->opaque;
 813     struct IscsiTask iTask;
 814     struct unmap_list list;
 815
 816     if (!is_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
 817         return -EINVAL;
 818     }
 819
 820     if (!iscsilun->lbp.lbpu) {
 821         /* UNMAP is not supported by the target */
 822         return 0;
 823     }
 824
 825     list.lba = sector_qemu2lun(sector_num, iscsilun);
 826     list.num = sector_qemu2lun(nb_sectors, iscsilun);
 827
 828     iscsi_co_init_iscsitask(iscsilun, &iTask);
 829 retry:
 830     if (iscsi_unmap_task(iscsilun->iscsi, iscsilun->lun, 0, 0, &list, 1,
 831                      iscsi_co_generic_cb, &iTask) == NULL) {
 832         return -ENOMEM;
 833     }
 834
 835     while (!iTask.complete) {
 836         iscsi_set_events(iscsilun);
 837         qemu_coroutine_yield();
 838     }
 839
 840     if (iTask.task != NULL) {
 841         scsi_free_scsi_task(iTask.task);
 842         iTask.task = NULL;
 843     }
 844
 845     if (iTask.do_retry) {
 846         iTask.complete = 0;
 847         goto retry;
 848     }
 849
 850     if (iTask.status == SCSI_STATUS_CHECK_CONDITION) {
 851         /* the target might fail with a check condition if it
 852            is not happy with the alignment of the UNMAP request
 853            we silently fail in this case */
 854         return 0;
 855     }
 856
 857     if (iTask.status != SCSI_STATUS_GOOD) {
 858         return -EIO;
 859     }
 860
 861     iscsi_allocationmap_clear(iscsilun, sector_num, nb_sectors);
 862
 863     return 0;
 864 }
 865
 866 static int
 867 coroutine_fn iscsi_co_write_zeroes(BlockDriverState *bs, int64_t sector_num,
 868                                    int nb_sectors, BdrvRequestFlags flags)
 869 {
 870     IscsiLun *iscsilun = bs->opaque;
 871     struct IscsiTask iTask;
 872     uint64_t lba;
 873     uint32_t nb_blocks;
 874     bool use_16_for_ws = iscsilun->use_16_for_rw;
 875
 876     if (!is_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
 877         return -EINVAL;
 878     }
 879
 880     if (flags & BDRV_REQ_MAY_UNMAP) {
 881         if (!use_16_for_ws && !iscsilun->lbp.lbpws10) {
 882             /* WRITESAME10 with UNMAP is unsupported try WRITESAME16 */
 883             use_16_for_ws = true;
 884         }
 885         if (use_16_for_ws && !iscsilun->lbp.lbpws) {
 886             /* WRITESAME16 with UNMAP is not supported by the target,
 887              * fall back and try WRITESAME10/16 without UNMAP */
 888             flags &= ~BDRV_REQ_MAY_UNMAP;
 889             use_16_for_ws = iscsilun->use_16_for_rw;
 890         }
 891     }
 892
 893     if (!(flags & BDRV_REQ_MAY_UNMAP) && !iscsilun->has_write_same) {
 894         /* WRITESAME without UNMAP is not supported by the target */
 895         return -ENOTSUP;
 896     }
 897
 898     lba = sector_qemu2lun(sector_num, iscsilun);
 899     nb_blocks = sector_qemu2lun(nb_sectors, iscsilun);
 900
 901     if (iscsilun->zeroblock == NULL) {
 902         iscsilun->zeroblock = g_malloc0(iscsilun->block_size);
 903     }
 904
 905     iscsi_co_init_iscsitask(iscsilun, &iTask);
 906 retry:
 907     if (use_16_for_ws) {
 908         iTask.task = iscsi_writesame16_task(iscsilun->iscsi, iscsilun->lun, lba,
 909                                             iscsilun->zeroblock, iscsilun->block_size,
 910                                             nb_blocks, 0, !!(flags & BDRV_REQ_MAY_UNMAP),
 911                                             0, 0, iscsi_co_generic_cb, &iTask);
 912     } else {
 913         iTask.task = iscsi_writesame10_task(iscsilun->iscsi, iscsilun->lun, lba,
 914                                             iscsilun->zeroblock, iscsilun->block_size,
 915                                             nb_blocks, 0, !!(flags & BDRV_REQ_MAY_UNMAP),
 916                                             0, 0, iscsi_co_generic_cb, &iTask);
 917     }
 918     if (iTask.task == NULL) {
 919         return -ENOMEM;
 920     }
 921
 922     while (!iTask.complete) {
 923         iscsi_set_events(iscsilun);
 924         qemu_coroutine_yield();
 925     }
 926
 927     if (iTask.status == SCSI_STATUS_CHECK_CONDITION &&
 928         iTask.task->sense.key == SCSI_SENSE_ILLEGAL_REQUEST &&
 929         (iTask.task->sense.ascq == SCSI_SENSE_ASCQ_INVALID_OPERATION_CODE ||
 930          iTask.task->sense.ascq == SCSI_SENSE_ASCQ_INVALID_FIELD_IN_CDB)) {
 931         /* WRITE SAME is not supported by the target */
 932         iscsilun->has_write_same = false;
 933         scsi_free_scsi_task(iTask.task);
 934         return -ENOTSUP;
 935     }
 936
 937     if (iTask.task != NULL) {
 938         scsi_free_scsi_task(iTask.task);
 939         iTask.task = NULL;
 940     }
 941
 942     if (iTask.do_retry) {
 943         iTask.complete = 0;
 944         goto retry;
 945     }
 946
 947     if (iTask.status != SCSI_STATUS_GOOD) {
 948         return -EIO;
 949     }
 950
 951     if (flags & BDRV_REQ_MAY_UNMAP) {
 952         iscsi_allocationmap_clear(iscsilun, sector_num, nb_sectors);
 953     } else {
 954         iscsi_allocationmap_set(iscsilun, sector_num, nb_sectors);
 955     }
 956
 957     return 0;
 958 }
 959
 960 static void parse_chap(struct iscsi_context *iscsi, const char *target,
 961                        Error **errp)
 962 {
 963     QemuOptsList *list;
 964     QemuOpts *opts;
 965     const char *user = NULL;
 966     const char *password = NULL;
 967
 968     list = qemu_find_opts("iscsi");
 969     if (!list) {
 970         return;
 971     }
 972
 973     opts = qemu_opts_find(list, target);
 974     if (opts == NULL) {
 975         opts = QTAILQ_FIRST(&list->head);
 976         if (!opts) {
 977             return;
 978         }
 979     }
 980
 981     user = qemu_opt_get(opts, "user");
 982     if (!user) {
 983         return;
 984     }
 985
 986     password = qemu_opt_get(opts, "password");
 987     if (!password) {
 988         error_setg(errp, "CHAP username specified but no password was given");
 989         return;
 990     }
 991
 992     if (iscsi_set_initiator_username_pwd(iscsi, user, password)) {
 993         error_setg(errp, "Failed to set initiator username and password");
 994     }
 995 }
 996
 997 static void parse_header_digest(struct iscsi_context *iscsi, const char *target,
 998                                 Error **errp)
 999 {
1000     QemuOptsList *list;
1001     QemuOpts *opts;
1002     const char *digest = NULL;
1003
1004     list = qemu_find_opts("iscsi");
1005     if (!list) {
1006         return;
1007     }
1008
1009     opts = qemu_opts_find(list, target);
1010     if (opts == NULL) {
1011         opts = QTAILQ_FIRST(&list->head);
1012         if (!opts) {
1013             return;
1014         }
1015     }
1016
1017     digest = qemu_opt_get(opts, "header-digest");
1018     if (!digest) {
1019         return;
1020     }
1021
1022     if (!strcmp(digest, "CRC32C")) {
1023         iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_CRC32C);
1024     } else if (!strcmp(digest, "NONE")) {
1025         iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_NONE);
1026     } else if (!strcmp(digest, "CRC32C-NONE")) {
1027         iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_CRC32C_NONE);
1028     } else if (!strcmp(digest, "NONE-CRC32C")) {
1029         iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_NONE_CRC32C);
1030     } else {
1031         error_setg(errp, "Invalid header-digest setting : %s", digest);
1032     }
1033 }
1034
1035 static char *parse_initiator_name(const char *target)
1036 {
1037     QemuOptsList *list;
1038     QemuOpts *opts;
1039     const char *name;
1040     char *iscsi_name;
1041     UuidInfo *uuid_info;
1042
1043     list = qemu_find_opts("iscsi");
1044     if (list) {
1045         opts = qemu_opts_find(list, target);
1046         if (!opts) {
1047             opts = QTAILQ_FIRST(&list->head);
1048         }
1049         if (opts) {
1050             name = qemu_opt_get(opts, "initiator-name");
1051             if (name) {
1052                 return g_strdup(name);
1053             }
1054         }
1055     }
1056
1057     uuid_info = qmp_query_uuid(NULL);
1058     if (strcmp(uuid_info->UUID, UUID_NONE) == 0) {
1059         name = qemu_get_vm_name();
1060     } else {
1061         name = uuid_info->UUID;
1062     }
1063     iscsi_name = g_strdup_printf("iqn.2008-11.org.linux-kvm%s%s",
1064                                  name ? ":" : "", name ? name : "");
1065     qapi_free_UuidInfo(uuid_info);
1066     return iscsi_name;
1067 }
1068
1069 static void iscsi_nop_timed_event(void *opaque)
1070 {
1071     IscsiLun *iscsilun = opaque;
1072
1073     if (iscsi_get_nops_in_flight(iscsilun->iscsi) > MAX_NOP_FAILURES) {
1074         error_report("iSCSI: NOP timeout. Reconnecting...");
1075         iscsi_reconnect(iscsilun->iscsi);
1076     }
1077
1078     if (iscsi_nop_out_async(iscsilun->iscsi, NULL, NULL, 0, NULL) != 0) {
1079         error_report("iSCSI: failed to sent NOP-Out. Disabling NOP messages.");
1080         return;
1081     }
1082
1083     timer_mod(iscsilun->nop_timer, qemu_clock_get_ms(QEMU_CLOCK_REALTIME) + NOP_INTERVAL);
1084     iscsi_set_events(iscsilun);
1085 }
1086
1087 static void iscsi_readcapacity_sync(IscsiLun *iscsilun, Error **errp)
1088 {
1089     struct scsi_task *task = NULL;
1090     struct scsi_readcapacity10 *rc10 = NULL;
1091     struct scsi_readcapacity16 *rc16 = NULL;
1092     int retries = ISCSI_CMD_RETRIES;
1093
1094     do {
1095         if (task != NULL) {
1096             scsi_free_scsi_task(task);
1097             task = NULL;
1098         }
1099
1100         switch (iscsilun->type) {
1101         case TYPE_DISK:
1102             task = iscsi_readcapacity16_sync(iscsilun->iscsi, iscsilun->lun);
1103             if (task != NULL && task->status == SCSI_STATUS_GOOD) {
1104                 rc16 = scsi_datain_unmarshall(task);
1105                 if (rc16 == NULL) {
1106                     error_setg(errp, "iSCSI: Failed to unmarshall readcapacity16 data.");
1107                 } else {
1108                     iscsilun->block_size = rc16->block_length;
1109                     iscsilun->num_blocks = rc16->returned_lba + 1;
1110                     iscsilun->lbpme = rc16->lbpme;
1111                     iscsilun->lbprz = rc16->lbprz;
1112                     iscsilun->use_16_for_rw = (rc16->returned_lba > 0xffffffff);
1113                 }
1114             }
1115             break;
1116         case TYPE_ROM:
1117             task = iscsi_readcapacity10_sync(iscsilun->iscsi, iscsilun->lun, 0, 0);
1118             if (task != NULL && task->status == SCSI_STATUS_GOOD) {
1119                 rc10 = scsi_datain_unmarshall(task);
1120                 if (rc10 == NULL) {
1121                     error_setg(errp, "iSCSI: Failed to unmarshall readcapacity10 data.");
1122                 } else {
1123                     iscsilun->block_size = rc10->block_size;
1124                     if (rc10->lba == 0) {
1125                         /* blank disk loaded */
1126                         iscsilun->num_blocks = 0;
1127                     } else {
1128                         iscsilun->num_blocks = rc10->lba + 1;
1129                     }
1130                 }
1131             }
1132             break;
1133         default:
1134             return;
1135         }
1136     } while (task != NULL && task->status == SCSI_STATUS_CHECK_CONDITION
1137              && task->sense.key == SCSI_SENSE_UNIT_ATTENTION
1138              && retries-- > 0);
1139
1140     if (task == NULL || task->status != SCSI_STATUS_GOOD) {
1141         error_setg(errp, "iSCSI: failed to send readcapacity10 command.");
1142     }
1143     if (task) {
1144         scsi_free_scsi_task(task);
1145     }
1146 }
1147
1148 /* TODO Convert to fine grained options */
1149 static QemuOptsList runtime_opts = {
1150     .name = "iscsi",
1151     .head = QTAILQ_HEAD_INITIALIZER(runtime_opts.head),
1152     .desc = {
1153         {
1154             .name = "filename",
1155             .type = QEMU_OPT_STRING,
1156             .help = "URL to the iscsi image",
1157         },
1158         { /* end of list */ }
1159     },
1160 };
1161
1162 static struct scsi_task *iscsi_do_inquiry(struct iscsi_context *iscsi, int lun,
1163                                           int evpd, int pc, void **inq, Error **errp)
1164 {
1165     int full_size;
1166     struct scsi_task *task = NULL;
1167     task = iscsi_inquiry_sync(iscsi, lun, evpd, pc, 64);
1168     if (task == NULL || task->status != SCSI_STATUS_GOOD) {
1169         goto fail;
1170     }
1171     full_size = scsi_datain_getfullsize(task);
1172     if (full_size > task->datain.size) {
1173         scsi_free_scsi_task(task);
1174
1175         /* we need more data for the full list */
1176         task = iscsi_inquiry_sync(iscsi, lun, evpd, pc, full_size);
1177         if (task == NULL || task->status != SCSI_STATUS_GOOD) {
1178             goto fail;
1179         }
1180     }
1181
1182     *inq = scsi_datain_unmarshall(task);
1183     if (*inq == NULL) {
1184         error_setg(errp, "iSCSI: failed to unmarshall inquiry datain blob");
1185         goto fail_with_err;
1186     }
1187
1188     return task;
1189
1190 fail:
1191     error_setg(errp, "iSCSI: Inquiry command failed : %s",
1192                iscsi_get_error(iscsi));
1193 fail_with_err:
1194     if (task != NULL) {
1195         scsi_free_scsi_task(task);
1196     }
1197     return NULL;
1198 }
1199
1200 static void iscsi_detach_aio_context(BlockDriverState *bs)
1201 {
1202     IscsiLun *iscsilun = bs->opaque;
1203
1204     aio_set_fd_handler(iscsilun->aio_context,
1205                        iscsi_get_fd(iscsilun->iscsi),
1206                        NULL, NULL, NULL);
1207     iscsilun->events = 0;
1208
1209     if (iscsilun->nop_timer) {
1210         timer_del(iscsilun->nop_timer);
1211         timer_free(iscsilun->nop_timer);
1212         iscsilun->nop_timer = NULL;
1213     }
1214 }
1215
1216 static void iscsi_attach_aio_context(BlockDriverState *bs,
1217                                      AioContext *new_context)
1218 {
1219     IscsiLun *iscsilun = bs->opaque;
1220
1221     iscsilun->aio_context = new_context;
1222     iscsi_set_events(iscsilun);
1223
1224     /* Set up a timer for sending out iSCSI NOPs */
1225     iscsilun->nop_timer = aio_timer_new(iscsilun->aio_context,
1226                                         QEMU_CLOCK_REALTIME, SCALE_MS,
1227                                         iscsi_nop_timed_event, iscsilun);
1228     timer_mod(iscsilun->nop_timer,
1229               qemu_clock_get_ms(QEMU_CLOCK_REALTIME) + NOP_INTERVAL);
1230 }
1231
1232 /*
1233  * We support iscsi url's on the form
1234  * iscsi://[<username>%<password>@]<host>[:<port>]/<targetname>/<lun>
1235  *
1236  * Note: flags are currently not used by iscsi_open.  If this function
1237  * is changed such that flags are used, please examine iscsi_reopen_prepare()
1238  * to see if needs to be changed as well.
1239  */
1240 static int iscsi_open(BlockDriverState *bs, QDict *options, int flags,
1241                       Error **errp)
1242 {
1243     IscsiLun *iscsilun = bs->opaque;
1244     struct iscsi_context *iscsi = NULL;
1245     struct iscsi_url *iscsi_url = NULL;
1246     struct scsi_task *task = NULL;
1247     struct scsi_inquiry_standard *inq = NULL;
1248     struct scsi_inquiry_supported_pages *inq_vpd;
1249     char *initiator_name = NULL;
1250     QemuOpts *opts;
1251     Error *local_err = NULL;
1252     const char *filename;
1253     int i, ret;
1254
1255     if ((BDRV_SECTOR_SIZE % 512) != 0) {
1256         error_setg(errp, "iSCSI: Invalid BDRV_SECTOR_SIZE. "
1257                    "BDRV_SECTOR_SIZE(%lld) is not a multiple "
1258                    "of 512", BDRV_SECTOR_SIZE);
1259         return -EINVAL;
1260     }
1261
1262     opts = qemu_opts_create(&runtime_opts, NULL, 0, &error_abort);
1263     qemu_opts_absorb_qdict(opts, options, &local_err);
1264     if (local_err) {
1265         error_propagate(errp, local_err);
1266         ret = -EINVAL;
1267         goto out;
1268     }
1269
1270     filename = qemu_opt_get(opts, "filename");
1271
1272     iscsi_url = iscsi_parse_full_url(iscsi, filename);
1273     if (iscsi_url == NULL) {
1274         error_setg(errp, "Failed to parse URL : %s", filename);
1275         ret = -EINVAL;
1276         goto out;
1277     }
1278
1279     memset(iscsilun, 0, sizeof(IscsiLun));
1280
1281     initiator_name = parse_initiator_name(iscsi_url->target);
1282
1283     iscsi = iscsi_create_context(initiator_name);
1284     if (iscsi == NULL) {
1285         error_setg(errp, "iSCSI: Failed to create iSCSI context.");
1286         ret = -ENOMEM;
1287         goto out;
1288     }
1289
1290     if (iscsi_set_targetname(iscsi, iscsi_url->target)) {
1291         error_setg(errp, "iSCSI: Failed to set target name.");
1292         ret = -EINVAL;
1293         goto out;
1294     }
1295
1296     if (iscsi_url->user != NULL) {
1297         ret = iscsi_set_initiator_username_pwd(iscsi, iscsi_url->user,
1298                                               iscsi_url->passwd);
1299         if (ret != 0) {
1300             error_setg(errp, "Failed to set initiator username and password");
1301             ret = -EINVAL;
1302             goto out;
1303         }
1304     }
1305
1306     /* check if we got CHAP username/password via the options */
1307     parse_chap(iscsi, iscsi_url->target, &local_err);
1308     if (local_err != NULL) {
1309         error_propagate(errp, local_err);
1310         ret = -EINVAL;
1311         goto out;
1312     }
1313
1314     if (iscsi_set_session_type(iscsi, ISCSI_SESSION_NORMAL) != 0) {
1315         error_setg(errp, "iSCSI: Failed to set session type to normal.");
1316         ret = -EINVAL;
1317         goto out;
1318     }
1319
1320     iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_NONE_CRC32C);
1321
1322     /* check if we got HEADER_DIGEST via the options */
1323     parse_header_digest(iscsi, iscsi_url->target, &local_err);
1324     if (local_err != NULL) {
1325         error_propagate(errp, local_err);
1326         ret = -EINVAL;
1327         goto out;
1328     }
1329
1330     if (iscsi_full_connect_sync(iscsi, iscsi_url->portal, iscsi_url->lun) != 0) {
1331         error_setg(errp, "iSCSI: Failed to connect to LUN : %s",
1332             iscsi_get_error(iscsi));
1333         ret = -EINVAL;
1334         goto out;
1335     }
1336
1337     iscsilun->iscsi = iscsi;
1338     iscsilun->aio_context = bdrv_get_aio_context(bs);
1339     iscsilun->lun   = iscsi_url->lun;
1340     iscsilun->has_write_same = true;
1341
1342     task = iscsi_do_inquiry(iscsilun->iscsi, iscsilun->lun, 0, 0,
1343                             (void **) &inq, errp);
1344     if (task == NULL) {
1345         ret = -EINVAL;
1346         goto out;
1347     }
1348     iscsilun->type = inq->periperal_device_type;
1349     scsi_free_scsi_task(task);
1350     task = NULL;
1351
1352     iscsi_readcapacity_sync(iscsilun, &local_err);
1353     if (local_err != NULL) {
1354         error_propagate(errp, local_err);
1355         ret = -EINVAL;
1356         goto out;
1357     }
1358     bs->total_sectors = sector_lun2qemu(iscsilun->num_blocks, iscsilun);
1359     bs->request_alignment = iscsilun->block_size;
1360
1361     /* We don't have any emulation for devices other than disks and CD-ROMs, so
1362      * this must be sg ioctl compatible. We force it to be sg, otherwise qemu
1363      * will try to read from the device to guess the image format.
1364      */
1365     if (iscsilun->type != TYPE_DISK && iscsilun->type != TYPE_ROM) {
1366         bs->sg = 1;
1367     }
1368
1369     task = iscsi_do_inquiry(iscsilun->iscsi, iscsilun->lun, 1,
1370                             SCSI_INQUIRY_PAGECODE_SUPPORTED_VPD_PAGES,
1371                             (void **) &inq_vpd, errp);
1372     if (task == NULL) {
1373         ret = -EINVAL;
1374         goto out;
1375     }
1376     for (i = 0; i < inq_vpd->num_pages; i++) {
1377         struct scsi_task *inq_task;
1378         struct scsi_inquiry_logical_block_provisioning *inq_lbp;
1379         struct scsi_inquiry_block_limits *inq_bl;
1380         switch (inq_vpd->pages[i]) {
1381         case SCSI_INQUIRY_PAGECODE_LOGICAL_BLOCK_PROVISIONING:
1382             inq_task = iscsi_do_inquiry(iscsilun->iscsi, iscsilun->lun, 1,
1383                                         SCSI_INQUIRY_PAGECODE_LOGICAL_BLOCK_PROVISIONING,
1384                                         (void **) &inq_lbp, errp);
1385             if (inq_task == NULL) {
1386                 ret = -EINVAL;
1387                 goto out;
1388             }
1389             memcpy(&iscsilun->lbp, inq_lbp,
1390                    sizeof(struct scsi_inquiry_logical_block_provisioning));
1391             scsi_free_scsi_task(inq_task);
1392             break;
1393         case SCSI_INQUIRY_PAGECODE_BLOCK_LIMITS:
1394             inq_task = iscsi_do_inquiry(iscsilun->iscsi, iscsilun->lun, 1,
1395                                     SCSI_INQUIRY_PAGECODE_BLOCK_LIMITS,
1396                                     (void **) &inq_bl, errp);
1397             if (inq_task == NULL) {
1398                 ret = -EINVAL;
1399                 goto out;
1400             }
1401             memcpy(&iscsilun->bl, inq_bl,
1402                    sizeof(struct scsi_inquiry_block_limits));
1403             scsi_free_scsi_task(inq_task);
1404             break;
1405         default:
1406             break;
1407         }
1408     }
1409     scsi_free_scsi_task(task);
1410     task = NULL;
1411
1412     iscsi_attach_aio_context(bs, iscsilun->aio_context);
1413
1414     /* Guess the internal cluster (page) size of the iscsi target by the means
1415      * of opt_unmap_gran. Transfer the unmap granularity only if it has a
1416      * reasonable size */
1417     if (iscsilun->bl.opt_unmap_gran * iscsilun->block_size >= 4 * 1024 &&
1418         iscsilun->bl.opt_unmap_gran * iscsilun->block_size <= 16 * 1024 * 1024) {
1419         iscsilun->cluster_sectors = (iscsilun->bl.opt_unmap_gran *
1420                                      iscsilun->block_size) >> BDRV_SECTOR_BITS;
1421         if (iscsilun->lbprz && !(bs->open_flags & BDRV_O_NOCACHE)) {
1422             iscsilun->allocationmap =
1423                 bitmap_new(DIV_ROUND_UP(bs->total_sectors,
1424                                         iscsilun->cluster_sectors));
1425         }
1426     }
1427
1428 out:
1429     qemu_opts_del(opts);
1430     g_free(initiator_name);
1431     if (iscsi_url != NULL) {
1432         iscsi_destroy_url(iscsi_url);
1433     }
1434     if (task != NULL) {
1435         scsi_free_scsi_task(task);
1436     }
1437
1438     if (ret) {
1439         if (iscsi != NULL) {
1440             iscsi_destroy_context(iscsi);
1441         }
1442         memset(iscsilun, 0, sizeof(IscsiLun));
1443     }
1444     return ret;
1445 }
1446
1447 static void iscsi_close(BlockDriverState *bs)
1448 {
1449     IscsiLun *iscsilun = bs->opaque;
1450     struct iscsi_context *iscsi = iscsilun->iscsi;
1451
1452     iscsi_detach_aio_context(bs);
1453     iscsi_destroy_context(iscsi);
1454     g_free(iscsilun->zeroblock);
1455     g_free(iscsilun->allocationmap);
1456     memset(iscsilun, 0, sizeof(IscsiLun));
1457 }
1458
1459 static int iscsi_refresh_limits(BlockDriverState *bs)
1460 {
1461     IscsiLun *iscsilun = bs->opaque;
1462
1463     /* We don't actually refresh here, but just return data queried in
1464      * iscsi_open(): iscsi targets don't change their limits. */
1465     if (iscsilun->lbp.lbpu) {
1466         if (iscsilun->bl.max_unmap < 0xffffffff) {
1467             bs->bl.max_discard = sector_lun2qemu(iscsilun->bl.max_unmap,
1468                                                  iscsilun);
1469         }
1470         bs->bl.discard_alignment = sector_lun2qemu(iscsilun->bl.opt_unmap_gran,
1471                                                    iscsilun);
1472     }
1473
1474     if (iscsilun->bl.max_ws_len < 0xffffffff) {
1475         bs->bl.max_write_zeroes = sector_lun2qemu(iscsilun->bl.max_ws_len,
1476                                                   iscsilun);
1477     }
1478     if (iscsilun->lbp.lbpws) {
1479         bs->bl.write_zeroes_alignment = sector_lun2qemu(iscsilun->bl.opt_unmap_gran,
1480                                                         iscsilun);
1481     }
1482     bs->bl.opt_transfer_length = sector_lun2qemu(iscsilun->bl.opt_xfer_len,
1483                                                  iscsilun);
1484     return 0;
1485 }
1486
1487 /* Since iscsi_open() ignores bdrv_flags, there is nothing to do here in
1488  * prepare.  Note that this will not re-establish a connection with an iSCSI
1489  * target - it is effectively a NOP.  */
1490 static int iscsi_reopen_prepare(BDRVReopenState *state,
1491                                 BlockReopenQueue *queue, Error **errp)
1492 {
1493     /* NOP */
1494     return 0;
1495 }
1496
1497 static int iscsi_truncate(BlockDriverState *bs, int64_t offset)
1498 {
1499     IscsiLun *iscsilun = bs->opaque;
1500     Error *local_err = NULL;
1501
1502     if (iscsilun->type != TYPE_DISK) {
1503         return -ENOTSUP;
1504     }
1505
1506     iscsi_readcapacity_sync(iscsilun, &local_err);
1507     if (local_err != NULL) {
1508         error_free(local_err);
1509         return -EIO;
1510     }
1511
1512     if (offset > iscsi_getlength(bs)) {
1513         return -EINVAL;
1514     }
1515
1516     if (iscsilun->allocationmap != NULL) {
1517         g_free(iscsilun->allocationmap);
1518         iscsilun->allocationmap =
1519             bitmap_new(DIV_ROUND_UP(bs->total_sectors,
1520                                     iscsilun->cluster_sectors));
1521     }
1522
1523     return 0;
1524 }
1525
1526 static int iscsi_create(const char *filename, QemuOpts *opts, Error **errp)
1527 {
1528     int ret = 0;
1529     int64_t total_size = 0;
1530     BlockDriverState *bs;
1531     IscsiLun *iscsilun = NULL;
1532     QDict *bs_options;
1533
1534     bs = bdrv_new("", &error_abort);
1535
1536     /* Read out options */
1537     total_size =
1538         qemu_opt_get_size_del(opts, BLOCK_OPT_SIZE, 0) / BDRV_SECTOR_SIZE;
1539     bs->opaque = g_malloc0(sizeof(struct IscsiLun));
1540     iscsilun = bs->opaque;
1541
1542     bs_options = qdict_new();
1543     qdict_put(bs_options, "filename", qstring_from_str(filename));
1544     ret = iscsi_open(bs, bs_options, 0, NULL);
1545     QDECREF(bs_options);
1546
1547     if (ret != 0) {
1548         goto out;
1549     }
1550     iscsi_detach_aio_context(bs);
1551     if (iscsilun->type != TYPE_DISK) {
1552         ret = -ENODEV;
1553         goto out;
1554     }
1555     if (bs->total_sectors < total_size) {
1556         ret = -ENOSPC;
1557         goto out;
1558     }
1559
1560     ret = 0;
1561 out:
1562     if (iscsilun->iscsi != NULL) {
1563         iscsi_destroy_context(iscsilun->iscsi);
1564     }
1565     g_free(bs->opaque);
1566     bs->opaque = NULL;
1567     bdrv_unref(bs);
1568     return ret;
1569 }
1570
1571 static int iscsi_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
1572 {
1573     IscsiLun *iscsilun = bs->opaque;
1574     bdi->unallocated_blocks_are_zero = !!iscsilun->lbprz;
1575     bdi->can_write_zeroes_with_unmap = iscsilun->lbprz && iscsilun->lbp.lbpws;
1576     bdi->cluster_size = iscsilun->cluster_sectors * BDRV_SECTOR_SIZE;
1577     return 0;
1578 }
1579
1580 static QemuOptsList iscsi_create_opts = {
1581     .name = "iscsi-create-opts",
1582     .head = QTAILQ_HEAD_INITIALIZER(iscsi_create_opts.head),
1583     .desc = {
1584         {
1585             .name = BLOCK_OPT_SIZE,
1586             .type = QEMU_OPT_SIZE,
1587             .help = "Virtual disk size"
1588         },
1589         { /* end of list */ }
1590     }
1591 };
1592
1593 static BlockDriver bdrv_iscsi = {
1594     .format_name     = "iscsi",
1595     .protocol_name   = "iscsi",
1596
1597     .instance_size   = sizeof(IscsiLun),
1598     .bdrv_needs_filename = true,
1599     .bdrv_file_open  = iscsi_open,
1600     .bdrv_close      = iscsi_close,
1601     .bdrv_create     = iscsi_create,
1602     .create_opts     = &iscsi_create_opts,
1603     .bdrv_reopen_prepare  = iscsi_reopen_prepare,
1604
1605     .bdrv_getlength  = iscsi_getlength,
1606     .bdrv_get_info   = iscsi_get_info,
1607     .bdrv_truncate   = iscsi_truncate,
1608     .bdrv_refresh_limits = iscsi_refresh_limits,
1609
1610     .bdrv_co_get_block_status = iscsi_co_get_block_status,
1611     .bdrv_co_discard      = iscsi_co_discard,
1612     .bdrv_co_write_zeroes = iscsi_co_write_zeroes,
1613     .bdrv_co_readv         = iscsi_co_readv,
1614     .bdrv_co_writev        = iscsi_co_writev,
1615     .bdrv_co_flush_to_disk = iscsi_co_flush,
1616
1617 #ifdef __linux__
1618     .bdrv_ioctl       = iscsi_ioctl,
1619     .bdrv_aio_ioctl   = iscsi_aio_ioctl,
1620 #endif
1621
1622     .bdrv_detach_aio_context = iscsi_detach_aio_context,
1623     .bdrv_attach_aio_context = iscsi_attach_aio_context,
1624 };
1625
1626 static QemuOptsList qemu_iscsi_opts = {
1627     .name = "iscsi",
1628     .head = QTAILQ_HEAD_INITIALIZER(qemu_iscsi_opts.head),
1629     .desc = {
1630         {
1631             .name = "user",
1632             .type = QEMU_OPT_STRING,
1633             .help = "username for CHAP authentication to target",
1634         },{
1635             .name = "password",
1636             .type = QEMU_OPT_STRING,
1637             .help = "password for CHAP authentication to target",
1638         },{
1639             .name = "header-digest",
1640             .type = QEMU_OPT_STRING,
1641             .help = "HeaderDigest setting. "
1642                     "{CRC32C|CRC32C-NONE|NONE-CRC32C|NONE}",
1643         },{
1644             .name = "initiator-name",
1645             .type = QEMU_OPT_STRING,
1646             .help = "Initiator iqn name to use when connecting",
1647         },
1648         { /* end of list */ }
1649     },
1650 };
1651
1652 static void iscsi_block_init(void)
1653 {
1654     bdrv_register(&bdrv_iscsi);
1655     qemu_add_opts(&qemu_iscsi_opts);
1656 }
1657
1658 block_init(iscsi_block_init);