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 "block/scsi.h"
  38 #include "qemu/iov.h"
  39 #include "sysemu/sysemu.h"
  40 #include "qmp-commands.h"
  41
  42 #include <iscsi/iscsi.h>
  43 #include <iscsi/scsi-lowlevel.h>
  44
  45 #ifdef __linux__
  46 #include <scsi/sg.h>
  47 #include <block/scsi.h>
  48 #endif
  49
  50 typedef struct IscsiLun {
  51     struct iscsi_context *iscsi;
  52     AioContext *aio_context;
  53     int lun;
  54     enum scsi_inquiry_peripheral_device_type type;
  55     int block_size;
  56     uint64_t num_blocks;
  57     int events;
  58     QEMUTimer *nop_timer;
  59     uint8_t lbpme;
  60     uint8_t lbprz;
  61     uint8_t has_write_same;
  62     struct scsi_inquiry_logical_block_provisioning lbp;
  63     struct scsi_inquiry_block_limits bl;
  64     unsigned char *zeroblock;
  65     unsigned long *allocationmap;
  66     int cluster_sectors;
  67     bool use_16_for_rw;
  68 } IscsiLun;
  69
  70 typedef struct IscsiTask {
  71     int status;
  72     int complete;
  73     int retries;
  74     int do_retry;
  75     struct scsi_task *task;
  76     Coroutine *co;
  77     QEMUBH *bh;
  78     IscsiLun *iscsilun;
  79     QEMUTimer retry_timer;
  80 } IscsiTask;
  81
  82 typedef struct IscsiAIOCB {
  83     BlockDriverAIOCB common;
  84     QEMUIOVector *qiov;
  85     QEMUBH *bh;
  86     IscsiLun *iscsilun;
  87     struct scsi_task *task;
  88     uint8_t *buf;
  89     int status;
  90     int64_t sector_num;
  91     int nb_sectors;
  92 #ifdef __linux__
  93     sg_io_hdr_t *ioh;
  94 #endif
  95 } IscsiAIOCB;
  96
  97 #define NOP_INTERVAL 5000
  98 #define MAX_NOP_FAILURES 3
  99 #define ISCSI_CMD_RETRIES ARRAY_SIZE(iscsi_retry_times)
 100 static const unsigned iscsi_retry_times[] = {8, 32, 128, 512, 2048};
 101
 102 /* this threshold is a trade-off knob to choose between
 103  * the potential additional overhead of an extra GET_LBA_STATUS request
 104  * vs. unnecessarily reading a lot of zero sectors over the wire.
 105  * If a read request is greater or equal than ISCSI_CHECKALLOC_THRES
 106  * sectors we check the allocation status of the area covered by the
 107  * request first if the allocationmap indicates that the area might be
 108  * unallocated. */
 109 #define ISCSI_CHECKALLOC_THRES 64
 110
 111 static void
 112 iscsi_bh_cb(void *p)
 113 {
 114     IscsiAIOCB *acb = p;
 115
 116     qemu_bh_delete(acb->bh);
 117
 118     g_free(acb->buf);
 119     acb->buf = NULL;
 120
 121     acb->common.cb(acb->common.opaque, acb->status);
 122
 123     if (acb->task != NULL) {
 124         scsi_free_scsi_task(acb->task);
 125         acb->task = NULL;
 126     }
 127
 128     qemu_aio_unref(acb);
 129 }
 130
 131 static void
 132 iscsi_schedule_bh(IscsiAIOCB *acb)
 133 {
 134     if (acb->bh) {
 135         return;
 136     }
 137     acb->bh = aio_bh_new(acb->iscsilun->aio_context, iscsi_bh_cb, acb);
 138     qemu_bh_schedule(acb->bh);
 139 }
 140
 141 static void iscsi_co_generic_bh_cb(void *opaque)
 142 {
 143     struct IscsiTask *iTask = opaque;
 144     iTask->complete = 1;
 145     qemu_bh_delete(iTask->bh);
 146     qemu_coroutine_enter(iTask->co, NULL);
 147 }
 148
 149 static void iscsi_retry_timer_expired(void *opaque)
 150 {
 151     struct IscsiTask *iTask = opaque;
 152     iTask->complete = 1;
 153     if (iTask->co) {
 154         qemu_coroutine_enter(iTask->co, NULL);
 155     }
 156 }
 157
 158 static inline unsigned exp_random(double mean)
 159 {
 160     return -mean * log((double)rand() / RAND_MAX);
 161 }
 162
 163 static void
 164 iscsi_co_generic_cb(struct iscsi_context *iscsi, int status,
 165                         void *command_data, void *opaque)
 166 {
 167     struct IscsiTask *iTask = opaque;
 168     struct scsi_task *task = command_data;
 169
 170     iTask->status = status;
 171     iTask->do_retry = 0;
 172     iTask->task = task;
 173
 174     if (status != SCSI_STATUS_GOOD) {
 175         if (iTask->retries++ < ISCSI_CMD_RETRIES) {
 176             if (status == SCSI_STATUS_CHECK_CONDITION
 177                 && task->sense.key == SCSI_SENSE_UNIT_ATTENTION) {
 178                 error_report("iSCSI CheckCondition: %s",
 179                              iscsi_get_error(iscsi));
 180                 iTask->do_retry = 1;
 181                 goto out;
 182             }
 183             if (status == SCSI_STATUS_BUSY) {
 184                 unsigned retry_time =
 185                     exp_random(iscsi_retry_times[iTask->retries - 1]);
 186                 error_report("iSCSI Busy (retry #%u in %u ms): %s",
 187                              iTask->retries, retry_time,
 188                              iscsi_get_error(iscsi));
 189                 aio_timer_init(iTask->iscsilun->aio_context,
 190                                &iTask->retry_timer, QEMU_CLOCK_REALTIME,
 191                                SCALE_MS, iscsi_retry_timer_expired, iTask);
 192                 timer_mod(&iTask->retry_timer,
 193                           qemu_clock_get_ms(QEMU_CLOCK_REALTIME) + retry_time);
 194                 iTask->do_retry = 1;
 195                 return;
 196             }
 197         }
 198         error_report("iSCSI Failure: %s", iscsi_get_error(iscsi));
 199     }
 200
 201 out:
 202     if (iTask->co) {
 203         iTask->bh = aio_bh_new(iTask->iscsilun->aio_context,
 204                                iscsi_co_generic_bh_cb, iTask);
 205         qemu_bh_schedule(iTask->bh);
 206     } else {
 207         iTask->complete = 1;
 208     }
 209 }
 210
 211 static void iscsi_co_init_iscsitask(IscsiLun *iscsilun, struct IscsiTask *iTask)
 212 {
 213     *iTask = (struct IscsiTask) {
 214         .co         = qemu_coroutine_self(),
 215         .iscsilun   = iscsilun,
 216     };
 217 }
 218
 219 static void
 220 iscsi_abort_task_cb(struct iscsi_context *iscsi, int status, void *command_data,
 221                     void *private_data)
 222 {
 223     IscsiAIOCB *acb = private_data;
 224
 225     acb->status = -ECANCELED;
 226     iscsi_schedule_bh(acb);
 227 }
 228
 229 static void
 230 iscsi_aio_cancel(BlockDriverAIOCB *blockacb)
 231 {
 232     IscsiAIOCB *acb = (IscsiAIOCB *)blockacb;
 233     IscsiLun *iscsilun = acb->iscsilun;
 234
 235     if (acb->status != -EINPROGRESS) {
 236         return;
 237     }
 238
 239     /* send a task mgmt call to the target to cancel the task on the target */
 240     iscsi_task_mgmt_abort_task_async(iscsilun->iscsi, acb->task,
 241                                      iscsi_abort_task_cb, acb);
 242
 243 }
 244
 245 static const AIOCBInfo iscsi_aiocb_info = {
 246     .aiocb_size         = sizeof(IscsiAIOCB),
 247     .cancel_async       = iscsi_aio_cancel,
 248 };
 249
 250
 251 static void iscsi_process_read(void *arg);
 252 static void iscsi_process_write(void *arg);
 253
 254 static void
 255 iscsi_set_events(IscsiLun *iscsilun)
 256 {
 257     struct iscsi_context *iscsi = iscsilun->iscsi;
 258     int ev;
 259
 260     /* We always register a read handler.  */
 261     ev = POLLIN;
 262     ev |= iscsi_which_events(iscsi);
 263     if (ev != iscsilun->events) {
 264         aio_set_fd_handler(iscsilun->aio_context,
 265                            iscsi_get_fd(iscsi),
 266                            iscsi_process_read,
 267                            (ev & POLLOUT) ? iscsi_process_write : NULL,
 268                            iscsilun);
 269
 270     }
 271
 272     iscsilun->events = ev;
 273 }
 274
 275 static void
 276 iscsi_process_read(void *arg)
 277 {
 278     IscsiLun *iscsilun = arg;
 279     struct iscsi_context *iscsi = iscsilun->iscsi;
 280
 281     iscsi_service(iscsi, POLLIN);
 282     iscsi_set_events(iscsilun);
 283 }
 284
 285 static void
 286 iscsi_process_write(void *arg)
 287 {
 288     IscsiLun *iscsilun = arg;
 289     struct iscsi_context *iscsi = iscsilun->iscsi;
 290
 291     iscsi_service(iscsi, POLLOUT);
 292     iscsi_set_events(iscsilun);
 293 }
 294
 295 static int64_t sector_lun2qemu(int64_t sector, IscsiLun *iscsilun)
 296 {
 297     return sector * iscsilun->block_size / BDRV_SECTOR_SIZE;
 298 }
 299
 300 static int64_t sector_qemu2lun(int64_t sector, IscsiLun *iscsilun)
 301 {
 302     return sector * BDRV_SECTOR_SIZE / iscsilun->block_size;
 303 }
 304
 305 static bool is_request_lun_aligned(int64_t sector_num, int nb_sectors,
 306                                       IscsiLun *iscsilun)
 307 {
 308     if ((sector_num * BDRV_SECTOR_SIZE) % iscsilun->block_size ||
 309         (nb_sectors * BDRV_SECTOR_SIZE) % iscsilun->block_size) {
 310             error_report("iSCSI misaligned request: "
 311                          "iscsilun->block_size %u, sector_num %" PRIi64
 312                          ", nb_sectors %d",
 313                          iscsilun->block_size, sector_num, nb_sectors);
 314             return 0;
 315     }
 316     return 1;
 317 }
 318
 319 static void iscsi_allocationmap_set(IscsiLun *iscsilun, int64_t sector_num,
 320                                     int nb_sectors)
 321 {
 322     if (iscsilun->allocationmap == NULL) {
 323         return;
 324     }
 325     bitmap_set(iscsilun->allocationmap,
 326                sector_num / iscsilun->cluster_sectors,
 327                DIV_ROUND_UP(nb_sectors, iscsilun->cluster_sectors));
 328 }
 329
 330 static void iscsi_allocationmap_clear(IscsiLun *iscsilun, int64_t sector_num,
 331                                       int nb_sectors)
 332 {
 333     int64_t cluster_num, nb_clusters;
 334     if (iscsilun->allocationmap == NULL) {
 335         return;
 336     }
 337     cluster_num = DIV_ROUND_UP(sector_num, iscsilun->cluster_sectors);
 338     nb_clusters = (sector_num + nb_sectors) / iscsilun->cluster_sectors
 339                   - cluster_num;
 340     if (nb_clusters > 0) {
 341         bitmap_clear(iscsilun->allocationmap, cluster_num, nb_clusters);
 342     }
 343 }
 344
 345 static int coroutine_fn iscsi_co_writev(BlockDriverState *bs,
 346                                         int64_t sector_num, int nb_sectors,
 347                                         QEMUIOVector *iov)
 348 {
 349     IscsiLun *iscsilun = bs->opaque;
 350     struct IscsiTask iTask;
 351     uint64_t lba;
 352     uint32_t num_sectors;
 353
 354     if (!is_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
 355         return -EINVAL;
 356     }
 357
 358     lba = sector_qemu2lun(sector_num, iscsilun);
 359     num_sectors = sector_qemu2lun(nb_sectors, iscsilun);
 360     iscsi_co_init_iscsitask(iscsilun, &iTask);
 361 retry:
 362     if (iscsilun->use_16_for_rw) {
 363         iTask.task = iscsi_write16_task(iscsilun->iscsi, iscsilun->lun, lba,
 364                                         NULL, num_sectors * iscsilun->block_size,
 365                                         iscsilun->block_size, 0, 0, 0, 0, 0,
 366                                         iscsi_co_generic_cb, &iTask);
 367     } else {
 368         iTask.task = iscsi_write10_task(iscsilun->iscsi, iscsilun->lun, lba,
 369                                         NULL, num_sectors * iscsilun->block_size,
 370                                         iscsilun->block_size, 0, 0, 0, 0, 0,
 371                                         iscsi_co_generic_cb, &iTask);
 372     }
 373     if (iTask.task == NULL) {
 374         return -ENOMEM;
 375     }
 376     scsi_task_set_iov_out(iTask.task, (struct scsi_iovec *) iov->iov,
 377                           iov->niov);
 378     while (!iTask.complete) {
 379         iscsi_set_events(iscsilun);
 380         qemu_coroutine_yield();
 381     }
 382
 383     if (iTask.task != NULL) {
 384         scsi_free_scsi_task(iTask.task);
 385         iTask.task = NULL;
 386     }
 387
 388     if (iTask.do_retry) {
 389         iTask.complete = 0;
 390         goto retry;
 391     }
 392
 393     if (iTask.status != SCSI_STATUS_GOOD) {
 394         return -EIO;
 395     }
 396
 397     iscsi_allocationmap_set(iscsilun, sector_num, nb_sectors);
 398
 399     return 0;
 400 }
 401
 402
 403 static bool iscsi_allocationmap_is_allocated(IscsiLun *iscsilun,
 404                                              int64_t sector_num, int nb_sectors)
 405 {
 406     unsigned long size;
 407     if (iscsilun->allocationmap == NULL) {
 408         return true;
 409     }
 410     size = DIV_ROUND_UP(sector_num + nb_sectors, iscsilun->cluster_sectors);
 411     return !(find_next_bit(iscsilun->allocationmap, size,
 412                            sector_num / iscsilun->cluster_sectors) == size);
 413 }
 414
 415 static int64_t coroutine_fn iscsi_co_get_block_status(BlockDriverState *bs,
 416                                                   int64_t sector_num,
 417                                                   int nb_sectors, int *pnum)
 418 {
 419     IscsiLun *iscsilun = bs->opaque;
 420     struct scsi_get_lba_status *lbas = NULL;
 421     struct scsi_lba_status_descriptor *lbasd = NULL;
 422     struct IscsiTask iTask;
 423     int64_t ret;
 424
 425     iscsi_co_init_iscsitask(iscsilun, &iTask);
 426
 427     if (!is_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
 428         ret = -EINVAL;
 429         goto out;
 430     }
 431
 432     /* default to all sectors allocated */
 433     ret = BDRV_BLOCK_DATA;
 434     ret |= (sector_num << BDRV_SECTOR_BITS) | BDRV_BLOCK_OFFSET_VALID;
 435     *pnum = nb_sectors;
 436
 437     /* LUN does not support logical block provisioning */
 438     if (iscsilun->lbpme == 0) {
 439         goto out;
 440     }
 441
 442 retry:
 443     if (iscsi_get_lba_status_task(iscsilun->iscsi, iscsilun->lun,
 444                                   sector_qemu2lun(sector_num, iscsilun),
 445                                   8 + 16, iscsi_co_generic_cb,
 446                                   &iTask) == NULL) {
 447         ret = -ENOMEM;
 448         goto out;
 449     }
 450
 451     while (!iTask.complete) {
 452         iscsi_set_events(iscsilun);
 453         qemu_coroutine_yield();
 454     }
 455
 456     if (iTask.do_retry) {
 457         if (iTask.task != NULL) {
 458             scsi_free_scsi_task(iTask.task);
 459             iTask.task = NULL;
 460         }
 461         iTask.complete = 0;
 462         goto retry;
 463     }
 464
 465     if (iTask.status != SCSI_STATUS_GOOD) {
 466         /* in case the get_lba_status_callout fails (i.e.
 467          * because the device is busy or the cmd is not
 468          * supported) we pretend all blocks are allocated
 469          * for backwards compatibility */
 470         goto out;
 471     }
 472
 473     lbas = scsi_datain_unmarshall(iTask.task);
 474     if (lbas == NULL) {
 475         ret = -EIO;
 476         goto out;
 477     }
 478
 479     lbasd = &lbas->descriptors[0];
 480
 481     if (sector_qemu2lun(sector_num, iscsilun) != lbasd->lba) {
 482         ret = -EIO;
 483         goto out;
 484     }
 485
 486     *pnum = sector_lun2qemu(lbasd->num_blocks, iscsilun);
 487
 488     if (lbasd->provisioning == SCSI_PROVISIONING_TYPE_DEALLOCATED ||
 489         lbasd->provisioning == SCSI_PROVISIONING_TYPE_ANCHORED) {
 490         ret &= ~BDRV_BLOCK_DATA;
 491         if (iscsilun->lbprz) {
 492             ret |= BDRV_BLOCK_ZERO;
 493         }
 494     }
 495
 496     if (ret & BDRV_BLOCK_ZERO) {
 497         iscsi_allocationmap_clear(iscsilun, sector_num, *pnum);
 498     } else {
 499         iscsi_allocationmap_set(iscsilun, sector_num, *pnum);
 500     }
 501
 502     if (*pnum > nb_sectors) {
 503         *pnum = nb_sectors;
 504     }
 505 out:
 506     if (iTask.task != NULL) {
 507         scsi_free_scsi_task(iTask.task);
 508     }
 509     return ret;
 510 }
 511
 512 static int coroutine_fn iscsi_co_readv(BlockDriverState *bs,
 513                                        int64_t sector_num, int nb_sectors,
 514                                        QEMUIOVector *iov)
 515 {
 516     IscsiLun *iscsilun = bs->opaque;
 517     struct IscsiTask iTask;
 518     uint64_t lba;
 519     uint32_t num_sectors;
 520
 521     if (!is_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
 522         return -EINVAL;
 523     }
 524
 525     if (iscsilun->lbprz && nb_sectors >= ISCSI_CHECKALLOC_THRES &&
 526         !iscsi_allocationmap_is_allocated(iscsilun, sector_num, nb_sectors)) {
 527         int64_t ret;
 528         int pnum;
 529         ret = iscsi_co_get_block_status(bs, sector_num, INT_MAX, &pnum);
 530         if (ret < 0) {
 531             return ret;
 532         }
 533         if (ret & BDRV_BLOCK_ZERO && pnum >= nb_sectors) {
 534             qemu_iovec_memset(iov, 0, 0x00, iov->size);
 535             return 0;
 536         }
 537     }
 538
 539     lba = sector_qemu2lun(sector_num, iscsilun);
 540     num_sectors = sector_qemu2lun(nb_sectors, iscsilun);
 541
 542     iscsi_co_init_iscsitask(iscsilun, &iTask);
 543 retry:
 544     if (iscsilun->use_16_for_rw) {
 545         iTask.task = iscsi_read16_task(iscsilun->iscsi, iscsilun->lun, lba,
 546                                        num_sectors * iscsilun->block_size,
 547                                        iscsilun->block_size, 0, 0, 0, 0, 0,
 548                                        iscsi_co_generic_cb, &iTask);
 549     } else {
 550         iTask.task = iscsi_read10_task(iscsilun->iscsi, iscsilun->lun, lba,
 551                                        num_sectors * iscsilun->block_size,
 552                                        iscsilun->block_size,
 553                                        0, 0, 0, 0, 0,
 554                                        iscsi_co_generic_cb, &iTask);
 555     }
 556     if (iTask.task == NULL) {
 557         return -ENOMEM;
 558     }
 559     scsi_task_set_iov_in(iTask.task, (struct scsi_iovec *) iov->iov, iov->niov);
 560
 561     while (!iTask.complete) {
 562         iscsi_set_events(iscsilun);
 563         qemu_coroutine_yield();
 564     }
 565
 566     if (iTask.task != NULL) {
 567         scsi_free_scsi_task(iTask.task);
 568         iTask.task = NULL;
 569     }
 570
 571     if (iTask.do_retry) {
 572         iTask.complete = 0;
 573         goto retry;
 574     }
 575
 576     if (iTask.status != SCSI_STATUS_GOOD) {
 577         return -EIO;
 578     }
 579
 580     return 0;
 581 }
 582
 583 static int coroutine_fn iscsi_co_flush(BlockDriverState *bs)
 584 {
 585     IscsiLun *iscsilun = bs->opaque;
 586     struct IscsiTask iTask;
 587
 588     if (bs->sg) {
 589         return 0;
 590     }
 591
 592     iscsi_co_init_iscsitask(iscsilun, &iTask);
 593
 594 retry:
 595     if (iscsi_synchronizecache10_task(iscsilun->iscsi, iscsilun->lun, 0, 0, 0,
 596                                       0, iscsi_co_generic_cb, &iTask) == NULL) {
 597         return -ENOMEM;
 598     }
 599
 600     while (!iTask.complete) {
 601         iscsi_set_events(iscsilun);
 602         qemu_coroutine_yield();
 603     }
 604
 605     if (iTask.task != NULL) {
 606         scsi_free_scsi_task(iTask.task);
 607         iTask.task = NULL;
 608     }
 609
 610     if (iTask.do_retry) {
 611         iTask.complete = 0;
 612         goto retry;
 613     }
 614
 615     if (iTask.status != SCSI_STATUS_GOOD) {
 616         return -EIO;
 617     }
 618
 619     return 0;
 620 }
 621
 622 #ifdef __linux__
 623 static void
 624 iscsi_aio_ioctl_cb(struct iscsi_context *iscsi, int status,
 625                      void *command_data, void *opaque)
 626 {
 627     IscsiAIOCB *acb = opaque;
 628
 629     g_free(acb->buf);
 630     acb->buf = NULL;
 631
 632     acb->status = 0;
 633     if (status < 0) {
 634         error_report("Failed to ioctl(SG_IO) to iSCSI lun. %s",
 635                      iscsi_get_error(iscsi));
 636         acb->status = -EIO;
 637     }
 638
 639     acb->ioh->driver_status = 0;
 640     acb->ioh->host_status   = 0;
 641     acb->ioh->resid         = 0;
 642
 643 #define SG_ERR_DRIVER_SENSE    0x08
 644
 645     if (status == SCSI_STATUS_CHECK_CONDITION && acb->task->datain.size >= 2) {
 646         int ss;
 647
 648         acb->ioh->driver_status |= SG_ERR_DRIVER_SENSE;
 649
 650         acb->ioh->sb_len_wr = acb->task->datain.size - 2;
 651         ss = (acb->ioh->mx_sb_len >= acb->ioh->sb_len_wr) ?
 652              acb->ioh->mx_sb_len : acb->ioh->sb_len_wr;
 653         memcpy(acb->ioh->sbp, &acb->task->datain.data[2], ss);
 654     }
 655
 656     iscsi_schedule_bh(acb);
 657 }
 658
 659 static BlockDriverAIOCB *iscsi_aio_ioctl(BlockDriverState *bs,
 660         unsigned long int req, void *buf,
 661         BlockDriverCompletionFunc *cb, void *opaque)
 662 {
 663     IscsiLun *iscsilun = bs->opaque;
 664     struct iscsi_context *iscsi = iscsilun->iscsi;
 665     struct iscsi_data data;
 666     IscsiAIOCB *acb;
 667
 668     assert(req == SG_IO);
 669
 670     acb = qemu_aio_get(&iscsi_aiocb_info, bs, cb, opaque);
 671
 672     acb->iscsilun = iscsilun;
 673     acb->bh          = NULL;
 674     acb->status      = -EINPROGRESS;
 675     acb->buf         = NULL;
 676     acb->ioh         = buf;
 677
 678     acb->task = malloc(sizeof(struct scsi_task));
 679     if (acb->task == NULL) {
 680         error_report("iSCSI: Failed to allocate task for scsi command. %s",
 681                      iscsi_get_error(iscsi));
 682         qemu_aio_unref(acb);
 683         return NULL;
 684     }
 685     memset(acb->task, 0, sizeof(struct scsi_task));
 686
 687     switch (acb->ioh->dxfer_direction) {
 688     case SG_DXFER_TO_DEV:
 689         acb->task->xfer_dir = SCSI_XFER_WRITE;
 690         break;
 691     case SG_DXFER_FROM_DEV:
 692         acb->task->xfer_dir = SCSI_XFER_READ;
 693         break;
 694     default:
 695         acb->task->xfer_dir = SCSI_XFER_NONE;
 696         break;
 697     }
 698
 699     acb->task->cdb_size = acb->ioh->cmd_len;
 700     memcpy(&acb->task->cdb[0], acb->ioh->cmdp, acb->ioh->cmd_len);
 701     acb->task->expxferlen = acb->ioh->dxfer_len;
 702
 703     data.size = 0;
 704     if (acb->task->xfer_dir == SCSI_XFER_WRITE) {
 705         if (acb->ioh->iovec_count == 0) {
 706             data.data = acb->ioh->dxferp;
 707             data.size = acb->ioh->dxfer_len;
 708         } else {
 709             scsi_task_set_iov_out(acb->task,
 710                                  (struct scsi_iovec *) acb->ioh->dxferp,
 711                                  acb->ioh->iovec_count);
 712         }
 713     }
 714
 715     if (iscsi_scsi_command_async(iscsi, iscsilun->lun, acb->task,
 716                                  iscsi_aio_ioctl_cb,
 717                                  (data.size > 0) ? &data : NULL,
 718                                  acb) != 0) {
 719         scsi_free_scsi_task(acb->task);
 720         qemu_aio_unref(acb);
 721         return NULL;
 722     }
 723
 724     /* tell libiscsi to read straight into the buffer we got from ioctl */
 725     if (acb->task->xfer_dir == SCSI_XFER_READ) {
 726         if (acb->ioh->iovec_count == 0) {
 727             scsi_task_add_data_in_buffer(acb->task,
 728                                          acb->ioh->dxfer_len,
 729                                          acb->ioh->dxferp);
 730         } else {
 731             scsi_task_set_iov_in(acb->task,
 732                                  (struct scsi_iovec *) acb->ioh->dxferp,
 733                                  acb->ioh->iovec_count);
 734         }
 735     }
 736
 737     iscsi_set_events(iscsilun);
 738
 739     return &acb->common;
 740 }
 741
 742 static void ioctl_cb(void *opaque, int status)
 743 {
 744     int *p_status = opaque;
 745     *p_status = status;
 746 }
 747
 748 static int iscsi_ioctl(BlockDriverState *bs, unsigned long int req, void *buf)
 749 {
 750     IscsiLun *iscsilun = bs->opaque;
 751     int status;
 752
 753     switch (req) {
 754     case SG_GET_VERSION_NUM:
 755         *(int *)buf = 30000;
 756         break;
 757     case SG_GET_SCSI_ID:
 758         ((struct sg_scsi_id *)buf)->scsi_type = iscsilun->type;
 759         break;
 760     case SG_IO:
 761         status = -EINPROGRESS;
 762         iscsi_aio_ioctl(bs, req, buf, ioctl_cb, &status);
 763
 764         while (status == -EINPROGRESS) {
 765             aio_poll(iscsilun->aio_context, true);
 766         }
 767
 768         return 0;
 769     default:
 770         return -1;
 771     }
 772     return 0;
 773 }
 774 #endif
 775
 776 static int64_t
 777 iscsi_getlength(BlockDriverState *bs)
 778 {
 779     IscsiLun *iscsilun = bs->opaque;
 780     int64_t len;
 781
 782     len  = iscsilun->num_blocks;
 783     len *= iscsilun->block_size;
 784
 785     return len;
 786 }
 787
 788 static int
 789 coroutine_fn iscsi_co_discard(BlockDriverState *bs, int64_t sector_num,
 790                                    int nb_sectors)
 791 {
 792     IscsiLun *iscsilun = bs->opaque;
 793     struct IscsiTask iTask;
 794     struct unmap_list list;
 795
 796     if (!is_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
 797         return -EINVAL;
 798     }
 799
 800     if (!iscsilun->lbp.lbpu) {
 801         /* UNMAP is not supported by the target */
 802         return 0;
 803     }
 804
 805     list.lba = sector_qemu2lun(sector_num, iscsilun);
 806     list.num = sector_qemu2lun(nb_sectors, iscsilun);
 807
 808     iscsi_co_init_iscsitask(iscsilun, &iTask);
 809 retry:
 810     if (iscsi_unmap_task(iscsilun->iscsi, iscsilun->lun, 0, 0, &list, 1,
 811                      iscsi_co_generic_cb, &iTask) == NULL) {
 812         return -ENOMEM;
 813     }
 814
 815     while (!iTask.complete) {
 816         iscsi_set_events(iscsilun);
 817         qemu_coroutine_yield();
 818     }
 819
 820     if (iTask.task != NULL) {
 821         scsi_free_scsi_task(iTask.task);
 822         iTask.task = NULL;
 823     }
 824
 825     if (iTask.do_retry) {
 826         iTask.complete = 0;
 827         goto retry;
 828     }
 829
 830     if (iTask.status == SCSI_STATUS_CHECK_CONDITION) {
 831         /* the target might fail with a check condition if it
 832            is not happy with the alignment of the UNMAP request
 833            we silently fail in this case */
 834         return 0;
 835     }
 836
 837     if (iTask.status != SCSI_STATUS_GOOD) {
 838         return -EIO;
 839     }
 840
 841     iscsi_allocationmap_clear(iscsilun, sector_num, nb_sectors);
 842
 843     return 0;
 844 }
 845
 846 static int
 847 coroutine_fn iscsi_co_write_zeroes(BlockDriverState *bs, int64_t sector_num,
 848                                    int nb_sectors, BdrvRequestFlags flags)
 849 {
 850     IscsiLun *iscsilun = bs->opaque;
 851     struct IscsiTask iTask;
 852     uint64_t lba;
 853     uint32_t nb_blocks;
 854     bool use_16_for_ws = iscsilun->use_16_for_rw;
 855
 856     if (!is_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
 857         return -EINVAL;
 858     }
 859
 860     if (flags & BDRV_REQ_MAY_UNMAP) {
 861         if (!use_16_for_ws && !iscsilun->lbp.lbpws10) {
 862             /* WRITESAME10 with UNMAP is unsupported try WRITESAME16 */
 863             use_16_for_ws = true;
 864         }
 865         if (use_16_for_ws && !iscsilun->lbp.lbpws) {
 866             /* WRITESAME16 with UNMAP is not supported by the target,
 867              * fall back and try WRITESAME10/16 without UNMAP */
 868             flags &= ~BDRV_REQ_MAY_UNMAP;
 869             use_16_for_ws = iscsilun->use_16_for_rw;
 870         }
 871     }
 872
 873     if (!(flags & BDRV_REQ_MAY_UNMAP) && !iscsilun->has_write_same) {
 874         /* WRITESAME without UNMAP is not supported by the target */
 875         return -ENOTSUP;
 876     }
 877
 878     lba = sector_qemu2lun(sector_num, iscsilun);
 879     nb_blocks = sector_qemu2lun(nb_sectors, iscsilun);
 880
 881     if (iscsilun->zeroblock == NULL) {
 882         iscsilun->zeroblock = g_try_malloc0(iscsilun->block_size);
 883         if (iscsilun->zeroblock == NULL) {
 884             return -ENOMEM;
 885         }
 886     }
 887
 888     iscsi_co_init_iscsitask(iscsilun, &iTask);
 889 retry:
 890     if (use_16_for_ws) {
 891         iTask.task = iscsi_writesame16_task(iscsilun->iscsi, iscsilun->lun, lba,
 892                                             iscsilun->zeroblock, iscsilun->block_size,
 893                                             nb_blocks, 0, !!(flags & BDRV_REQ_MAY_UNMAP),
 894                                             0, 0, iscsi_co_generic_cb, &iTask);
 895     } else {
 896         iTask.task = iscsi_writesame10_task(iscsilun->iscsi, iscsilun->lun, lba,
 897                                             iscsilun->zeroblock, iscsilun->block_size,
 898                                             nb_blocks, 0, !!(flags & BDRV_REQ_MAY_UNMAP),
 899                                             0, 0, iscsi_co_generic_cb, &iTask);
 900     }
 901     if (iTask.task == NULL) {
 902         return -ENOMEM;
 903     }
 904
 905     while (!iTask.complete) {
 906         iscsi_set_events(iscsilun);
 907         qemu_coroutine_yield();
 908     }
 909
 910     if (iTask.status == SCSI_STATUS_CHECK_CONDITION &&
 911         iTask.task->sense.key == SCSI_SENSE_ILLEGAL_REQUEST &&
 912         (iTask.task->sense.ascq == SCSI_SENSE_ASCQ_INVALID_OPERATION_CODE ||
 913          iTask.task->sense.ascq == SCSI_SENSE_ASCQ_INVALID_FIELD_IN_CDB)) {
 914         /* WRITE SAME is not supported by the target */
 915         iscsilun->has_write_same = false;
 916         scsi_free_scsi_task(iTask.task);
 917         return -ENOTSUP;
 918     }
 919
 920     if (iTask.task != NULL) {
 921         scsi_free_scsi_task(iTask.task);
 922         iTask.task = NULL;
 923     }
 924
 925     if (iTask.do_retry) {
 926         iTask.complete = 0;
 927         goto retry;
 928     }
 929
 930     if (iTask.status != SCSI_STATUS_GOOD) {
 931         return -EIO;
 932     }
 933
 934     if (flags & BDRV_REQ_MAY_UNMAP) {
 935         iscsi_allocationmap_clear(iscsilun, sector_num, nb_sectors);
 936     } else {
 937         iscsi_allocationmap_set(iscsilun, sector_num, nb_sectors);
 938     }
 939
 940     return 0;
 941 }
 942
 943 static void parse_chap(struct iscsi_context *iscsi, const char *target,
 944                        Error **errp)
 945 {
 946     QemuOptsList *list;
 947     QemuOpts *opts;
 948     const char *user = NULL;
 949     const char *password = NULL;
 950
 951     list = qemu_find_opts("iscsi");
 952     if (!list) {
 953         return;
 954     }
 955
 956     opts = qemu_opts_find(list, target);
 957     if (opts == NULL) {
 958         opts = QTAILQ_FIRST(&list->head);
 959         if (!opts) {
 960             return;
 961         }
 962     }
 963
 964     user = qemu_opt_get(opts, "user");
 965     if (!user) {
 966         return;
 967     }
 968
 969     password = qemu_opt_get(opts, "password");
 970     if (!password) {
 971         error_setg(errp, "CHAP username specified but no password was given");
 972         return;
 973     }
 974
 975     if (iscsi_set_initiator_username_pwd(iscsi, user, password)) {
 976         error_setg(errp, "Failed to set initiator username and password");
 977     }
 978 }
 979
 980 static void parse_header_digest(struct iscsi_context *iscsi, const char *target,
 981                                 Error **errp)
 982 {
 983     QemuOptsList *list;
 984     QemuOpts *opts;
 985     const char *digest = NULL;
 986
 987     list = qemu_find_opts("iscsi");
 988     if (!list) {
 989         return;
 990     }
 991
 992     opts = qemu_opts_find(list, target);
 993     if (opts == NULL) {
 994         opts = QTAILQ_FIRST(&list->head);
 995         if (!opts) {
 996             return;
 997         }
 998     }
 999
1000     digest = qemu_opt_get(opts, "header-digest");
1001     if (!digest) {
1002         return;
1003     }
1004
1005     if (!strcmp(digest, "CRC32C")) {
1006         iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_CRC32C);
1007     } else if (!strcmp(digest, "NONE")) {
1008         iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_NONE);
1009     } else if (!strcmp(digest, "CRC32C-NONE")) {
1010         iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_CRC32C_NONE);
1011     } else if (!strcmp(digest, "NONE-CRC32C")) {
1012         iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_NONE_CRC32C);
1013     } else {
1014         error_setg(errp, "Invalid header-digest setting : %s", digest);
1015     }
1016 }
1017
1018 static char *parse_initiator_name(const char *target)
1019 {
1020     QemuOptsList *list;
1021     QemuOpts *opts;
1022     const char *name;
1023     char *iscsi_name;
1024     UuidInfo *uuid_info;
1025
1026     list = qemu_find_opts("iscsi");
1027     if (list) {
1028         opts = qemu_opts_find(list, target);
1029         if (!opts) {
1030             opts = QTAILQ_FIRST(&list->head);
1031         }
1032         if (opts) {
1033             name = qemu_opt_get(opts, "initiator-name");
1034             if (name) {
1035                 return g_strdup(name);
1036             }
1037         }
1038     }
1039
1040     uuid_info = qmp_query_uuid(NULL);
1041     if (strcmp(uuid_info->UUID, UUID_NONE) == 0) {
1042         name = qemu_get_vm_name();
1043     } else {
1044         name = uuid_info->UUID;
1045     }
1046     iscsi_name = g_strdup_printf("iqn.2008-11.org.linux-kvm%s%s",
1047                                  name ? ":" : "", name ? name : "");
1048     qapi_free_UuidInfo(uuid_info);
1049     return iscsi_name;
1050 }
1051
1052 static void iscsi_nop_timed_event(void *opaque)
1053 {
1054     IscsiLun *iscsilun = opaque;
1055
1056     if (iscsi_get_nops_in_flight(iscsilun->iscsi) > MAX_NOP_FAILURES) {
1057         error_report("iSCSI: NOP timeout. Reconnecting...");
1058         iscsi_reconnect(iscsilun->iscsi);
1059     }
1060
1061     if (iscsi_nop_out_async(iscsilun->iscsi, NULL, NULL, 0, NULL) != 0) {
1062         error_report("iSCSI: failed to sent NOP-Out. Disabling NOP messages.");
1063         return;
1064     }
1065
1066     timer_mod(iscsilun->nop_timer, qemu_clock_get_ms(QEMU_CLOCK_REALTIME) + NOP_INTERVAL);
1067     iscsi_set_events(iscsilun);
1068 }
1069
1070 static void iscsi_readcapacity_sync(IscsiLun *iscsilun, Error **errp)
1071 {
1072     struct scsi_task *task = NULL;
1073     struct scsi_readcapacity10 *rc10 = NULL;
1074     struct scsi_readcapacity16 *rc16 = NULL;
1075     int retries = ISCSI_CMD_RETRIES;
1076
1077     do {
1078         if (task != NULL) {
1079             scsi_free_scsi_task(task);
1080             task = NULL;
1081         }
1082
1083         switch (iscsilun->type) {
1084         case TYPE_DISK:
1085             task = iscsi_readcapacity16_sync(iscsilun->iscsi, iscsilun->lun);
1086             if (task != NULL && task->status == SCSI_STATUS_GOOD) {
1087                 rc16 = scsi_datain_unmarshall(task);
1088                 if (rc16 == NULL) {
1089                     error_setg(errp, "iSCSI: Failed to unmarshall readcapacity16 data.");
1090                 } else {
1091                     iscsilun->block_size = rc16->block_length;
1092                     iscsilun->num_blocks = rc16->returned_lba + 1;
1093                     iscsilun->lbpme = rc16->lbpme;
1094                     iscsilun->lbprz = rc16->lbprz;
1095                     iscsilun->use_16_for_rw = (rc16->returned_lba > 0xffffffff);
1096                 }
1097             }
1098             break;
1099         case TYPE_ROM:
1100             task = iscsi_readcapacity10_sync(iscsilun->iscsi, iscsilun->lun, 0, 0);
1101             if (task != NULL && task->status == SCSI_STATUS_GOOD) {
1102                 rc10 = scsi_datain_unmarshall(task);
1103                 if (rc10 == NULL) {
1104                     error_setg(errp, "iSCSI: Failed to unmarshall readcapacity10 data.");
1105                 } else {
1106                     iscsilun->block_size = rc10->block_size;
1107                     if (rc10->lba == 0) {
1108                         /* blank disk loaded */
1109                         iscsilun->num_blocks = 0;
1110                     } else {
1111                         iscsilun->num_blocks = rc10->lba + 1;
1112                     }
1113                 }
1114             }
1115             break;
1116         default:
1117             return;
1118         }
1119     } while (task != NULL && task->status == SCSI_STATUS_CHECK_CONDITION
1120              && task->sense.key == SCSI_SENSE_UNIT_ATTENTION
1121              && retries-- > 0);
1122
1123     if (task == NULL || task->status != SCSI_STATUS_GOOD) {
1124         error_setg(errp, "iSCSI: failed to send readcapacity10 command.");
1125     }
1126     if (task) {
1127         scsi_free_scsi_task(task);
1128     }
1129 }
1130
1131 /* TODO Convert to fine grained options */
1132 static QemuOptsList runtime_opts = {
1133     .name = "iscsi",
1134     .head = QTAILQ_HEAD_INITIALIZER(runtime_opts.head),
1135     .desc = {
1136         {
1137             .name = "filename",
1138             .type = QEMU_OPT_STRING,
1139             .help = "URL to the iscsi image",
1140         },
1141         { /* end of list */ }
1142     },
1143 };
1144
1145 static struct scsi_task *iscsi_do_inquiry(struct iscsi_context *iscsi, int lun,
1146                                           int evpd, int pc, void **inq, Error **errp)
1147 {
1148     int full_size;
1149     struct scsi_task *task = NULL;
1150     task = iscsi_inquiry_sync(iscsi, lun, evpd, pc, 64);
1151     if (task == NULL || task->status != SCSI_STATUS_GOOD) {
1152         goto fail;
1153     }
1154     full_size = scsi_datain_getfullsize(task);
1155     if (full_size > task->datain.size) {
1156         scsi_free_scsi_task(task);
1157
1158         /* we need more data for the full list */
1159         task = iscsi_inquiry_sync(iscsi, lun, evpd, pc, full_size);
1160         if (task == NULL || task->status != SCSI_STATUS_GOOD) {
1161             goto fail;
1162         }
1163     }
1164
1165     *inq = scsi_datain_unmarshall(task);
1166     if (*inq == NULL) {
1167         error_setg(errp, "iSCSI: failed to unmarshall inquiry datain blob");
1168         goto fail_with_err;
1169     }
1170
1171     return task;
1172
1173 fail:
1174     error_setg(errp, "iSCSI: Inquiry command failed : %s",
1175                iscsi_get_error(iscsi));
1176 fail_with_err:
1177     if (task != NULL) {
1178         scsi_free_scsi_task(task);
1179     }
1180     return NULL;
1181 }
1182
1183 static void iscsi_detach_aio_context(BlockDriverState *bs)
1184 {
1185     IscsiLun *iscsilun = bs->opaque;
1186
1187     aio_set_fd_handler(iscsilun->aio_context,
1188                        iscsi_get_fd(iscsilun->iscsi),
1189                        NULL, NULL, NULL);
1190     iscsilun->events = 0;
1191
1192     if (iscsilun->nop_timer) {
1193         timer_del(iscsilun->nop_timer);
1194         timer_free(iscsilun->nop_timer);
1195         iscsilun->nop_timer = NULL;
1196     }
1197 }
1198
1199 static void iscsi_attach_aio_context(BlockDriverState *bs,
1200                                      AioContext *new_context)
1201 {
1202     IscsiLun *iscsilun = bs->opaque;
1203
1204     iscsilun->aio_context = new_context;
1205     iscsi_set_events(iscsilun);
1206
1207     /* Set up a timer for sending out iSCSI NOPs */
1208     iscsilun->nop_timer = aio_timer_new(iscsilun->aio_context,
1209                                         QEMU_CLOCK_REALTIME, SCALE_MS,
1210                                         iscsi_nop_timed_event, iscsilun);
1211     timer_mod(iscsilun->nop_timer,
1212               qemu_clock_get_ms(QEMU_CLOCK_REALTIME) + NOP_INTERVAL);
1213 }
1214
1215 /*
1216  * We support iscsi url's on the form
1217  * iscsi://[<username>%<password>@]<host>[:<port>]/<targetname>/<lun>
1218  *
1219  * Note: flags are currently not used by iscsi_open.  If this function
1220  * is changed such that flags are used, please examine iscsi_reopen_prepare()
1221  * to see if needs to be changed as well.
1222  */
1223 static int iscsi_open(BlockDriverState *bs, QDict *options, int flags,
1224                       Error **errp)
1225 {
1226     IscsiLun *iscsilun = bs->opaque;
1227     struct iscsi_context *iscsi = NULL;
1228     struct iscsi_url *iscsi_url = NULL;
1229     struct scsi_task *task = NULL;
1230     struct scsi_inquiry_standard *inq = NULL;
1231     struct scsi_inquiry_supported_pages *inq_vpd;
1232     char *initiator_name = NULL;
1233     QemuOpts *opts;
1234     Error *local_err = NULL;
1235     const char *filename;
1236     int i, ret;
1237
1238     if ((BDRV_SECTOR_SIZE % 512) != 0) {
1239         error_setg(errp, "iSCSI: Invalid BDRV_SECTOR_SIZE. "
1240                    "BDRV_SECTOR_SIZE(%lld) is not a multiple "
1241                    "of 512", BDRV_SECTOR_SIZE);
1242         return -EINVAL;
1243     }
1244
1245     opts = qemu_opts_create(&runtime_opts, NULL, 0, &error_abort);
1246     qemu_opts_absorb_qdict(opts, options, &local_err);
1247     if (local_err) {
1248         error_propagate(errp, local_err);
1249         ret = -EINVAL;
1250         goto out;
1251     }
1252
1253     filename = qemu_opt_get(opts, "filename");
1254
1255     iscsi_url = iscsi_parse_full_url(iscsi, filename);
1256     if (iscsi_url == NULL) {
1257         error_setg(errp, "Failed to parse URL : %s", filename);
1258         ret = -EINVAL;
1259         goto out;
1260     }
1261
1262     memset(iscsilun, 0, sizeof(IscsiLun));
1263
1264     initiator_name = parse_initiator_name(iscsi_url->target);
1265
1266     iscsi = iscsi_create_context(initiator_name);
1267     if (iscsi == NULL) {
1268         error_setg(errp, "iSCSI: Failed to create iSCSI context.");
1269         ret = -ENOMEM;
1270         goto out;
1271     }
1272
1273     if (iscsi_set_targetname(iscsi, iscsi_url->target)) {
1274         error_setg(errp, "iSCSI: Failed to set target name.");
1275         ret = -EINVAL;
1276         goto out;
1277     }
1278
1279     if (iscsi_url->user != NULL) {
1280         ret = iscsi_set_initiator_username_pwd(iscsi, iscsi_url->user,
1281                                               iscsi_url->passwd);
1282         if (ret != 0) {
1283             error_setg(errp, "Failed to set initiator username and password");
1284             ret = -EINVAL;
1285             goto out;
1286         }
1287     }
1288
1289     /* check if we got CHAP username/password via the options */
1290     parse_chap(iscsi, iscsi_url->target, &local_err);
1291     if (local_err != NULL) {
1292         error_propagate(errp, local_err);
1293         ret = -EINVAL;
1294         goto out;
1295     }
1296
1297     if (iscsi_set_session_type(iscsi, ISCSI_SESSION_NORMAL) != 0) {
1298         error_setg(errp, "iSCSI: Failed to set session type to normal.");
1299         ret = -EINVAL;
1300         goto out;
1301     }
1302
1303     iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_NONE_CRC32C);
1304
1305     /* check if we got HEADER_DIGEST via the options */
1306     parse_header_digest(iscsi, iscsi_url->target, &local_err);
1307     if (local_err != NULL) {
1308         error_propagate(errp, local_err);
1309         ret = -EINVAL;
1310         goto out;
1311     }
1312
1313     if (iscsi_full_connect_sync(iscsi, iscsi_url->portal, iscsi_url->lun) != 0) {
1314         error_setg(errp, "iSCSI: Failed to connect to LUN : %s",
1315             iscsi_get_error(iscsi));
1316         ret = -EINVAL;
1317         goto out;
1318     }
1319
1320     iscsilun->iscsi = iscsi;
1321     iscsilun->aio_context = bdrv_get_aio_context(bs);
1322     iscsilun->lun   = iscsi_url->lun;
1323     iscsilun->has_write_same = true;
1324
1325     task = iscsi_do_inquiry(iscsilun->iscsi, iscsilun->lun, 0, 0,
1326                             (void **) &inq, errp);
1327     if (task == NULL) {
1328         ret = -EINVAL;
1329         goto out;
1330     }
1331     iscsilun->type = inq->periperal_device_type;
1332     scsi_free_scsi_task(task);
1333     task = NULL;
1334
1335     iscsi_readcapacity_sync(iscsilun, &local_err);
1336     if (local_err != NULL) {
1337         error_propagate(errp, local_err);
1338         ret = -EINVAL;
1339         goto out;
1340     }
1341     bs->total_sectors = sector_lun2qemu(iscsilun->num_blocks, iscsilun);
1342     bs->request_alignment = iscsilun->block_size;
1343
1344     /* We don't have any emulation for devices other than disks and CD-ROMs, so
1345      * this must be sg ioctl compatible. We force it to be sg, otherwise qemu
1346      * will try to read from the device to guess the image format.
1347      */
1348     if (iscsilun->type != TYPE_DISK && iscsilun->type != TYPE_ROM) {
1349         bs->sg = 1;
1350     }
1351
1352     task = iscsi_do_inquiry(iscsilun->iscsi, iscsilun->lun, 1,
1353                             SCSI_INQUIRY_PAGECODE_SUPPORTED_VPD_PAGES,
1354                             (void **) &inq_vpd, errp);
1355     if (task == NULL) {
1356         ret = -EINVAL;
1357         goto out;
1358     }
1359     for (i = 0; i < inq_vpd->num_pages; i++) {
1360         struct scsi_task *inq_task;
1361         struct scsi_inquiry_logical_block_provisioning *inq_lbp;
1362         struct scsi_inquiry_block_limits *inq_bl;
1363         switch (inq_vpd->pages[i]) {
1364         case SCSI_INQUIRY_PAGECODE_LOGICAL_BLOCK_PROVISIONING:
1365             inq_task = iscsi_do_inquiry(iscsilun->iscsi, iscsilun->lun, 1,
1366                                         SCSI_INQUIRY_PAGECODE_LOGICAL_BLOCK_PROVISIONING,
1367                                         (void **) &inq_lbp, errp);
1368             if (inq_task == NULL) {
1369                 ret = -EINVAL;
1370                 goto out;
1371             }
1372             memcpy(&iscsilun->lbp, inq_lbp,
1373                    sizeof(struct scsi_inquiry_logical_block_provisioning));
1374             scsi_free_scsi_task(inq_task);
1375             break;
1376         case SCSI_INQUIRY_PAGECODE_BLOCK_LIMITS:
1377             inq_task = iscsi_do_inquiry(iscsilun->iscsi, iscsilun->lun, 1,
1378                                     SCSI_INQUIRY_PAGECODE_BLOCK_LIMITS,
1379                                     (void **) &inq_bl, errp);
1380             if (inq_task == NULL) {
1381                 ret = -EINVAL;
1382                 goto out;
1383             }
1384             memcpy(&iscsilun->bl, inq_bl,
1385                    sizeof(struct scsi_inquiry_block_limits));
1386             scsi_free_scsi_task(inq_task);
1387             break;
1388         default:
1389             break;
1390         }
1391     }
1392     scsi_free_scsi_task(task);
1393     task = NULL;
1394
1395     iscsi_attach_aio_context(bs, iscsilun->aio_context);
1396
1397     /* Guess the internal cluster (page) size of the iscsi target by the means
1398      * of opt_unmap_gran. Transfer the unmap granularity only if it has a
1399      * reasonable size */
1400     if (iscsilun->bl.opt_unmap_gran * iscsilun->block_size >= 4 * 1024 &&
1401         iscsilun->bl.opt_unmap_gran * iscsilun->block_size <= 16 * 1024 * 1024) {
1402         iscsilun->cluster_sectors = (iscsilun->bl.opt_unmap_gran *
1403                                      iscsilun->block_size) >> BDRV_SECTOR_BITS;
1404         if (iscsilun->lbprz && !(bs->open_flags & BDRV_O_NOCACHE)) {
1405             iscsilun->allocationmap =
1406                 bitmap_new(DIV_ROUND_UP(bs->total_sectors,
1407                                         iscsilun->cluster_sectors));
1408         }
1409     }
1410
1411 out:
1412     qemu_opts_del(opts);
1413     g_free(initiator_name);
1414     if (iscsi_url != NULL) {
1415         iscsi_destroy_url(iscsi_url);
1416     }
1417     if (task != NULL) {
1418         scsi_free_scsi_task(task);
1419     }
1420
1421     if (ret) {
1422         if (iscsi != NULL) {
1423             iscsi_destroy_context(iscsi);
1424         }
1425         memset(iscsilun, 0, sizeof(IscsiLun));
1426     }
1427     return ret;
1428 }
1429
1430 static void iscsi_close(BlockDriverState *bs)
1431 {
1432     IscsiLun *iscsilun = bs->opaque;
1433     struct iscsi_context *iscsi = iscsilun->iscsi;
1434
1435     iscsi_detach_aio_context(bs);
1436     iscsi_destroy_context(iscsi);
1437     g_free(iscsilun->zeroblock);
1438     g_free(iscsilun->allocationmap);
1439     memset(iscsilun, 0, sizeof(IscsiLun));
1440 }
1441
1442 static void iscsi_refresh_limits(BlockDriverState *bs, Error **errp)
1443 {
1444     IscsiLun *iscsilun = bs->opaque;
1445
1446     /* We don't actually refresh here, but just return data queried in
1447      * iscsi_open(): iscsi targets don't change their limits. */
1448     if (iscsilun->lbp.lbpu) {
1449         if (iscsilun->bl.max_unmap < 0xffffffff) {
1450             bs->bl.max_discard = sector_lun2qemu(iscsilun->bl.max_unmap,
1451                                                  iscsilun);
1452         }
1453         bs->bl.discard_alignment = sector_lun2qemu(iscsilun->bl.opt_unmap_gran,
1454                                                    iscsilun);
1455     }
1456
1457     if (iscsilun->bl.max_ws_len < 0xffffffff) {
1458         bs->bl.max_write_zeroes = sector_lun2qemu(iscsilun->bl.max_ws_len,
1459                                                   iscsilun);
1460     }
1461     if (iscsilun->lbp.lbpws) {
1462         bs->bl.write_zeroes_alignment = sector_lun2qemu(iscsilun->bl.opt_unmap_gran,
1463                                                         iscsilun);
1464     }
1465     bs->bl.opt_transfer_length = sector_lun2qemu(iscsilun->bl.opt_xfer_len,
1466                                                  iscsilun);
1467 }
1468
1469 /* Since iscsi_open() ignores bdrv_flags, there is nothing to do here in
1470  * prepare.  Note that this will not re-establish a connection with an iSCSI
1471  * target - it is effectively a NOP.  */
1472 static int iscsi_reopen_prepare(BDRVReopenState *state,
1473                                 BlockReopenQueue *queue, Error **errp)
1474 {
1475     /* NOP */
1476     return 0;
1477 }
1478
1479 static int iscsi_truncate(BlockDriverState *bs, int64_t offset)
1480 {
1481     IscsiLun *iscsilun = bs->opaque;
1482     Error *local_err = NULL;
1483
1484     if (iscsilun->type != TYPE_DISK) {
1485         return -ENOTSUP;
1486     }
1487
1488     iscsi_readcapacity_sync(iscsilun, &local_err);
1489     if (local_err != NULL) {
1490         error_free(local_err);
1491         return -EIO;
1492     }
1493
1494     if (offset > iscsi_getlength(bs)) {
1495         return -EINVAL;
1496     }
1497
1498     if (iscsilun->allocationmap != NULL) {
1499         g_free(iscsilun->allocationmap);
1500         iscsilun->allocationmap =
1501             bitmap_new(DIV_ROUND_UP(sector_lun2qemu(iscsilun->num_blocks,
1502                                                     iscsilun),
1503                                     iscsilun->cluster_sectors));
1504     }
1505
1506     return 0;
1507 }
1508
1509 static int iscsi_create(const char *filename, QemuOpts *opts, Error **errp)
1510 {
1511     int ret = 0;
1512     int64_t total_size = 0;
1513     BlockDriverState *bs;
1514     IscsiLun *iscsilun = NULL;
1515     QDict *bs_options;
1516
1517     bs = bdrv_new("", &error_abort);
1518
1519     /* Read out options */
1520     total_size = DIV_ROUND_UP(qemu_opt_get_size_del(opts, BLOCK_OPT_SIZE, 0),
1521                               BDRV_SECTOR_SIZE);
1522     bs->opaque = g_new0(struct IscsiLun, 1);
1523     iscsilun = bs->opaque;
1524
1525     bs_options = qdict_new();
1526     qdict_put(bs_options, "filename", qstring_from_str(filename));
1527     ret = iscsi_open(bs, bs_options, 0, NULL);
1528     QDECREF(bs_options);
1529
1530     if (ret != 0) {
1531         goto out;
1532     }
1533     iscsi_detach_aio_context(bs);
1534     if (iscsilun->type != TYPE_DISK) {
1535         ret = -ENODEV;
1536         goto out;
1537     }
1538     if (bs->total_sectors < total_size) {
1539         ret = -ENOSPC;
1540         goto out;
1541     }
1542
1543     ret = 0;
1544 out:
1545     if (iscsilun->iscsi != NULL) {
1546         iscsi_destroy_context(iscsilun->iscsi);
1547     }
1548     g_free(bs->opaque);
1549     bs->opaque = NULL;
1550     bdrv_unref(bs);
1551     return ret;
1552 }
1553
1554 static int iscsi_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
1555 {
1556     IscsiLun *iscsilun = bs->opaque;
1557     bdi->unallocated_blocks_are_zero = !!iscsilun->lbprz;
1558     bdi->can_write_zeroes_with_unmap = iscsilun->lbprz && iscsilun->lbp.lbpws;
1559     bdi->cluster_size = iscsilun->cluster_sectors * BDRV_SECTOR_SIZE;
1560     return 0;
1561 }
1562
1563 static QemuOptsList iscsi_create_opts = {
1564     .name = "iscsi-create-opts",
1565     .head = QTAILQ_HEAD_INITIALIZER(iscsi_create_opts.head),
1566     .desc = {
1567         {
1568             .name = BLOCK_OPT_SIZE,
1569             .type = QEMU_OPT_SIZE,
1570             .help = "Virtual disk size"
1571         },
1572         { /* end of list */ }
1573     }
1574 };
1575
1576 static BlockDriver bdrv_iscsi = {
1577     .format_name     = "iscsi",
1578     .protocol_name   = "iscsi",
1579
1580     .instance_size   = sizeof(IscsiLun),
1581     .bdrv_needs_filename = true,
1582     .bdrv_file_open  = iscsi_open,
1583     .bdrv_close      = iscsi_close,
1584     .bdrv_create     = iscsi_create,
1585     .create_opts     = &iscsi_create_opts,
1586     .bdrv_reopen_prepare  = iscsi_reopen_prepare,
1587
1588     .bdrv_getlength  = iscsi_getlength,
1589     .bdrv_get_info   = iscsi_get_info,
1590     .bdrv_truncate   = iscsi_truncate,
1591     .bdrv_refresh_limits = iscsi_refresh_limits,
1592
1593     .bdrv_co_get_block_status = iscsi_co_get_block_status,
1594     .bdrv_co_discard      = iscsi_co_discard,
1595     .bdrv_co_write_zeroes = iscsi_co_write_zeroes,
1596     .bdrv_co_readv         = iscsi_co_readv,
1597     .bdrv_co_writev        = iscsi_co_writev,
1598     .bdrv_co_flush_to_disk = iscsi_co_flush,
1599
1600 #ifdef __linux__
1601     .bdrv_ioctl       = iscsi_ioctl,
1602     .bdrv_aio_ioctl   = iscsi_aio_ioctl,
1603 #endif
1604
1605     .bdrv_detach_aio_context = iscsi_detach_aio_context,
1606     .bdrv_attach_aio_context = iscsi_attach_aio_context,
1607 };
1608
1609 static QemuOptsList qemu_iscsi_opts = {
1610     .name = "iscsi",
1611     .head = QTAILQ_HEAD_INITIALIZER(qemu_iscsi_opts.head),
1612     .desc = {
1613         {
1614             .name = "user",
1615             .type = QEMU_OPT_STRING,
1616             .help = "username for CHAP authentication to target",
1617         },{
1618             .name = "password",
1619             .type = QEMU_OPT_STRING,
1620             .help = "password for CHAP authentication to target",
1621         },{
1622             .name = "header-digest",
1623             .type = QEMU_OPT_STRING,
1624             .help = "HeaderDigest setting. "
1625                     "{CRC32C|CRC32C-NONE|NONE-CRC32C|NONE}",
1626         },{
1627             .name = "initiator-name",
1628             .type = QEMU_OPT_STRING,
1629             .help = "Initiator iqn name to use when connecting",
1630         },
1631         { /* end of list */ }
1632     },
1633 };
1634
1635 static void iscsi_block_init(void)
1636 {
1637     bdrv_register(&bdrv_iscsi);
1638     qemu_add_opts(&qemu_iscsi_opts);
1639 }
1640
1641 block_init(iscsi_block_init);