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     QEMUTimer *event_timer;
  60     uint8_t lbpme;
  61     uint8_t lbprz;
  62     uint8_t has_write_same;
  63     struct scsi_inquiry_logical_block_provisioning lbp;
  64     struct scsi_inquiry_block_limits bl;
  65     unsigned char *zeroblock;
  66     unsigned long *allocationmap;
  67     int cluster_sectors;
  68     bool use_16_for_rw;
  69     bool write_protected;
  70 } IscsiLun;
  71
  72 typedef struct IscsiTask {
  73     int status;
  74     int complete;
  75     int retries;
  76     int do_retry;
  77     struct scsi_task *task;
  78     Coroutine *co;
  79     QEMUBH *bh;
  80     IscsiLun *iscsilun;
  81     QEMUTimer retry_timer;
  82 } IscsiTask;
  83
  84 typedef struct IscsiAIOCB {
  85     BlockAIOCB common;
  86     QEMUIOVector *qiov;
  87     QEMUBH *bh;
  88     IscsiLun *iscsilun;
  89     struct scsi_task *task;
  90     uint8_t *buf;
  91     int status;
  92     int64_t sector_num;
  93     int nb_sectors;
  94 #ifdef __linux__
  95     sg_io_hdr_t *ioh;
  96 #endif
  97 } IscsiAIOCB;
  98
  99 #define EVENT_INTERVAL 250
 100 #define NOP_INTERVAL 5000
 101 #define MAX_NOP_FAILURES 3
 102 #define ISCSI_CMD_RETRIES ARRAY_SIZE(iscsi_retry_times)
 103 static const unsigned iscsi_retry_times[] = {8, 32, 128, 512, 2048};
 104
 105 /* this threshold is a trade-off knob to choose between
 106  * the potential additional overhead of an extra GET_LBA_STATUS request
 107  * vs. unnecessarily reading a lot of zero sectors over the wire.
 108  * If a read request is greater or equal than ISCSI_CHECKALLOC_THRES
 109  * sectors we check the allocation status of the area covered by the
 110  * request first if the allocationmap indicates that the area might be
 111  * unallocated. */
 112 #define ISCSI_CHECKALLOC_THRES 64
 113
 114 static void
 115 iscsi_bh_cb(void *p)
 116 {
 117     IscsiAIOCB *acb = p;
 118
 119     qemu_bh_delete(acb->bh);
 120
 121     g_free(acb->buf);
 122     acb->buf = NULL;
 123
 124     acb->common.cb(acb->common.opaque, acb->status);
 125
 126     if (acb->task != NULL) {
 127         scsi_free_scsi_task(acb->task);
 128         acb->task = NULL;
 129     }
 130
 131     qemu_aio_unref(acb);
 132 }
 133
 134 static void
 135 iscsi_schedule_bh(IscsiAIOCB *acb)
 136 {
 137     if (acb->bh) {
 138         return;
 139     }
 140     acb->bh = aio_bh_new(acb->iscsilun->aio_context, iscsi_bh_cb, acb);
 141     qemu_bh_schedule(acb->bh);
 142 }
 143
 144 static void iscsi_co_generic_bh_cb(void *opaque)
 145 {
 146     struct IscsiTask *iTask = opaque;
 147     iTask->complete = 1;
 148     qemu_bh_delete(iTask->bh);
 149     qemu_coroutine_enter(iTask->co, NULL);
 150 }
 151
 152 static void iscsi_retry_timer_expired(void *opaque)
 153 {
 154     struct IscsiTask *iTask = opaque;
 155     iTask->complete = 1;
 156     if (iTask->co) {
 157         qemu_coroutine_enter(iTask->co, NULL);
 158     }
 159 }
 160
 161 static inline unsigned exp_random(double mean)
 162 {
 163     return -mean * log((double)rand() / RAND_MAX);
 164 }
 165
 166 static void
 167 iscsi_co_generic_cb(struct iscsi_context *iscsi, int status,
 168                         void *command_data, void *opaque)
 169 {
 170     struct IscsiTask *iTask = opaque;
 171     struct scsi_task *task = command_data;
 172
 173     iTask->status = status;
 174     iTask->do_retry = 0;
 175     iTask->task = task;
 176
 177     if (status != SCSI_STATUS_GOOD) {
 178         if (iTask->retries++ < ISCSI_CMD_RETRIES) {
 179             if (status == SCSI_STATUS_CHECK_CONDITION
 180                 && task->sense.key == SCSI_SENSE_UNIT_ATTENTION) {
 181                 error_report("iSCSI CheckCondition: %s",
 182                              iscsi_get_error(iscsi));
 183                 iTask->do_retry = 1;
 184                 goto out;
 185             }
 186             if (status == SCSI_STATUS_BUSY) {
 187                 unsigned retry_time =
 188                     exp_random(iscsi_retry_times[iTask->retries - 1]);
 189                 error_report("iSCSI Busy (retry #%u in %u ms): %s",
 190                              iTask->retries, retry_time,
 191                              iscsi_get_error(iscsi));
 192                 aio_timer_init(iTask->iscsilun->aio_context,
 193                                &iTask->retry_timer, QEMU_CLOCK_REALTIME,
 194                                SCALE_MS, iscsi_retry_timer_expired, iTask);
 195                 timer_mod(&iTask->retry_timer,
 196                           qemu_clock_get_ms(QEMU_CLOCK_REALTIME) + retry_time);
 197                 iTask->do_retry = 1;
 198                 return;
 199             }
 200         }
 201         error_report("iSCSI Failure: %s", iscsi_get_error(iscsi));
 202     }
 203
 204 out:
 205     if (iTask->co) {
 206         iTask->bh = aio_bh_new(iTask->iscsilun->aio_context,
 207                                iscsi_co_generic_bh_cb, iTask);
 208         qemu_bh_schedule(iTask->bh);
 209     } else {
 210         iTask->complete = 1;
 211     }
 212 }
 213
 214 static void iscsi_co_init_iscsitask(IscsiLun *iscsilun, struct IscsiTask *iTask)
 215 {
 216     *iTask = (struct IscsiTask) {
 217         .co         = qemu_coroutine_self(),
 218         .iscsilun   = iscsilun,
 219     };
 220 }
 221
 222 static void
 223 iscsi_abort_task_cb(struct iscsi_context *iscsi, int status, void *command_data,
 224                     void *private_data)
 225 {
 226     IscsiAIOCB *acb = private_data;
 227
 228     acb->status = -ECANCELED;
 229     iscsi_schedule_bh(acb);
 230 }
 231
 232 static void
 233 iscsi_aio_cancel(BlockAIOCB *blockacb)
 234 {
 235     IscsiAIOCB *acb = (IscsiAIOCB *)blockacb;
 236     IscsiLun *iscsilun = acb->iscsilun;
 237
 238     if (acb->status != -EINPROGRESS) {
 239         return;
 240     }
 241
 242     /* send a task mgmt call to the target to cancel the task on the target */
 243     iscsi_task_mgmt_abort_task_async(iscsilun->iscsi, acb->task,
 244                                      iscsi_abort_task_cb, acb);
 245
 246 }
 247
 248 static const AIOCBInfo iscsi_aiocb_info = {
 249     .aiocb_size         = sizeof(IscsiAIOCB),
 250     .cancel_async       = iscsi_aio_cancel,
 251 };
 252
 253
 254 static void iscsi_process_read(void *arg);
 255 static void iscsi_process_write(void *arg);
 256
 257 static void
 258 iscsi_set_events(IscsiLun *iscsilun)
 259 {
 260     struct iscsi_context *iscsi = iscsilun->iscsi;
 261     int ev = iscsi_which_events(iscsi);
 262
 263     if (ev != iscsilun->events) {
 264         aio_set_fd_handler(iscsilun->aio_context,
 265                            iscsi_get_fd(iscsi),
 266                            (ev & POLLIN) ? iscsi_process_read : NULL,
 267                            (ev & POLLOUT) ? iscsi_process_write : NULL,
 268                            iscsilun);
 269         iscsilun->events = ev;
 270     }
 271
 272     /* newer versions of libiscsi may return zero events. In this
 273      * case start a timer to ensure we are able to return to service
 274      * once this situation changes. */
 275     if (!ev) {
 276         timer_mod(iscsilun->event_timer,
 277                   qemu_clock_get_ms(QEMU_CLOCK_REALTIME) + EVENT_INTERVAL);
 278     }
 279 }
 280
 281 static void iscsi_timed_set_events(void *opaque)
 282 {
 283     IscsiLun *iscsilun = opaque;
 284     iscsi_set_events(iscsilun);
 285 }
 286
 287 static void
 288 iscsi_process_read(void *arg)
 289 {
 290     IscsiLun *iscsilun = arg;
 291     struct iscsi_context *iscsi = iscsilun->iscsi;
 292
 293     iscsi_service(iscsi, POLLIN);
 294     iscsi_set_events(iscsilun);
 295 }
 296
 297 static void
 298 iscsi_process_write(void *arg)
 299 {
 300     IscsiLun *iscsilun = arg;
 301     struct iscsi_context *iscsi = iscsilun->iscsi;
 302
 303     iscsi_service(iscsi, POLLOUT);
 304     iscsi_set_events(iscsilun);
 305 }
 306
 307 static int64_t sector_lun2qemu(int64_t sector, IscsiLun *iscsilun)
 308 {
 309     return sector * iscsilun->block_size / BDRV_SECTOR_SIZE;
 310 }
 311
 312 static int64_t sector_qemu2lun(int64_t sector, IscsiLun *iscsilun)
 313 {
 314     return sector * BDRV_SECTOR_SIZE / iscsilun->block_size;
 315 }
 316
 317 static bool is_request_lun_aligned(int64_t sector_num, int nb_sectors,
 318                                       IscsiLun *iscsilun)
 319 {
 320     if ((sector_num * BDRV_SECTOR_SIZE) % iscsilun->block_size ||
 321         (nb_sectors * BDRV_SECTOR_SIZE) % iscsilun->block_size) {
 322             error_report("iSCSI misaligned request: "
 323                          "iscsilun->block_size %u, sector_num %" PRIi64
 324                          ", nb_sectors %d",
 325                          iscsilun->block_size, sector_num, nb_sectors);
 326             return 0;
 327     }
 328     return 1;
 329 }
 330
 331 static unsigned long *iscsi_allocationmap_init(IscsiLun *iscsilun)
 332 {
 333     return bitmap_try_new(DIV_ROUND_UP(sector_lun2qemu(iscsilun->num_blocks,
 334                                                        iscsilun),
 335                                        iscsilun->cluster_sectors));
 336 }
 337
 338 static void iscsi_allocationmap_set(IscsiLun *iscsilun, int64_t sector_num,
 339                                     int nb_sectors)
 340 {
 341     if (iscsilun->allocationmap == NULL) {
 342         return;
 343     }
 344     bitmap_set(iscsilun->allocationmap,
 345                sector_num / iscsilun->cluster_sectors,
 346                DIV_ROUND_UP(nb_sectors, iscsilun->cluster_sectors));
 347 }
 348
 349 static void iscsi_allocationmap_clear(IscsiLun *iscsilun, int64_t sector_num,
 350                                       int nb_sectors)
 351 {
 352     int64_t cluster_num, nb_clusters;
 353     if (iscsilun->allocationmap == NULL) {
 354         return;
 355     }
 356     cluster_num = DIV_ROUND_UP(sector_num, iscsilun->cluster_sectors);
 357     nb_clusters = (sector_num + nb_sectors) / iscsilun->cluster_sectors
 358                   - cluster_num;
 359     if (nb_clusters > 0) {
 360         bitmap_clear(iscsilun->allocationmap, cluster_num, nb_clusters);
 361     }
 362 }
 363
 364 static int coroutine_fn iscsi_co_writev(BlockDriverState *bs,
 365                                         int64_t sector_num, int nb_sectors,
 366                                         QEMUIOVector *iov)
 367 {
 368     IscsiLun *iscsilun = bs->opaque;
 369     struct IscsiTask iTask;
 370     uint64_t lba;
 371     uint32_t num_sectors;
 372
 373     if (!is_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
 374         return -EINVAL;
 375     }
 376
 377     if (bs->bl.max_transfer_length && nb_sectors > bs->bl.max_transfer_length) {
 378         error_report("iSCSI Error: Write of %d sectors exceeds max_xfer_len "
 379                      "of %d sectors", nb_sectors, bs->bl.max_transfer_length);
 380         return -EINVAL;
 381     }
 382
 383     lba = sector_qemu2lun(sector_num, iscsilun);
 384     num_sectors = sector_qemu2lun(nb_sectors, iscsilun);
 385     iscsi_co_init_iscsitask(iscsilun, &iTask);
 386 retry:
 387     if (iscsilun->use_16_for_rw) {
 388         iTask.task = iscsi_write16_task(iscsilun->iscsi, iscsilun->lun, lba,
 389                                         NULL, num_sectors * iscsilun->block_size,
 390                                         iscsilun->block_size, 0, 0, 0, 0, 0,
 391                                         iscsi_co_generic_cb, &iTask);
 392     } else {
 393         iTask.task = iscsi_write10_task(iscsilun->iscsi, iscsilun->lun, lba,
 394                                         NULL, num_sectors * iscsilun->block_size,
 395                                         iscsilun->block_size, 0, 0, 0, 0, 0,
 396                                         iscsi_co_generic_cb, &iTask);
 397     }
 398     if (iTask.task == NULL) {
 399         return -ENOMEM;
 400     }
 401     scsi_task_set_iov_out(iTask.task, (struct scsi_iovec *) iov->iov,
 402                           iov->niov);
 403     while (!iTask.complete) {
 404         iscsi_set_events(iscsilun);
 405         qemu_coroutine_yield();
 406     }
 407
 408     if (iTask.task != NULL) {
 409         scsi_free_scsi_task(iTask.task);
 410         iTask.task = NULL;
 411     }
 412
 413     if (iTask.do_retry) {
 414         iTask.complete = 0;
 415         goto retry;
 416     }
 417
 418     if (iTask.status != SCSI_STATUS_GOOD) {
 419         return -EIO;
 420     }
 421
 422     iscsi_allocationmap_set(iscsilun, sector_num, nb_sectors);
 423
 424     return 0;
 425 }
 426
 427
 428 static bool iscsi_allocationmap_is_allocated(IscsiLun *iscsilun,
 429                                              int64_t sector_num, int nb_sectors)
 430 {
 431     unsigned long size;
 432     if (iscsilun->allocationmap == NULL) {
 433         return true;
 434     }
 435     size = DIV_ROUND_UP(sector_num + nb_sectors, iscsilun->cluster_sectors);
 436     return !(find_next_bit(iscsilun->allocationmap, size,
 437                            sector_num / iscsilun->cluster_sectors) == size);
 438 }
 439
 440 static int64_t coroutine_fn iscsi_co_get_block_status(BlockDriverState *bs,
 441                                                   int64_t sector_num,
 442                                                   int nb_sectors, int *pnum)
 443 {
 444     IscsiLun *iscsilun = bs->opaque;
 445     struct scsi_get_lba_status *lbas = NULL;
 446     struct scsi_lba_status_descriptor *lbasd = NULL;
 447     struct IscsiTask iTask;
 448     int64_t ret;
 449
 450     iscsi_co_init_iscsitask(iscsilun, &iTask);
 451
 452     if (!is_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
 453         ret = -EINVAL;
 454         goto out;
 455     }
 456
 457     /* default to all sectors allocated */
 458     ret = BDRV_BLOCK_DATA;
 459     ret |= (sector_num << BDRV_SECTOR_BITS) | BDRV_BLOCK_OFFSET_VALID;
 460     *pnum = nb_sectors;
 461
 462     /* LUN does not support logical block provisioning */
 463     if (iscsilun->lbpme == 0) {
 464         goto out;
 465     }
 466
 467 retry:
 468     if (iscsi_get_lba_status_task(iscsilun->iscsi, iscsilun->lun,
 469                                   sector_qemu2lun(sector_num, iscsilun),
 470                                   8 + 16, iscsi_co_generic_cb,
 471                                   &iTask) == NULL) {
 472         ret = -ENOMEM;
 473         goto out;
 474     }
 475
 476     while (!iTask.complete) {
 477         iscsi_set_events(iscsilun);
 478         qemu_coroutine_yield();
 479     }
 480
 481     if (iTask.do_retry) {
 482         if (iTask.task != NULL) {
 483             scsi_free_scsi_task(iTask.task);
 484             iTask.task = NULL;
 485         }
 486         iTask.complete = 0;
 487         goto retry;
 488     }
 489
 490     if (iTask.status != SCSI_STATUS_GOOD) {
 491         /* in case the get_lba_status_callout fails (i.e.
 492          * because the device is busy or the cmd is not
 493          * supported) we pretend all blocks are allocated
 494          * for backwards compatibility */
 495         goto out;
 496     }
 497
 498     lbas = scsi_datain_unmarshall(iTask.task);
 499     if (lbas == NULL) {
 500         ret = -EIO;
 501         goto out;
 502     }
 503
 504     lbasd = &lbas->descriptors[0];
 505
 506     if (sector_qemu2lun(sector_num, iscsilun) != lbasd->lba) {
 507         ret = -EIO;
 508         goto out;
 509     }
 510
 511     *pnum = sector_lun2qemu(lbasd->num_blocks, iscsilun);
 512
 513     if (lbasd->provisioning == SCSI_PROVISIONING_TYPE_DEALLOCATED ||
 514         lbasd->provisioning == SCSI_PROVISIONING_TYPE_ANCHORED) {
 515         ret &= ~BDRV_BLOCK_DATA;
 516         if (iscsilun->lbprz) {
 517             ret |= BDRV_BLOCK_ZERO;
 518         }
 519     }
 520
 521     if (ret & BDRV_BLOCK_ZERO) {
 522         iscsi_allocationmap_clear(iscsilun, sector_num, *pnum);
 523     } else {
 524         iscsi_allocationmap_set(iscsilun, sector_num, *pnum);
 525     }
 526
 527     if (*pnum > nb_sectors) {
 528         *pnum = nb_sectors;
 529     }
 530 out:
 531     if (iTask.task != NULL) {
 532         scsi_free_scsi_task(iTask.task);
 533     }
 534     return ret;
 535 }
 536
 537 static int coroutine_fn iscsi_co_readv(BlockDriverState *bs,
 538                                        int64_t sector_num, int nb_sectors,
 539                                        QEMUIOVector *iov)
 540 {
 541     IscsiLun *iscsilun = bs->opaque;
 542     struct IscsiTask iTask;
 543     uint64_t lba;
 544     uint32_t num_sectors;
 545
 546     if (!is_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
 547         return -EINVAL;
 548     }
 549
 550     if (bs->bl.max_transfer_length && nb_sectors > bs->bl.max_transfer_length) {
 551         error_report("iSCSI Error: Read of %d sectors exceeds max_xfer_len "
 552                      "of %d sectors", nb_sectors, bs->bl.max_transfer_length);
 553         return -EINVAL;
 554     }
 555
 556     if (iscsilun->lbprz && nb_sectors >= ISCSI_CHECKALLOC_THRES &&
 557         !iscsi_allocationmap_is_allocated(iscsilun, sector_num, nb_sectors)) {
 558         int64_t ret;
 559         int pnum;
 560         ret = iscsi_co_get_block_status(bs, sector_num, INT_MAX, &pnum);
 561         if (ret < 0) {
 562             return ret;
 563         }
 564         if (ret & BDRV_BLOCK_ZERO && pnum >= nb_sectors) {
 565             qemu_iovec_memset(iov, 0, 0x00, iov->size);
 566             return 0;
 567         }
 568     }
 569
 570     lba = sector_qemu2lun(sector_num, iscsilun);
 571     num_sectors = sector_qemu2lun(nb_sectors, iscsilun);
 572
 573     iscsi_co_init_iscsitask(iscsilun, &iTask);
 574 retry:
 575     if (iscsilun->use_16_for_rw) {
 576         iTask.task = iscsi_read16_task(iscsilun->iscsi, iscsilun->lun, lba,
 577                                        num_sectors * iscsilun->block_size,
 578                                        iscsilun->block_size, 0, 0, 0, 0, 0,
 579                                        iscsi_co_generic_cb, &iTask);
 580     } else {
 581         iTask.task = iscsi_read10_task(iscsilun->iscsi, iscsilun->lun, lba,
 582                                        num_sectors * iscsilun->block_size,
 583                                        iscsilun->block_size,
 584                                        0, 0, 0, 0, 0,
 585                                        iscsi_co_generic_cb, &iTask);
 586     }
 587     if (iTask.task == NULL) {
 588         return -ENOMEM;
 589     }
 590     scsi_task_set_iov_in(iTask.task, (struct scsi_iovec *) iov->iov, iov->niov);
 591
 592     while (!iTask.complete) {
 593         iscsi_set_events(iscsilun);
 594         qemu_coroutine_yield();
 595     }
 596
 597     if (iTask.task != NULL) {
 598         scsi_free_scsi_task(iTask.task);
 599         iTask.task = NULL;
 600     }
 601
 602     if (iTask.do_retry) {
 603         iTask.complete = 0;
 604         goto retry;
 605     }
 606
 607     if (iTask.status != SCSI_STATUS_GOOD) {
 608         return -EIO;
 609     }
 610
 611     return 0;
 612 }
 613
 614 static int coroutine_fn iscsi_co_flush(BlockDriverState *bs)
 615 {
 616     IscsiLun *iscsilun = bs->opaque;
 617     struct IscsiTask iTask;
 618
 619     if (bs->sg) {
 620         return 0;
 621     }
 622
 623     iscsi_co_init_iscsitask(iscsilun, &iTask);
 624
 625 retry:
 626     if (iscsi_synchronizecache10_task(iscsilun->iscsi, iscsilun->lun, 0, 0, 0,
 627                                       0, iscsi_co_generic_cb, &iTask) == NULL) {
 628         return -ENOMEM;
 629     }
 630
 631     while (!iTask.complete) {
 632         iscsi_set_events(iscsilun);
 633         qemu_coroutine_yield();
 634     }
 635
 636     if (iTask.task != NULL) {
 637         scsi_free_scsi_task(iTask.task);
 638         iTask.task = NULL;
 639     }
 640
 641     if (iTask.do_retry) {
 642         iTask.complete = 0;
 643         goto retry;
 644     }
 645
 646     if (iTask.status != SCSI_STATUS_GOOD) {
 647         return -EIO;
 648     }
 649
 650     return 0;
 651 }
 652
 653 #ifdef __linux__
 654 static void
 655 iscsi_aio_ioctl_cb(struct iscsi_context *iscsi, int status,
 656                      void *command_data, void *opaque)
 657 {
 658     IscsiAIOCB *acb = opaque;
 659
 660     g_free(acb->buf);
 661     acb->buf = NULL;
 662
 663     acb->status = 0;
 664     if (status < 0) {
 665         error_report("Failed to ioctl(SG_IO) to iSCSI lun. %s",
 666                      iscsi_get_error(iscsi));
 667         acb->status = -EIO;
 668     }
 669
 670     acb->ioh->driver_status = 0;
 671     acb->ioh->host_status   = 0;
 672     acb->ioh->resid         = 0;
 673
 674 #define SG_ERR_DRIVER_SENSE    0x08
 675
 676     if (status == SCSI_STATUS_CHECK_CONDITION && acb->task->datain.size >= 2) {
 677         int ss;
 678
 679         acb->ioh->driver_status |= SG_ERR_DRIVER_SENSE;
 680
 681         acb->ioh->sb_len_wr = acb->task->datain.size - 2;
 682         ss = (acb->ioh->mx_sb_len >= acb->ioh->sb_len_wr) ?
 683              acb->ioh->mx_sb_len : acb->ioh->sb_len_wr;
 684         memcpy(acb->ioh->sbp, &acb->task->datain.data[2], ss);
 685     }
 686
 687     iscsi_schedule_bh(acb);
 688 }
 689
 690 static BlockAIOCB *iscsi_aio_ioctl(BlockDriverState *bs,
 691         unsigned long int req, void *buf,
 692         BlockCompletionFunc *cb, void *opaque)
 693 {
 694     IscsiLun *iscsilun = bs->opaque;
 695     struct iscsi_context *iscsi = iscsilun->iscsi;
 696     struct iscsi_data data;
 697     IscsiAIOCB *acb;
 698
 699     assert(req == SG_IO);
 700
 701     acb = qemu_aio_get(&iscsi_aiocb_info, bs, cb, opaque);
 702
 703     acb->iscsilun = iscsilun;
 704     acb->bh          = NULL;
 705     acb->status      = -EINPROGRESS;
 706     acb->buf         = NULL;
 707     acb->ioh         = buf;
 708
 709     acb->task = malloc(sizeof(struct scsi_task));
 710     if (acb->task == NULL) {
 711         error_report("iSCSI: Failed to allocate task for scsi command. %s",
 712                      iscsi_get_error(iscsi));
 713         qemu_aio_unref(acb);
 714         return NULL;
 715     }
 716     memset(acb->task, 0, sizeof(struct scsi_task));
 717
 718     switch (acb->ioh->dxfer_direction) {
 719     case SG_DXFER_TO_DEV:
 720         acb->task->xfer_dir = SCSI_XFER_WRITE;
 721         break;
 722     case SG_DXFER_FROM_DEV:
 723         acb->task->xfer_dir = SCSI_XFER_READ;
 724         break;
 725     default:
 726         acb->task->xfer_dir = SCSI_XFER_NONE;
 727         break;
 728     }
 729
 730     acb->task->cdb_size = acb->ioh->cmd_len;
 731     memcpy(&acb->task->cdb[0], acb->ioh->cmdp, acb->ioh->cmd_len);
 732     acb->task->expxferlen = acb->ioh->dxfer_len;
 733
 734     data.size = 0;
 735     if (acb->task->xfer_dir == SCSI_XFER_WRITE) {
 736         if (acb->ioh->iovec_count == 0) {
 737             data.data = acb->ioh->dxferp;
 738             data.size = acb->ioh->dxfer_len;
 739         } else {
 740             scsi_task_set_iov_out(acb->task,
 741                                  (struct scsi_iovec *) acb->ioh->dxferp,
 742                                  acb->ioh->iovec_count);
 743         }
 744     }
 745
 746     if (iscsi_scsi_command_async(iscsi, iscsilun->lun, acb->task,
 747                                  iscsi_aio_ioctl_cb,
 748                                  (data.size > 0) ? &data : NULL,
 749                                  acb) != 0) {
 750         scsi_free_scsi_task(acb->task);
 751         qemu_aio_unref(acb);
 752         return NULL;
 753     }
 754
 755     /* tell libiscsi to read straight into the buffer we got from ioctl */
 756     if (acb->task->xfer_dir == SCSI_XFER_READ) {
 757         if (acb->ioh->iovec_count == 0) {
 758             scsi_task_add_data_in_buffer(acb->task,
 759                                          acb->ioh->dxfer_len,
 760                                          acb->ioh->dxferp);
 761         } else {
 762             scsi_task_set_iov_in(acb->task,
 763                                  (struct scsi_iovec *) acb->ioh->dxferp,
 764                                  acb->ioh->iovec_count);
 765         }
 766     }
 767
 768     iscsi_set_events(iscsilun);
 769
 770     return &acb->common;
 771 }
 772
 773 static void ioctl_cb(void *opaque, int status)
 774 {
 775     int *p_status = opaque;
 776     *p_status = status;
 777 }
 778
 779 static int iscsi_ioctl(BlockDriverState *bs, unsigned long int req, void *buf)
 780 {
 781     IscsiLun *iscsilun = bs->opaque;
 782     int status;
 783
 784     switch (req) {
 785     case SG_GET_VERSION_NUM:
 786         *(int *)buf = 30000;
 787         break;
 788     case SG_GET_SCSI_ID:
 789         ((struct sg_scsi_id *)buf)->scsi_type = iscsilun->type;
 790         break;
 791     case SG_IO:
 792         status = -EINPROGRESS;
 793         iscsi_aio_ioctl(bs, req, buf, ioctl_cb, &status);
 794
 795         while (status == -EINPROGRESS) {
 796             aio_poll(iscsilun->aio_context, true);
 797         }
 798
 799         return 0;
 800     default:
 801         return -1;
 802     }
 803     return 0;
 804 }
 805 #endif
 806
 807 static int64_t
 808 iscsi_getlength(BlockDriverState *bs)
 809 {
 810     IscsiLun *iscsilun = bs->opaque;
 811     int64_t len;
 812
 813     len  = iscsilun->num_blocks;
 814     len *= iscsilun->block_size;
 815
 816     return len;
 817 }
 818
 819 static int
 820 coroutine_fn iscsi_co_discard(BlockDriverState *bs, int64_t sector_num,
 821                                    int nb_sectors)
 822 {
 823     IscsiLun *iscsilun = bs->opaque;
 824     struct IscsiTask iTask;
 825     struct unmap_list list;
 826
 827     if (!is_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
 828         return -EINVAL;
 829     }
 830
 831     if (!iscsilun->lbp.lbpu) {
 832         /* UNMAP is not supported by the target */
 833         return 0;
 834     }
 835
 836     list.lba = sector_qemu2lun(sector_num, iscsilun);
 837     list.num = sector_qemu2lun(nb_sectors, iscsilun);
 838
 839     iscsi_co_init_iscsitask(iscsilun, &iTask);
 840 retry:
 841     if (iscsi_unmap_task(iscsilun->iscsi, iscsilun->lun, 0, 0, &list, 1,
 842                      iscsi_co_generic_cb, &iTask) == NULL) {
 843         return -ENOMEM;
 844     }
 845
 846     while (!iTask.complete) {
 847         iscsi_set_events(iscsilun);
 848         qemu_coroutine_yield();
 849     }
 850
 851     if (iTask.task != NULL) {
 852         scsi_free_scsi_task(iTask.task);
 853         iTask.task = NULL;
 854     }
 855
 856     if (iTask.do_retry) {
 857         iTask.complete = 0;
 858         goto retry;
 859     }
 860
 861     if (iTask.status == SCSI_STATUS_CHECK_CONDITION) {
 862         /* the target might fail with a check condition if it
 863            is not happy with the alignment of the UNMAP request
 864            we silently fail in this case */
 865         return 0;
 866     }
 867
 868     if (iTask.status != SCSI_STATUS_GOOD) {
 869         return -EIO;
 870     }
 871
 872     iscsi_allocationmap_clear(iscsilun, sector_num, nb_sectors);
 873
 874     return 0;
 875 }
 876
 877 static int
 878 coroutine_fn iscsi_co_write_zeroes(BlockDriverState *bs, int64_t sector_num,
 879                                    int nb_sectors, BdrvRequestFlags flags)
 880 {
 881     IscsiLun *iscsilun = bs->opaque;
 882     struct IscsiTask iTask;
 883     uint64_t lba;
 884     uint32_t nb_blocks;
 885     bool use_16_for_ws = iscsilun->use_16_for_rw;
 886
 887     if (!is_request_lun_aligned(sector_num, nb_sectors, iscsilun)) {
 888         return -EINVAL;
 889     }
 890
 891     if (flags & BDRV_REQ_MAY_UNMAP) {
 892         if (!use_16_for_ws && !iscsilun->lbp.lbpws10) {
 893             /* WRITESAME10 with UNMAP is unsupported try WRITESAME16 */
 894             use_16_for_ws = true;
 895         }
 896         if (use_16_for_ws && !iscsilun->lbp.lbpws) {
 897             /* WRITESAME16 with UNMAP is not supported by the target,
 898              * fall back and try WRITESAME10/16 without UNMAP */
 899             flags &= ~BDRV_REQ_MAY_UNMAP;
 900             use_16_for_ws = iscsilun->use_16_for_rw;
 901         }
 902     }
 903
 904     if (!(flags & BDRV_REQ_MAY_UNMAP) && !iscsilun->has_write_same) {
 905         /* WRITESAME without UNMAP is not supported by the target */
 906         return -ENOTSUP;
 907     }
 908
 909     lba = sector_qemu2lun(sector_num, iscsilun);
 910     nb_blocks = sector_qemu2lun(nb_sectors, iscsilun);
 911
 912     if (iscsilun->zeroblock == NULL) {
 913         iscsilun->zeroblock = g_try_malloc0(iscsilun->block_size);
 914         if (iscsilun->zeroblock == NULL) {
 915             return -ENOMEM;
 916         }
 917     }
 918
 919     iscsi_co_init_iscsitask(iscsilun, &iTask);
 920 retry:
 921     if (use_16_for_ws) {
 922         iTask.task = iscsi_writesame16_task(iscsilun->iscsi, iscsilun->lun, lba,
 923                                             iscsilun->zeroblock, iscsilun->block_size,
 924                                             nb_blocks, 0, !!(flags & BDRV_REQ_MAY_UNMAP),
 925                                             0, 0, iscsi_co_generic_cb, &iTask);
 926     } else {
 927         iTask.task = iscsi_writesame10_task(iscsilun->iscsi, iscsilun->lun, lba,
 928                                             iscsilun->zeroblock, iscsilun->block_size,
 929                                             nb_blocks, 0, !!(flags & BDRV_REQ_MAY_UNMAP),
 930                                             0, 0, iscsi_co_generic_cb, &iTask);
 931     }
 932     if (iTask.task == NULL) {
 933         return -ENOMEM;
 934     }
 935
 936     while (!iTask.complete) {
 937         iscsi_set_events(iscsilun);
 938         qemu_coroutine_yield();
 939     }
 940
 941     if (iTask.status == SCSI_STATUS_CHECK_CONDITION &&
 942         iTask.task->sense.key == SCSI_SENSE_ILLEGAL_REQUEST &&
 943         (iTask.task->sense.ascq == SCSI_SENSE_ASCQ_INVALID_OPERATION_CODE ||
 944          iTask.task->sense.ascq == SCSI_SENSE_ASCQ_INVALID_FIELD_IN_CDB)) {
 945         /* WRITE SAME is not supported by the target */
 946         iscsilun->has_write_same = false;
 947         scsi_free_scsi_task(iTask.task);
 948         return -ENOTSUP;
 949     }
 950
 951     if (iTask.task != NULL) {
 952         scsi_free_scsi_task(iTask.task);
 953         iTask.task = NULL;
 954     }
 955
 956     if (iTask.do_retry) {
 957         iTask.complete = 0;
 958         goto retry;
 959     }
 960
 961     if (iTask.status != SCSI_STATUS_GOOD) {
 962         return -EIO;
 963     }
 964
 965     if (flags & BDRV_REQ_MAY_UNMAP) {
 966         iscsi_allocationmap_clear(iscsilun, sector_num, nb_sectors);
 967     } else {
 968         iscsi_allocationmap_set(iscsilun, sector_num, nb_sectors);
 969     }
 970
 971     return 0;
 972 }
 973
 974 static void parse_chap(struct iscsi_context *iscsi, const char *target,
 975                        Error **errp)
 976 {
 977     QemuOptsList *list;
 978     QemuOpts *opts;
 979     const char *user = NULL;
 980     const char *password = NULL;
 981
 982     list = qemu_find_opts("iscsi");
 983     if (!list) {
 984         return;
 985     }
 986
 987     opts = qemu_opts_find(list, target);
 988     if (opts == NULL) {
 989         opts = QTAILQ_FIRST(&list->head);
 990         if (!opts) {
 991             return;
 992         }
 993     }
 994
 995     user = qemu_opt_get(opts, "user");
 996     if (!user) {
 997         return;
 998     }
 999
1000     password = qemu_opt_get(opts, "password");
1001     if (!password) {
1002         error_setg(errp, "CHAP username specified but no password was given");
1003         return;
1004     }
1005
1006     if (iscsi_set_initiator_username_pwd(iscsi, user, password)) {
1007         error_setg(errp, "Failed to set initiator username and password");
1008     }
1009 }
1010
1011 static void parse_header_digest(struct iscsi_context *iscsi, const char *target,
1012                                 Error **errp)
1013 {
1014     QemuOptsList *list;
1015     QemuOpts *opts;
1016     const char *digest = NULL;
1017
1018     list = qemu_find_opts("iscsi");
1019     if (!list) {
1020         return;
1021     }
1022
1023     opts = qemu_opts_find(list, target);
1024     if (opts == NULL) {
1025         opts = QTAILQ_FIRST(&list->head);
1026         if (!opts) {
1027             return;
1028         }
1029     }
1030
1031     digest = qemu_opt_get(opts, "header-digest");
1032     if (!digest) {
1033         return;
1034     }
1035
1036     if (!strcmp(digest, "CRC32C")) {
1037         iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_CRC32C);
1038     } else if (!strcmp(digest, "NONE")) {
1039         iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_NONE);
1040     } else if (!strcmp(digest, "CRC32C-NONE")) {
1041         iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_CRC32C_NONE);
1042     } else if (!strcmp(digest, "NONE-CRC32C")) {
1043         iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_NONE_CRC32C);
1044     } else {
1045         error_setg(errp, "Invalid header-digest setting : %s", digest);
1046     }
1047 }
1048
1049 static char *parse_initiator_name(const char *target)
1050 {
1051     QemuOptsList *list;
1052     QemuOpts *opts;
1053     const char *name;
1054     char *iscsi_name;
1055     UuidInfo *uuid_info;
1056
1057     list = qemu_find_opts("iscsi");
1058     if (list) {
1059         opts = qemu_opts_find(list, target);
1060         if (!opts) {
1061             opts = QTAILQ_FIRST(&list->head);
1062         }
1063         if (opts) {
1064             name = qemu_opt_get(opts, "initiator-name");
1065             if (name) {
1066                 return g_strdup(name);
1067             }
1068         }
1069     }
1070
1071     uuid_info = qmp_query_uuid(NULL);
1072     if (strcmp(uuid_info->UUID, UUID_NONE) == 0) {
1073         name = qemu_get_vm_name();
1074     } else {
1075         name = uuid_info->UUID;
1076     }
1077     iscsi_name = g_strdup_printf("iqn.2008-11.org.linux-kvm%s%s",
1078                                  name ? ":" : "", name ? name : "");
1079     qapi_free_UuidInfo(uuid_info);
1080     return iscsi_name;
1081 }
1082
1083 static void iscsi_nop_timed_event(void *opaque)
1084 {
1085     IscsiLun *iscsilun = opaque;
1086
1087     if (iscsi_get_nops_in_flight(iscsilun->iscsi) > MAX_NOP_FAILURES) {
1088         error_report("iSCSI: NOP timeout. Reconnecting...");
1089         iscsi_reconnect(iscsilun->iscsi);
1090     }
1091
1092     if (iscsi_nop_out_async(iscsilun->iscsi, NULL, NULL, 0, NULL) != 0) {
1093         error_report("iSCSI: failed to sent NOP-Out. Disabling NOP messages.");
1094         return;
1095     }
1096
1097     timer_mod(iscsilun->nop_timer, qemu_clock_get_ms(QEMU_CLOCK_REALTIME) + NOP_INTERVAL);
1098     iscsi_set_events(iscsilun);
1099 }
1100
1101 static void iscsi_readcapacity_sync(IscsiLun *iscsilun, Error **errp)
1102 {
1103     struct scsi_task *task = NULL;
1104     struct scsi_readcapacity10 *rc10 = NULL;
1105     struct scsi_readcapacity16 *rc16 = NULL;
1106     int retries = ISCSI_CMD_RETRIES;
1107
1108     do {
1109         if (task != NULL) {
1110             scsi_free_scsi_task(task);
1111             task = NULL;
1112         }
1113
1114         switch (iscsilun->type) {
1115         case TYPE_DISK:
1116             task = iscsi_readcapacity16_sync(iscsilun->iscsi, iscsilun->lun);
1117             if (task != NULL && task->status == SCSI_STATUS_GOOD) {
1118                 rc16 = scsi_datain_unmarshall(task);
1119                 if (rc16 == NULL) {
1120                     error_setg(errp, "iSCSI: Failed to unmarshall readcapacity16 data.");
1121                 } else {
1122                     iscsilun->block_size = rc16->block_length;
1123                     iscsilun->num_blocks = rc16->returned_lba + 1;
1124                     iscsilun->lbpme = rc16->lbpme;
1125                     iscsilun->lbprz = rc16->lbprz;
1126                     iscsilun->use_16_for_rw = (rc16->returned_lba > 0xffffffff);
1127                 }
1128             }
1129             break;
1130         case TYPE_ROM:
1131             task = iscsi_readcapacity10_sync(iscsilun->iscsi, iscsilun->lun, 0, 0);
1132             if (task != NULL && task->status == SCSI_STATUS_GOOD) {
1133                 rc10 = scsi_datain_unmarshall(task);
1134                 if (rc10 == NULL) {
1135                     error_setg(errp, "iSCSI: Failed to unmarshall readcapacity10 data.");
1136                 } else {
1137                     iscsilun->block_size = rc10->block_size;
1138                     if (rc10->lba == 0) {
1139                         /* blank disk loaded */
1140                         iscsilun->num_blocks = 0;
1141                     } else {
1142                         iscsilun->num_blocks = rc10->lba + 1;
1143                     }
1144                 }
1145             }
1146             break;
1147         default:
1148             return;
1149         }
1150     } while (task != NULL && task->status == SCSI_STATUS_CHECK_CONDITION
1151              && task->sense.key == SCSI_SENSE_UNIT_ATTENTION
1152              && retries-- > 0);
1153
1154     if (task == NULL || task->status != SCSI_STATUS_GOOD) {
1155         error_setg(errp, "iSCSI: failed to send readcapacity10 command.");
1156     }
1157     if (task) {
1158         scsi_free_scsi_task(task);
1159     }
1160 }
1161
1162 /* TODO Convert to fine grained options */
1163 static QemuOptsList runtime_opts = {
1164     .name = "iscsi",
1165     .head = QTAILQ_HEAD_INITIALIZER(runtime_opts.head),
1166     .desc = {
1167         {
1168             .name = "filename",
1169             .type = QEMU_OPT_STRING,
1170             .help = "URL to the iscsi image",
1171         },
1172         { /* end of list */ }
1173     },
1174 };
1175
1176 static struct scsi_task *iscsi_do_inquiry(struct iscsi_context *iscsi, int lun,
1177                                           int evpd, int pc, void **inq, Error **errp)
1178 {
1179     int full_size;
1180     struct scsi_task *task = NULL;
1181     task = iscsi_inquiry_sync(iscsi, lun, evpd, pc, 64);
1182     if (task == NULL || task->status != SCSI_STATUS_GOOD) {
1183         goto fail;
1184     }
1185     full_size = scsi_datain_getfullsize(task);
1186     if (full_size > task->datain.size) {
1187         scsi_free_scsi_task(task);
1188
1189         /* we need more data for the full list */
1190         task = iscsi_inquiry_sync(iscsi, lun, evpd, pc, full_size);
1191         if (task == NULL || task->status != SCSI_STATUS_GOOD) {
1192             goto fail;
1193         }
1194     }
1195
1196     *inq = scsi_datain_unmarshall(task);
1197     if (*inq == NULL) {
1198         error_setg(errp, "iSCSI: failed to unmarshall inquiry datain blob");
1199         goto fail_with_err;
1200     }
1201
1202     return task;
1203
1204 fail:
1205     error_setg(errp, "iSCSI: Inquiry command failed : %s",
1206                iscsi_get_error(iscsi));
1207 fail_with_err:
1208     if (task != NULL) {
1209         scsi_free_scsi_task(task);
1210     }
1211     return NULL;
1212 }
1213
1214 static void iscsi_detach_aio_context(BlockDriverState *bs)
1215 {
1216     IscsiLun *iscsilun = bs->opaque;
1217
1218     aio_set_fd_handler(iscsilun->aio_context,
1219                        iscsi_get_fd(iscsilun->iscsi),
1220                        NULL, NULL, NULL);
1221     iscsilun->events = 0;
1222
1223     if (iscsilun->nop_timer) {
1224         timer_del(iscsilun->nop_timer);
1225         timer_free(iscsilun->nop_timer);
1226         iscsilun->nop_timer = NULL;
1227     }
1228     if (iscsilun->event_timer) {
1229         timer_del(iscsilun->event_timer);
1230         timer_free(iscsilun->event_timer);
1231         iscsilun->event_timer = NULL;
1232     }
1233 }
1234
1235 static void iscsi_attach_aio_context(BlockDriverState *bs,
1236                                      AioContext *new_context)
1237 {
1238     IscsiLun *iscsilun = bs->opaque;
1239
1240     iscsilun->aio_context = new_context;
1241     iscsi_set_events(iscsilun);
1242
1243     /* Set up a timer for sending out iSCSI NOPs */
1244     iscsilun->nop_timer = aio_timer_new(iscsilun->aio_context,
1245                                         QEMU_CLOCK_REALTIME, SCALE_MS,
1246                                         iscsi_nop_timed_event, iscsilun);
1247     timer_mod(iscsilun->nop_timer,
1248               qemu_clock_get_ms(QEMU_CLOCK_REALTIME) + NOP_INTERVAL);
1249
1250     /* Prepare a timer for a delayed call to iscsi_set_events */
1251     iscsilun->event_timer = aio_timer_new(iscsilun->aio_context,
1252                                           QEMU_CLOCK_REALTIME, SCALE_MS,
1253                                           iscsi_timed_set_events, iscsilun);
1254 }
1255
1256 static bool iscsi_is_write_protected(IscsiLun *iscsilun)
1257 {
1258     struct scsi_task *task;
1259     struct scsi_mode_sense *ms = NULL;
1260     bool wrprotected = false;
1261
1262     task = iscsi_modesense6_sync(iscsilun->iscsi, iscsilun->lun,
1263                                  1, SCSI_MODESENSE_PC_CURRENT,
1264                                  0x3F, 0, 255);
1265     if (task == NULL) {
1266         error_report("iSCSI: Failed to send MODE_SENSE(6) command: %s",
1267                      iscsi_get_error(iscsilun->iscsi));
1268         goto out;
1269     }
1270
1271     if (task->status != SCSI_STATUS_GOOD) {
1272         error_report("iSCSI: Failed MODE_SENSE(6), LUN assumed writable");
1273         goto out;
1274     }
1275     ms = scsi_datain_unmarshall(task);
1276     if (!ms) {
1277         error_report("iSCSI: Failed to unmarshall MODE_SENSE(6) data: %s",
1278                      iscsi_get_error(iscsilun->iscsi));
1279         goto out;
1280     }
1281     wrprotected = ms->device_specific_parameter & 0x80;
1282
1283 out:
1284     if (task) {
1285         scsi_free_scsi_task(task);
1286     }
1287     return wrprotected;
1288 }
1289
1290 /*
1291  * We support iscsi url's on the form
1292  * iscsi://[<username>%<password>@]<host>[:<port>]/<targetname>/<lun>
1293  */
1294 static int iscsi_open(BlockDriverState *bs, QDict *options, int flags,
1295                       Error **errp)
1296 {
1297     IscsiLun *iscsilun = bs->opaque;
1298     struct iscsi_context *iscsi = NULL;
1299     struct iscsi_url *iscsi_url = NULL;
1300     struct scsi_task *task = NULL;
1301     struct scsi_inquiry_standard *inq = NULL;
1302     struct scsi_inquiry_supported_pages *inq_vpd;
1303     char *initiator_name = NULL;
1304     QemuOpts *opts;
1305     Error *local_err = NULL;
1306     const char *filename;
1307     int i, ret = 0;
1308
1309     if ((BDRV_SECTOR_SIZE % 512) != 0) {
1310         error_setg(errp, "iSCSI: Invalid BDRV_SECTOR_SIZE. "
1311                    "BDRV_SECTOR_SIZE(%lld) is not a multiple "
1312                    "of 512", BDRV_SECTOR_SIZE);
1313         return -EINVAL;
1314     }
1315
1316     opts = qemu_opts_create(&runtime_opts, NULL, 0, &error_abort);
1317     qemu_opts_absorb_qdict(opts, options, &local_err);
1318     if (local_err) {
1319         error_propagate(errp, local_err);
1320         ret = -EINVAL;
1321         goto out;
1322     }
1323
1324     filename = qemu_opt_get(opts, "filename");
1325
1326     iscsi_url = iscsi_parse_full_url(iscsi, filename);
1327     if (iscsi_url == NULL) {
1328         error_setg(errp, "Failed to parse URL : %s", filename);
1329         ret = -EINVAL;
1330         goto out;
1331     }
1332
1333     memset(iscsilun, 0, sizeof(IscsiLun));
1334
1335     initiator_name = parse_initiator_name(iscsi_url->target);
1336
1337     iscsi = iscsi_create_context(initiator_name);
1338     if (iscsi == NULL) {
1339         error_setg(errp, "iSCSI: Failed to create iSCSI context.");
1340         ret = -ENOMEM;
1341         goto out;
1342     }
1343
1344     if (iscsi_set_targetname(iscsi, iscsi_url->target)) {
1345         error_setg(errp, "iSCSI: Failed to set target name.");
1346         ret = -EINVAL;
1347         goto out;
1348     }
1349
1350     if (iscsi_url->user[0] != '\0') {
1351         ret = iscsi_set_initiator_username_pwd(iscsi, iscsi_url->user,
1352                                               iscsi_url->passwd);
1353         if (ret != 0) {
1354             error_setg(errp, "Failed to set initiator username and password");
1355             ret = -EINVAL;
1356             goto out;
1357         }
1358     }
1359
1360     /* check if we got CHAP username/password via the options */
1361     parse_chap(iscsi, iscsi_url->target, &local_err);
1362     if (local_err != NULL) {
1363         error_propagate(errp, local_err);
1364         ret = -EINVAL;
1365         goto out;
1366     }
1367
1368     if (iscsi_set_session_type(iscsi, ISCSI_SESSION_NORMAL) != 0) {
1369         error_setg(errp, "iSCSI: Failed to set session type to normal.");
1370         ret = -EINVAL;
1371         goto out;
1372     }
1373
1374     iscsi_set_header_digest(iscsi, ISCSI_HEADER_DIGEST_NONE_CRC32C);
1375
1376     /* check if we got HEADER_DIGEST via the options */
1377     parse_header_digest(iscsi, iscsi_url->target, &local_err);
1378     if (local_err != NULL) {
1379         error_propagate(errp, local_err);
1380         ret = -EINVAL;
1381         goto out;
1382     }
1383
1384     if (iscsi_full_connect_sync(iscsi, iscsi_url->portal, iscsi_url->lun) != 0) {
1385         error_setg(errp, "iSCSI: Failed to connect to LUN : %s",
1386             iscsi_get_error(iscsi));
1387         ret = -EINVAL;
1388         goto out;
1389     }
1390
1391     iscsilun->iscsi = iscsi;
1392     iscsilun->aio_context = bdrv_get_aio_context(bs);
1393     iscsilun->lun   = iscsi_url->lun;
1394     iscsilun->has_write_same = true;
1395
1396     task = iscsi_do_inquiry(iscsilun->iscsi, iscsilun->lun, 0, 0,
1397                             (void **) &inq, errp);
1398     if (task == NULL) {
1399         ret = -EINVAL;
1400         goto out;
1401     }
1402     iscsilun->type = inq->periperal_device_type;
1403     scsi_free_scsi_task(task);
1404     task = NULL;
1405
1406     iscsilun->write_protected = iscsi_is_write_protected(iscsilun);
1407     /* Check the write protect flag of the LUN if we want to write */
1408     if (iscsilun->type == TYPE_DISK && (flags & BDRV_O_RDWR) &&
1409         iscsilun->write_protected) {
1410         error_setg(errp, "Cannot open a write protected LUN as read-write");
1411         ret = -EACCES;
1412         goto out;
1413     }
1414
1415     iscsi_readcapacity_sync(iscsilun, &local_err);
1416     if (local_err != NULL) {
1417         error_propagate(errp, local_err);
1418         ret = -EINVAL;
1419         goto out;
1420     }
1421     bs->total_sectors = sector_lun2qemu(iscsilun->num_blocks, iscsilun);
1422     bs->request_alignment = iscsilun->block_size;
1423
1424     /* We don't have any emulation for devices other than disks and CD-ROMs, so
1425      * this must be sg ioctl compatible. We force it to be sg, otherwise qemu
1426      * will try to read from the device to guess the image format.
1427      */
1428     if (iscsilun->type != TYPE_DISK && iscsilun->type != TYPE_ROM) {
1429         bs->sg = 1;
1430     }
1431
1432     task = iscsi_do_inquiry(iscsilun->iscsi, iscsilun->lun, 1,
1433                             SCSI_INQUIRY_PAGECODE_SUPPORTED_VPD_PAGES,
1434                             (void **) &inq_vpd, errp);
1435     if (task == NULL) {
1436         ret = -EINVAL;
1437         goto out;
1438     }
1439     for (i = 0; i < inq_vpd->num_pages; i++) {
1440         struct scsi_task *inq_task;
1441         struct scsi_inquiry_logical_block_provisioning *inq_lbp;
1442         struct scsi_inquiry_block_limits *inq_bl;
1443         switch (inq_vpd->pages[i]) {
1444         case SCSI_INQUIRY_PAGECODE_LOGICAL_BLOCK_PROVISIONING:
1445             inq_task = iscsi_do_inquiry(iscsilun->iscsi, iscsilun->lun, 1,
1446                                         SCSI_INQUIRY_PAGECODE_LOGICAL_BLOCK_PROVISIONING,
1447                                         (void **) &inq_lbp, errp);
1448             if (inq_task == NULL) {
1449                 ret = -EINVAL;
1450                 goto out;
1451             }
1452             memcpy(&iscsilun->lbp, inq_lbp,
1453                    sizeof(struct scsi_inquiry_logical_block_provisioning));
1454             scsi_free_scsi_task(inq_task);
1455             break;
1456         case SCSI_INQUIRY_PAGECODE_BLOCK_LIMITS:
1457             inq_task = iscsi_do_inquiry(iscsilun->iscsi, iscsilun->lun, 1,
1458                                     SCSI_INQUIRY_PAGECODE_BLOCK_LIMITS,
1459                                     (void **) &inq_bl, errp);
1460             if (inq_task == NULL) {
1461                 ret = -EINVAL;
1462                 goto out;
1463             }
1464             memcpy(&iscsilun->bl, inq_bl,
1465                    sizeof(struct scsi_inquiry_block_limits));
1466             scsi_free_scsi_task(inq_task);
1467             break;
1468         default:
1469             break;
1470         }
1471     }
1472     scsi_free_scsi_task(task);
1473     task = NULL;
1474
1475     iscsi_attach_aio_context(bs, iscsilun->aio_context);
1476
1477     /* Guess the internal cluster (page) size of the iscsi target by the means
1478      * of opt_unmap_gran. Transfer the unmap granularity only if it has a
1479      * reasonable size */
1480     if (iscsilun->bl.opt_unmap_gran * iscsilun->block_size >= 4 * 1024 &&
1481         iscsilun->bl.opt_unmap_gran * iscsilun->block_size <= 16 * 1024 * 1024) {
1482         iscsilun->cluster_sectors = (iscsilun->bl.opt_unmap_gran *
1483                                      iscsilun->block_size) >> BDRV_SECTOR_BITS;
1484         if (iscsilun->lbprz && !(bs->open_flags & BDRV_O_NOCACHE)) {
1485             iscsilun->allocationmap = iscsi_allocationmap_init(iscsilun);
1486             if (iscsilun->allocationmap == NULL) {
1487                 ret = -ENOMEM;
1488             }
1489         }
1490     }
1491
1492 out:
1493     qemu_opts_del(opts);
1494     g_free(initiator_name);
1495     if (iscsi_url != NULL) {
1496         iscsi_destroy_url(iscsi_url);
1497     }
1498     if (task != NULL) {
1499         scsi_free_scsi_task(task);
1500     }
1501
1502     if (ret) {
1503         if (iscsi != NULL) {
1504             if (iscsi_is_logged_in(iscsi)) {
1505                 iscsi_logout_sync(iscsi);
1506             }
1507             iscsi_destroy_context(iscsi);
1508         }
1509         memset(iscsilun, 0, sizeof(IscsiLun));
1510     }
1511     return ret;
1512 }
1513
1514 static void iscsi_close(BlockDriverState *bs)
1515 {
1516     IscsiLun *iscsilun = bs->opaque;
1517     struct iscsi_context *iscsi = iscsilun->iscsi;
1518
1519     iscsi_detach_aio_context(bs);
1520     if (iscsi_is_logged_in(iscsi)) {
1521         iscsi_logout_sync(iscsi);
1522     }
1523     iscsi_destroy_context(iscsi);
1524     g_free(iscsilun->zeroblock);
1525     g_free(iscsilun->allocationmap);
1526     memset(iscsilun, 0, sizeof(IscsiLun));
1527 }
1528
1529 static int sector_limits_lun2qemu(int64_t sector, IscsiLun *iscsilun)
1530 {
1531     return MIN(sector_lun2qemu(sector, iscsilun), INT_MAX / 2 + 1);
1532 }
1533
1534 static void iscsi_refresh_limits(BlockDriverState *bs, Error **errp)
1535 {
1536     /* We don't actually refresh here, but just return data queried in
1537      * iscsi_open(): iscsi targets don't change their limits. */
1538
1539     IscsiLun *iscsilun = bs->opaque;
1540     uint32_t max_xfer_len = iscsilun->use_16_for_rw ? 0xffffffff : 0xffff;
1541
1542     if (iscsilun->bl.max_xfer_len) {
1543         max_xfer_len = MIN(max_xfer_len, iscsilun->bl.max_xfer_len);
1544     }
1545
1546     bs->bl.max_transfer_length = sector_limits_lun2qemu(max_xfer_len, iscsilun);
1547
1548     if (iscsilun->lbp.lbpu) {
1549         if (iscsilun->bl.max_unmap < 0xffffffff) {
1550             bs->bl.max_discard =
1551                 sector_limits_lun2qemu(iscsilun->bl.max_unmap, iscsilun);
1552         }
1553         bs->bl.discard_alignment =
1554             sector_limits_lun2qemu(iscsilun->bl.opt_unmap_gran, iscsilun);
1555     }
1556
1557     if (iscsilun->bl.max_ws_len < 0xffffffff) {
1558         bs->bl.max_write_zeroes =
1559             sector_limits_lun2qemu(iscsilun->bl.max_ws_len, iscsilun);
1560     }
1561     if (iscsilun->lbp.lbpws) {
1562         bs->bl.write_zeroes_alignment =
1563             sector_limits_lun2qemu(iscsilun->bl.opt_unmap_gran, iscsilun);
1564     }
1565     bs->bl.opt_transfer_length =
1566         sector_limits_lun2qemu(iscsilun->bl.opt_xfer_len, iscsilun);
1567 }
1568
1569 /* Note that this will not re-establish a connection with an iSCSI target - it
1570  * is effectively a NOP.  */
1571 static int iscsi_reopen_prepare(BDRVReopenState *state,
1572                                 BlockReopenQueue *queue, Error **errp)
1573 {
1574     IscsiLun *iscsilun = state->bs->opaque;
1575
1576     if (state->flags & BDRV_O_RDWR && iscsilun->write_protected) {
1577         error_setg(errp, "Cannot open a write protected LUN as read-write");
1578         return -EACCES;
1579     }
1580     return 0;
1581 }
1582
1583 static int iscsi_truncate(BlockDriverState *bs, int64_t offset)
1584 {
1585     IscsiLun *iscsilun = bs->opaque;
1586     Error *local_err = NULL;
1587
1588     if (iscsilun->type != TYPE_DISK) {
1589         return -ENOTSUP;
1590     }
1591
1592     iscsi_readcapacity_sync(iscsilun, &local_err);
1593     if (local_err != NULL) {
1594         error_free(local_err);
1595         return -EIO;
1596     }
1597
1598     if (offset > iscsi_getlength(bs)) {
1599         return -EINVAL;
1600     }
1601
1602     if (iscsilun->allocationmap != NULL) {
1603         g_free(iscsilun->allocationmap);
1604         iscsilun->allocationmap = iscsi_allocationmap_init(iscsilun);
1605     }
1606
1607     return 0;
1608 }
1609
1610 static int iscsi_create(const char *filename, QemuOpts *opts, Error **errp)
1611 {
1612     int ret = 0;
1613     int64_t total_size = 0;
1614     BlockDriverState *bs;
1615     IscsiLun *iscsilun = NULL;
1616     QDict *bs_options;
1617
1618     bs = bdrv_new();
1619
1620     /* Read out options */
1621     total_size = DIV_ROUND_UP(qemu_opt_get_size_del(opts, BLOCK_OPT_SIZE, 0),
1622                               BDRV_SECTOR_SIZE);
1623     bs->opaque = g_new0(struct IscsiLun, 1);
1624     iscsilun = bs->opaque;
1625
1626     bs_options = qdict_new();
1627     qdict_put(bs_options, "filename", qstring_from_str(filename));
1628     ret = iscsi_open(bs, bs_options, 0, NULL);
1629     QDECREF(bs_options);
1630
1631     if (ret != 0) {
1632         goto out;
1633     }
1634     iscsi_detach_aio_context(bs);
1635     if (iscsilun->type != TYPE_DISK) {
1636         ret = -ENODEV;
1637         goto out;
1638     }
1639     if (bs->total_sectors < total_size) {
1640         ret = -ENOSPC;
1641         goto out;
1642     }
1643
1644     ret = 0;
1645 out:
1646     if (iscsilun->iscsi != NULL) {
1647         iscsi_destroy_context(iscsilun->iscsi);
1648     }
1649     g_free(bs->opaque);
1650     bs->opaque = NULL;
1651     bdrv_unref(bs);
1652     return ret;
1653 }
1654
1655 static int iscsi_get_info(BlockDriverState *bs, BlockDriverInfo *bdi)
1656 {
1657     IscsiLun *iscsilun = bs->opaque;
1658     bdi->unallocated_blocks_are_zero = !!iscsilun->lbprz;
1659     bdi->can_write_zeroes_with_unmap = iscsilun->lbprz && iscsilun->lbp.lbpws;
1660     bdi->cluster_size = iscsilun->cluster_sectors * BDRV_SECTOR_SIZE;
1661     return 0;
1662 }
1663
1664 static QemuOptsList iscsi_create_opts = {
1665     .name = "iscsi-create-opts",
1666     .head = QTAILQ_HEAD_INITIALIZER(iscsi_create_opts.head),
1667     .desc = {
1668         {
1669             .name = BLOCK_OPT_SIZE,
1670             .type = QEMU_OPT_SIZE,
1671             .help = "Virtual disk size"
1672         },
1673         { /* end of list */ }
1674     }
1675 };
1676
1677 static BlockDriver bdrv_iscsi = {
1678     .format_name     = "iscsi",
1679     .protocol_name   = "iscsi",
1680
1681     .instance_size   = sizeof(IscsiLun),
1682     .bdrv_needs_filename = true,
1683     .bdrv_file_open  = iscsi_open,
1684     .bdrv_close      = iscsi_close,
1685     .bdrv_create     = iscsi_create,
1686     .create_opts     = &iscsi_create_opts,
1687     .bdrv_reopen_prepare  = iscsi_reopen_prepare,
1688
1689     .bdrv_getlength  = iscsi_getlength,
1690     .bdrv_get_info   = iscsi_get_info,
1691     .bdrv_truncate   = iscsi_truncate,
1692     .bdrv_refresh_limits = iscsi_refresh_limits,
1693
1694     .bdrv_co_get_block_status = iscsi_co_get_block_status,
1695     .bdrv_co_discard      = iscsi_co_discard,
1696     .bdrv_co_write_zeroes = iscsi_co_write_zeroes,
1697     .bdrv_co_readv         = iscsi_co_readv,
1698     .bdrv_co_writev        = iscsi_co_writev,
1699     .bdrv_co_flush_to_disk = iscsi_co_flush,
1700
1701 #ifdef __linux__
1702     .bdrv_ioctl       = iscsi_ioctl,
1703     .bdrv_aio_ioctl   = iscsi_aio_ioctl,
1704 #endif
1705
1706     .bdrv_detach_aio_context = iscsi_detach_aio_context,
1707     .bdrv_attach_aio_context = iscsi_attach_aio_context,
1708 };
1709
1710 static QemuOptsList qemu_iscsi_opts = {
1711     .name = "iscsi",
1712     .head = QTAILQ_HEAD_INITIALIZER(qemu_iscsi_opts.head),
1713     .desc = {
1714         {
1715             .name = "user",
1716             .type = QEMU_OPT_STRING,
1717             .help = "username for CHAP authentication to target",
1718         },{
1719             .name = "password",
1720             .type = QEMU_OPT_STRING,
1721             .help = "password for CHAP authentication to target",
1722         },{
1723             .name = "header-digest",
1724             .type = QEMU_OPT_STRING,
1725             .help = "HeaderDigest setting. "
1726                     "{CRC32C|CRC32C-NONE|NONE-CRC32C|NONE}",
1727         },{
1728             .name = "initiator-name",
1729             .type = QEMU_OPT_STRING,
1730             .help = "Initiator iqn name to use when connecting",
1731         },
1732         { /* end of list */ }
1733     },
1734 };
1735
1736 static void iscsi_block_init(void)
1737 {
1738     bdrv_register(&bdrv_iscsi);
1739     qemu_add_opts(&qemu_iscsi_opts);
1740 }
1741
1742 block_init(iscsi_block_init);