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 threshold 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         .iscsilun   = iscsilun,
 220     };
 221 }
 222
 223 static void
 224 iscsi_abort_task_cb(struct iscsi_context *iscsi, int status, void *command_data,
 225                     void *private_data)
 226 {
 227     IscsiAIOCB *acb = private_data;
 228
 229     acb->status = -ECANCELED;
 230     iscsi_schedule_bh(acb);
 231 }
 232
 233 static void
 234 iscsi_aio_cancel(BlockDriverAIOCB *blockacb)
 235 {
 236     IscsiAIOCB *acb = (IscsiAIOCB *)blockacb;
 237     IscsiLun *iscsilun = acb->iscsilun;
 238
 239     if (acb->status != -EINPROGRESS) {
 240         return;
 241     }
 242
 243     acb->canceled = 1;
 244
 245     /* send a task mgmt call to the target to cancel the task on the target */
 246     iscsi_task_mgmt_abort_task_async(iscsilun->iscsi, acb->task,
 247                                      iscsi_abort_task_cb, acb);
 248
 249     while (acb->status == -EINPROGRESS) {
 250         aio_poll(iscsilun->aio_context, true);
 251     }
 252 }
 253
 254 static const AIOCBInfo iscsi_aiocb_info = {
 255     .aiocb_size         = sizeof(IscsiAIOCB),
 256     .cancel             = iscsi_aio_cancel,
 257 };
 258
 259
 260 static void iscsi_process_read(void *arg);
 261 static void iscsi_process_write(void *arg);
 262
 263 static void
 264 iscsi_set_events(IscsiLun *iscsilun)
 265 {
 266     struct iscsi_context *iscsi = iscsilun->iscsi;
 267     int ev;
 268
 269     /* We always register a read handler.  */
 270     ev = POLLIN;
 271     ev |= iscsi_which_events(iscsi);
 272     if (ev != iscsilun->events) {
 273         aio_set_fd_handler(iscsilun->aio_context,
 274                            iscsi_get_fd(iscsi),
 275                            iscsi_process_read,
 276                            (ev & POLLOUT) ? iscsi_process_write : NULL,
 277                            iscsilun);
 278
 279     }
 280
 281     iscsilun->events = ev;
 282 }
 283
 284 static void
 285 iscsi_process_read(void *arg)
 286 {
 287     IscsiLun *iscsilun = arg;
 288     struct iscsi_context *iscsi = iscsilun->iscsi;
 289
 290     iscsi_service(iscsi, POLLIN);
 291     iscsi_set_events(iscsilun);
 292 }
 293
 294 static void
 295 iscsi_process_write(void *arg)
 296 {
 297     IscsiLun *iscsilun = arg;
 298     struct iscsi_context *iscsi = iscsilun->iscsi;
 299
 300     iscsi_service(iscsi, POLLOUT);
 301     iscsi_set_events(iscsilun);
 302 }
 303
 304 static int64_t sector_lun2qemu(int64_t sector, IscsiLun *iscsilun)
 305 {
 306     return sector * iscsilun->block_size / BDRV_SECTOR_SIZE;
 307 }
 308
 309 static int64_t sector_qemu2lun(int64_t sector, IscsiLun *iscsilun)
 310 {
 311     return sector * BDRV_SECTOR_SIZE / iscsilun->block_size;
 312 }
 313
 314 static bool is_request_lun_aligned(int64_t sector_num, int nb_sectors,
 315                                       IscsiLun *iscsilun)
 316 {
 317     if ((sector_num * BDRV_SECTOR_SIZE) % iscsilun->block_size ||
 318         (nb_sectors * BDRV_SECTOR_SIZE) % iscsilun->block_size) {
 319             error_report("iSCSI misaligned request: "
 320                          "iscsilun->block_size %u, sector_num %" PRIi64
 321                          ", nb_sectors %d",
 322                          iscsilun->block_size, sector_num, nb_sectors);
 323             return 0;
 324     }
 325     return 1;
 326 }
 327
 328 static void iscsi_allocationmap_set(IscsiLun *iscsilun, int64_t sector_num,
 329                                     int nb_sectors)
 330 {
 331     if (iscsilun->allocationmap == NULL) {
 332         return;
 333     }
 334     bitmap_set(iscsilun->allocationmap,
 335                sector_num / iscsilun->cluster_sectors,
 336                DIV_ROUND_UP(nb_sectors, iscsilun->cluster_sectors));
 337 }
 338
 339 static void iscsi_allocationmap_clear(IscsiLun *iscsilun, int64_t sector_num,
 340                                       int nb_sectors)
 341 {
 342     int64_t cluster_num, nb_clusters;
 343     if (iscsilun->allocationmap == NULL) {
 344         return;
 345     }
 346     cluster_num = DIV_ROUND_UP(sector_num, iscsilun->cluster_sectors);
 347     nb_clusters = (sector_num + nb_sectors) / iscsilun->cluster_sectors
 348                   - cluster_num;
 349     if (nb_clusters > 0) {
 350         bitmap_clear(iscsilun->allocationmap, cluster_num, nb_clusters);
 351     }
 352 }
 353
 354 static int coroutine_fn iscsi_co_writev(BlockDriverState *bs,
 355                                         int64_t sector_num, int nb_sectors,
 356                                         QEMUIOVector *iov)
 357 {
 358     IscsiLun *iscsilun = bs->opaque;
 359     struct IscsiTask iTask;
 360     uint64_t lba;
 361     uint32_t num_sectors;
 362
 363     if (!is_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
 364         return -EINVAL;
 365     }
 366
 367     lba = sector_qemu2lun(sector_num, iscsilun);
 368     num_sectors = sector_qemu2lun(nb_sectors, iscsilun);
 369     iscsi_co_init_iscsitask(iscsilun, &iTask);
 370 retry:
 371     if (iscsilun->use_16_for_rw) {
 372         iTask.task = iscsi_write16_task(iscsilun->iscsi, iscsilun->lun, lba,
 373                                         NULL, num_sectors * iscsilun->block_size,
 374                                         iscsilun->block_size, 0, 0, 0, 0, 0,
 375                                         iscsi_co_generic_cb, &iTask);
 376     } else {
 377         iTask.task = iscsi_write10_task(iscsilun->iscsi, iscsilun->lun, lba,
 378                                         NULL, num_sectors * iscsilun->block_size,
 379                                         iscsilun->block_size, 0, 0, 0, 0, 0,
 380                                         iscsi_co_generic_cb, &iTask);
 381     }
 382     if (iTask.task == NULL) {
 383         return -ENOMEM;
 384     }
 385     scsi_task_set_iov_out(iTask.task, (struct scsi_iovec *) iov->iov,
 386                           iov->niov);
 387     while (!iTask.complete) {
 388         iscsi_set_events(iscsilun);
 389         qemu_coroutine_yield();
 390     }
 391
 392     if (iTask.task != NULL) {
 393         scsi_free_scsi_task(iTask.task);
 394         iTask.task = NULL;
 395     }
 396
 397     if (iTask.do_retry) {
 398         iTask.complete = 0;
 399         goto retry;
 400     }
 401
 402     if (iTask.status != SCSI_STATUS_GOOD) {
 403         return -EIO;
 404     }
 405
 406     iscsi_allocationmap_set(iscsilun, sector_num, nb_sectors);
 407
 408     return 0;
 409 }
 410
 411
 412 static bool iscsi_allocationmap_is_allocated(IscsiLun *iscsilun,
 413                                              int64_t sector_num, int nb_sectors)
 414 {
 415     unsigned long size;
 416     if (iscsilun->allocationmap == NULL) {
 417         return true;
 418     }
 419     size = DIV_ROUND_UP(sector_num + nb_sectors, iscsilun->cluster_sectors);
 420     return !(find_next_bit(iscsilun->allocationmap, size,
 421                            sector_num / iscsilun->cluster_sectors) == size);
 422 }
 423
 424 static int64_t coroutine_fn iscsi_co_get_block_status(BlockDriverState *bs,
 425                                                   int64_t sector_num,
 426                                                   int nb_sectors, int *pnum)
 427 {
 428     IscsiLun *iscsilun = bs->opaque;
 429     struct scsi_get_lba_status *lbas = NULL;
 430     struct scsi_lba_status_descriptor *lbasd = NULL;
 431     struct IscsiTask iTask;
 432     int64_t ret;
 433
 434     iscsi_co_init_iscsitask(iscsilun, &iTask);
 435
 436     if (!is_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
 437         ret = -EINVAL;
 438         goto out;
 439     }
 440
 441     /* default to all sectors allocated */
 442     ret = BDRV_BLOCK_DATA;
 443     ret |= (sector_num << BDRV_SECTOR_BITS) | BDRV_BLOCK_OFFSET_VALID;
 444     *pnum = nb_sectors;
 445
 446     /* LUN does not support logical block provisioning */
 447     if (iscsilun->lbpme == 0) {
 448         goto out;
 449     }
 450
 451 retry:
 452     if (iscsi_get_lba_status_task(iscsilun->iscsi, iscsilun->lun,
 453                                   sector_qemu2lun(sector_num, iscsilun),
 454                                   8 + 16, iscsi_co_generic_cb,
 455                                   &iTask) == NULL) {
 456         ret = -ENOMEM;
 457         goto out;
 458     }
 459
 460     while (!iTask.complete) {
 461         iscsi_set_events(iscsilun);
 462         qemu_coroutine_yield();
 463     }
 464
 465     if (iTask.do_retry) {
 466         if (iTask.task != NULL) {
 467             scsi_free_scsi_task(iTask.task);
 468             iTask.task = NULL;
 469         }
 470         iTask.complete = 0;
 471         goto retry;
 472     }
 473
 474     if (iTask.status != SCSI_STATUS_GOOD) {
 475         /* in case the get_lba_status_callout fails (i.e.
 476          * because the device is busy or the cmd is not
 477          * supported) we pretend all blocks are allocated
 478          * for backwards compatibility */
 479         goto out;
 480     }
 481
 482     lbas = scsi_datain_unmarshall(iTask.task);
 483     if (lbas == NULL) {
 484         ret = -EIO;
 485         goto out;
 486     }
 487
 488     lbasd = &lbas->descriptors[0];
 489
 490     if (sector_qemu2lun(sector_num, iscsilun) != lbasd->lba) {
 491         ret = -EIO;
 492         goto out;
 493     }
 494
 495     *pnum = sector_lun2qemu(lbasd->num_blocks, iscsilun);
 496
 497     if (lbasd->provisioning == SCSI_PROVISIONING_TYPE_DEALLOCATED ||
 498         lbasd->provisioning == SCSI_PROVISIONING_TYPE_ANCHORED) {
 499         ret &= ~BDRV_BLOCK_DATA;
 500         if (iscsilun->lbprz) {
 501             ret |= BDRV_BLOCK_ZERO;
 502         }
 503     }
 504
 505     if (ret & BDRV_BLOCK_ZERO) {
 506         iscsi_allocationmap_clear(iscsilun, sector_num, *pnum);
 507     } else {
 508         iscsi_allocationmap_set(iscsilun, sector_num, *pnum);
 509     }
 510
 511     if (*pnum > nb_sectors) {
 512         *pnum = nb_sectors;
 513     }
 514 out:
 515     if (iTask.task != NULL) {
 516         scsi_free_scsi_task(iTask.task);
 517     }
 518     return ret;
 519 }
 520
 521 static int coroutine_fn iscsi_co_readv(BlockDriverState *bs,
 522                                        int64_t sector_num, int nb_sectors,
 523                                        QEMUIOVector *iov)
 524 {
 525     IscsiLun *iscsilun = bs->opaque;
 526     struct IscsiTask iTask;
 527     uint64_t lba;
 528     uint32_t num_sectors;
 529
 530     if (!is_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
 531         return -EINVAL;
 532     }
 533
 534     if (iscsilun->lbprz && nb_sectors >= ISCSI_CHECKALLOC_THRES &&
 535         !iscsi_allocationmap_is_allocated(iscsilun, sector_num, nb_sectors)) {
 536         int64_t ret;
 537         int pnum;
 538         ret = iscsi_co_get_block_status(bs, sector_num, INT_MAX, &pnum);
 539         if (ret < 0) {
 540             return ret;
 541         }
 542         if (ret & BDRV_BLOCK_ZERO && pnum >= nb_sectors) {
 543             qemu_iovec_memset(iov, 0, 0x00, iov->size);
 544             return 0;
 545         }
 546     }
 547
 548     lba = sector_qemu2lun(sector_num, iscsilun);
 549     num_sectors = sector_qemu2lun(nb_sectors, iscsilun);
 550
 551     iscsi_co_init_iscsitask(iscsilun, &iTask);
 552 retry:
 553     if (iscsilun->use_16_for_rw) {
 554         iTask.task = iscsi_read16_task(iscsilun->iscsi, iscsilun->lun, lba,
 555                                        num_sectors * iscsilun->block_size,
 556                                        iscsilun->block_size, 0, 0, 0, 0, 0,
 557                                        iscsi_co_generic_cb, &iTask);
 558     } else {
 559         iTask.task = iscsi_read10_task(iscsilun->iscsi, iscsilun->lun, lba,
 560                                        num_sectors * iscsilun->block_size,
 561                                        iscsilun->block_size,
 562                                        0, 0, 0, 0, 0,
 563                                        iscsi_co_generic_cb, &iTask);
 564     }
 565     if (iTask.task == NULL) {
 566         return -ENOMEM;
 567     }
 568     scsi_task_set_iov_in(iTask.task, (struct scsi_iovec *) iov->iov, iov->niov);
 569
 570     while (!iTask.complete) {
 571         iscsi_set_events(iscsilun);
 572         qemu_coroutine_yield();
 573     }
 574
 575     if (iTask.task != NULL) {
 576         scsi_free_scsi_task(iTask.task);
 577         iTask.task = NULL;
 578     }
 579
 580     if (iTask.do_retry) {
 581         iTask.complete = 0;
 582         goto retry;
 583     }
 584
 585     if (iTask.status != SCSI_STATUS_GOOD) {
 586         return -EIO;
 587     }
 588
 589     return 0;
 590 }
 591
 592 static int coroutine_fn iscsi_co_flush(BlockDriverState *bs)
 593 {
 594     IscsiLun *iscsilun = bs->opaque;
 595     struct IscsiTask iTask;
 596
 597     if (bs->sg) {
 598         return 0;
 599     }
 600
 601     iscsi_co_init_iscsitask(iscsilun, &iTask);
 602
 603 retry:
 604     if (iscsi_synchronizecache10_task(iscsilun->iscsi, iscsilun->lun, 0, 0, 0,
 605                                       0, iscsi_co_generic_cb, &iTask) == NULL) {
 606         return -ENOMEM;
 607     }
 608
 609     while (!iTask.complete) {
 610         iscsi_set_events(iscsilun);
 611         qemu_coroutine_yield();
 612     }
 613
 614     if (iTask.task != NULL) {
 615         scsi_free_scsi_task(iTask.task);
 616         iTask.task = NULL;
 617     }
 618
 619     if (iTask.do_retry) {
 620         iTask.complete = 0;
 621         goto retry;
 622     }
 623
 624     if (iTask.status != SCSI_STATUS_GOOD) {
 625         return -EIO;
 626     }
 627
 628     return 0;
 629 }
 630
 631 #ifdef __linux__
 632 static void
 633 iscsi_aio_ioctl_cb(struct iscsi_context *iscsi, int status,
 634                      void *command_data, void *opaque)
 635 {
 636     IscsiAIOCB *acb = opaque;
 637
 638     g_free(acb->buf);
 639     acb->buf = NULL;
 640
 641     if (acb->canceled != 0) {
 642         return;
 643     }
 644
 645     acb->status = 0;
 646     if (status < 0) {
 647         error_report("Failed to ioctl(SG_IO) to iSCSI lun. %s",
 648                      iscsi_get_error(iscsi));
 649         acb->status = -EIO;
 650     }
 651
 652     acb->ioh->driver_status = 0;
 653     acb->ioh->host_status   = 0;
 654     acb->ioh->resid         = 0;
 655
 656 #define SG_ERR_DRIVER_SENSE    0x08
 657
 658     if (status == SCSI_STATUS_CHECK_CONDITION && acb->task->datain.size >= 2) {
 659         int ss;
 660
 661         acb->ioh->driver_status |= SG_ERR_DRIVER_SENSE;
 662
 663         acb->ioh->sb_len_wr = acb->task->datain.size - 2;
 664         ss = (acb->ioh->mx_sb_len >= acb->ioh->sb_len_wr) ?
 665              acb->ioh->mx_sb_len : acb->ioh->sb_len_wr;
 666         memcpy(acb->ioh->sbp, &acb->task->datain.data[2], ss);
 667     }
 668
 669     iscsi_schedule_bh(acb);
 670 }
 671
 672 static BlockDriverAIOCB *iscsi_aio_ioctl(BlockDriverState *bs,
 673         unsigned long int req, void *buf,
 674         BlockDriverCompletionFunc *cb, void *opaque)
 675 {
 676     IscsiLun *iscsilun = bs->opaque;
 677     struct iscsi_context *iscsi = iscsilun->iscsi;
 678     struct iscsi_data data;
 679     IscsiAIOCB *acb;
 680
 681     assert(req == SG_IO);
 682
 683     acb = qemu_aio_get(&iscsi_aiocb_info, bs, cb, opaque);
 684
 685     acb->iscsilun = iscsilun;
 686     acb->canceled    = 0;
 687     acb->bh          = NULL;
 688     acb->status      = -EINPROGRESS;
 689     acb->buf         = NULL;
 690     acb->ioh         = buf;
 691
 692     acb->task = malloc(sizeof(struct scsi_task));
 693     if (acb->task == NULL) {
 694         error_report("iSCSI: Failed to allocate task for scsi command. %s",
 695                      iscsi_get_error(iscsi));
 696         qemu_aio_release(acb);
 697         return NULL;
 698     }
 699     memset(acb->task, 0, sizeof(struct scsi_task));
 700
 701     switch (acb->ioh->dxfer_direction) {
 702     case SG_DXFER_TO_DEV:
 703         acb->task->xfer_dir = SCSI_XFER_WRITE;
 704         break;
 705     case SG_DXFER_FROM_DEV:
 706         acb->task->xfer_dir = SCSI_XFER_READ;
 707         break;
 708     default:
 709         acb->task->xfer_dir = SCSI_XFER_NONE;
 710         break;
 711     }
 712
 713     acb->task->cdb_size = acb->ioh->cmd_len;
 714     memcpy(&acb->task->cdb[0], acb->ioh->cmdp, acb->ioh->cmd_len);
 715     acb->task->expxferlen = acb->ioh->dxfer_len;
 716
 717     data.size = 0;
 718     if (acb->task->xfer_dir == SCSI_XFER_WRITE) {
 719         if (acb->ioh->iovec_count == 0) {
 720             data.data = acb->ioh->dxferp;
 721             data.size = acb->ioh->dxfer_len;
 722         } else {
 723             scsi_task_set_iov_out(acb->task,
 724                                  (struct scsi_iovec *) acb->ioh->dxferp,
 725                                  acb->ioh->iovec_count);
 726         }
 727     }
 728
 729     if (iscsi_scsi_command_async(iscsi, iscsilun->lun, acb->task,
 730                                  iscsi_aio_ioctl_cb,
 731                                  (data.size > 0) ? &data : NULL,
 732                                  acb) != 0) {
 733         scsi_free_scsi_task(acb->task);
 734         qemu_aio_release(acb);
 735         return NULL;
 736     }
 737
 738     /* tell libiscsi to read straight into the buffer we got from ioctl */
 739     if (acb->task->xfer_dir == SCSI_XFER_READ) {
 740         if (acb->ioh->iovec_count == 0) {
 741             scsi_task_add_data_in_buffer(acb->task,
 742                                          acb->ioh->dxfer_len,
 743                                          acb->ioh->dxferp);
 744         } else {
 745             scsi_task_set_iov_in(acb->task,
 746                                  (struct scsi_iovec *) acb->ioh->dxferp,
 747                                  acb->ioh->iovec_count);
 748         }
 749     }
 750
 751     iscsi_set_events(iscsilun);
 752
 753     return &acb->common;
 754 }
 755
 756 static void ioctl_cb(void *opaque, int status)
 757 {
 758     int *p_status = opaque;
 759     *p_status = status;
 760 }
 761
 762 static int iscsi_ioctl(BlockDriverState *bs, unsigned long int req, void *buf)
 763 {
 764     IscsiLun *iscsilun = bs->opaque;
 765     int status;
 766
 767     switch (req) {
 768     case SG_GET_VERSION_NUM:
 769         *(int *)buf = 30000;
 770         break;
 771     case SG_GET_SCSI_ID:
 772         ((struct sg_scsi_id *)buf)->scsi_type = iscsilun->type;
 773         break;
 774     case SG_IO:
 775         status = -EINPROGRESS;
 776         iscsi_aio_ioctl(bs, req, buf, ioctl_cb, &status);
 777
 778         while (status == -EINPROGRESS) {
 779             aio_poll(iscsilun->aio_context, true);
 780         }
 781
 782         return 0;
 783     default:
 784         return -1;
 785     }
 786     return 0;
 787 }
 788 #endif
 789
 790 static int64_t
 791 iscsi_getlength(BlockDriverState *bs)
 792 {
 793     IscsiLun *iscsilun = bs->opaque;
 794     int64_t len;
 795
 796     len  = iscsilun->num_blocks;
 797     len *= iscsilun->block_size;
 798
 799     return len;
 800 }
 801
 802 static int
 803 coroutine_fn iscsi_co_discard(BlockDriverState *bs, int64_t sector_num,
 804                                    int nb_sectors)
 805 {
 806     IscsiLun *iscsilun = bs->opaque;
 807     struct IscsiTask iTask;
 808     struct unmap_list list;
 809
 810     if (!is_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
 811         return -EINVAL;
 812     }
 813
 814     if (!iscsilun->lbp.lbpu) {
 815         /* UNMAP is not supported by the target */
 816         return 0;
 817     }
 818
 819     list.lba = sector_qemu2lun(sector_num, iscsilun);
 820     list.num = sector_qemu2lun(nb_sectors, iscsilun);
 821
 822     iscsi_co_init_iscsitask(iscsilun, &iTask);
 823 retry:
 824     if (iscsi_unmap_task(iscsilun->iscsi, iscsilun->lun, 0, 0, &list, 1,
 825                      iscsi_co_generic_cb, &iTask) == NULL) {
 826         return -ENOMEM;
 827     }
 828
 829     while (!iTask.complete) {
 830         iscsi_set_events(iscsilun);
 831         qemu_coroutine_yield();
 832     }
 833
 834     if (iTask.task != NULL) {
 835         scsi_free_scsi_task(iTask.task);
 836         iTask.task = NULL;
 837     }
 838
 839     if (iTask.do_retry) {
 840         iTask.complete = 0;
 841         goto retry;
 842     }
 843
 844     if (iTask.status == SCSI_STATUS_CHECK_CONDITION) {
 845         /* the target might fail with a check condition if it
 846            is not happy with the alignment of the UNMAP request
 847            we silently fail in this case */
 848         return 0;
 849     }
 850
 851     if (iTask.status != SCSI_STATUS_GOOD) {
 852         return -EIO;
 853     }
 854
 855     iscsi_allocationmap_clear(iscsilun, sector_num, nb_sectors);
 856
 857     return 0;
 858 }
 859
 860 static int
 861 coroutine_fn iscsi_co_write_zeroes(BlockDriverState *bs, int64_t sector_num,
 862                                    int nb_sectors, BdrvRequestFlags flags)
 863 {
 864     IscsiLun *iscsilun = bs->opaque;
 865     struct IscsiTask iTask;
 866     uint64_t lba;
 867     uint32_t nb_blocks;
 868     bool use_16_for_ws = iscsilun->use_16_for_rw;
 869
 870     if (!is_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
 871         return -EINVAL;
 872     }
 873
 874     if (flags & BDRV_REQ_MAY_UNMAP) {
 875         if (!use_16_for_ws && !iscsilun->lbp.lbpws10) {
 876             /* WRITESAME10 with UNMAP is unsupported try WRITESAME16 */
 877             use_16_for_ws = true;
 878         }
 879         if (use_16_for_ws && !iscsilun->lbp.lbpws) {
 880             /* WRITESAME16 with UNMAP is not supported by the target,
 881              * fall back and try WRITESAME10/16 without UNMAP */
 882             flags &= ~BDRV_REQ_MAY_UNMAP;
 883             use_16_for_ws = iscsilun->use_16_for_rw;
 884         }
 885     }
 886
 887     if (!(flags & BDRV_REQ_MAY_UNMAP) && !iscsilun->has_write_same) {
 888         /* WRITESAME without UNMAP is not supported by the target */
 889         return -ENOTSUP;
 890     }
 891
 892     lba = sector_qemu2lun(sector_num, iscsilun);
 893     nb_blocks = sector_qemu2lun(nb_sectors, iscsilun);
 894
 895     if (iscsilun->zeroblock == NULL) {
 896         iscsilun->zeroblock = g_try_malloc0(iscsilun->block_size);
 897         if (iscsilun->zeroblock == NULL) {
 898             return -ENOMEM;
 899         }
 900     }
 901
 902     iscsi_co_init_iscsitask(iscsilun, &iTask);
 903 retry:
 904     if (use_16_for_ws) {
 905         iTask.task = iscsi_writesame16_task(iscsilun->iscsi, iscsilun->lun, lba,
 906                                             iscsilun->zeroblock, iscsilun->block_size,
 907                                             nb_blocks, 0, !!(flags & BDRV_REQ_MAY_UNMAP),
 908                                             0, 0, iscsi_co_generic_cb, &iTask);
 909     } else {
 910         iTask.task = iscsi_writesame10_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);
 914     }
 915     if (iTask.task == NULL) {
 916         return -ENOMEM;
 917     }
 918
 919     while (!iTask.complete) {
 920         iscsi_set_events(iscsilun);
 921         qemu_coroutine_yield();
 922     }
 923
 924     if (iTask.status == SCSI_STATUS_CHECK_CONDITION &&
 925         iTask.task->sense.key == SCSI_SENSE_ILLEGAL_REQUEST &&
 926         (iTask.task->sense.ascq == SCSI_SENSE_ASCQ_INVALID_OPERATION_CODE ||
 927          iTask.task->sense.ascq == SCSI_SENSE_ASCQ_INVALID_FIELD_IN_CDB)) {
 928         /* WRITE SAME is not supported by the target */
 929         iscsilun->has_write_same = false;
 930         scsi_free_scsi_task(iTask.task);
 931         return -ENOTSUP;
 932     }
 933
 934     if (iTask.task != NULL) {
 935         scsi_free_scsi_task(iTask.task);
 936         iTask.task = NULL;
 937     }
 938
 939     if (iTask.do_retry) {
 940         iTask.complete = 0;
 941         goto retry;
 942     }
 943
 944     if (iTask.status != SCSI_STATUS_GOOD) {
 945         return -EIO;
 946     }
 947
 948     if (flags & BDRV_REQ_MAY_UNMAP) {
 949         iscsi_allocationmap_clear(iscsilun, sector_num, nb_sectors);
 950     } else {
 951         iscsi_allocationmap_set(iscsilun, sector_num, nb_sectors);
 952     }
 953
 954     return 0;
 955 }
 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 static void iscsi_nop_timed_event(void *opaque)
1067 {
1068     IscsiLun *iscsilun = opaque;
1069
1070     if (iscsi_get_nops_in_flight(iscsilun->iscsi) > MAX_NOP_FAILURES) {
1071         error_report("iSCSI: NOP timeout. Reconnecting...");
1072         iscsi_reconnect(iscsilun->iscsi);
1073     }
1074
1075     if (iscsi_nop_out_async(iscsilun->iscsi, NULL, NULL, 0, NULL) != 0) {
1076         error_report("iSCSI: failed to sent NOP-Out. Disabling NOP messages.");
1077         return;
1078     }
1079
1080     timer_mod(iscsilun->nop_timer, qemu_clock_get_ms(QEMU_CLOCK_REALTIME) + NOP_INTERVAL);
1081     iscsi_set_events(iscsilun);
1082 }
1083
1084 static void iscsi_readcapacity_sync(IscsiLun *iscsilun, Error **errp)
1085 {
1086     struct scsi_task *task = NULL;
1087     struct scsi_readcapacity10 *rc10 = NULL;
1088     struct scsi_readcapacity16 *rc16 = NULL;
1089     int retries = ISCSI_CMD_RETRIES;
1090
1091     do {
1092         if (task != NULL) {
1093             scsi_free_scsi_task(task);
1094             task = NULL;
1095         }
1096
1097         switch (iscsilun->type) {
1098         case TYPE_DISK:
1099             task = iscsi_readcapacity16_sync(iscsilun->iscsi, iscsilun->lun);
1100             if (task != NULL && task->status == SCSI_STATUS_GOOD) {
1101                 rc16 = scsi_datain_unmarshall(task);
1102                 if (rc16 == NULL) {
1103                     error_setg(errp, "iSCSI: Failed to unmarshall readcapacity16 data.");
1104                 } else {
1105                     iscsilun->block_size = rc16->block_length;
1106                     iscsilun->num_blocks = rc16->returned_lba + 1;
1107                     iscsilun->lbpme = rc16->lbpme;
1108                     iscsilun->lbprz = rc16->lbprz;
1109                     iscsilun->use_16_for_rw = (rc16->returned_lba > 0xffffffff);
1110                 }
1111             }
1112             break;
1113         case TYPE_ROM:
1114             task = iscsi_readcapacity10_sync(iscsilun->iscsi, iscsilun->lun, 0, 0);
1115             if (task != NULL && task->status == SCSI_STATUS_GOOD) {
1116                 rc10 = scsi_datain_unmarshall(task);
1117                 if (rc10 == NULL) {
1118                     error_setg(errp, "iSCSI: Failed to unmarshall readcapacity10 data.");
1119                 } else {
1120                     iscsilun->block_size = rc10->block_size;
1121                     if (rc10->lba == 0) {
1122                         /* blank disk loaded */
1123                         iscsilun->num_blocks = 0;
1124                     } else {
1125                         iscsilun->num_blocks = rc10->lba + 1;
1126                     }
1127                 }
1128             }
1129             break;
1130         default:
1131             return;
1132         }
1133     } while (task != NULL && task->status == SCSI_STATUS_CHECK_CONDITION
1134              && task->sense.key == SCSI_SENSE_UNIT_ATTENTION
1135              && retries-- > 0);
1136
1137     if (task == NULL || task->status != SCSI_STATUS_GOOD) {
1138         error_setg(errp, "iSCSI: failed to send readcapacity10 command.");
1139     }
1140     if (task) {
1141         scsi_free_scsi_task(task);
1142     }
1143 }
1144
1145 /* TODO Convert to fine grained options */
1146 static QemuOptsList runtime_opts = {
1147     .name = "iscsi",
1148     .head = QTAILQ_HEAD_INITIALIZER(runtime_opts.head),
1149     .desc = {
1150         {
1151             .name = "filename",
1152             .type = QEMU_OPT_STRING,
1153             .help = "URL to the iscsi image",
1154         },
1155         { /* end of list */ }
1156     },
1157 };
1158
1159 static struct scsi_task *iscsi_do_inquiry(struct iscsi_context *iscsi, int lun,
1160                                           int evpd, int pc, void **inq, Error **errp)
1161 {
1162     int full_size;
1163     struct scsi_task *task = NULL;
1164     task = iscsi_inquiry_sync(iscsi, lun, evpd, pc, 64);
1165     if (task == NULL || task->status != SCSI_STATUS_GOOD) {
1166         goto fail;
1167     }
1168     full_size = scsi_datain_getfullsize(task);
1169     if (full_size > task->datain.size) {
1170         scsi_free_scsi_task(task);
1171
1172         /* we need more data for the full list */
1173         task = iscsi_inquiry_sync(iscsi, lun, evpd, pc, full_size);
1174         if (task == NULL || task->status != SCSI_STATUS_GOOD) {
1175             goto fail;
1176         }
1177     }
1178
1179     *inq = scsi_datain_unmarshall(task);
1180     if (*inq == NULL) {
1181         error_setg(errp, "iSCSI: failed to unmarshall inquiry datain blob");
1182         goto fail_with_err;
1183     }
1184
1185     return task;
1186
1187 fail:
1188     error_setg(errp, "iSCSI: Inquiry command failed : %s",
1189                iscsi_get_error(iscsi));
1190 fail_with_err:
1191     if (task != NULL) {
1192         scsi_free_scsi_task(task);
1193     }
1194     return NULL;
1195 }
1196
1197 static void iscsi_detach_aio_context(BlockDriverState *bs)
1198 {
1199     IscsiLun *iscsilun = bs->opaque;
1200
1201     aio_set_fd_handler(iscsilun->aio_context,
1202                        iscsi_get_fd(iscsilun->iscsi),
1203                        NULL, NULL, NULL);
1204     iscsilun->events = 0;
1205
1206     if (iscsilun->nop_timer) {
1207         timer_del(iscsilun->nop_timer);
1208         timer_free(iscsilun->nop_timer);
1209         iscsilun->nop_timer = NULL;
1210     }
1211 }
1212
1213 static void iscsi_attach_aio_context(BlockDriverState *bs,
1214                                      AioContext *new_context)
1215 {
1216     IscsiLun *iscsilun = bs->opaque;
1217
1218     iscsilun->aio_context = new_context;
1219     iscsi_set_events(iscsilun);
1220
1221     /* Set up a timer for sending out iSCSI NOPs */
1222     iscsilun->nop_timer = aio_timer_new(iscsilun->aio_context,
1223                                         QEMU_CLOCK_REALTIME, SCALE_MS,
1224                                         iscsi_nop_timed_event, iscsilun);
1225     timer_mod(iscsilun->nop_timer,
1226               qemu_clock_get_ms(QEMU_CLOCK_REALTIME) + NOP_INTERVAL);
1227 }
1228
1229 /*
1230  * We support iscsi url's on the form
1231  * iscsi://[<username>%<password>@]<host>[:<port>]/<targetname>/<lun>
1232  *
1233  * Note: flags are currently not used by iscsi_open.  If this function
1234  * is changed such that flags are used, please examine iscsi_reopen_prepare()
1235  * to see if needs to be changed as well.
1236  */
1237 static int iscsi_open(BlockDriverState *bs, QDict *options, int flags,
1238                       Error **errp)
1239 {
1240     IscsiLun *iscsilun = bs->opaque;
1241     struct iscsi_context *iscsi = NULL;
1242     struct iscsi_url *iscsi_url = NULL;
1243     struct scsi_task *task = NULL;
1244     struct scsi_inquiry_standard *inq = NULL;
1245     struct scsi_inquiry_supported_pages *inq_vpd;
1246     char *initiator_name = NULL;
1247     QemuOpts *opts;
1248     Error *local_err = NULL;
1249     const char *filename;
1250     int i, ret;
1251
1252     if ((BDRV_SECTOR_SIZE % 512) != 0) {
1253         error_setg(errp, "iSCSI: Invalid BDRV_SECTOR_SIZE. "
1254                    "BDRV_SECTOR_SIZE(%lld) is not a multiple "
1255                    "of 512", BDRV_SECTOR_SIZE);
1256         return -EINVAL;
1257     }
1258
1259     opts = qemu_opts_create(&runtime_opts, NULL, 0, &error_abort);
1260     qemu_opts_absorb_qdict(opts, options, &local_err);
1261     if (local_err) {
1262         error_propagate(errp, local_err);
1263         ret = -EINVAL;
1264         goto out;
1265     }
1266
1267     filename = qemu_opt_get(opts, "filename");
1268
1269     iscsi_url = iscsi_parse_full_url(iscsi, filename);
1270     if (iscsi_url == NULL) {
1271         error_setg(errp, "Failed to parse URL : %s", filename);
1272         ret = -EINVAL;
1273         goto out;
1274     }
1275
1276     memset(iscsilun, 0, sizeof(IscsiLun));
1277
1278     initiator_name = parse_initiator_name(iscsi_url->target);
1279
1280     iscsi = iscsi_create_context(initiator_name);
1281     if (iscsi == NULL) {
1282         error_setg(errp, "iSCSI: Failed to create iSCSI context.");
1283         ret = -ENOMEM;
1284         goto out;
1285     }
1286
1287     if (iscsi_set_targetname(iscsi, iscsi_url->target)) {
1288         error_setg(errp, "iSCSI: Failed to set target name.");
1289         ret = -EINVAL;
1290         goto out;
1291     }
1292
1293     if (iscsi_url->user != NULL) {
1294         ret = iscsi_set_initiator_username_pwd(iscsi, iscsi_url->user,
1295                                               iscsi_url->passwd);
1296         if (ret != 0) {
1297             error_setg(errp, "Failed to set initiator username and password");
1298             ret = -EINVAL;
1299             goto out;
1300         }
1301     }
1302
1303     /* check if we got CHAP username/password via the options */
1304     parse_chap(iscsi, iscsi_url->target, &local_err);
1305     if (local_err != NULL) {
1306         error_propagate(errp, local_err);
1307         ret = -EINVAL;
1308         goto out;
1309     }
1310
1311     if (iscsi_set_session_type(iscsi, ISCSI_SESSION_NORMAL) != 0) {
1312         error_setg(errp, "iSCSI: Failed to set session type to normal.");
1313         ret = -EINVAL;
1314         goto out;
1315     }
1316
1317     iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_NONE_CRC32C);
1318
1319     /* check if we got HEADER_DIGEST via the options */
1320     parse_header_digest(iscsi, iscsi_url->target, &local_err);
1321     if (local_err != NULL) {
1322         error_propagate(errp, local_err);
1323         ret = -EINVAL;
1324         goto out;
1325     }
1326
1327     if (iscsi_full_connect_sync(iscsi, iscsi_url->portal, iscsi_url->lun) != 0) {
1328         error_setg(errp, "iSCSI: Failed to connect to LUN : %s",
1329             iscsi_get_error(iscsi));
1330         ret = -EINVAL;
1331         goto out;
1332     }
1333
1334     iscsilun->iscsi = iscsi;
1335     iscsilun->aio_context = bdrv_get_aio_context(bs);
1336     iscsilun->lun   = iscsi_url->lun;
1337     iscsilun->has_write_same = true;
1338
1339     task = iscsi_do_inquiry(iscsilun->iscsi, iscsilun->lun, 0, 0,
1340                             (void **) &inq, errp);
1341     if (task == NULL) {
1342         ret = -EINVAL;
1343         goto out;
1344     }
1345     iscsilun->type = inq->periperal_device_type;
1346     scsi_free_scsi_task(task);
1347     task = NULL;
1348
1349     iscsi_readcapacity_sync(iscsilun, &local_err);
1350     if (local_err != NULL) {
1351         error_propagate(errp, local_err);
1352         ret = -EINVAL;
1353         goto out;
1354     }
1355     bs->total_sectors = sector_lun2qemu(iscsilun->num_blocks, iscsilun);
1356     bs->request_alignment = iscsilun->block_size;
1357
1358     /* We don't have any emulation for devices other than disks and CD-ROMs, so
1359      * this must be sg ioctl compatible. We force it to be sg, otherwise qemu
1360      * will try to read from the device to guess the image format.
1361      */
1362     if (iscsilun->type != TYPE_DISK && iscsilun->type != TYPE_ROM) {
1363         bs->sg = 1;
1364     }
1365
1366     task = iscsi_do_inquiry(iscsilun->iscsi, iscsilun->lun, 1,
1367                             SCSI_INQUIRY_PAGECODE_SUPPORTED_VPD_PAGES,
1368                             (void **) &inq_vpd, errp);
1369     if (task == NULL) {
1370         ret = -EINVAL;
1371         goto out;
1372     }
1373     for (i = 0; i < inq_vpd->num_pages; i++) {
1374         struct scsi_task *inq_task;
1375         struct scsi_inquiry_logical_block_provisioning *inq_lbp;
1376         struct scsi_inquiry_block_limits *inq_bl;
1377         switch (inq_vpd->pages[i]) {
1378         case SCSI_INQUIRY_PAGECODE_LOGICAL_BLOCK_PROVISIONING:
1379             inq_task = iscsi_do_inquiry(iscsilun->iscsi, iscsilun->lun, 1,
1380                                         SCSI_INQUIRY_PAGECODE_LOGICAL_BLOCK_PROVISIONING,
1381                                         (void **) &inq_lbp, errp);
1382             if (inq_task == NULL) {
1383                 ret = -EINVAL;
1384                 goto out;
1385             }
1386             memcpy(&iscsilun->lbp, inq_lbp,
1387                    sizeof(struct scsi_inquiry_logical_block_provisioning));
1388             scsi_free_scsi_task(inq_task);
1389             break;
1390         case SCSI_INQUIRY_PAGECODE_BLOCK_LIMITS:
1391             inq_task = iscsi_do_inquiry(iscsilun->iscsi, iscsilun->lun, 1,
1392                                     SCSI_INQUIRY_PAGECODE_BLOCK_LIMITS,
1393                                     (void **) &inq_bl, errp);
1394             if (inq_task == NULL) {
1395                 ret = -EINVAL;
1396                 goto out;
1397             }
1398             memcpy(&iscsilun->bl, inq_bl,
1399                    sizeof(struct scsi_inquiry_block_limits));
1400             scsi_free_scsi_task(inq_task);
1401             break;
1402         default:
1403             break;
1404         }
1405     }
1406     scsi_free_scsi_task(task);
1407     task = NULL;
1408
1409     iscsi_attach_aio_context(bs, iscsilun->aio_context);
1410
1411     /* Guess the internal cluster (page) size of the iscsi target by the means
1412      * of opt_unmap_gran. Transfer the unmap granularity only if it has a
1413      * reasonable size */
1414     if (iscsilun->bl.opt_unmap_gran * iscsilun->block_size >= 4 * 1024 &&
1415         iscsilun->bl.opt_unmap_gran * iscsilun->block_size <= 16 * 1024 * 1024) {
1416         iscsilun->cluster_sectors = (iscsilun->bl.opt_unmap_gran *
1417                                      iscsilun->block_size) >> BDRV_SECTOR_BITS;
1418         if (iscsilun->lbprz && !(bs->open_flags & BDRV_O_NOCACHE)) {
1419             iscsilun->allocationmap =
1420                 bitmap_new(DIV_ROUND_UP(bs->total_sectors,
1421                                         iscsilun->cluster_sectors));
1422         }
1423     }
1424
1425 out:
1426     qemu_opts_del(opts);
1427     g_free(initiator_name);
1428     if (iscsi_url != NULL) {
1429         iscsi_destroy_url(iscsi_url);
1430     }
1431     if (task != NULL) {
1432         scsi_free_scsi_task(task);
1433     }
1434
1435     if (ret) {
1436         if (iscsi != NULL) {
1437             iscsi_destroy_context(iscsi);
1438         }
1439         memset(iscsilun, 0, sizeof(IscsiLun));
1440     }
1441     return ret;
1442 }
1443
1444 static void iscsi_close(BlockDriverState *bs)
1445 {
1446     IscsiLun *iscsilun = bs->opaque;
1447     struct iscsi_context *iscsi = iscsilun->iscsi;
1448
1449     iscsi_detach_aio_context(bs);
1450     iscsi_destroy_context(iscsi);
1451     g_free(iscsilun->zeroblock);
1452     g_free(iscsilun->allocationmap);
1453     memset(iscsilun, 0, sizeof(IscsiLun));
1454 }
1455
1456 static void iscsi_refresh_limits(BlockDriverState *bs, Error **errp)
1457 {
1458     IscsiLun *iscsilun = bs->opaque;
1459
1460     /* We don't actually refresh here, but just return data queried in
1461      * iscsi_open(): iscsi targets don't change their limits. */
1462     if (iscsilun->lbp.lbpu) {
1463         if (iscsilun->bl.max_unmap < 0xffffffff) {
1464             bs->bl.max_discard = sector_lun2qemu(iscsilun->bl.max_unmap,
1465                                                  iscsilun);
1466         }
1467         bs->bl.discard_alignment = sector_lun2qemu(iscsilun->bl.opt_unmap_gran,
1468                                                    iscsilun);
1469     }
1470
1471     if (iscsilun->bl.max_ws_len < 0xffffffff) {
1472         bs->bl.max_write_zeroes = sector_lun2qemu(iscsilun->bl.max_ws_len,
1473                                                   iscsilun);
1474     }
1475     if (iscsilun->lbp.lbpws) {
1476         bs->bl.write_zeroes_alignment = sector_lun2qemu(iscsilun->bl.opt_unmap_gran,
1477                                                         iscsilun);
1478     }
1479     bs->bl.opt_transfer_length = sector_lun2qemu(iscsilun->bl.opt_xfer_len,
1480                                                  iscsilun);
1481 }
1482
1483 /* Since iscsi_open() ignores bdrv_flags, there is nothing to do here in
1484  * prepare.  Note that this will not re-establish a connection with an iSCSI
1485  * target - it is effectively a NOP.  */
1486 static int iscsi_reopen_prepare(BDRVReopenState *state,
1487                                 BlockReopenQueue *queue, Error **errp)
1488 {
1489     /* NOP */
1490     return 0;
1491 }
1492
1493 static int iscsi_truncate(BlockDriverState *bs, int64_t offset)
1494 {
1495     IscsiLun *iscsilun = bs->opaque;
1496     Error *local_err = NULL;
1497
1498     if (iscsilun->type != TYPE_DISK) {
1499         return -ENOTSUP;
1500     }
1501
1502     iscsi_readcapacity_sync(iscsilun, &local_err);
1503     if (local_err != NULL) {
1504         error_free(local_err);
1505         return -EIO;
1506     }
1507
1508     if (offset > iscsi_getlength(bs)) {
1509         return -EINVAL;
1510     }
1511
1512     if (iscsilun->allocationmap != NULL) {
1513         g_free(iscsilun->allocationmap);
1514         iscsilun->allocationmap =
1515             bitmap_new(DIV_ROUND_UP(sector_lun2qemu(iscsilun->num_blocks,
1516                                                     iscsilun),
1517                                     iscsilun->cluster_sectors));
1518     }
1519
1520     return 0;
1521 }
1522
1523 static int iscsi_create(const char *filename, QemuOpts *opts, Error **errp)
1524 {
1525     int ret = 0;
1526     int64_t total_size = 0;
1527     BlockDriverState *bs;
1528     IscsiLun *iscsilun = NULL;
1529     QDict *bs_options;
1530
1531     bs = bdrv_new("", &error_abort);
1532
1533     /* Read out options */
1534     total_size =
1535         qemu_opt_get_size_del(opts, BLOCK_OPT_SIZE, 0) / BDRV_SECTOR_SIZE;
1536     bs->opaque = g_new0(struct IscsiLun, 1);
1537     iscsilun = bs->opaque;
1538
1539     bs_options = qdict_new();
1540     qdict_put(bs_options, "filename", qstring_from_str(filename));
1541     ret = iscsi_open(bs, bs_options, 0, NULL);
1542     QDECREF(bs_options);
1543
1544     if (ret != 0) {
1545         goto out;
1546     }
1547     iscsi_detach_aio_context(bs);
1548     if (iscsilun->type != TYPE_DISK) {
1549         ret = -ENODEV;
1550         goto out;
1551     }
1552     if (bs->total_sectors < total_size) {
1553         ret = -ENOSPC;
1554         goto out;
1555     }
1556
1557     ret = 0;
1558 out:
1559     if (iscsilun->iscsi != NULL) {
1560         iscsi_destroy_context(iscsilun->iscsi);
1561     }
1562     g_free(bs->opaque);
1563     bs->opaque = NULL;
1564     bdrv_unref(bs);
1565     return ret;
1566 }
1567
1568 static int iscsi_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
1569 {
1570     IscsiLun *iscsilun = bs->opaque;
1571     bdi->unallocated_blocks_are_zero = !!iscsilun->lbprz;
1572     bdi->can_write_zeroes_with_unmap = iscsilun->lbprz && iscsilun->lbp.lbpws;
1573     bdi->cluster_size = iscsilun->cluster_sectors * BDRV_SECTOR_SIZE;
1574     return 0;
1575 }
1576
1577 static QemuOptsList iscsi_create_opts = {
1578     .name = "iscsi-create-opts",
1579     .head = QTAILQ_HEAD_INITIALIZER(iscsi_create_opts.head),
1580     .desc = {
1581         {
1582             .name = BLOCK_OPT_SIZE,
1583             .type = QEMU_OPT_SIZE,
1584             .help = "Virtual disk size"
1585         },
1586         { /* end of list */ }
1587     }
1588 };
1589
1590 static BlockDriver bdrv_iscsi = {
1591     .format_name     = "iscsi",
1592     .protocol_name   = "iscsi",
1593
1594     .instance_size   = sizeof(IscsiLun),
1595     .bdrv_needs_filename = true,
1596     .bdrv_file_open  = iscsi_open,
1597     .bdrv_close      = iscsi_close,
1598     .bdrv_create     = iscsi_create,
1599     .create_opts     = &iscsi_create_opts,
1600     .bdrv_reopen_prepare  = iscsi_reopen_prepare,
1601
1602     .bdrv_getlength  = iscsi_getlength,
1603     .bdrv_get_info   = iscsi_get_info,
1604     .bdrv_truncate   = iscsi_truncate,
1605     .bdrv_refresh_limits = iscsi_refresh_limits,
1606
1607     .bdrv_co_get_block_status = iscsi_co_get_block_status,
1608     .bdrv_co_discard      = iscsi_co_discard,
1609     .bdrv_co_write_zeroes = iscsi_co_write_zeroes,
1610     .bdrv_co_readv         = iscsi_co_readv,
1611     .bdrv_co_writev        = iscsi_co_writev,
1612     .bdrv_co_flush_to_disk = iscsi_co_flush,
1613
1614 #ifdef __linux__
1615     .bdrv_ioctl       = iscsi_ioctl,
1616     .bdrv_aio_ioctl   = iscsi_aio_ioctl,
1617 #endif
1618
1619     .bdrv_detach_aio_context = iscsi_detach_aio_context,
1620     .bdrv_attach_aio_context = iscsi_attach_aio_context,
1621 };
1622
1623 static QemuOptsList qemu_iscsi_opts = {
1624     .name = "iscsi",
1625     .head = QTAILQ_HEAD_INITIALIZER(qemu_iscsi_opts.head),
1626     .desc = {
1627         {
1628             .name = "user",
1629             .type = QEMU_OPT_STRING,
1630             .help = "username for CHAP authentication to target",
1631         },{
1632             .name = "password",
1633             .type = QEMU_OPT_STRING,
1634             .help = "password for CHAP authentication to target",
1635         },{
1636             .name = "header-digest",
1637             .type = QEMU_OPT_STRING,
1638             .help = "HeaderDigest setting. "
1639                     "{CRC32C|CRC32C-NONE|NONE-CRC32C|NONE}",
1640         },{
1641             .name = "initiator-name",
1642             .type = QEMU_OPT_STRING,
1643             .help = "Initiator iqn name to use when connecting",
1644         },
1645         { /* end of list */ }
1646     },
1647 };
1648
1649 static void iscsi_block_init(void)
1650 {
1651     bdrv_register(&bdrv_iscsi);
1652     qemu_add_opts(&qemu_iscsi_opts);
1653 }
1654
1655 block_init(iscsi_block_init);