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     BlockAIOCB 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(BlockAIOCB *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 unsigned long *iscsi_allocationmap_init(IscsiLun *iscsilun)
 320 {
 321     return bitmap_try_new(DIV_ROUND_UP(sector_lun2qemu(iscsilun->num_blocks,
 322                                                        iscsilun),
 323                                        iscsilun->cluster_sectors));
 324 }
 325
 326 static void iscsi_allocationmap_set(IscsiLun *iscsilun, int64_t sector_num,
 327                                     int nb_sectors)
 328 {
 329     if (iscsilun->allocationmap == NULL) {
 330         return;
 331     }
 332     bitmap_set(iscsilun->allocationmap,
 333                sector_num / iscsilun->cluster_sectors,
 334                DIV_ROUND_UP(nb_sectors, iscsilun->cluster_sectors));
 335 }
 336
 337 static void iscsi_allocationmap_clear(IscsiLun *iscsilun, int64_t sector_num,
 338                                       int nb_sectors)
 339 {
 340     int64_t cluster_num, nb_clusters;
 341     if (iscsilun->allocationmap == NULL) {
 342         return;
 343     }
 344     cluster_num = DIV_ROUND_UP(sector_num, iscsilun->cluster_sectors);
 345     nb_clusters = (sector_num + nb_sectors) / iscsilun->cluster_sectors
 346                   - cluster_num;
 347     if (nb_clusters > 0) {
 348         bitmap_clear(iscsilun->allocationmap, cluster_num, nb_clusters);
 349     }
 350 }
 351
 352 static int coroutine_fn iscsi_co_writev(BlockDriverState *bs,
 353                                         int64_t sector_num, int nb_sectors,
 354                                         QEMUIOVector *iov)
 355 {
 356     IscsiLun *iscsilun = bs->opaque;
 357     struct IscsiTask iTask;
 358     uint64_t lba;
 359     uint32_t num_sectors;
 360
 361     if (!is_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
 362         return -EINVAL;
 363     }
 364
 365     lba = sector_qemu2lun(sector_num, iscsilun);
 366     num_sectors = sector_qemu2lun(nb_sectors, iscsilun);
 367     iscsi_co_init_iscsitask(iscsilun, &iTask);
 368 retry:
 369     if (iscsilun->use_16_for_rw) {
 370         iTask.task = iscsi_write16_task(iscsilun->iscsi, iscsilun->lun, lba,
 371                                         NULL, num_sectors * iscsilun->block_size,
 372                                         iscsilun->block_size, 0, 0, 0, 0, 0,
 373                                         iscsi_co_generic_cb, &iTask);
 374     } else {
 375         iTask.task = iscsi_write10_task(iscsilun->iscsi, iscsilun->lun, lba,
 376                                         NULL, num_sectors * iscsilun->block_size,
 377                                         iscsilun->block_size, 0, 0, 0, 0, 0,
 378                                         iscsi_co_generic_cb, &iTask);
 379     }
 380     if (iTask.task == NULL) {
 381         return -ENOMEM;
 382     }
 383     scsi_task_set_iov_out(iTask.task, (struct scsi_iovec *) iov->iov,
 384                           iov->niov);
 385     while (!iTask.complete) {
 386         iscsi_set_events(iscsilun);
 387         qemu_coroutine_yield();
 388     }
 389
 390     if (iTask.task != NULL) {
 391         scsi_free_scsi_task(iTask.task);
 392         iTask.task = NULL;
 393     }
 394
 395     if (iTask.do_retry) {
 396         iTask.complete = 0;
 397         goto retry;
 398     }
 399
 400     if (iTask.status != SCSI_STATUS_GOOD) {
 401         return -EIO;
 402     }
 403
 404     iscsi_allocationmap_set(iscsilun, sector_num, nb_sectors);
 405
 406     return 0;
 407 }
 408
 409
 410 static bool iscsi_allocationmap_is_allocated(IscsiLun *iscsilun,
 411                                              int64_t sector_num, int nb_sectors)
 412 {
 413     unsigned long size;
 414     if (iscsilun->allocationmap == NULL) {
 415         return true;
 416     }
 417     size = DIV_ROUND_UP(sector_num + nb_sectors, iscsilun->cluster_sectors);
 418     return !(find_next_bit(iscsilun->allocationmap, size,
 419                            sector_num / iscsilun->cluster_sectors) == size);
 420 }
 421
 422 static int64_t coroutine_fn iscsi_co_get_block_status(BlockDriverState *bs,
 423                                                   int64_t sector_num,
 424                                                   int nb_sectors, int *pnum)
 425 {
 426     IscsiLun *iscsilun = bs->opaque;
 427     struct scsi_get_lba_status *lbas = NULL;
 428     struct scsi_lba_status_descriptor *lbasd = NULL;
 429     struct IscsiTask iTask;
 430     int64_t ret;
 431
 432     iscsi_co_init_iscsitask(iscsilun, &iTask);
 433
 434     if (!is_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
 435         ret = -EINVAL;
 436         goto out;
 437     }
 438
 439     /* default to all sectors allocated */
 440     ret = BDRV_BLOCK_DATA;
 441     ret |= (sector_num << BDRV_SECTOR_BITS) | BDRV_BLOCK_OFFSET_VALID;
 442     *pnum = nb_sectors;
 443
 444     /* LUN does not support logical block provisioning */
 445     if (iscsilun->lbpme == 0) {
 446         goto out;
 447     }
 448
 449 retry:
 450     if (iscsi_get_lba_status_task(iscsilun->iscsi, iscsilun->lun,
 451                                   sector_qemu2lun(sector_num, iscsilun),
 452                                   8 + 16, iscsi_co_generic_cb,
 453                                   &iTask) == NULL) {
 454         ret = -ENOMEM;
 455         goto out;
 456     }
 457
 458     while (!iTask.complete) {
 459         iscsi_set_events(iscsilun);
 460         qemu_coroutine_yield();
 461     }
 462
 463     if (iTask.do_retry) {
 464         if (iTask.task != NULL) {
 465             scsi_free_scsi_task(iTask.task);
 466             iTask.task = NULL;
 467         }
 468         iTask.complete = 0;
 469         goto retry;
 470     }
 471
 472     if (iTask.status != SCSI_STATUS_GOOD) {
 473         /* in case the get_lba_status_callout fails (i.e.
 474          * because the device is busy or the cmd is not
 475          * supported) we pretend all blocks are allocated
 476          * for backwards compatibility */
 477         goto out;
 478     }
 479
 480     lbas = scsi_datain_unmarshall(iTask.task);
 481     if (lbas == NULL) {
 482         ret = -EIO;
 483         goto out;
 484     }
 485
 486     lbasd = &lbas->descriptors[0];
 487
 488     if (sector_qemu2lun(sector_num, iscsilun) != lbasd->lba) {
 489         ret = -EIO;
 490         goto out;
 491     }
 492
 493     *pnum = sector_lun2qemu(lbasd->num_blocks, iscsilun);
 494
 495     if (lbasd->provisioning == SCSI_PROVISIONING_TYPE_DEALLOCATED ||
 496         lbasd->provisioning == SCSI_PROVISIONING_TYPE_ANCHORED) {
 497         ret &= ~BDRV_BLOCK_DATA;
 498         if (iscsilun->lbprz) {
 499             ret |= BDRV_BLOCK_ZERO;
 500         }
 501     }
 502
 503     if (ret & BDRV_BLOCK_ZERO) {
 504         iscsi_allocationmap_clear(iscsilun, sector_num, *pnum);
 505     } else {
 506         iscsi_allocationmap_set(iscsilun, sector_num, *pnum);
 507     }
 508
 509     if (*pnum > nb_sectors) {
 510         *pnum = nb_sectors;
 511     }
 512 out:
 513     if (iTask.task != NULL) {
 514         scsi_free_scsi_task(iTask.task);
 515     }
 516     return ret;
 517 }
 518
 519 static int coroutine_fn iscsi_co_readv(BlockDriverState *bs,
 520                                        int64_t sector_num, int nb_sectors,
 521                                        QEMUIOVector *iov)
 522 {
 523     IscsiLun *iscsilun = bs->opaque;
 524     struct IscsiTask iTask;
 525     uint64_t lba;
 526     uint32_t num_sectors;
 527
 528     if (!is_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
 529         return -EINVAL;
 530     }
 531
 532     if (iscsilun->lbprz && nb_sectors >= ISCSI_CHECKALLOC_THRES &&
 533         !iscsi_allocationmap_is_allocated(iscsilun, sector_num, nb_sectors)) {
 534         int64_t ret;
 535         int pnum;
 536         ret = iscsi_co_get_block_status(bs, sector_num, INT_MAX, &pnum);
 537         if (ret < 0) {
 538             return ret;
 539         }
 540         if (ret & BDRV_BLOCK_ZERO && pnum >= nb_sectors) {
 541             qemu_iovec_memset(iov, 0, 0x00, iov->size);
 542             return 0;
 543         }
 544     }
 545
 546     lba = sector_qemu2lun(sector_num, iscsilun);
 547     num_sectors = sector_qemu2lun(nb_sectors, iscsilun);
 548
 549     iscsi_co_init_iscsitask(iscsilun, &iTask);
 550 retry:
 551     if (iscsilun->use_16_for_rw) {
 552         iTask.task = iscsi_read16_task(iscsilun->iscsi, iscsilun->lun, lba,
 553                                        num_sectors * iscsilun->block_size,
 554                                        iscsilun->block_size, 0, 0, 0, 0, 0,
 555                                        iscsi_co_generic_cb, &iTask);
 556     } else {
 557         iTask.task = iscsi_read10_task(iscsilun->iscsi, iscsilun->lun, lba,
 558                                        num_sectors * iscsilun->block_size,
 559                                        iscsilun->block_size,
 560                                        0, 0, 0, 0, 0,
 561                                        iscsi_co_generic_cb, &iTask);
 562     }
 563     if (iTask.task == NULL) {
 564         return -ENOMEM;
 565     }
 566     scsi_task_set_iov_in(iTask.task, (struct scsi_iovec *) iov->iov, iov->niov);
 567
 568     while (!iTask.complete) {
 569         iscsi_set_events(iscsilun);
 570         qemu_coroutine_yield();
 571     }
 572
 573     if (iTask.task != NULL) {
 574         scsi_free_scsi_task(iTask.task);
 575         iTask.task = NULL;
 576     }
 577
 578     if (iTask.do_retry) {
 579         iTask.complete = 0;
 580         goto retry;
 581     }
 582
 583     if (iTask.status != SCSI_STATUS_GOOD) {
 584         return -EIO;
 585     }
 586
 587     return 0;
 588 }
 589
 590 static int coroutine_fn iscsi_co_flush(BlockDriverState *bs)
 591 {
 592     IscsiLun *iscsilun = bs->opaque;
 593     struct IscsiTask iTask;
 594
 595     if (bs->sg) {
 596         return 0;
 597     }
 598
 599     iscsi_co_init_iscsitask(iscsilun, &iTask);
 600
 601 retry:
 602     if (iscsi_synchronizecache10_task(iscsilun->iscsi, iscsilun->lun, 0, 0, 0,
 603                                       0, iscsi_co_generic_cb, &iTask) == NULL) {
 604         return -ENOMEM;
 605     }
 606
 607     while (!iTask.complete) {
 608         iscsi_set_events(iscsilun);
 609         qemu_coroutine_yield();
 610     }
 611
 612     if (iTask.task != NULL) {
 613         scsi_free_scsi_task(iTask.task);
 614         iTask.task = NULL;
 615     }
 616
 617     if (iTask.do_retry) {
 618         iTask.complete = 0;
 619         goto retry;
 620     }
 621
 622     if (iTask.status != SCSI_STATUS_GOOD) {
 623         return -EIO;
 624     }
 625
 626     return 0;
 627 }
 628
 629 #ifdef __linux__
 630 static void
 631 iscsi_aio_ioctl_cb(struct iscsi_context *iscsi, int status,
 632                      void *command_data, void *opaque)
 633 {
 634     IscsiAIOCB *acb = opaque;
 635
 636     g_free(acb->buf);
 637     acb->buf = NULL;
 638
 639     acb->status = 0;
 640     if (status < 0) {
 641         error_report("Failed to ioctl(SG_IO) to iSCSI lun. %s",
 642                      iscsi_get_error(iscsi));
 643         acb->status = -EIO;
 644     }
 645
 646     acb->ioh->driver_status = 0;
 647     acb->ioh->host_status   = 0;
 648     acb->ioh->resid         = 0;
 649
 650 #define SG_ERR_DRIVER_SENSE    0x08
 651
 652     if (status == SCSI_STATUS_CHECK_CONDITION && acb->task->datain.size >= 2) {
 653         int ss;
 654
 655         acb->ioh->driver_status |= SG_ERR_DRIVER_SENSE;
 656
 657         acb->ioh->sb_len_wr = acb->task->datain.size - 2;
 658         ss = (acb->ioh->mx_sb_len >= acb->ioh->sb_len_wr) ?
 659              acb->ioh->mx_sb_len : acb->ioh->sb_len_wr;
 660         memcpy(acb->ioh->sbp, &acb->task->datain.data[2], ss);
 661     }
 662
 663     iscsi_schedule_bh(acb);
 664 }
 665
 666 static BlockAIOCB *iscsi_aio_ioctl(BlockDriverState *bs,
 667         unsigned long int req, void *buf,
 668         BlockCompletionFunc *cb, void *opaque)
 669 {
 670     IscsiLun *iscsilun = bs->opaque;
 671     struct iscsi_context *iscsi = iscsilun->iscsi;
 672     struct iscsi_data data;
 673     IscsiAIOCB *acb;
 674
 675     assert(req == SG_IO);
 676
 677     acb = qemu_aio_get(&iscsi_aiocb_info, bs, cb, opaque);
 678
 679     acb->iscsilun = iscsilun;
 680     acb->bh          = NULL;
 681     acb->status      = -EINPROGRESS;
 682     acb->buf         = NULL;
 683     acb->ioh         = buf;
 684
 685     acb->task = malloc(sizeof(struct scsi_task));
 686     if (acb->task == NULL) {
 687         error_report("iSCSI: Failed to allocate task for scsi command. %s",
 688                      iscsi_get_error(iscsi));
 689         qemu_aio_unref(acb);
 690         return NULL;
 691     }
 692     memset(acb->task, 0, sizeof(struct scsi_task));
 693
 694     switch (acb->ioh->dxfer_direction) {
 695     case SG_DXFER_TO_DEV:
 696         acb->task->xfer_dir = SCSI_XFER_WRITE;
 697         break;
 698     case SG_DXFER_FROM_DEV:
 699         acb->task->xfer_dir = SCSI_XFER_READ;
 700         break;
 701     default:
 702         acb->task->xfer_dir = SCSI_XFER_NONE;
 703         break;
 704     }
 705
 706     acb->task->cdb_size = acb->ioh->cmd_len;
 707     memcpy(&acb->task->cdb[0], acb->ioh->cmdp, acb->ioh->cmd_len);
 708     acb->task->expxferlen = acb->ioh->dxfer_len;
 709
 710     data.size = 0;
 711     if (acb->task->xfer_dir == SCSI_XFER_WRITE) {
 712         if (acb->ioh->iovec_count == 0) {
 713             data.data = acb->ioh->dxferp;
 714             data.size = acb->ioh->dxfer_len;
 715         } else {
 716             scsi_task_set_iov_out(acb->task,
 717                                  (struct scsi_iovec *) acb->ioh->dxferp,
 718                                  acb->ioh->iovec_count);
 719         }
 720     }
 721
 722     if (iscsi_scsi_command_async(iscsi, iscsilun->lun, acb->task,
 723                                  iscsi_aio_ioctl_cb,
 724                                  (data.size > 0) ? &data : NULL,
 725                                  acb) != 0) {
 726         scsi_free_scsi_task(acb->task);
 727         qemu_aio_unref(acb);
 728         return NULL;
 729     }
 730
 731     /* tell libiscsi to read straight into the buffer we got from ioctl */
 732     if (acb->task->xfer_dir == SCSI_XFER_READ) {
 733         if (acb->ioh->iovec_count == 0) {
 734             scsi_task_add_data_in_buffer(acb->task,
 735                                          acb->ioh->dxfer_len,
 736                                          acb->ioh->dxferp);
 737         } else {
 738             scsi_task_set_iov_in(acb->task,
 739                                  (struct scsi_iovec *) acb->ioh->dxferp,
 740                                  acb->ioh->iovec_count);
 741         }
 742     }
 743
 744     iscsi_set_events(iscsilun);
 745
 746     return &acb->common;
 747 }
 748
 749 static void ioctl_cb(void *opaque, int status)
 750 {
 751     int *p_status = opaque;
 752     *p_status = status;
 753 }
 754
 755 static int iscsi_ioctl(BlockDriverState *bs, unsigned long int req, void *buf)
 756 {
 757     IscsiLun *iscsilun = bs->opaque;
 758     int status;
 759
 760     switch (req) {
 761     case SG_GET_VERSION_NUM:
 762         *(int *)buf = 30000;
 763         break;
 764     case SG_GET_SCSI_ID:
 765         ((struct sg_scsi_id *)buf)->scsi_type = iscsilun->type;
 766         break;
 767     case SG_IO:
 768         status = -EINPROGRESS;
 769         iscsi_aio_ioctl(bs, req, buf, ioctl_cb, &status);
 770
 771         while (status == -EINPROGRESS) {
 772             aio_poll(iscsilun->aio_context, true);
 773         }
 774
 775         return 0;
 776     default:
 777         return -1;
 778     }
 779     return 0;
 780 }
 781 #endif
 782
 783 static int64_t
 784 iscsi_getlength(BlockDriverState *bs)
 785 {
 786     IscsiLun *iscsilun = bs->opaque;
 787     int64_t len;
 788
 789     len  = iscsilun->num_blocks;
 790     len *= iscsilun->block_size;
 791
 792     return len;
 793 }
 794
 795 static int
 796 coroutine_fn iscsi_co_discard(BlockDriverState *bs, int64_t sector_num,
 797                                    int nb_sectors)
 798 {
 799     IscsiLun *iscsilun = bs->opaque;
 800     struct IscsiTask iTask;
 801     struct unmap_list list;
 802
 803     if (!is_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
 804         return -EINVAL;
 805     }
 806
 807     if (!iscsilun->lbp.lbpu) {
 808         /* UNMAP is not supported by the target */
 809         return 0;
 810     }
 811
 812     list.lba = sector_qemu2lun(sector_num, iscsilun);
 813     list.num = sector_qemu2lun(nb_sectors, iscsilun);
 814
 815     iscsi_co_init_iscsitask(iscsilun, &iTask);
 816 retry:
 817     if (iscsi_unmap_task(iscsilun->iscsi, iscsilun->lun, 0, 0, &list, 1,
 818                      iscsi_co_generic_cb, &iTask) == NULL) {
 819         return -ENOMEM;
 820     }
 821
 822     while (!iTask.complete) {
 823         iscsi_set_events(iscsilun);
 824         qemu_coroutine_yield();
 825     }
 826
 827     if (iTask.task != NULL) {
 828         scsi_free_scsi_task(iTask.task);
 829         iTask.task = NULL;
 830     }
 831
 832     if (iTask.do_retry) {
 833         iTask.complete = 0;
 834         goto retry;
 835     }
 836
 837     if (iTask.status == SCSI_STATUS_CHECK_CONDITION) {
 838         /* the target might fail with a check condition if it
 839            is not happy with the alignment of the UNMAP request
 840            we silently fail in this case */
 841         return 0;
 842     }
 843
 844     if (iTask.status != SCSI_STATUS_GOOD) {
 845         return -EIO;
 846     }
 847
 848     iscsi_allocationmap_clear(iscsilun, sector_num, nb_sectors);
 849
 850     return 0;
 851 }
 852
 853 static int
 854 coroutine_fn iscsi_co_write_zeroes(BlockDriverState *bs, int64_t sector_num,
 855                                    int nb_sectors, BdrvRequestFlags flags)
 856 {
 857     IscsiLun *iscsilun = bs->opaque;
 858     struct IscsiTask iTask;
 859     uint64_t lba;
 860     uint32_t nb_blocks;
 861     bool use_16_for_ws = iscsilun->use_16_for_rw;
 862
 863     if (!is_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
 864         return -EINVAL;
 865     }
 866
 867     if (flags & BDRV_REQ_MAY_UNMAP) {
 868         if (!use_16_for_ws && !iscsilun->lbp.lbpws10) {
 869             /* WRITESAME10 with UNMAP is unsupported try WRITESAME16 */
 870             use_16_for_ws = true;
 871         }
 872         if (use_16_for_ws && !iscsilun->lbp.lbpws) {
 873             /* WRITESAME16 with UNMAP is not supported by the target,
 874              * fall back and try WRITESAME10/16 without UNMAP */
 875             flags &= ~BDRV_REQ_MAY_UNMAP;
 876             use_16_for_ws = iscsilun->use_16_for_rw;
 877         }
 878     }
 879
 880     if (!(flags & BDRV_REQ_MAY_UNMAP) && !iscsilun->has_write_same) {
 881         /* WRITESAME without UNMAP is not supported by the target */
 882         return -ENOTSUP;
 883     }
 884
 885     lba = sector_qemu2lun(sector_num, iscsilun);
 886     nb_blocks = sector_qemu2lun(nb_sectors, iscsilun);
 887
 888     if (iscsilun->zeroblock == NULL) {
 889         iscsilun->zeroblock = g_try_malloc0(iscsilun->block_size);
 890         if (iscsilun->zeroblock == NULL) {
 891             return -ENOMEM;
 892         }
 893     }
 894
 895     iscsi_co_init_iscsitask(iscsilun, &iTask);
 896 retry:
 897     if (use_16_for_ws) {
 898         iTask.task = iscsi_writesame16_task(iscsilun->iscsi, iscsilun->lun, lba,
 899                                             iscsilun->zeroblock, iscsilun->block_size,
 900                                             nb_blocks, 0, !!(flags & BDRV_REQ_MAY_UNMAP),
 901                                             0, 0, iscsi_co_generic_cb, &iTask);
 902     } else {
 903         iTask.task = iscsi_writesame10_task(iscsilun->iscsi, iscsilun->lun, lba,
 904                                             iscsilun->zeroblock, iscsilun->block_size,
 905                                             nb_blocks, 0, !!(flags & BDRV_REQ_MAY_UNMAP),
 906                                             0, 0, iscsi_co_generic_cb, &iTask);
 907     }
 908     if (iTask.task == NULL) {
 909         return -ENOMEM;
 910     }
 911
 912     while (!iTask.complete) {
 913         iscsi_set_events(iscsilun);
 914         qemu_coroutine_yield();
 915     }
 916
 917     if (iTask.status == SCSI_STATUS_CHECK_CONDITION &&
 918         iTask.task->sense.key == SCSI_SENSE_ILLEGAL_REQUEST &&
 919         (iTask.task->sense.ascq == SCSI_SENSE_ASCQ_INVALID_OPERATION_CODE ||
 920          iTask.task->sense.ascq == SCSI_SENSE_ASCQ_INVALID_FIELD_IN_CDB)) {
 921         /* WRITE SAME is not supported by the target */
 922         iscsilun->has_write_same = false;
 923         scsi_free_scsi_task(iTask.task);
 924         return -ENOTSUP;
 925     }
 926
 927     if (iTask.task != NULL) {
 928         scsi_free_scsi_task(iTask.task);
 929         iTask.task = NULL;
 930     }
 931
 932     if (iTask.do_retry) {
 933         iTask.complete = 0;
 934         goto retry;
 935     }
 936
 937     if (iTask.status != SCSI_STATUS_GOOD) {
 938         return -EIO;
 939     }
 940
 941     if (flags & BDRV_REQ_MAY_UNMAP) {
 942         iscsi_allocationmap_clear(iscsilun, sector_num, nb_sectors);
 943     } else {
 944         iscsi_allocationmap_set(iscsilun, sector_num, nb_sectors);
 945     }
 946
 947     return 0;
 948 }
 949
 950 static void parse_chap(struct iscsi_context *iscsi, const char *target,
 951                        Error **errp)
 952 {
 953     QemuOptsList *list;
 954     QemuOpts *opts;
 955     const char *user = NULL;
 956     const char *password = NULL;
 957
 958     list = qemu_find_opts("iscsi");
 959     if (!list) {
 960         return;
 961     }
 962
 963     opts = qemu_opts_find(list, target);
 964     if (opts == NULL) {
 965         opts = QTAILQ_FIRST(&list->head);
 966         if (!opts) {
 967             return;
 968         }
 969     }
 970
 971     user = qemu_opt_get(opts, "user");
 972     if (!user) {
 973         return;
 974     }
 975
 976     password = qemu_opt_get(opts, "password");
 977     if (!password) {
 978         error_setg(errp, "CHAP username specified but no password was given");
 979         return;
 980     }
 981
 982     if (iscsi_set_initiator_username_pwd(iscsi, user, password)) {
 983         error_setg(errp, "Failed to set initiator username and password");
 984     }
 985 }
 986
 987 static void parse_header_digest(struct iscsi_context *iscsi, const char *target,
 988                                 Error **errp)
 989 {
 990     QemuOptsList *list;
 991     QemuOpts *opts;
 992     const char *digest = NULL;
 993
 994     list = qemu_find_opts("iscsi");
 995     if (!list) {
 996         return;
 997     }
 998
 999     opts = qemu_opts_find(list, target);
1000     if (opts == NULL) {
1001         opts = QTAILQ_FIRST(&list->head);
1002         if (!opts) {
1003             return;
1004         }
1005     }
1006
1007     digest = qemu_opt_get(opts, "header-digest");
1008     if (!digest) {
1009         return;
1010     }
1011
1012     if (!strcmp(digest, "CRC32C")) {
1013         iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_CRC32C);
1014     } else if (!strcmp(digest, "NONE")) {
1015         iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_NONE);
1016     } else if (!strcmp(digest, "CRC32C-NONE")) {
1017         iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_CRC32C_NONE);
1018     } else if (!strcmp(digest, "NONE-CRC32C")) {
1019         iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_NONE_CRC32C);
1020     } else {
1021         error_setg(errp, "Invalid header-digest setting : %s", digest);
1022     }
1023 }
1024
1025 static char *parse_initiator_name(const char *target)
1026 {
1027     QemuOptsList *list;
1028     QemuOpts *opts;
1029     const char *name;
1030     char *iscsi_name;
1031     UuidInfo *uuid_info;
1032
1033     list = qemu_find_opts("iscsi");
1034     if (list) {
1035         opts = qemu_opts_find(list, target);
1036         if (!opts) {
1037             opts = QTAILQ_FIRST(&list->head);
1038         }
1039         if (opts) {
1040             name = qemu_opt_get(opts, "initiator-name");
1041             if (name) {
1042                 return g_strdup(name);
1043             }
1044         }
1045     }
1046
1047     uuid_info = qmp_query_uuid(NULL);
1048     if (strcmp(uuid_info->UUID, UUID_NONE) == 0) {
1049         name = qemu_get_vm_name();
1050     } else {
1051         name = uuid_info->UUID;
1052     }
1053     iscsi_name = g_strdup_printf("iqn.2008-11.org.linux-kvm%s%s",
1054                                  name ? ":" : "", name ? name : "");
1055     qapi_free_UuidInfo(uuid_info);
1056     return iscsi_name;
1057 }
1058
1059 static void iscsi_nop_timed_event(void *opaque)
1060 {
1061     IscsiLun *iscsilun = opaque;
1062
1063     if (iscsi_get_nops_in_flight(iscsilun->iscsi) > MAX_NOP_FAILURES) {
1064         error_report("iSCSI: NOP timeout. Reconnecting...");
1065         iscsi_reconnect(iscsilun->iscsi);
1066     }
1067
1068     if (iscsi_nop_out_async(iscsilun->iscsi, NULL, NULL, 0, NULL) != 0) {
1069         error_report("iSCSI: failed to sent NOP-Out. Disabling NOP messages.");
1070         return;
1071     }
1072
1073     timer_mod(iscsilun->nop_timer, qemu_clock_get_ms(QEMU_CLOCK_REALTIME) + NOP_INTERVAL);
1074     iscsi_set_events(iscsilun);
1075 }
1076
1077 static void iscsi_readcapacity_sync(IscsiLun *iscsilun, Error **errp)
1078 {
1079     struct scsi_task *task = NULL;
1080     struct scsi_readcapacity10 *rc10 = NULL;
1081     struct scsi_readcapacity16 *rc16 = NULL;
1082     int retries = ISCSI_CMD_RETRIES;
1083
1084     do {
1085         if (task != NULL) {
1086             scsi_free_scsi_task(task);
1087             task = NULL;
1088         }
1089
1090         switch (iscsilun->type) {
1091         case TYPE_DISK:
1092             task = iscsi_readcapacity16_sync(iscsilun->iscsi, iscsilun->lun);
1093             if (task != NULL && task->status == SCSI_STATUS_GOOD) {
1094                 rc16 = scsi_datain_unmarshall(task);
1095                 if (rc16 == NULL) {
1096                     error_setg(errp, "iSCSI: Failed to unmarshall readcapacity16 data.");
1097                 } else {
1098                     iscsilun->block_size = rc16->block_length;
1099                     iscsilun->num_blocks = rc16->returned_lba + 1;
1100                     iscsilun->lbpme = rc16->lbpme;
1101                     iscsilun->lbprz = rc16->lbprz;
1102                     iscsilun->use_16_for_rw = (rc16->returned_lba > 0xffffffff);
1103                 }
1104             }
1105             break;
1106         case TYPE_ROM:
1107             task = iscsi_readcapacity10_sync(iscsilun->iscsi, iscsilun->lun, 0, 0);
1108             if (task != NULL && task->status == SCSI_STATUS_GOOD) {
1109                 rc10 = scsi_datain_unmarshall(task);
1110                 if (rc10 == NULL) {
1111                     error_setg(errp, "iSCSI: Failed to unmarshall readcapacity10 data.");
1112                 } else {
1113                     iscsilun->block_size = rc10->block_size;
1114                     if (rc10->lba == 0) {
1115                         /* blank disk loaded */
1116                         iscsilun->num_blocks = 0;
1117                     } else {
1118                         iscsilun->num_blocks = rc10->lba + 1;
1119                     }
1120                 }
1121             }
1122             break;
1123         default:
1124             return;
1125         }
1126     } while (task != NULL && task->status == SCSI_STATUS_CHECK_CONDITION
1127              && task->sense.key == SCSI_SENSE_UNIT_ATTENTION
1128              && retries-- > 0);
1129
1130     if (task == NULL || task->status != SCSI_STATUS_GOOD) {
1131         error_setg(errp, "iSCSI: failed to send readcapacity10 command.");
1132     }
1133     if (task) {
1134         scsi_free_scsi_task(task);
1135     }
1136 }
1137
1138 /* TODO Convert to fine grained options */
1139 static QemuOptsList runtime_opts = {
1140     .name = "iscsi",
1141     .head = QTAILQ_HEAD_INITIALIZER(runtime_opts.head),
1142     .desc = {
1143         {
1144             .name = "filename",
1145             .type = QEMU_OPT_STRING,
1146             .help = "URL to the iscsi image",
1147         },
1148         { /* end of list */ }
1149     },
1150 };
1151
1152 static struct scsi_task *iscsi_do_inquiry(struct iscsi_context *iscsi, int lun,
1153                                           int evpd, int pc, void **inq, Error **errp)
1154 {
1155     int full_size;
1156     struct scsi_task *task = NULL;
1157     task = iscsi_inquiry_sync(iscsi, lun, evpd, pc, 64);
1158     if (task == NULL || task->status != SCSI_STATUS_GOOD) {
1159         goto fail;
1160     }
1161     full_size = scsi_datain_getfullsize(task);
1162     if (full_size > task->datain.size) {
1163         scsi_free_scsi_task(task);
1164
1165         /* we need more data for the full list */
1166         task = iscsi_inquiry_sync(iscsi, lun, evpd, pc, full_size);
1167         if (task == NULL || task->status != SCSI_STATUS_GOOD) {
1168             goto fail;
1169         }
1170     }
1171
1172     *inq = scsi_datain_unmarshall(task);
1173     if (*inq == NULL) {
1174         error_setg(errp, "iSCSI: failed to unmarshall inquiry datain blob");
1175         goto fail_with_err;
1176     }
1177
1178     return task;
1179
1180 fail:
1181     error_setg(errp, "iSCSI: Inquiry command failed : %s",
1182                iscsi_get_error(iscsi));
1183 fail_with_err:
1184     if (task != NULL) {
1185         scsi_free_scsi_task(task);
1186     }
1187     return NULL;
1188 }
1189
1190 static void iscsi_detach_aio_context(BlockDriverState *bs)
1191 {
1192     IscsiLun *iscsilun = bs->opaque;
1193
1194     aio_set_fd_handler(iscsilun->aio_context,
1195                        iscsi_get_fd(iscsilun->iscsi),
1196                        NULL, NULL, NULL);
1197     iscsilun->events = 0;
1198
1199     if (iscsilun->nop_timer) {
1200         timer_del(iscsilun->nop_timer);
1201         timer_free(iscsilun->nop_timer);
1202         iscsilun->nop_timer = NULL;
1203     }
1204 }
1205
1206 static void iscsi_attach_aio_context(BlockDriverState *bs,
1207                                      AioContext *new_context)
1208 {
1209     IscsiLun *iscsilun = bs->opaque;
1210
1211     iscsilun->aio_context = new_context;
1212     iscsi_set_events(iscsilun);
1213
1214     /* Set up a timer for sending out iSCSI NOPs */
1215     iscsilun->nop_timer = aio_timer_new(iscsilun->aio_context,
1216                                         QEMU_CLOCK_REALTIME, SCALE_MS,
1217                                         iscsi_nop_timed_event, iscsilun);
1218     timer_mod(iscsilun->nop_timer,
1219               qemu_clock_get_ms(QEMU_CLOCK_REALTIME) + NOP_INTERVAL);
1220 }
1221
1222 /*
1223  * We support iscsi url's on the form
1224  * iscsi://[<username>%<password>@]<host>[:<port>]/<targetname>/<lun>
1225  *
1226  * Note: flags are currently not used by iscsi_open.  If this function
1227  * is changed such that flags are used, please examine iscsi_reopen_prepare()
1228  * to see if needs to be changed as well.
1229  */
1230 static int iscsi_open(BlockDriverState *bs, QDict *options, int flags,
1231                       Error **errp)
1232 {
1233     IscsiLun *iscsilun = bs->opaque;
1234     struct iscsi_context *iscsi = NULL;
1235     struct iscsi_url *iscsi_url = NULL;
1236     struct scsi_task *task = NULL;
1237     struct scsi_inquiry_standard *inq = NULL;
1238     struct scsi_inquiry_supported_pages *inq_vpd;
1239     char *initiator_name = NULL;
1240     QemuOpts *opts;
1241     Error *local_err = NULL;
1242     const char *filename;
1243     int i, ret;
1244
1245     if ((BDRV_SECTOR_SIZE % 512) != 0) {
1246         error_setg(errp, "iSCSI: Invalid BDRV_SECTOR_SIZE. "
1247                    "BDRV_SECTOR_SIZE(%lld) is not a multiple "
1248                    "of 512", BDRV_SECTOR_SIZE);
1249         return -EINVAL;
1250     }
1251
1252     opts = qemu_opts_create(&runtime_opts, NULL, 0, &error_abort);
1253     qemu_opts_absorb_qdict(opts, options, &local_err);
1254     if (local_err) {
1255         error_propagate(errp, local_err);
1256         ret = -EINVAL;
1257         goto out;
1258     }
1259
1260     filename = qemu_opt_get(opts, "filename");
1261
1262     iscsi_url = iscsi_parse_full_url(iscsi, filename);
1263     if (iscsi_url == NULL) {
1264         error_setg(errp, "Failed to parse URL : %s", filename);
1265         ret = -EINVAL;
1266         goto out;
1267     }
1268
1269     memset(iscsilun, 0, sizeof(IscsiLun));
1270
1271     initiator_name = parse_initiator_name(iscsi_url->target);
1272
1273     iscsi = iscsi_create_context(initiator_name);
1274     if (iscsi == NULL) {
1275         error_setg(errp, "iSCSI: Failed to create iSCSI context.");
1276         ret = -ENOMEM;
1277         goto out;
1278     }
1279
1280     if (iscsi_set_targetname(iscsi, iscsi_url->target)) {
1281         error_setg(errp, "iSCSI: Failed to set target name.");
1282         ret = -EINVAL;
1283         goto out;
1284     }
1285
1286     if (iscsi_url->user != NULL) {
1287         ret = iscsi_set_initiator_username_pwd(iscsi, iscsi_url->user,
1288                                               iscsi_url->passwd);
1289         if (ret != 0) {
1290             error_setg(errp, "Failed to set initiator username and password");
1291             ret = -EINVAL;
1292             goto out;
1293         }
1294     }
1295
1296     /* check if we got CHAP username/password via the options */
1297     parse_chap(iscsi, iscsi_url->target, &local_err);
1298     if (local_err != NULL) {
1299         error_propagate(errp, local_err);
1300         ret = -EINVAL;
1301         goto out;
1302     }
1303
1304     if (iscsi_set_session_type(iscsi, ISCSI_SESSION_NORMAL) != 0) {
1305         error_setg(errp, "iSCSI: Failed to set session type to normal.");
1306         ret = -EINVAL;
1307         goto out;
1308     }
1309
1310     iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_NONE_CRC32C);
1311
1312     /* check if we got HEADER_DIGEST via the options */
1313     parse_header_digest(iscsi, iscsi_url->target, &local_err);
1314     if (local_err != NULL) {
1315         error_propagate(errp, local_err);
1316         ret = -EINVAL;
1317         goto out;
1318     }
1319
1320     if (iscsi_full_connect_sync(iscsi, iscsi_url->portal, iscsi_url->lun) != 0) {
1321         error_setg(errp, "iSCSI: Failed to connect to LUN : %s",
1322             iscsi_get_error(iscsi));
1323         ret = -EINVAL;
1324         goto out;
1325     }
1326
1327     iscsilun->iscsi = iscsi;
1328     iscsilun->aio_context = bdrv_get_aio_context(bs);
1329     iscsilun->lun   = iscsi_url->lun;
1330     iscsilun->has_write_same = true;
1331
1332     task = iscsi_do_inquiry(iscsilun->iscsi, iscsilun->lun, 0, 0,
1333                             (void **) &inq, errp);
1334     if (task == NULL) {
1335         ret = -EINVAL;
1336         goto out;
1337     }
1338     iscsilun->type = inq->periperal_device_type;
1339     scsi_free_scsi_task(task);
1340     task = NULL;
1341
1342     iscsi_readcapacity_sync(iscsilun, &local_err);
1343     if (local_err != NULL) {
1344         error_propagate(errp, local_err);
1345         ret = -EINVAL;
1346         goto out;
1347     }
1348     bs->total_sectors = sector_lun2qemu(iscsilun->num_blocks, iscsilun);
1349     bs->request_alignment = iscsilun->block_size;
1350
1351     /* We don't have any emulation for devices other than disks and CD-ROMs, so
1352      * this must be sg ioctl compatible. We force it to be sg, otherwise qemu
1353      * will try to read from the device to guess the image format.
1354      */
1355     if (iscsilun->type != TYPE_DISK && iscsilun->type != TYPE_ROM) {
1356         bs->sg = 1;
1357     }
1358
1359     task = iscsi_do_inquiry(iscsilun->iscsi, iscsilun->lun, 1,
1360                             SCSI_INQUIRY_PAGECODE_SUPPORTED_VPD_PAGES,
1361                             (void **) &inq_vpd, errp);
1362     if (task == NULL) {
1363         ret = -EINVAL;
1364         goto out;
1365     }
1366     for (i = 0; i < inq_vpd->num_pages; i++) {
1367         struct scsi_task *inq_task;
1368         struct scsi_inquiry_logical_block_provisioning *inq_lbp;
1369         struct scsi_inquiry_block_limits *inq_bl;
1370         switch (inq_vpd->pages[i]) {
1371         case SCSI_INQUIRY_PAGECODE_LOGICAL_BLOCK_PROVISIONING:
1372             inq_task = iscsi_do_inquiry(iscsilun->iscsi, iscsilun->lun, 1,
1373                                         SCSI_INQUIRY_PAGECODE_LOGICAL_BLOCK_PROVISIONING,
1374                                         (void **) &inq_lbp, errp);
1375             if (inq_task == NULL) {
1376                 ret = -EINVAL;
1377                 goto out;
1378             }
1379             memcpy(&iscsilun->lbp, inq_lbp,
1380                    sizeof(struct scsi_inquiry_logical_block_provisioning));
1381             scsi_free_scsi_task(inq_task);
1382             break;
1383         case SCSI_INQUIRY_PAGECODE_BLOCK_LIMITS:
1384             inq_task = iscsi_do_inquiry(iscsilun->iscsi, iscsilun->lun, 1,
1385                                     SCSI_INQUIRY_PAGECODE_BLOCK_LIMITS,
1386                                     (void **) &inq_bl, errp);
1387             if (inq_task == NULL) {
1388                 ret = -EINVAL;
1389                 goto out;
1390             }
1391             memcpy(&iscsilun->bl, inq_bl,
1392                    sizeof(struct scsi_inquiry_block_limits));
1393             scsi_free_scsi_task(inq_task);
1394             break;
1395         default:
1396             break;
1397         }
1398     }
1399     scsi_free_scsi_task(task);
1400     task = NULL;
1401
1402     iscsi_attach_aio_context(bs, iscsilun->aio_context);
1403
1404     /* Guess the internal cluster (page) size of the iscsi target by the means
1405      * of opt_unmap_gran. Transfer the unmap granularity only if it has a
1406      * reasonable size */
1407     if (iscsilun->bl.opt_unmap_gran * iscsilun->block_size >= 4 * 1024 &&
1408         iscsilun->bl.opt_unmap_gran * iscsilun->block_size <= 16 * 1024 * 1024) {
1409         iscsilun->cluster_sectors = (iscsilun->bl.opt_unmap_gran *
1410                                      iscsilun->block_size) >> BDRV_SECTOR_BITS;
1411         if (iscsilun->lbprz && !(bs->open_flags & BDRV_O_NOCACHE)) {
1412             iscsilun->allocationmap = iscsi_allocationmap_init(iscsilun);
1413             if (iscsilun->allocationmap == NULL) {
1414                 ret = -ENOMEM;
1415             }
1416         }
1417     }
1418
1419 out:
1420     qemu_opts_del(opts);
1421     g_free(initiator_name);
1422     if (iscsi_url != NULL) {
1423         iscsi_destroy_url(iscsi_url);
1424     }
1425     if (task != NULL) {
1426         scsi_free_scsi_task(task);
1427     }
1428
1429     if (ret) {
1430         if (iscsi != NULL) {
1431             iscsi_destroy_context(iscsi);
1432         }
1433         memset(iscsilun, 0, sizeof(IscsiLun));
1434     }
1435     return ret;
1436 }
1437
1438 static void iscsi_close(BlockDriverState *bs)
1439 {
1440     IscsiLun *iscsilun = bs->opaque;
1441     struct iscsi_context *iscsi = iscsilun->iscsi;
1442
1443     iscsi_detach_aio_context(bs);
1444     iscsi_destroy_context(iscsi);
1445     g_free(iscsilun->zeroblock);
1446     g_free(iscsilun->allocationmap);
1447     memset(iscsilun, 0, sizeof(IscsiLun));
1448 }
1449
1450 static void iscsi_refresh_limits(BlockDriverState *bs, Error **errp)
1451 {
1452     IscsiLun *iscsilun = bs->opaque;
1453
1454     /* We don't actually refresh here, but just return data queried in
1455      * iscsi_open(): iscsi targets don't change their limits. */
1456     if (iscsilun->lbp.lbpu) {
1457         if (iscsilun->bl.max_unmap < 0xffffffff) {
1458             bs->bl.max_discard = sector_lun2qemu(iscsilun->bl.max_unmap,
1459                                                  iscsilun);
1460         }
1461         bs->bl.discard_alignment = sector_lun2qemu(iscsilun->bl.opt_unmap_gran,
1462                                                    iscsilun);
1463     }
1464
1465     if (iscsilun->bl.max_ws_len < 0xffffffff) {
1466         bs->bl.max_write_zeroes = sector_lun2qemu(iscsilun->bl.max_ws_len,
1467                                                   iscsilun);
1468     }
1469     if (iscsilun->lbp.lbpws) {
1470         bs->bl.write_zeroes_alignment = sector_lun2qemu(iscsilun->bl.opt_unmap_gran,
1471                                                         iscsilun);
1472     }
1473     bs->bl.opt_transfer_length = sector_lun2qemu(iscsilun->bl.opt_xfer_len,
1474                                                  iscsilun);
1475 }
1476
1477 /* Since iscsi_open() ignores bdrv_flags, there is nothing to do here in
1478  * prepare.  Note that this will not re-establish a connection with an iSCSI
1479  * target - it is effectively a NOP.  */
1480 static int iscsi_reopen_prepare(BDRVReopenState *state,
1481                                 BlockReopenQueue *queue, Error **errp)
1482 {
1483     /* NOP */
1484     return 0;
1485 }
1486
1487 static int iscsi_truncate(BlockDriverState *bs, int64_t offset)
1488 {
1489     IscsiLun *iscsilun = bs->opaque;
1490     Error *local_err = NULL;
1491
1492     if (iscsilun->type != TYPE_DISK) {
1493         return -ENOTSUP;
1494     }
1495
1496     iscsi_readcapacity_sync(iscsilun, &local_err);
1497     if (local_err != NULL) {
1498         error_free(local_err);
1499         return -EIO;
1500     }
1501
1502     if (offset > iscsi_getlength(bs)) {
1503         return -EINVAL;
1504     }
1505
1506     if (iscsilun->allocationmap != NULL) {
1507         g_free(iscsilun->allocationmap);
1508         iscsilun->allocationmap = iscsi_allocationmap_init(iscsilun);
1509     }
1510
1511     return 0;
1512 }
1513
1514 static int iscsi_create(const char *filename, QemuOpts *opts, Error **errp)
1515 {
1516     int ret = 0;
1517     int64_t total_size = 0;
1518     BlockDriverState *bs;
1519     IscsiLun *iscsilun = NULL;
1520     QDict *bs_options;
1521
1522     bs = bdrv_new();
1523
1524     /* Read out options */
1525     total_size = DIV_ROUND_UP(qemu_opt_get_size_del(opts, BLOCK_OPT_SIZE, 0),
1526                               BDRV_SECTOR_SIZE);
1527     bs->opaque = g_new0(struct IscsiLun, 1);
1528     iscsilun = bs->opaque;
1529
1530     bs_options = qdict_new();
1531     qdict_put(bs_options, "filename", qstring_from_str(filename));
1532     ret = iscsi_open(bs, bs_options, 0, NULL);
1533     QDECREF(bs_options);
1534
1535     if (ret != 0) {
1536         goto out;
1537     }
1538     iscsi_detach_aio_context(bs);
1539     if (iscsilun->type != TYPE_DISK) {
1540         ret = -ENODEV;
1541         goto out;
1542     }
1543     if (bs->total_sectors < total_size) {
1544         ret = -ENOSPC;
1545         goto out;
1546     }
1547
1548     ret = 0;
1549 out:
1550     if (iscsilun->iscsi != NULL) {
1551         iscsi_destroy_context(iscsilun->iscsi);
1552     }
1553     g_free(bs->opaque);
1554     bs->opaque = NULL;
1555     bdrv_unref(bs);
1556     return ret;
1557 }
1558
1559 static int iscsi_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
1560 {
1561     IscsiLun *iscsilun = bs->opaque;
1562     bdi->unallocated_blocks_are_zero = !!iscsilun->lbprz;
1563     bdi->can_write_zeroes_with_unmap = iscsilun->lbprz && iscsilun->lbp.lbpws;
1564     bdi->cluster_size = iscsilun->cluster_sectors * BDRV_SECTOR_SIZE;
1565     return 0;
1566 }
1567
1568 static QemuOptsList iscsi_create_opts = {
1569     .name = "iscsi-create-opts",
1570     .head = QTAILQ_HEAD_INITIALIZER(iscsi_create_opts.head),
1571     .desc = {
1572         {
1573             .name = BLOCK_OPT_SIZE,
1574             .type = QEMU_OPT_SIZE,
1575             .help = "Virtual disk size"
1576         },
1577         { /* end of list */ }
1578     }
1579 };
1580
1581 static BlockDriver bdrv_iscsi = {
1582     .format_name     = "iscsi",
1583     .protocol_name   = "iscsi",
1584
1585     .instance_size   = sizeof(IscsiLun),
1586     .bdrv_needs_filename = true,
1587     .bdrv_file_open  = iscsi_open,
1588     .bdrv_close      = iscsi_close,
1589     .bdrv_create     = iscsi_create,
1590     .create_opts     = &iscsi_create_opts,
1591     .bdrv_reopen_prepare  = iscsi_reopen_prepare,
1592
1593     .bdrv_getlength  = iscsi_getlength,
1594     .bdrv_get_info   = iscsi_get_info,
1595     .bdrv_truncate   = iscsi_truncate,
1596     .bdrv_refresh_limits = iscsi_refresh_limits,
1597
1598     .bdrv_co_get_block_status = iscsi_co_get_block_status,
1599     .bdrv_co_discard      = iscsi_co_discard,
1600     .bdrv_co_write_zeroes = iscsi_co_write_zeroes,
1601     .bdrv_co_readv         = iscsi_co_readv,
1602     .bdrv_co_writev        = iscsi_co_writev,
1603     .bdrv_co_flush_to_disk = iscsi_co_flush,
1604
1605 #ifdef __linux__
1606     .bdrv_ioctl       = iscsi_ioctl,
1607     .bdrv_aio_ioctl   = iscsi_aio_ioctl,
1608 #endif
1609
1610     .bdrv_detach_aio_context = iscsi_detach_aio_context,
1611     .bdrv_attach_aio_context = iscsi_attach_aio_context,
1612 };
1613
1614 static QemuOptsList qemu_iscsi_opts = {
1615     .name = "iscsi",
1616     .head = QTAILQ_HEAD_INITIALIZER(qemu_iscsi_opts.head),
1617     .desc = {
1618         {
1619             .name = "user",
1620             .type = QEMU_OPT_STRING,
1621             .help = "username for CHAP authentication to target",
1622         },{
1623             .name = "password",
1624             .type = QEMU_OPT_STRING,
1625             .help = "password for CHAP authentication to target",
1626         },{
1627             .name = "header-digest",
1628             .type = QEMU_OPT_STRING,
1629             .help = "HeaderDigest setting. "
1630                     "{CRC32C|CRC32C-NONE|NONE-CRC32C|NONE}",
1631         },{
1632             .name = "initiator-name",
1633             .type = QEMU_OPT_STRING,
1634             .help = "Initiator iqn name to use when connecting",
1635         },
1636         { /* end of list */ }
1637     },
1638 };
1639
1640 static void iscsi_block_init(void)
1641 {
1642     bdrv_register(&bdrv_iscsi);
1643     qemu_add_opts(&qemu_iscsi_opts);
1644 }
1645
1646 block_init(iscsi_block_init);