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     bool write_protected;
  69 } IscsiLun;
  70
  71 typedef struct IscsiTask {
  72     int status;
  73     int complete;
  74     int retries;
  75     int do_retry;
  76     struct scsi_task *task;
  77     Coroutine *co;
  78     QEMUBH *bh;
  79     IscsiLun *iscsilun;
  80     QEMUTimer retry_timer;
  81 } IscsiTask;
  82
  83 typedef struct IscsiAIOCB {
  84     BlockAIOCB common;
  85     QEMUIOVector *qiov;
  86     QEMUBH *bh;
  87     IscsiLun *iscsilun;
  88     struct scsi_task *task;
  89     uint8_t *buf;
  90     int status;
  91     int64_t sector_num;
  92     int nb_sectors;
  93 #ifdef __linux__
  94     sg_io_hdr_t *ioh;
  95 #endif
  96 } IscsiAIOCB;
  97
  98 #define NOP_INTERVAL 5000
  99 #define MAX_NOP_FAILURES 3
 100 #define ISCSI_CMD_RETRIES ARRAY_SIZE(iscsi_retry_times)
 101 static const unsigned iscsi_retry_times[] = {8, 32, 128, 512, 2048};
 102
 103 /* this threshold is a trade-off knob to choose between
 104  * the potential additional overhead of an extra GET_LBA_STATUS request
 105  * vs. unnecessarily reading a lot of zero sectors over the wire.
 106  * If a read request is greater or equal than ISCSI_CHECKALLOC_THRES
 107  * sectors we check the allocation status of the area covered by the
 108  * request first if the allocationmap indicates that the area might be
 109  * unallocated. */
 110 #define ISCSI_CHECKALLOC_THRES 64
 111
 112 static void
 113 iscsi_bh_cb(void *p)
 114 {
 115     IscsiAIOCB *acb = p;
 116
 117     qemu_bh_delete(acb->bh);
 118
 119     g_free(acb->buf);
 120     acb->buf = NULL;
 121
 122     acb->common.cb(acb->common.opaque, acb->status);
 123
 124     if (acb->task != NULL) {
 125         scsi_free_scsi_task(acb->task);
 126         acb->task = NULL;
 127     }
 128
 129     qemu_aio_unref(acb);
 130 }
 131
 132 static void
 133 iscsi_schedule_bh(IscsiAIOCB *acb)
 134 {
 135     if (acb->bh) {
 136         return;
 137     }
 138     acb->bh = aio_bh_new(acb->iscsilun->aio_context, iscsi_bh_cb, acb);
 139     qemu_bh_schedule(acb->bh);
 140 }
 141
 142 static void iscsi_co_generic_bh_cb(void *opaque)
 143 {
 144     struct IscsiTask *iTask = opaque;
 145     iTask->complete = 1;
 146     qemu_bh_delete(iTask->bh);
 147     qemu_coroutine_enter(iTask->co, NULL);
 148 }
 149
 150 static void iscsi_retry_timer_expired(void *opaque)
 151 {
 152     struct IscsiTask *iTask = opaque;
 153     iTask->complete = 1;
 154     if (iTask->co) {
 155         qemu_coroutine_enter(iTask->co, NULL);
 156     }
 157 }
 158
 159 static inline unsigned exp_random(double mean)
 160 {
 161     return -mean * log((double)rand() / RAND_MAX);
 162 }
 163
 164 static void
 165 iscsi_co_generic_cb(struct iscsi_context *iscsi, int status,
 166                         void *command_data, void *opaque)
 167 {
 168     struct IscsiTask *iTask = opaque;
 169     struct scsi_task *task = command_data;
 170
 171     iTask->status = status;
 172     iTask->do_retry = 0;
 173     iTask->task = task;
 174
 175     if (status != SCSI_STATUS_GOOD) {
 176         if (iTask->retries++ < ISCSI_CMD_RETRIES) {
 177             if (status == SCSI_STATUS_CHECK_CONDITION
 178                 && task->sense.key == SCSI_SENSE_UNIT_ATTENTION) {
 179                 error_report("iSCSI CheckCondition: %s",
 180                              iscsi_get_error(iscsi));
 181                 iTask->do_retry = 1;
 182                 goto out;
 183             }
 184             if (status == SCSI_STATUS_BUSY) {
 185                 unsigned retry_time =
 186                     exp_random(iscsi_retry_times[iTask->retries - 1]);
 187                 error_report("iSCSI Busy (retry #%u in %u ms): %s",
 188                              iTask->retries, retry_time,
 189                              iscsi_get_error(iscsi));
 190                 aio_timer_init(iTask->iscsilun->aio_context,
 191                                &iTask->retry_timer, QEMU_CLOCK_REALTIME,
 192                                SCALE_MS, iscsi_retry_timer_expired, iTask);
 193                 timer_mod(&iTask->retry_timer,
 194                           qemu_clock_get_ms(QEMU_CLOCK_REALTIME) + retry_time);
 195                 iTask->do_retry = 1;
 196                 return;
 197             }
 198         }
 199         error_report("iSCSI Failure: %s", iscsi_get_error(iscsi));
 200     }
 201
 202 out:
 203     if (iTask->co) {
 204         iTask->bh = aio_bh_new(iTask->iscsilun->aio_context,
 205                                iscsi_co_generic_bh_cb, iTask);
 206         qemu_bh_schedule(iTask->bh);
 207     } else {
 208         iTask->complete = 1;
 209     }
 210 }
 211
 212 static void iscsi_co_init_iscsitask(IscsiLun *iscsilun, struct IscsiTask *iTask)
 213 {
 214     *iTask = (struct IscsiTask) {
 215         .co         = qemu_coroutine_self(),
 216         .iscsilun   = iscsilun,
 217     };
 218 }
 219
 220 static void
 221 iscsi_abort_task_cb(struct iscsi_context *iscsi, int status, void *command_data,
 222                     void *private_data)
 223 {
 224     IscsiAIOCB *acb = private_data;
 225
 226     acb->status = -ECANCELED;
 227     iscsi_schedule_bh(acb);
 228 }
 229
 230 static void
 231 iscsi_aio_cancel(BlockAIOCB *blockacb)
 232 {
 233     IscsiAIOCB *acb = (IscsiAIOCB *)blockacb;
 234     IscsiLun *iscsilun = acb->iscsilun;
 235
 236     if (acb->status != -EINPROGRESS) {
 237         return;
 238     }
 239
 240     /* send a task mgmt call to the target to cancel the task on the target */
 241     iscsi_task_mgmt_abort_task_async(iscsilun->iscsi, acb->task,
 242                                      iscsi_abort_task_cb, acb);
 243
 244 }
 245
 246 static const AIOCBInfo iscsi_aiocb_info = {
 247     .aiocb_size         = sizeof(IscsiAIOCB),
 248     .cancel_async       = iscsi_aio_cancel,
 249 };
 250
 251
 252 static void iscsi_process_read(void *arg);
 253 static void iscsi_process_write(void *arg);
 254
 255 static void
 256 iscsi_set_events(IscsiLun *iscsilun)
 257 {
 258     struct iscsi_context *iscsi = iscsilun->iscsi;
 259     int ev;
 260
 261     /* We always register a read handler.  */
 262     ev = POLLIN;
 263     ev |= iscsi_which_events(iscsi);
 264     if (ev != iscsilun->events) {
 265         aio_set_fd_handler(iscsilun->aio_context,
 266                            iscsi_get_fd(iscsi),
 267                            iscsi_process_read,
 268                            (ev & POLLOUT) ? iscsi_process_write : NULL,
 269                            iscsilun);
 270
 271     }
 272
 273     iscsilun->events = ev;
 274 }
 275
 276 static void
 277 iscsi_process_read(void *arg)
 278 {
 279     IscsiLun *iscsilun = arg;
 280     struct iscsi_context *iscsi = iscsilun->iscsi;
 281
 282     iscsi_service(iscsi, POLLIN);
 283     iscsi_set_events(iscsilun);
 284 }
 285
 286 static void
 287 iscsi_process_write(void *arg)
 288 {
 289     IscsiLun *iscsilun = arg;
 290     struct iscsi_context *iscsi = iscsilun->iscsi;
 291
 292     iscsi_service(iscsi, POLLOUT);
 293     iscsi_set_events(iscsilun);
 294 }
 295
 296 static int64_t sector_lun2qemu(int64_t sector, IscsiLun *iscsilun)
 297 {
 298     return sector * iscsilun->block_size / BDRV_SECTOR_SIZE;
 299 }
 300
 301 static int64_t sector_qemu2lun(int64_t sector, IscsiLun *iscsilun)
 302 {
 303     return sector * BDRV_SECTOR_SIZE / iscsilun->block_size;
 304 }
 305
 306 static bool is_request_lun_aligned(int64_t sector_num, int nb_sectors,
 307                                       IscsiLun *iscsilun)
 308 {
 309     if ((sector_num * BDRV_SECTOR_SIZE) % iscsilun->block_size ||
 310         (nb_sectors * BDRV_SECTOR_SIZE) % iscsilun->block_size) {
 311             error_report("iSCSI misaligned request: "
 312                          "iscsilun->block_size %u, sector_num %" PRIi64
 313                          ", nb_sectors %d",
 314                          iscsilun->block_size, sector_num, nb_sectors);
 315             return 0;
 316     }
 317     return 1;
 318 }
 319
 320 static unsigned long *iscsi_allocationmap_init(IscsiLun *iscsilun)
 321 {
 322     return bitmap_try_new(DIV_ROUND_UP(sector_lun2qemu(iscsilun->num_blocks,
 323                                                        iscsilun),
 324                                        iscsilun->cluster_sectors));
 325 }
 326
 327 static void iscsi_allocationmap_set(IscsiLun *iscsilun, int64_t sector_num,
 328                                     int nb_sectors)
 329 {
 330     if (iscsilun->allocationmap == NULL) {
 331         return;
 332     }
 333     bitmap_set(iscsilun->allocationmap,
 334                sector_num / iscsilun->cluster_sectors,
 335                DIV_ROUND_UP(nb_sectors, iscsilun->cluster_sectors));
 336 }
 337
 338 static void iscsi_allocationmap_clear(IscsiLun *iscsilun, int64_t sector_num,
 339                                       int nb_sectors)
 340 {
 341     int64_t cluster_num, nb_clusters;
 342     if (iscsilun->allocationmap == NULL) {
 343         return;
 344     }
 345     cluster_num = DIV_ROUND_UP(sector_num, iscsilun->cluster_sectors);
 346     nb_clusters = (sector_num + nb_sectors) / iscsilun->cluster_sectors
 347                   - cluster_num;
 348     if (nb_clusters > 0) {
 349         bitmap_clear(iscsilun->allocationmap, cluster_num, nb_clusters);
 350     }
 351 }
 352
 353 static int coroutine_fn iscsi_co_writev(BlockDriverState *bs,
 354                                         int64_t sector_num, int nb_sectors,
 355                                         QEMUIOVector *iov)
 356 {
 357     IscsiLun *iscsilun = bs->opaque;
 358     struct IscsiTask iTask;
 359     uint64_t lba;
 360     uint32_t num_sectors;
 361
 362     if (!is_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
 363         return -EINVAL;
 364     }
 365
 366     if (bs->bl.max_transfer_length && nb_sectors > bs->bl.max_transfer_length) {
 367         error_report("iSCSI Error: Write of %d sectors exceeds max_xfer_len "
 368                      "of %d sectors", nb_sectors, bs->bl.max_transfer_length);
 369         return -EINVAL;
 370     }
 371
 372     lba = sector_qemu2lun(sector_num, iscsilun);
 373     num_sectors = sector_qemu2lun(nb_sectors, iscsilun);
 374     iscsi_co_init_iscsitask(iscsilun, &iTask);
 375 retry:
 376     if (iscsilun->use_16_for_rw) {
 377         iTask.task = iscsi_write16_task(iscsilun->iscsi, iscsilun->lun, lba,
 378                                         NULL, num_sectors * iscsilun->block_size,
 379                                         iscsilun->block_size, 0, 0, 0, 0, 0,
 380                                         iscsi_co_generic_cb, &iTask);
 381     } else {
 382         iTask.task = iscsi_write10_task(iscsilun->iscsi, iscsilun->lun, lba,
 383                                         NULL, num_sectors * iscsilun->block_size,
 384                                         iscsilun->block_size, 0, 0, 0, 0, 0,
 385                                         iscsi_co_generic_cb, &iTask);
 386     }
 387     if (iTask.task == NULL) {
 388         return -ENOMEM;
 389     }
 390     scsi_task_set_iov_out(iTask.task, (struct scsi_iovec *) iov->iov,
 391                           iov->niov);
 392     while (!iTask.complete) {
 393         iscsi_set_events(iscsilun);
 394         qemu_coroutine_yield();
 395     }
 396
 397     if (iTask.task != NULL) {
 398         scsi_free_scsi_task(iTask.task);
 399         iTask.task = NULL;
 400     }
 401
 402     if (iTask.do_retry) {
 403         iTask.complete = 0;
 404         goto retry;
 405     }
 406
 407     if (iTask.status != SCSI_STATUS_GOOD) {
 408         return -EIO;
 409     }
 410
 411     iscsi_allocationmap_set(iscsilun, sector_num, nb_sectors);
 412
 413     return 0;
 414 }
 415
 416
 417 static bool iscsi_allocationmap_is_allocated(IscsiLun *iscsilun,
 418                                              int64_t sector_num, int nb_sectors)
 419 {
 420     unsigned long size;
 421     if (iscsilun->allocationmap == NULL) {
 422         return true;
 423     }
 424     size = DIV_ROUND_UP(sector_num + nb_sectors, iscsilun->cluster_sectors);
 425     return !(find_next_bit(iscsilun->allocationmap, size,
 426                            sector_num / iscsilun->cluster_sectors) == size);
 427 }
 428
 429 static int64_t coroutine_fn iscsi_co_get_block_status(BlockDriverState *bs,
 430                                                   int64_t sector_num,
 431                                                   int nb_sectors, int *pnum)
 432 {
 433     IscsiLun *iscsilun = bs->opaque;
 434     struct scsi_get_lba_status *lbas = NULL;
 435     struct scsi_lba_status_descriptor *lbasd = NULL;
 436     struct IscsiTask iTask;
 437     int64_t ret;
 438
 439     iscsi_co_init_iscsitask(iscsilun, &iTask);
 440
 441     if (!is_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
 442         ret = -EINVAL;
 443         goto out;
 444     }
 445
 446     /* default to all sectors allocated */
 447     ret = BDRV_BLOCK_DATA;
 448     ret |= (sector_num << BDRV_SECTOR_BITS) | BDRV_BLOCK_OFFSET_VALID;
 449     *pnum = nb_sectors;
 450
 451     /* LUN does not support logical block provisioning */
 452     if (iscsilun->lbpme == 0) {
 453         goto out;
 454     }
 455
 456 retry:
 457     if (iscsi_get_lba_status_task(iscsilun->iscsi, iscsilun->lun,
 458                                   sector_qemu2lun(sector_num, iscsilun),
 459                                   8 + 16, iscsi_co_generic_cb,
 460                                   &iTask) == NULL) {
 461         ret = -ENOMEM;
 462         goto out;
 463     }
 464
 465     while (!iTask.complete) {
 466         iscsi_set_events(iscsilun);
 467         qemu_coroutine_yield();
 468     }
 469
 470     if (iTask.do_retry) {
 471         if (iTask.task != NULL) {
 472             scsi_free_scsi_task(iTask.task);
 473             iTask.task = NULL;
 474         }
 475         iTask.complete = 0;
 476         goto retry;
 477     }
 478
 479     if (iTask.status != SCSI_STATUS_GOOD) {
 480         /* in case the get_lba_status_callout fails (i.e.
 481          * because the device is busy or the cmd is not
 482          * supported) we pretend all blocks are allocated
 483          * for backwards compatibility */
 484         goto out;
 485     }
 486
 487     lbas = scsi_datain_unmarshall(iTask.task);
 488     if (lbas == NULL) {
 489         ret = -EIO;
 490         goto out;
 491     }
 492
 493     lbasd = &lbas->descriptors[0];
 494
 495     if (sector_qemu2lun(sector_num, iscsilun) != lbasd->lba) {
 496         ret = -EIO;
 497         goto out;
 498     }
 499
 500     *pnum = sector_lun2qemu(lbasd->num_blocks, iscsilun);
 501
 502     if (lbasd->provisioning == SCSI_PROVISIONING_TYPE_DEALLOCATED ||
 503         lbasd->provisioning == SCSI_PROVISIONING_TYPE_ANCHORED) {
 504         ret &= ~BDRV_BLOCK_DATA;
 505         if (iscsilun->lbprz) {
 506             ret |= BDRV_BLOCK_ZERO;
 507         }
 508     }
 509
 510     if (ret & BDRV_BLOCK_ZERO) {
 511         iscsi_allocationmap_clear(iscsilun, sector_num, *pnum);
 512     } else {
 513         iscsi_allocationmap_set(iscsilun, sector_num, *pnum);
 514     }
 515
 516     if (*pnum > nb_sectors) {
 517         *pnum = nb_sectors;
 518     }
 519 out:
 520     if (iTask.task != NULL) {
 521         scsi_free_scsi_task(iTask.task);
 522     }
 523     return ret;
 524 }
 525
 526 static int coroutine_fn iscsi_co_readv(BlockDriverState *bs,
 527                                        int64_t sector_num, int nb_sectors,
 528                                        QEMUIOVector *iov)
 529 {
 530     IscsiLun *iscsilun = bs->opaque;
 531     struct IscsiTask iTask;
 532     uint64_t lba;
 533     uint32_t num_sectors;
 534
 535     if (!is_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
 536         return -EINVAL;
 537     }
 538
 539     if (bs->bl.max_transfer_length && nb_sectors > bs->bl.max_transfer_length) {
 540         error_report("iSCSI Error: Read of %d sectors exceeds max_xfer_len "
 541                      "of %d sectors", nb_sectors, bs->bl.max_transfer_length);
 542         return -EINVAL;
 543     }
 544
 545     if (iscsilun->lbprz && nb_sectors >= ISCSI_CHECKALLOC_THRES &&
 546         !iscsi_allocationmap_is_allocated(iscsilun, sector_num, nb_sectors)) {
 547         int64_t ret;
 548         int pnum;
 549         ret = iscsi_co_get_block_status(bs, sector_num, INT_MAX, &pnum);
 550         if (ret < 0) {
 551             return ret;
 552         }
 553         if (ret & BDRV_BLOCK_ZERO && pnum >= nb_sectors) {
 554             qemu_iovec_memset(iov, 0, 0x00, iov->size);
 555             return 0;
 556         }
 557     }
 558
 559     lba = sector_qemu2lun(sector_num, iscsilun);
 560     num_sectors = sector_qemu2lun(nb_sectors, iscsilun);
 561
 562     iscsi_co_init_iscsitask(iscsilun, &iTask);
 563 retry:
 564     if (iscsilun->use_16_for_rw) {
 565         iTask.task = iscsi_read16_task(iscsilun->iscsi, iscsilun->lun, lba,
 566                                        num_sectors * iscsilun->block_size,
 567                                        iscsilun->block_size, 0, 0, 0, 0, 0,
 568                                        iscsi_co_generic_cb, &iTask);
 569     } else {
 570         iTask.task = iscsi_read10_task(iscsilun->iscsi, iscsilun->lun, lba,
 571                                        num_sectors * iscsilun->block_size,
 572                                        iscsilun->block_size,
 573                                        0, 0, 0, 0, 0,
 574                                        iscsi_co_generic_cb, &iTask);
 575     }
 576     if (iTask.task == NULL) {
 577         return -ENOMEM;
 578     }
 579     scsi_task_set_iov_in(iTask.task, (struct scsi_iovec *) iov->iov, iov->niov);
 580
 581     while (!iTask.complete) {
 582         iscsi_set_events(iscsilun);
 583         qemu_coroutine_yield();
 584     }
 585
 586     if (iTask.task != NULL) {
 587         scsi_free_scsi_task(iTask.task);
 588         iTask.task = NULL;
 589     }
 590
 591     if (iTask.do_retry) {
 592         iTask.complete = 0;
 593         goto retry;
 594     }
 595
 596     if (iTask.status != SCSI_STATUS_GOOD) {
 597         return -EIO;
 598     }
 599
 600     return 0;
 601 }
 602
 603 static int coroutine_fn iscsi_co_flush(BlockDriverState *bs)
 604 {
 605     IscsiLun *iscsilun = bs->opaque;
 606     struct IscsiTask iTask;
 607
 608     if (bs->sg) {
 609         return 0;
 610     }
 611
 612     iscsi_co_init_iscsitask(iscsilun, &iTask);
 613
 614 retry:
 615     if (iscsi_synchronizecache10_task(iscsilun->iscsi, iscsilun->lun, 0, 0, 0,
 616                                       0, iscsi_co_generic_cb, &iTask) == NULL) {
 617         return -ENOMEM;
 618     }
 619
 620     while (!iTask.complete) {
 621         iscsi_set_events(iscsilun);
 622         qemu_coroutine_yield();
 623     }
 624
 625     if (iTask.task != NULL) {
 626         scsi_free_scsi_task(iTask.task);
 627         iTask.task = NULL;
 628     }
 629
 630     if (iTask.do_retry) {
 631         iTask.complete = 0;
 632         goto retry;
 633     }
 634
 635     if (iTask.status != SCSI_STATUS_GOOD) {
 636         return -EIO;
 637     }
 638
 639     return 0;
 640 }
 641
 642 #ifdef __linux__
 643 static void
 644 iscsi_aio_ioctl_cb(struct iscsi_context *iscsi, int status,
 645                      void *command_data, void *opaque)
 646 {
 647     IscsiAIOCB *acb = opaque;
 648
 649     g_free(acb->buf);
 650     acb->buf = NULL;
 651
 652     acb->status = 0;
 653     if (status < 0) {
 654         error_report("Failed to ioctl(SG_IO) to iSCSI lun. %s",
 655                      iscsi_get_error(iscsi));
 656         acb->status = -EIO;
 657     }
 658
 659     acb->ioh->driver_status = 0;
 660     acb->ioh->host_status   = 0;
 661     acb->ioh->resid         = 0;
 662
 663 #define SG_ERR_DRIVER_SENSE    0x08
 664
 665     if (status == SCSI_STATUS_CHECK_CONDITION && acb->task->datain.size >= 2) {
 666         int ss;
 667
 668         acb->ioh->driver_status |= SG_ERR_DRIVER_SENSE;
 669
 670         acb->ioh->sb_len_wr = acb->task->datain.size - 2;
 671         ss = (acb->ioh->mx_sb_len >= acb->ioh->sb_len_wr) ?
 672              acb->ioh->mx_sb_len : acb->ioh->sb_len_wr;
 673         memcpy(acb->ioh->sbp, &acb->task->datain.data[2], ss);
 674     }
 675
 676     iscsi_schedule_bh(acb);
 677 }
 678
 679 static BlockAIOCB *iscsi_aio_ioctl(BlockDriverState *bs,
 680         unsigned long int req, void *buf,
 681         BlockCompletionFunc *cb, void *opaque)
 682 {
 683     IscsiLun *iscsilun = bs->opaque;
 684     struct iscsi_context *iscsi = iscsilun->iscsi;
 685     struct iscsi_data data;
 686     IscsiAIOCB *acb;
 687
 688     assert(req == SG_IO);
 689
 690     acb = qemu_aio_get(&iscsi_aiocb_info, bs, cb, opaque);
 691
 692     acb->iscsilun = iscsilun;
 693     acb->bh          = NULL;
 694     acb->status      = -EINPROGRESS;
 695     acb->buf         = NULL;
 696     acb->ioh         = buf;
 697
 698     acb->task = malloc(sizeof(struct scsi_task));
 699     if (acb->task == NULL) {
 700         error_report("iSCSI: Failed to allocate task for scsi command. %s",
 701                      iscsi_get_error(iscsi));
 702         qemu_aio_unref(acb);
 703         return NULL;
 704     }
 705     memset(acb->task, 0, sizeof(struct scsi_task));
 706
 707     switch (acb->ioh->dxfer_direction) {
 708     case SG_DXFER_TO_DEV:
 709         acb->task->xfer_dir = SCSI_XFER_WRITE;
 710         break;
 711     case SG_DXFER_FROM_DEV:
 712         acb->task->xfer_dir = SCSI_XFER_READ;
 713         break;
 714     default:
 715         acb->task->xfer_dir = SCSI_XFER_NONE;
 716         break;
 717     }
 718
 719     acb->task->cdb_size = acb->ioh->cmd_len;
 720     memcpy(&acb->task->cdb[0], acb->ioh->cmdp, acb->ioh->cmd_len);
 721     acb->task->expxferlen = acb->ioh->dxfer_len;
 722
 723     data.size = 0;
 724     if (acb->task->xfer_dir == SCSI_XFER_WRITE) {
 725         if (acb->ioh->iovec_count == 0) {
 726             data.data = acb->ioh->dxferp;
 727             data.size = acb->ioh->dxfer_len;
 728         } else {
 729             scsi_task_set_iov_out(acb->task,
 730                                  (struct scsi_iovec *) acb->ioh->dxferp,
 731                                  acb->ioh->iovec_count);
 732         }
 733     }
 734
 735     if (iscsi_scsi_command_async(iscsi, iscsilun->lun, acb->task,
 736                                  iscsi_aio_ioctl_cb,
 737                                  (data.size > 0) ? &data : NULL,
 738                                  acb) != 0) {
 739         scsi_free_scsi_task(acb->task);
 740         qemu_aio_unref(acb);
 741         return NULL;
 742     }
 743
 744     /* tell libiscsi to read straight into the buffer we got from ioctl */
 745     if (acb->task->xfer_dir == SCSI_XFER_READ) {
 746         if (acb->ioh->iovec_count == 0) {
 747             scsi_task_add_data_in_buffer(acb->task,
 748                                          acb->ioh->dxfer_len,
 749                                          acb->ioh->dxferp);
 750         } else {
 751             scsi_task_set_iov_in(acb->task,
 752                                  (struct scsi_iovec *) acb->ioh->dxferp,
 753                                  acb->ioh->iovec_count);
 754         }
 755     }
 756
 757     iscsi_set_events(iscsilun);
 758
 759     return &acb->common;
 760 }
 761
 762 static void ioctl_cb(void *opaque, int status)
 763 {
 764     int *p_status = opaque;
 765     *p_status = status;
 766 }
 767
 768 static int iscsi_ioctl(BlockDriverState *bs, unsigned long int req, void *buf)
 769 {
 770     IscsiLun *iscsilun = bs->opaque;
 771     int status;
 772
 773     switch (req) {
 774     case SG_GET_VERSION_NUM:
 775         *(int *)buf = 30000;
 776         break;
 777     case SG_GET_SCSI_ID:
 778         ((struct sg_scsi_id *)buf)->scsi_type = iscsilun->type;
 779         break;
 780     case SG_IO:
 781         status = -EINPROGRESS;
 782         iscsi_aio_ioctl(bs, req, buf, ioctl_cb, &status);
 783
 784         while (status == -EINPROGRESS) {
 785             aio_poll(iscsilun->aio_context, true);
 786         }
 787
 788         return 0;
 789     default:
 790         return -1;
 791     }
 792     return 0;
 793 }
 794 #endif
 795
 796 static int64_t
 797 iscsi_getlength(BlockDriverState *bs)
 798 {
 799     IscsiLun *iscsilun = bs->opaque;
 800     int64_t len;
 801
 802     len  = iscsilun->num_blocks;
 803     len *= iscsilun->block_size;
 804
 805     return len;
 806 }
 807
 808 static int
 809 coroutine_fn iscsi_co_discard(BlockDriverState *bs, int64_t sector_num,
 810                                    int nb_sectors)
 811 {
 812     IscsiLun *iscsilun = bs->opaque;
 813     struct IscsiTask iTask;
 814     struct unmap_list list;
 815
 816     if (!is_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
 817         return -EINVAL;
 818     }
 819
 820     if (!iscsilun->lbp.lbpu) {
 821         /* UNMAP is not supported by the target */
 822         return 0;
 823     }
 824
 825     list.lba = sector_qemu2lun(sector_num, iscsilun);
 826     list.num = sector_qemu2lun(nb_sectors, iscsilun);
 827
 828     iscsi_co_init_iscsitask(iscsilun, &iTask);
 829 retry:
 830     if (iscsi_unmap_task(iscsilun->iscsi, iscsilun->lun, 0, 0, &list, 1,
 831                      iscsi_co_generic_cb, &iTask) == NULL) {
 832         return -ENOMEM;
 833     }
 834
 835     while (!iTask.complete) {
 836         iscsi_set_events(iscsilun);
 837         qemu_coroutine_yield();
 838     }
 839
 840     if (iTask.task != NULL) {
 841         scsi_free_scsi_task(iTask.task);
 842         iTask.task = NULL;
 843     }
 844
 845     if (iTask.do_retry) {
 846         iTask.complete = 0;
 847         goto retry;
 848     }
 849
 850     if (iTask.status == SCSI_STATUS_CHECK_CONDITION) {
 851         /* the target might fail with a check condition if it
 852            is not happy with the alignment of the UNMAP request
 853            we silently fail in this case */
 854         return 0;
 855     }
 856
 857     if (iTask.status != SCSI_STATUS_GOOD) {
 858         return -EIO;
 859     }
 860
 861     iscsi_allocationmap_clear(iscsilun, sector_num, nb_sectors);
 862
 863     return 0;
 864 }
 865
 866 static int
 867 coroutine_fn iscsi_co_write_zeroes(BlockDriverState *bs, int64_t sector_num,
 868                                    int nb_sectors, BdrvRequestFlags flags)
 869 {
 870     IscsiLun *iscsilun = bs->opaque;
 871     struct IscsiTask iTask;
 872     uint64_t lba;
 873     uint32_t nb_blocks;
 874     bool use_16_for_ws = iscsilun->use_16_for_rw;
 875
 876     if (!is_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
 877         return -EINVAL;
 878     }
 879
 880     if (flags & BDRV_REQ_MAY_UNMAP) {
 881         if (!use_16_for_ws && !iscsilun->lbp.lbpws10) {
 882             /* WRITESAME10 with UNMAP is unsupported try WRITESAME16 */
 883             use_16_for_ws = true;
 884         }
 885         if (use_16_for_ws && !iscsilun->lbp.lbpws) {
 886             /* WRITESAME16 with UNMAP is not supported by the target,
 887              * fall back and try WRITESAME10/16 without UNMAP */
 888             flags &= ~BDRV_REQ_MAY_UNMAP;
 889             use_16_for_ws = iscsilun->use_16_for_rw;
 890         }
 891     }
 892
 893     if (!(flags & BDRV_REQ_MAY_UNMAP) && !iscsilun->has_write_same) {
 894         /* WRITESAME without UNMAP is not supported by the target */
 895         return -ENOTSUP;
 896     }
 897
 898     lba = sector_qemu2lun(sector_num, iscsilun);
 899     nb_blocks = sector_qemu2lun(nb_sectors, iscsilun);
 900
 901     if (iscsilun->zeroblock == NULL) {
 902         iscsilun->zeroblock = g_try_malloc0(iscsilun->block_size);
 903         if (iscsilun->zeroblock == NULL) {
 904             return -ENOMEM;
 905         }
 906     }
 907
 908     iscsi_co_init_iscsitask(iscsilun, &iTask);
 909 retry:
 910     if (use_16_for_ws) {
 911         iTask.task = iscsi_writesame16_task(iscsilun->iscsi, iscsilun->lun, lba,
 912                                             iscsilun->zeroblock, iscsilun->block_size,
 913                                             nb_blocks, 0, !!(flags & BDRV_REQ_MAY_UNMAP),
 914                                             0, 0, iscsi_co_generic_cb, &iTask);
 915     } else {
 916         iTask.task = iscsi_writesame10_task(iscsilun->iscsi, iscsilun->lun, lba,
 917                                             iscsilun->zeroblock, iscsilun->block_size,
 918                                             nb_blocks, 0, !!(flags & BDRV_REQ_MAY_UNMAP),
 919                                             0, 0, iscsi_co_generic_cb, &iTask);
 920     }
 921     if (iTask.task == NULL) {
 922         return -ENOMEM;
 923     }
 924
 925     while (!iTask.complete) {
 926         iscsi_set_events(iscsilun);
 927         qemu_coroutine_yield();
 928     }
 929
 930     if (iTask.status == SCSI_STATUS_CHECK_CONDITION &&
 931         iTask.task->sense.key == SCSI_SENSE_ILLEGAL_REQUEST &&
 932         (iTask.task->sense.ascq == SCSI_SENSE_ASCQ_INVALID_OPERATION_CODE ||
 933          iTask.task->sense.ascq == SCSI_SENSE_ASCQ_INVALID_FIELD_IN_CDB)) {
 934         /* WRITE SAME is not supported by the target */
 935         iscsilun->has_write_same = false;
 936         scsi_free_scsi_task(iTask.task);
 937         return -ENOTSUP;
 938     }
 939
 940     if (iTask.task != NULL) {
 941         scsi_free_scsi_task(iTask.task);
 942         iTask.task = NULL;
 943     }
 944
 945     if (iTask.do_retry) {
 946         iTask.complete = 0;
 947         goto retry;
 948     }
 949
 950     if (iTask.status != SCSI_STATUS_GOOD) {
 951         return -EIO;
 952     }
 953
 954     if (flags & BDRV_REQ_MAY_UNMAP) {
 955         iscsi_allocationmap_clear(iscsilun, sector_num, nb_sectors);
 956     } else {
 957         iscsi_allocationmap_set(iscsilun, sector_num, nb_sectors);
 958     }
 959
 960     return 0;
 961 }
 962
 963 static void parse_chap(struct iscsi_context *iscsi, const char *target,
 964                        Error **errp)
 965 {
 966     QemuOptsList *list;
 967     QemuOpts *opts;
 968     const char *user = NULL;
 969     const char *password = NULL;
 970
 971     list = qemu_find_opts("iscsi");
 972     if (!list) {
 973         return;
 974     }
 975
 976     opts = qemu_opts_find(list, target);
 977     if (opts == NULL) {
 978         opts = QTAILQ_FIRST(&list->head);
 979         if (!opts) {
 980             return;
 981         }
 982     }
 983
 984     user = qemu_opt_get(opts, "user");
 985     if (!user) {
 986         return;
 987     }
 988
 989     password = qemu_opt_get(opts, "password");
 990     if (!password) {
 991         error_setg(errp, "CHAP username specified but no password was given");
 992         return;
 993     }
 994
 995     if (iscsi_set_initiator_username_pwd(iscsi, user, password)) {
 996         error_setg(errp, "Failed to set initiator username and password");
 997     }
 998 }
 999
1000 static void parse_header_digest(struct iscsi_context *iscsi, const char *target,
1001                                 Error **errp)
1002 {
1003     QemuOptsList *list;
1004     QemuOpts *opts;
1005     const char *digest = NULL;
1006
1007     list = qemu_find_opts("iscsi");
1008     if (!list) {
1009         return;
1010     }
1011
1012     opts = qemu_opts_find(list, target);
1013     if (opts == NULL) {
1014         opts = QTAILQ_FIRST(&list->head);
1015         if (!opts) {
1016             return;
1017         }
1018     }
1019
1020     digest = qemu_opt_get(opts, "header-digest");
1021     if (!digest) {
1022         return;
1023     }
1024
1025     if (!strcmp(digest, "CRC32C")) {
1026         iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_CRC32C);
1027     } else if (!strcmp(digest, "NONE")) {
1028         iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_NONE);
1029     } else if (!strcmp(digest, "CRC32C-NONE")) {
1030         iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_CRC32C_NONE);
1031     } else if (!strcmp(digest, "NONE-CRC32C")) {
1032         iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_NONE_CRC32C);
1033     } else {
1034         error_setg(errp, "Invalid header-digest setting : %s", digest);
1035     }
1036 }
1037
1038 static char *parse_initiator_name(const char *target)
1039 {
1040     QemuOptsList *list;
1041     QemuOpts *opts;
1042     const char *name;
1043     char *iscsi_name;
1044     UuidInfo *uuid_info;
1045
1046     list = qemu_find_opts("iscsi");
1047     if (list) {
1048         opts = qemu_opts_find(list, target);
1049         if (!opts) {
1050             opts = QTAILQ_FIRST(&list->head);
1051         }
1052         if (opts) {
1053             name = qemu_opt_get(opts, "initiator-name");
1054             if (name) {
1055                 return g_strdup(name);
1056             }
1057         }
1058     }
1059
1060     uuid_info = qmp_query_uuid(NULL);
1061     if (strcmp(uuid_info->UUID, UUID_NONE) == 0) {
1062         name = qemu_get_vm_name();
1063     } else {
1064         name = uuid_info->UUID;
1065     }
1066     iscsi_name = g_strdup_printf("iqn.2008-11.org.linux-kvm%s%s",
1067                                  name ? ":" : "", name ? name : "");
1068     qapi_free_UuidInfo(uuid_info);
1069     return iscsi_name;
1070 }
1071
1072 static void iscsi_nop_timed_event(void *opaque)
1073 {
1074     IscsiLun *iscsilun = opaque;
1075
1076     if (iscsi_get_nops_in_flight(iscsilun->iscsi) > MAX_NOP_FAILURES) {
1077         error_report("iSCSI: NOP timeout. Reconnecting...");
1078         iscsi_reconnect(iscsilun->iscsi);
1079     }
1080
1081     if (iscsi_nop_out_async(iscsilun->iscsi, NULL, NULL, 0, NULL) != 0) {
1082         error_report("iSCSI: failed to sent NOP-Out. Disabling NOP messages.");
1083         return;
1084     }
1085
1086     timer_mod(iscsilun->nop_timer, qemu_clock_get_ms(QEMU_CLOCK_REALTIME) + NOP_INTERVAL);
1087     iscsi_set_events(iscsilun);
1088 }
1089
1090 static void iscsi_readcapacity_sync(IscsiLun *iscsilun, Error **errp)
1091 {
1092     struct scsi_task *task = NULL;
1093     struct scsi_readcapacity10 *rc10 = NULL;
1094     struct scsi_readcapacity16 *rc16 = NULL;
1095     int retries = ISCSI_CMD_RETRIES;
1096
1097     do {
1098         if (task != NULL) {
1099             scsi_free_scsi_task(task);
1100             task = NULL;
1101         }
1102
1103         switch (iscsilun->type) {
1104         case TYPE_DISK:
1105             task = iscsi_readcapacity16_sync(iscsilun->iscsi, iscsilun->lun);
1106             if (task != NULL && task->status == SCSI_STATUS_GOOD) {
1107                 rc16 = scsi_datain_unmarshall(task);
1108                 if (rc16 == NULL) {
1109                     error_setg(errp, "iSCSI: Failed to unmarshall readcapacity16 data.");
1110                 } else {
1111                     iscsilun->block_size = rc16->block_length;
1112                     iscsilun->num_blocks = rc16->returned_lba + 1;
1113                     iscsilun->lbpme = rc16->lbpme;
1114                     iscsilun->lbprz = rc16->lbprz;
1115                     iscsilun->use_16_for_rw = (rc16->returned_lba > 0xffffffff);
1116                 }
1117             }
1118             break;
1119         case TYPE_ROM:
1120             task = iscsi_readcapacity10_sync(iscsilun->iscsi, iscsilun->lun, 0, 0);
1121             if (task != NULL && task->status == SCSI_STATUS_GOOD) {
1122                 rc10 = scsi_datain_unmarshall(task);
1123                 if (rc10 == NULL) {
1124                     error_setg(errp, "iSCSI: Failed to unmarshall readcapacity10 data.");
1125                 } else {
1126                     iscsilun->block_size = rc10->block_size;
1127                     if (rc10->lba == 0) {
1128                         /* blank disk loaded */
1129                         iscsilun->num_blocks = 0;
1130                     } else {
1131                         iscsilun->num_blocks = rc10->lba + 1;
1132                     }
1133                 }
1134             }
1135             break;
1136         default:
1137             return;
1138         }
1139     } while (task != NULL && task->status == SCSI_STATUS_CHECK_CONDITION
1140              && task->sense.key == SCSI_SENSE_UNIT_ATTENTION
1141              && retries-- > 0);
1142
1143     if (task == NULL || task->status != SCSI_STATUS_GOOD) {
1144         error_setg(errp, "iSCSI: failed to send readcapacity10 command.");
1145     }
1146     if (task) {
1147         scsi_free_scsi_task(task);
1148     }
1149 }
1150
1151 /* TODO Convert to fine grained options */
1152 static QemuOptsList runtime_opts = {
1153     .name = "iscsi",
1154     .head = QTAILQ_HEAD_INITIALIZER(runtime_opts.head),
1155     .desc = {
1156         {
1157             .name = "filename",
1158             .type = QEMU_OPT_STRING,
1159             .help = "URL to the iscsi image",
1160         },
1161         { /* end of list */ }
1162     },
1163 };
1164
1165 static struct scsi_task *iscsi_do_inquiry(struct iscsi_context *iscsi, int lun,
1166                                           int evpd, int pc, void **inq, Error **errp)
1167 {
1168     int full_size;
1169     struct scsi_task *task = NULL;
1170     task = iscsi_inquiry_sync(iscsi, lun, evpd, pc, 64);
1171     if (task == NULL || task->status != SCSI_STATUS_GOOD) {
1172         goto fail;
1173     }
1174     full_size = scsi_datain_getfullsize(task);
1175     if (full_size > task->datain.size) {
1176         scsi_free_scsi_task(task);
1177
1178         /* we need more data for the full list */
1179         task = iscsi_inquiry_sync(iscsi, lun, evpd, pc, full_size);
1180         if (task == NULL || task->status != SCSI_STATUS_GOOD) {
1181             goto fail;
1182         }
1183     }
1184
1185     *inq = scsi_datain_unmarshall(task);
1186     if (*inq == NULL) {
1187         error_setg(errp, "iSCSI: failed to unmarshall inquiry datain blob");
1188         goto fail_with_err;
1189     }
1190
1191     return task;
1192
1193 fail:
1194     error_setg(errp, "iSCSI: Inquiry command failed : %s",
1195                iscsi_get_error(iscsi));
1196 fail_with_err:
1197     if (task != NULL) {
1198         scsi_free_scsi_task(task);
1199     }
1200     return NULL;
1201 }
1202
1203 static void iscsi_detach_aio_context(BlockDriverState *bs)
1204 {
1205     IscsiLun *iscsilun = bs->opaque;
1206
1207     aio_set_fd_handler(iscsilun->aio_context,
1208                        iscsi_get_fd(iscsilun->iscsi),
1209                        NULL, NULL, NULL);
1210     iscsilun->events = 0;
1211
1212     if (iscsilun->nop_timer) {
1213         timer_del(iscsilun->nop_timer);
1214         timer_free(iscsilun->nop_timer);
1215         iscsilun->nop_timer = NULL;
1216     }
1217 }
1218
1219 static void iscsi_attach_aio_context(BlockDriverState *bs,
1220                                      AioContext *new_context)
1221 {
1222     IscsiLun *iscsilun = bs->opaque;
1223
1224     iscsilun->aio_context = new_context;
1225     iscsi_set_events(iscsilun);
1226
1227     /* Set up a timer for sending out iSCSI NOPs */
1228     iscsilun->nop_timer = aio_timer_new(iscsilun->aio_context,
1229                                         QEMU_CLOCK_REALTIME, SCALE_MS,
1230                                         iscsi_nop_timed_event, iscsilun);
1231     timer_mod(iscsilun->nop_timer,
1232               qemu_clock_get_ms(QEMU_CLOCK_REALTIME) + NOP_INTERVAL);
1233 }
1234
1235 static bool iscsi_is_write_protected(IscsiLun *iscsilun)
1236 {
1237     struct scsi_task *task;
1238     struct scsi_mode_sense *ms = NULL;
1239     bool wrprotected = false;
1240
1241     task = iscsi_modesense6_sync(iscsilun->iscsi, iscsilun->lun,
1242                                  1, SCSI_MODESENSE_PC_CURRENT,
1243                                  0x3F, 0, 255);
1244     if (task == NULL) {
1245         error_report("iSCSI: Failed to send MODE_SENSE(6) command: %s",
1246                      iscsi_get_error(iscsilun->iscsi));
1247         goto out;
1248     }
1249
1250     if (task->status != SCSI_STATUS_GOOD) {
1251         error_report("iSCSI: Failed MODE_SENSE(6), LUN assumed writable");
1252         goto out;
1253     }
1254     ms = scsi_datain_unmarshall(task);
1255     if (!ms) {
1256         error_report("iSCSI: Failed to unmarshall MODE_SENSE(6) data: %s",
1257                      iscsi_get_error(iscsilun->iscsi));
1258         goto out;
1259     }
1260     wrprotected = ms->device_specific_parameter & 0x80;
1261
1262 out:
1263     if (task) {
1264         scsi_free_scsi_task(task);
1265     }
1266     return wrprotected;
1267 }
1268
1269 /*
1270  * We support iscsi url's on the form
1271  * iscsi://[<username>%<password>@]<host>[:<port>]/<targetname>/<lun>
1272  */
1273 static int iscsi_open(BlockDriverState *bs, QDict *options, int flags,
1274                       Error **errp)
1275 {
1276     IscsiLun *iscsilun = bs->opaque;
1277     struct iscsi_context *iscsi = NULL;
1278     struct iscsi_url *iscsi_url = NULL;
1279     struct scsi_task *task = NULL;
1280     struct scsi_inquiry_standard *inq = NULL;
1281     struct scsi_inquiry_supported_pages *inq_vpd;
1282     char *initiator_name = NULL;
1283     QemuOpts *opts;
1284     Error *local_err = NULL;
1285     const char *filename;
1286     int i, ret = 0;
1287
1288     if ((BDRV_SECTOR_SIZE % 512) != 0) {
1289         error_setg(errp, "iSCSI: Invalid BDRV_SECTOR_SIZE. "
1290                    "BDRV_SECTOR_SIZE(%lld) is not a multiple "
1291                    "of 512", BDRV_SECTOR_SIZE);
1292         return -EINVAL;
1293     }
1294
1295     opts = qemu_opts_create(&runtime_opts, NULL, 0, &error_abort);
1296     qemu_opts_absorb_qdict(opts, options, &local_err);
1297     if (local_err) {
1298         error_propagate(errp, local_err);
1299         ret = -EINVAL;
1300         goto out;
1301     }
1302
1303     filename = qemu_opt_get(opts, "filename");
1304
1305     iscsi_url = iscsi_parse_full_url(iscsi, filename);
1306     if (iscsi_url == NULL) {
1307         error_setg(errp, "Failed to parse URL : %s", filename);
1308         ret = -EINVAL;
1309         goto out;
1310     }
1311
1312     memset(iscsilun, 0, sizeof(IscsiLun));
1313
1314     initiator_name = parse_initiator_name(iscsi_url->target);
1315
1316     iscsi = iscsi_create_context(initiator_name);
1317     if (iscsi == NULL) {
1318         error_setg(errp, "iSCSI: Failed to create iSCSI context.");
1319         ret = -ENOMEM;
1320         goto out;
1321     }
1322
1323     if (iscsi_set_targetname(iscsi, iscsi_url->target)) {
1324         error_setg(errp, "iSCSI: Failed to set target name.");
1325         ret = -EINVAL;
1326         goto out;
1327     }
1328
1329     if (iscsi_url->user != NULL) {
1330         ret = iscsi_set_initiator_username_pwd(iscsi, iscsi_url->user,
1331                                               iscsi_url->passwd);
1332         if (ret != 0) {
1333             error_setg(errp, "Failed to set initiator username and password");
1334             ret = -EINVAL;
1335             goto out;
1336         }
1337     }
1338
1339     /* check if we got CHAP username/password via the options */
1340     parse_chap(iscsi, iscsi_url->target, &local_err);
1341     if (local_err != NULL) {
1342         error_propagate(errp, local_err);
1343         ret = -EINVAL;
1344         goto out;
1345     }
1346
1347     if (iscsi_set_session_type(iscsi, ISCSI_SESSION_NORMAL) != 0) {
1348         error_setg(errp, "iSCSI: Failed to set session type to normal.");
1349         ret = -EINVAL;
1350         goto out;
1351     }
1352
1353     iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_NONE_CRC32C);
1354
1355     /* check if we got HEADER_DIGEST via the options */
1356     parse_header_digest(iscsi, iscsi_url->target, &local_err);
1357     if (local_err != NULL) {
1358         error_propagate(errp, local_err);
1359         ret = -EINVAL;
1360         goto out;
1361     }
1362
1363     if (iscsi_full_connect_sync(iscsi, iscsi_url->portal, iscsi_url->lun) != 0) {
1364         error_setg(errp, "iSCSI: Failed to connect to LUN : %s",
1365             iscsi_get_error(iscsi));
1366         ret = -EINVAL;
1367         goto out;
1368     }
1369
1370     iscsilun->iscsi = iscsi;
1371     iscsilun->aio_context = bdrv_get_aio_context(bs);
1372     iscsilun->lun   = iscsi_url->lun;
1373     iscsilun->has_write_same = true;
1374
1375     task = iscsi_do_inquiry(iscsilun->iscsi, iscsilun->lun, 0, 0,
1376                             (void **) &inq, errp);
1377     if (task == NULL) {
1378         ret = -EINVAL;
1379         goto out;
1380     }
1381     iscsilun->type = inq->periperal_device_type;
1382     scsi_free_scsi_task(task);
1383     task = NULL;
1384
1385     iscsilun->write_protected = iscsi_is_write_protected(iscsilun);
1386     /* Check the write protect flag of the LUN if we want to write */
1387     if (iscsilun->type == TYPE_DISK && (flags & BDRV_O_RDWR) &&
1388         iscsilun->write_protected) {
1389         error_setg(errp, "Cannot open a write protected LUN as read-write");
1390         ret = -EACCES;
1391         goto out;
1392     }
1393
1394     iscsi_readcapacity_sync(iscsilun, &local_err);
1395     if (local_err != NULL) {
1396         error_propagate(errp, local_err);
1397         ret = -EINVAL;
1398         goto out;
1399     }
1400     bs->total_sectors = sector_lun2qemu(iscsilun->num_blocks, iscsilun);
1401     bs->request_alignment = iscsilun->block_size;
1402
1403     /* We don't have any emulation for devices other than disks and CD-ROMs, so
1404      * this must be sg ioctl compatible. We force it to be sg, otherwise qemu
1405      * will try to read from the device to guess the image format.
1406      */
1407     if (iscsilun->type != TYPE_DISK && iscsilun->type != TYPE_ROM) {
1408         bs->sg = 1;
1409     }
1410
1411     task = iscsi_do_inquiry(iscsilun->iscsi, iscsilun->lun, 1,
1412                             SCSI_INQUIRY_PAGECODE_SUPPORTED_VPD_PAGES,
1413                             (void **) &inq_vpd, errp);
1414     if (task == NULL) {
1415         ret = -EINVAL;
1416         goto out;
1417     }
1418     for (i = 0; i < inq_vpd->num_pages; i++) {
1419         struct scsi_task *inq_task;
1420         struct scsi_inquiry_logical_block_provisioning *inq_lbp;
1421         struct scsi_inquiry_block_limits *inq_bl;
1422         switch (inq_vpd->pages[i]) {
1423         case SCSI_INQUIRY_PAGECODE_LOGICAL_BLOCK_PROVISIONING:
1424             inq_task = iscsi_do_inquiry(iscsilun->iscsi, iscsilun->lun, 1,
1425                                         SCSI_INQUIRY_PAGECODE_LOGICAL_BLOCK_PROVISIONING,
1426                                         (void **) &inq_lbp, errp);
1427             if (inq_task == NULL) {
1428                 ret = -EINVAL;
1429                 goto out;
1430             }
1431             memcpy(&iscsilun->lbp, inq_lbp,
1432                    sizeof(struct scsi_inquiry_logical_block_provisioning));
1433             scsi_free_scsi_task(inq_task);
1434             break;
1435         case SCSI_INQUIRY_PAGECODE_BLOCK_LIMITS:
1436             inq_task = iscsi_do_inquiry(iscsilun->iscsi, iscsilun->lun, 1,
1437                                     SCSI_INQUIRY_PAGECODE_BLOCK_LIMITS,
1438                                     (void **) &inq_bl, errp);
1439             if (inq_task == NULL) {
1440                 ret = -EINVAL;
1441                 goto out;
1442             }
1443             memcpy(&iscsilun->bl, inq_bl,
1444                    sizeof(struct scsi_inquiry_block_limits));
1445             scsi_free_scsi_task(inq_task);
1446             break;
1447         default:
1448             break;
1449         }
1450     }
1451     scsi_free_scsi_task(task);
1452     task = NULL;
1453
1454     iscsi_attach_aio_context(bs, iscsilun->aio_context);
1455
1456     /* Guess the internal cluster (page) size of the iscsi target by the means
1457      * of opt_unmap_gran. Transfer the unmap granularity only if it has a
1458      * reasonable size */
1459     if (iscsilun->bl.opt_unmap_gran * iscsilun->block_size >= 4 * 1024 &&
1460         iscsilun->bl.opt_unmap_gran * iscsilun->block_size <= 16 * 1024 * 1024) {
1461         iscsilun->cluster_sectors = (iscsilun->bl.opt_unmap_gran *
1462                                      iscsilun->block_size) >> BDRV_SECTOR_BITS;
1463         if (iscsilun->lbprz && !(bs->open_flags & BDRV_O_NOCACHE)) {
1464             iscsilun->allocationmap = iscsi_allocationmap_init(iscsilun);
1465             if (iscsilun->allocationmap == NULL) {
1466                 ret = -ENOMEM;
1467             }
1468         }
1469     }
1470
1471 out:
1472     qemu_opts_del(opts);
1473     g_free(initiator_name);
1474     if (iscsi_url != NULL) {
1475         iscsi_destroy_url(iscsi_url);
1476     }
1477     if (task != NULL) {
1478         scsi_free_scsi_task(task);
1479     }
1480
1481     if (ret) {
1482         if (iscsi != NULL) {
1483             iscsi_destroy_context(iscsi);
1484         }
1485         memset(iscsilun, 0, sizeof(IscsiLun));
1486     }
1487     return ret;
1488 }
1489
1490 static void iscsi_close(BlockDriverState *bs)
1491 {
1492     IscsiLun *iscsilun = bs->opaque;
1493     struct iscsi_context *iscsi = iscsilun->iscsi;
1494
1495     iscsi_detach_aio_context(bs);
1496     iscsi_destroy_context(iscsi);
1497     g_free(iscsilun->zeroblock);
1498     g_free(iscsilun->allocationmap);
1499     memset(iscsilun, 0, sizeof(IscsiLun));
1500 }
1501
1502 static int sector_limits_lun2qemu(int64_t sector, IscsiLun *iscsilun)
1503 {
1504     return MIN(sector_lun2qemu(sector, iscsilun), INT_MAX / 2 + 1);
1505 }
1506
1507 static void iscsi_refresh_limits(BlockDriverState *bs, Error **errp)
1508 {
1509     /* We don't actually refresh here, but just return data queried in
1510      * iscsi_open(): iscsi targets don't change their limits. */
1511
1512     IscsiLun *iscsilun = bs->opaque;
1513     uint32_t max_xfer_len = iscsilun->use_16_for_rw ? 0xffffffff : 0xffff;
1514
1515     if (iscsilun->bl.max_xfer_len) {
1516         max_xfer_len = MIN(max_xfer_len, iscsilun->bl.max_xfer_len);
1517     }
1518
1519     bs->bl.max_transfer_length = sector_limits_lun2qemu(max_xfer_len, iscsilun);
1520
1521     if (iscsilun->lbp.lbpu) {
1522         if (iscsilun->bl.max_unmap < 0xffffffff) {
1523             bs->bl.max_discard =
1524                 sector_limits_lun2qemu(iscsilun->bl.max_unmap, iscsilun);
1525         }
1526         bs->bl.discard_alignment =
1527             sector_limits_lun2qemu(iscsilun->bl.opt_unmap_gran, iscsilun);
1528     }
1529
1530     if (iscsilun->bl.max_ws_len < 0xffffffff) {
1531         bs->bl.max_write_zeroes =
1532             sector_limits_lun2qemu(iscsilun->bl.max_ws_len, iscsilun);
1533     }
1534     if (iscsilun->lbp.lbpws) {
1535         bs->bl.write_zeroes_alignment =
1536             sector_limits_lun2qemu(iscsilun->bl.opt_unmap_gran, iscsilun);
1537     }
1538     bs->bl.opt_transfer_length =
1539         sector_limits_lun2qemu(iscsilun->bl.opt_xfer_len, iscsilun);
1540 }
1541
1542 /* Note that this will not re-establish a connection with an iSCSI target - it
1543  * is effectively a NOP.  */
1544 static int iscsi_reopen_prepare(BDRVReopenState *state,
1545                                 BlockReopenQueue *queue, Error **errp)
1546 {
1547     IscsiLun *iscsilun = state->bs->opaque;
1548
1549     if (state->flags & BDRV_O_RDWR && iscsilun->write_protected) {
1550         error_setg(errp, "Cannot open a write protected LUN as read-write");
1551         return -EACCES;
1552     }
1553     return 0;
1554 }
1555
1556 static int iscsi_truncate(BlockDriverState *bs, int64_t offset)
1557 {
1558     IscsiLun *iscsilun = bs->opaque;
1559     Error *local_err = NULL;
1560
1561     if (iscsilun->type != TYPE_DISK) {
1562         return -ENOTSUP;
1563     }
1564
1565     iscsi_readcapacity_sync(iscsilun, &local_err);
1566     if (local_err != NULL) {
1567         error_free(local_err);
1568         return -EIO;
1569     }
1570
1571     if (offset > iscsi_getlength(bs)) {
1572         return -EINVAL;
1573     }
1574
1575     if (iscsilun->allocationmap != NULL) {
1576         g_free(iscsilun->allocationmap);
1577         iscsilun->allocationmap = iscsi_allocationmap_init(iscsilun);
1578     }
1579
1580     return 0;
1581 }
1582
1583 static int iscsi_create(const char *filename, QemuOpts *opts, Error **errp)
1584 {
1585     int ret = 0;
1586     int64_t total_size = 0;
1587     BlockDriverState *bs;
1588     IscsiLun *iscsilun = NULL;
1589     QDict *bs_options;
1590
1591     bs = bdrv_new();
1592
1593     /* Read out options */
1594     total_size = DIV_ROUND_UP(qemu_opt_get_size_del(opts, BLOCK_OPT_SIZE, 0),
1595                               BDRV_SECTOR_SIZE);
1596     bs->opaque = g_new0(struct IscsiLun, 1);
1597     iscsilun = bs->opaque;
1598
1599     bs_options = qdict_new();
1600     qdict_put(bs_options, "filename", qstring_from_str(filename));
1601     ret = iscsi_open(bs, bs_options, 0, NULL);
1602     QDECREF(bs_options);
1603
1604     if (ret != 0) {
1605         goto out;
1606     }
1607     iscsi_detach_aio_context(bs);
1608     if (iscsilun->type != TYPE_DISK) {
1609         ret = -ENODEV;
1610         goto out;
1611     }
1612     if (bs->total_sectors < total_size) {
1613         ret = -ENOSPC;
1614         goto out;
1615     }
1616
1617     ret = 0;
1618 out:
1619     if (iscsilun->iscsi != NULL) {
1620         iscsi_destroy_context(iscsilun->iscsi);
1621     }
1622     g_free(bs->opaque);
1623     bs->opaque = NULL;
1624     bdrv_unref(bs);
1625     return ret;
1626 }
1627
1628 static int iscsi_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
1629 {
1630     IscsiLun *iscsilun = bs->opaque;
1631     bdi->unallocated_blocks_are_zero = !!iscsilun->lbprz;
1632     bdi->can_write_zeroes_with_unmap = iscsilun->lbprz && iscsilun->lbp.lbpws;
1633     bdi->cluster_size = iscsilun->cluster_sectors * BDRV_SECTOR_SIZE;
1634     return 0;
1635 }
1636
1637 static QemuOptsList iscsi_create_opts = {
1638     .name = "iscsi-create-opts",
1639     .head = QTAILQ_HEAD_INITIALIZER(iscsi_create_opts.head),
1640     .desc = {
1641         {
1642             .name = BLOCK_OPT_SIZE,
1643             .type = QEMU_OPT_SIZE,
1644             .help = "Virtual disk size"
1645         },
1646         { /* end of list */ }
1647     }
1648 };
1649
1650 static BlockDriver bdrv_iscsi = {
1651     .format_name     = "iscsi",
1652     .protocol_name   = "iscsi",
1653
1654     .instance_size   = sizeof(IscsiLun),
1655     .bdrv_needs_filename = true,
1656     .bdrv_file_open  = iscsi_open,
1657     .bdrv_close      = iscsi_close,
1658     .bdrv_create     = iscsi_create,
1659     .create_opts     = &iscsi_create_opts,
1660     .bdrv_reopen_prepare  = iscsi_reopen_prepare,
1661
1662     .bdrv_getlength  = iscsi_getlength,
1663     .bdrv_get_info   = iscsi_get_info,
1664     .bdrv_truncate   = iscsi_truncate,
1665     .bdrv_refresh_limits = iscsi_refresh_limits,
1666
1667     .bdrv_co_get_block_status = iscsi_co_get_block_status,
1668     .bdrv_co_discard      = iscsi_co_discard,
1669     .bdrv_co_write_zeroes = iscsi_co_write_zeroes,
1670     .bdrv_co_readv         = iscsi_co_readv,
1671     .bdrv_co_writev        = iscsi_co_writev,
1672     .bdrv_co_flush_to_disk = iscsi_co_flush,
1673
1674 #ifdef __linux__
1675     .bdrv_ioctl       = iscsi_ioctl,
1676     .bdrv_aio_ioctl   = iscsi_aio_ioctl,
1677 #endif
1678
1679     .bdrv_detach_aio_context = iscsi_detach_aio_context,
1680     .bdrv_attach_aio_context = iscsi_attach_aio_context,
1681 };
1682
1683 static QemuOptsList qemu_iscsi_opts = {
1684     .name = "iscsi",
1685     .head = QTAILQ_HEAD_INITIALIZER(qemu_iscsi_opts.head),
1686     .desc = {
1687         {
1688             .name = "user",
1689             .type = QEMU_OPT_STRING,
1690             .help = "username for CHAP authentication to target",
1691         },{
1692             .name = "password",
1693             .type = QEMU_OPT_STRING,
1694             .help = "password for CHAP authentication to target",
1695         },{
1696             .name = "header-digest",
1697             .type = QEMU_OPT_STRING,
1698             .help = "HeaderDigest setting. "
1699                     "{CRC32C|CRC32C-NONE|NONE-CRC32C|NONE}",
1700         },{
1701             .name = "initiator-name",
1702             .type = QEMU_OPT_STRING,
1703             .help = "Initiator iqn name to use when connecting",
1704         },
1705         { /* end of list */ }
1706     },
1707 };
1708
1709 static void iscsi_block_init(void)
1710 {
1711     bdrv_register(&bdrv_iscsi);
1712     qemu_add_opts(&qemu_iscsi_opts);
1713 }
1714
1715 block_init(iscsi_block_init);