drivers/scsi/sd.c

   1 /*
   2  *      sd.c Copyright (C) 1992 Drew Eckhardt
   3  *           Copyright (C) 1993, 1994, 1995, 1999 Eric Youngdale
   4  *
   5  *      Linux scsi disk driver
   6  *              Initial versions: Drew Eckhardt
   7  *              Subsequent revisions: Eric Youngdale
   8  *      Modification history:
   9  *       - Drew Eckhardt <drew@colorado.edu> original
  10  *       - Eric Youngdale <eric@andante.org> add scatter-gather, multiple
  11  *         outstanding request, and other enhancements.
  12  *         Support loadable low-level scsi drivers.
  13  *       - Jirka Hanika <geo@ff.cuni.cz> support more scsi disks using
  14  *         eight major numbers.
  15  *       - Richard Gooch <rgooch@atnf.csiro.au> support devfs.
  16  *       - Torben Mathiasen <tmm@image.dk> Resource allocation fixes in
  17  *         sd_init and cleanups.
  18  *       - Alex Davis <letmein@erols.com> Fix problem where partition info
  19  *         not being read in sd_open. Fix problem where removable media
  20  *         could be ejected after sd_open.
  21  *       - Douglas Gilbert <dgilbert@interlog.com> cleanup for lk 2.5.x
  22  *       - Badari Pulavarty <pbadari@us.ibm.com>, Matthew Wilcox
  23  *         <willy@debian.org>, Kurt Garloff <garloff@suse.de>:
  24  *         Support 32k/1M disks.
  25  *
  26  *      Logging policy (needs CONFIG_SCSI_LOGGING defined):
  27  *       - setting up transfer: SCSI_LOG_HLQUEUE levels 1 and 2
  28  *       - end of transfer (bh + scsi_lib): SCSI_LOG_HLCOMPLETE level 1
  29  *       - entering sd_ioctl: SCSI_LOG_IOCTL level 1
  30  *       - entering other commands: SCSI_LOG_HLQUEUE level 3
  31  *      Note: when the logging level is set by the user, it must be greater
  32  *      than the level indicated above to trigger output.
  33  */
  34
  35 #include <linux/module.h>
  36 #include <linux/fs.h>
  37 #include <linux/kernel.h>
  38 #include <linux/mm.h>
  39 #include <linux/bio.h>
  40 #include <linux/genhd.h>
  41 #include <linux/hdreg.h>
  42 #include <linux/errno.h>
  43 #include <linux/idr.h>
  44 #include <linux/interrupt.h>
  45 #include <linux/init.h>
  46 #include <linux/blkdev.h>
  47 #include <linux/blkpg.h>
  48 #include <linux/delay.h>
  49 #include <linux/mutex.h>
  50 #include <linux/string_helpers.h>
  51 #include <linux/async.h>
  52 #include <linux/slab.h>
  53 #include <linux/pm_runtime.h>
  54 #include <linux/pr.h>
  55 #include <linux/t10-pi.h>
  56 #include <linux/uaccess.h>
  57 #include <asm/unaligned.h>
  58
  59 #include <scsi/scsi.h>
  60 #include <scsi/scsi_cmnd.h>
  61 #include <scsi/scsi_dbg.h>
  62 #include <scsi/scsi_device.h>
  63 #include <scsi/scsi_driver.h>
  64 #include <scsi/scsi_eh.h>
  65 #include <scsi/scsi_host.h>
  66 #include <scsi/scsi_ioctl.h>
  67 #include <scsi/scsicam.h>
  68
  69 #include "sd.h"
  70 #include "scsi_priv.h"
  71 #include "scsi_logging.h"
  72
  73 MODULE_AUTHOR("Eric Youngdale");
  74 MODULE_DESCRIPTION("SCSI disk (sd) driver");
  75 MODULE_LICENSE("GPL");
  76
  77 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK0_MAJOR);
  78 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK1_MAJOR);
  79 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK2_MAJOR);
  80 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK3_MAJOR);
  81 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK4_MAJOR);
  82 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK5_MAJOR);
  83 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK6_MAJOR);
  84 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK7_MAJOR);
  85 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK8_MAJOR);
  86 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK9_MAJOR);
  87 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK10_MAJOR);
  88 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK11_MAJOR);
  89 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK12_MAJOR);
  90 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK13_MAJOR);
  91 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK14_MAJOR);
  92 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK15_MAJOR);
  93 MODULE_ALIAS_SCSI_DEVICE(TYPE_DISK);
  94 MODULE_ALIAS_SCSI_DEVICE(TYPE_MOD);
  95 MODULE_ALIAS_SCSI_DEVICE(TYPE_RBC);
  96 MODULE_ALIAS_SCSI_DEVICE(TYPE_ZBC);
  97
  98 #if !defined(CONFIG_DEBUG_BLOCK_EXT_DEVT)
  99 #define SD_MINORS       16
 100 #else
 101 #define SD_MINORS       0
 102 #endif
 103
 104 static void sd_config_discard(struct scsi_disk *, unsigned int);
 105 static void sd_config_write_same(struct scsi_disk *);
 106 static int  sd_revalidate_disk(struct gendisk *);
 107 static void sd_unlock_native_capacity(struct gendisk *disk);
 108 static int  sd_probe(struct device *);
 109 static int  sd_remove(struct device *);
 110 static void sd_shutdown(struct device *);
 111 static int sd_suspend_system(struct device *);
 112 static int sd_suspend_runtime(struct device *);
 113 static int sd_resume(struct device *);
 114 static void sd_rescan(struct device *);
 115 static int sd_init_command(struct scsi_cmnd *SCpnt);
 116 static void sd_uninit_command(struct scsi_cmnd *SCpnt);
 117 static int sd_done(struct scsi_cmnd *);
 118 static int sd_eh_action(struct scsi_cmnd *, int);
 119 static void sd_read_capacity(struct scsi_disk *sdkp, unsigned char *buffer);
 120 static void scsi_disk_release(struct device *cdev);
 121 static void sd_print_sense_hdr(struct scsi_disk *, struct scsi_sense_hdr *);
 122 static void sd_print_result(const struct scsi_disk *, const char *, int);
 123
 124 static DEFINE_SPINLOCK(sd_index_lock);
 125 static DEFINE_IDA(sd_index_ida);
 126
 127 /* This semaphore is used to mediate the 0->1 reference get in the
 128  * face of object destruction (i.e. we can't allow a get on an
 129  * object after last put) */
 130 static DEFINE_MUTEX(sd_ref_mutex);
 131
 132 static struct kmem_cache *sd_cdb_cache;
 133 static mempool_t *sd_cdb_pool;
 134
 135 static const char *sd_cache_types[] = {
 136         "write through", "none", "write back",
 137         "write back, no read (daft)"
 138 };
 139
 140 static void sd_set_flush_flag(struct scsi_disk *sdkp)
 141 {
 142         bool wc = false, fua = false;
 143
 144         if (sdkp->WCE) {
 145                 wc = true;
 146                 if (sdkp->DPOFUA)
 147                         fua = true;
 148         }
 149
 150         blk_queue_write_cache(sdkp->disk->queue, wc, fua);
 151 }
 152
 153 static ssize_t
 154 cache_type_store(struct device *dev, struct device_attribute *attr,
 155                  const char *buf, size_t count)
 156 {
 157         int i, ct = -1, rcd, wce, sp;
 158         struct scsi_disk *sdkp = to_scsi_disk(dev);
 159         struct scsi_device *sdp = sdkp->device;
 160         char buffer[64];
 161         char *buffer_data;
 162         struct scsi_mode_data data;
 163         struct scsi_sense_hdr sshdr;
 164         static const char temp[] = "temporary ";
 165         int len;
 166
 167         if (sdp->type != TYPE_DISK && sdp->type != TYPE_ZBC)
 168                 /* no cache control on RBC devices; theoretically they
 169                  * can do it, but there's probably so many exceptions
 170                  * it's not worth the risk */
 171                 return -EINVAL;
 172
 173         if (strncmp(buf, temp, sizeof(temp) - 1) == 0) {
 174                 buf += sizeof(temp) - 1;
 175                 sdkp->cache_override = 1;
 176         } else {
 177                 sdkp->cache_override = 0;
 178         }
 179
 180         for (i = 0; i < ARRAY_SIZE(sd_cache_types); i++) {
 181                 len = strlen(sd_cache_types[i]);
 182                 if (strncmp(sd_cache_types[i], buf, len) == 0 &&
 183                     buf[len] == '\n') {
 184                         ct = i;
 185                         break;
 186                 }
 187         }
 188         if (ct < 0)
 189                 return -EINVAL;
 190         rcd = ct & 0x01 ? 1 : 0;
 191         wce = (ct & 0x02) && !sdkp->write_prot ? 1 : 0;
 192
 193         if (sdkp->cache_override) {
 194                 sdkp->WCE = wce;
 195                 sdkp->RCD = rcd;
 196                 sd_set_flush_flag(sdkp);
 197                 return count;
 198         }
 199
 200         if (scsi_mode_sense(sdp, 0x08, 8, buffer, sizeof(buffer), SD_TIMEOUT,
 201                             SD_MAX_RETRIES, &data, NULL))
 202                 return -EINVAL;
 203         len = min_t(size_t, sizeof(buffer), data.length - data.header_length -
 204                   data.block_descriptor_length);
 205         buffer_data = buffer + data.header_length +
 206                 data.block_descriptor_length;
 207         buffer_data[2] &= ~0x05;
 208         buffer_data[2] |= wce << 2 | rcd;
 209         sp = buffer_data[0] & 0x80 ? 1 : 0;
 210         buffer_data[0] &= ~0x80;
 211
 212         if (scsi_mode_select(sdp, 1, sp, 8, buffer_data, len, SD_TIMEOUT,
 213                              SD_MAX_RETRIES, &data, &sshdr)) {
 214                 if (scsi_sense_valid(&sshdr))
 215                         sd_print_sense_hdr(sdkp, &sshdr);
 216                 return -EINVAL;
 217         }
 218         revalidate_disk(sdkp->disk);
 219         return count;
 220 }
 221
 222 static ssize_t
 223 manage_start_stop_show(struct device *dev, struct device_attribute *attr,
 224                        char *buf)
 225 {
 226         struct scsi_disk *sdkp = to_scsi_disk(dev);
 227         struct scsi_device *sdp = sdkp->device;
 228
 229         return snprintf(buf, 20, "%u\n", sdp->manage_start_stop);
 230 }
 231
 232 static ssize_t
 233 manage_start_stop_store(struct device *dev, struct device_attribute *attr,
 234                         const char *buf, size_t count)
 235 {
 236         struct scsi_disk *sdkp = to_scsi_disk(dev);
 237         struct scsi_device *sdp = sdkp->device;
 238
 239         if (!capable(CAP_SYS_ADMIN))
 240                 return -EACCES;
 241
 242         sdp->manage_start_stop = simple_strtoul(buf, NULL, 10);
 243
 244         return count;
 245 }
 246 static DEVICE_ATTR_RW(manage_start_stop);
 247
 248 static ssize_t
 249 allow_restart_show(struct device *dev, struct device_attribute *attr, char *buf)
 250 {
 251         struct scsi_disk *sdkp = to_scsi_disk(dev);
 252
 253         return snprintf(buf, 40, "%d\n", sdkp->device->allow_restart);
 254 }
 255
 256 static ssize_t
 257 allow_restart_store(struct device *dev, struct device_attribute *attr,
 258                     const char *buf, size_t count)
 259 {
 260         struct scsi_disk *sdkp = to_scsi_disk(dev);
 261         struct scsi_device *sdp = sdkp->device;
 262
 263         if (!capable(CAP_SYS_ADMIN))
 264                 return -EACCES;
 265
 266         if (sdp->type != TYPE_DISK && sdp->type != TYPE_ZBC)
 267                 return -EINVAL;
 268
 269         sdp->allow_restart = simple_strtoul(buf, NULL, 10);
 270
 271         return count;
 272 }
 273 static DEVICE_ATTR_RW(allow_restart);
 274
 275 static ssize_t
 276 cache_type_show(struct device *dev, struct device_attribute *attr, char *buf)
 277 {
 278         struct scsi_disk *sdkp = to_scsi_disk(dev);
 279         int ct = sdkp->RCD + 2*sdkp->WCE;
 280
 281         return snprintf(buf, 40, "%s\n", sd_cache_types[ct]);
 282 }
 283 static DEVICE_ATTR_RW(cache_type);
 284
 285 static ssize_t
 286 FUA_show(struct device *dev, struct device_attribute *attr, char *buf)
 287 {
 288         struct scsi_disk *sdkp = to_scsi_disk(dev);
 289
 290         return snprintf(buf, 20, "%u\n", sdkp->DPOFUA);
 291 }
 292 static DEVICE_ATTR_RO(FUA);
 293
 294 static ssize_t
 295 protection_type_show(struct device *dev, struct device_attribute *attr,
 296                      char *buf)
 297 {
 298         struct scsi_disk *sdkp = to_scsi_disk(dev);
 299
 300         return snprintf(buf, 20, "%u\n", sdkp->protection_type);
 301 }
 302
 303 static ssize_t
 304 protection_type_store(struct device *dev, struct device_attribute *attr,
 305                       const char *buf, size_t count)
 306 {
 307         struct scsi_disk *sdkp = to_scsi_disk(dev);
 308         unsigned int val;
 309         int err;
 310
 311         if (!capable(CAP_SYS_ADMIN))
 312                 return -EACCES;
 313
 314         err = kstrtouint(buf, 10, &val);
 315
 316         if (err)
 317                 return err;
 318
 319         if (val >= 0 && val <= T10_PI_TYPE3_PROTECTION)
 320                 sdkp->protection_type = val;
 321
 322         return count;
 323 }
 324 static DEVICE_ATTR_RW(protection_type);
 325
 326 static ssize_t
 327 protection_mode_show(struct device *dev, struct device_attribute *attr,
 328                      char *buf)
 329 {
 330         struct scsi_disk *sdkp = to_scsi_disk(dev);
 331         struct scsi_device *sdp = sdkp->device;
 332         unsigned int dif, dix;
 333
 334         dif = scsi_host_dif_capable(sdp->host, sdkp->protection_type);
 335         dix = scsi_host_dix_capable(sdp->host, sdkp->protection_type);
 336
 337         if (!dix && scsi_host_dix_capable(sdp->host, T10_PI_TYPE0_PROTECTION)) {
 338                 dif = 0;
 339                 dix = 1;
 340         }
 341
 342         if (!dif && !dix)
 343                 return snprintf(buf, 20, "none\n");
 344
 345         return snprintf(buf, 20, "%s%u\n", dix ? "dix" : "dif", dif);
 346 }
 347 static DEVICE_ATTR_RO(protection_mode);
 348
 349 static ssize_t
 350 app_tag_own_show(struct device *dev, struct device_attribute *attr, char *buf)
 351 {
 352         struct scsi_disk *sdkp = to_scsi_disk(dev);
 353
 354         return snprintf(buf, 20, "%u\n", sdkp->ATO);
 355 }
 356 static DEVICE_ATTR_RO(app_tag_own);
 357
 358 static ssize_t
 359 thin_provisioning_show(struct device *dev, struct device_attribute *attr,
 360                        char *buf)
 361 {
 362         struct scsi_disk *sdkp = to_scsi_disk(dev);
 363
 364         return snprintf(buf, 20, "%u\n", sdkp->lbpme);
 365 }
 366 static DEVICE_ATTR_RO(thin_provisioning);
 367
 368 static const char *lbp_mode[] = {
 369         [SD_LBP_FULL]           = "full",
 370         [SD_LBP_UNMAP]          = "unmap",
 371         [SD_LBP_WS16]           = "writesame_16",
 372         [SD_LBP_WS10]           = "writesame_10",
 373         [SD_LBP_ZERO]           = "writesame_zero",
 374         [SD_LBP_DISABLE]        = "disabled",
 375 };
 376
 377 static ssize_t
 378 provisioning_mode_show(struct device *dev, struct device_attribute *attr,
 379                        char *buf)
 380 {
 381         struct scsi_disk *sdkp = to_scsi_disk(dev);
 382
 383         return snprintf(buf, 20, "%s\n", lbp_mode[sdkp->provisioning_mode]);
 384 }
 385
 386 static ssize_t
 387 provisioning_mode_store(struct device *dev, struct device_attribute *attr,
 388                         const char *buf, size_t count)
 389 {
 390         struct scsi_disk *sdkp = to_scsi_disk(dev);
 391         struct scsi_device *sdp = sdkp->device;
 392
 393         if (!capable(CAP_SYS_ADMIN))
 394                 return -EACCES;
 395
 396         if (sd_is_zoned(sdkp)) {
 397                 sd_config_discard(sdkp, SD_LBP_DISABLE);
 398                 return count;
 399         }
 400
 401         if (sdp->type != TYPE_DISK)
 402                 return -EINVAL;
 403
 404         if (!strncmp(buf, lbp_mode[SD_LBP_UNMAP], 20))
 405                 sd_config_discard(sdkp, SD_LBP_UNMAP);
 406         else if (!strncmp(buf, lbp_mode[SD_LBP_WS16], 20))
 407                 sd_config_discard(sdkp, SD_LBP_WS16);
 408         else if (!strncmp(buf, lbp_mode[SD_LBP_WS10], 20))
 409                 sd_config_discard(sdkp, SD_LBP_WS10);
 410         else if (!strncmp(buf, lbp_mode[SD_LBP_ZERO], 20))
 411                 sd_config_discard(sdkp, SD_LBP_ZERO);
 412         else if (!strncmp(buf, lbp_mode[SD_LBP_DISABLE], 20))
 413                 sd_config_discard(sdkp, SD_LBP_DISABLE);
 414         else
 415                 return -EINVAL;
 416
 417         return count;
 418 }
 419 static DEVICE_ATTR_RW(provisioning_mode);
 420
 421 static const char *zeroing_mode[] = {
 422         [SD_ZERO_WRITE]         = "write",
 423         [SD_ZERO_WS]            = "writesame",
 424         [SD_ZERO_WS16_UNMAP]    = "writesame_16_unmap",
 425         [SD_ZERO_WS10_UNMAP]    = "writesame_10_unmap",
 426 };
 427
 428 static ssize_t
 429 zeroing_mode_show(struct device *dev, struct device_attribute *attr,
 430                   char *buf)
 431 {
 432         struct scsi_disk *sdkp = to_scsi_disk(dev);
 433
 434         return snprintf(buf, 20, "%s\n", zeroing_mode[sdkp->zeroing_mode]);
 435 }
 436
 437 static ssize_t
 438 zeroing_mode_store(struct device *dev, struct device_attribute *attr,
 439                    const char *buf, size_t count)
 440 {
 441         struct scsi_disk *sdkp = to_scsi_disk(dev);
 442
 443         if (!capable(CAP_SYS_ADMIN))
 444                 return -EACCES;
 445
 446         if (!strncmp(buf, zeroing_mode[SD_ZERO_WRITE], 20))
 447                 sdkp->zeroing_mode = SD_ZERO_WRITE;
 448         else if (!strncmp(buf, zeroing_mode[SD_ZERO_WS], 20))
 449                 sdkp->zeroing_mode = SD_ZERO_WS;
 450         else if (!strncmp(buf, zeroing_mode[SD_ZERO_WS16_UNMAP], 20))
 451                 sdkp->zeroing_mode = SD_ZERO_WS16_UNMAP;
 452         else if (!strncmp(buf, zeroing_mode[SD_ZERO_WS10_UNMAP], 20))
 453                 sdkp->zeroing_mode = SD_ZERO_WS10_UNMAP;
 454         else
 455                 return -EINVAL;
 456
 457         return count;
 458 }
 459 static DEVICE_ATTR_RW(zeroing_mode);
 460
 461 static ssize_t
 462 max_medium_access_timeouts_show(struct device *dev,
 463                                 struct device_attribute *attr, char *buf)
 464 {
 465         struct scsi_disk *sdkp = to_scsi_disk(dev);
 466
 467         return snprintf(buf, 20, "%u\n", sdkp->max_medium_access_timeouts);
 468 }
 469
 470 static ssize_t
 471 max_medium_access_timeouts_store(struct device *dev,
 472                                  struct device_attribute *attr, const char *buf,
 473                                  size_t count)
 474 {
 475         struct scsi_disk *sdkp = to_scsi_disk(dev);
 476         int err;
 477
 478         if (!capable(CAP_SYS_ADMIN))
 479                 return -EACCES;
 480
 481         err = kstrtouint(buf, 10, &sdkp->max_medium_access_timeouts);
 482
 483         return err ? err : count;
 484 }
 485 static DEVICE_ATTR_RW(max_medium_access_timeouts);
 486
 487 static ssize_t
 488 max_write_same_blocks_show(struct device *dev, struct device_attribute *attr,
 489                            char *buf)
 490 {
 491         struct scsi_disk *sdkp = to_scsi_disk(dev);
 492
 493         return snprintf(buf, 20, "%u\n", sdkp->max_ws_blocks);
 494 }
 495
 496 static ssize_t
 497 max_write_same_blocks_store(struct device *dev, struct device_attribute *attr,
 498                             const char *buf, size_t count)
 499 {
 500         struct scsi_disk *sdkp = to_scsi_disk(dev);
 501         struct scsi_device *sdp = sdkp->device;
 502         unsigned long max;
 503         int err;
 504
 505         if (!capable(CAP_SYS_ADMIN))
 506                 return -EACCES;
 507
 508         if (sdp->type != TYPE_DISK && sdp->type != TYPE_ZBC)
 509                 return -EINVAL;
 510
 511         err = kstrtoul(buf, 10, &max);
 512
 513         if (err)
 514                 return err;
 515
 516         if (max == 0)
 517                 sdp->no_write_same = 1;
 518         else if (max <= SD_MAX_WS16_BLOCKS) {
 519                 sdp->no_write_same = 0;
 520                 sdkp->max_ws_blocks = max;
 521         }
 522
 523         sd_config_write_same(sdkp);
 524
 525         return count;
 526 }
 527 static DEVICE_ATTR_RW(max_write_same_blocks);
 528
 529 static struct attribute *sd_disk_attrs[] = {
 530         &dev_attr_cache_type.attr,
 531         &dev_attr_FUA.attr,
 532         &dev_attr_allow_restart.attr,
 533         &dev_attr_manage_start_stop.attr,
 534         &dev_attr_protection_type.attr,
 535         &dev_attr_protection_mode.attr,
 536         &dev_attr_app_tag_own.attr,
 537         &dev_attr_thin_provisioning.attr,
 538         &dev_attr_provisioning_mode.attr,
 539         &dev_attr_zeroing_mode.attr,
 540         &dev_attr_max_write_same_blocks.attr,
 541         &dev_attr_max_medium_access_timeouts.attr,
 542         NULL,
 543 };
 544 ATTRIBUTE_GROUPS(sd_disk);
 545
 546 static struct class sd_disk_class = {
 547         .name           = "scsi_disk",
 548         .owner          = THIS_MODULE,
 549         .dev_release    = scsi_disk_release,
 550         .dev_groups     = sd_disk_groups,
 551 };
 552
 553 static const struct dev_pm_ops sd_pm_ops = {
 554         .suspend                = sd_suspend_system,
 555         .resume                 = sd_resume,
 556         .poweroff               = sd_suspend_system,
 557         .restore                = sd_resume,
 558         .runtime_suspend        = sd_suspend_runtime,
 559         .runtime_resume         = sd_resume,
 560 };
 561
 562 static struct scsi_driver sd_template = {
 563         .gendrv = {
 564                 .name           = "sd",
 565                 .owner          = THIS_MODULE,
 566                 .probe          = sd_probe,
 567                 .remove         = sd_remove,
 568                 .shutdown       = sd_shutdown,
 569                 .pm             = &sd_pm_ops,
 570         },
 571         .rescan                 = sd_rescan,
 572         .init_command           = sd_init_command,
 573         .uninit_command         = sd_uninit_command,
 574         .done                   = sd_done,
 575         .eh_action              = sd_eh_action,
 576 };
 577
 578 /*
 579  * Dummy kobj_map->probe function.
 580  * The default ->probe function will call modprobe, which is
 581  * pointless as this module is already loaded.
 582  */
 583 static struct kobject *sd_default_probe(dev_t devt, int *partno, void *data)
 584 {
 585         return NULL;
 586 }
 587
 588 /*
 589  * Device no to disk mapping:
 590  *
 591  *       major         disc2     disc  p1
 592  *   |............|.............|....|....| <- dev_t
 593  *    31        20 19          8 7  4 3  0
 594  *
 595  * Inside a major, we have 16k disks, however mapped non-
 596  * contiguously. The first 16 disks are for major0, the next
 597  * ones with major1, ... Disk 256 is for major0 again, disk 272
 598  * for major1, ...
 599  * As we stay compatible with our numbering scheme, we can reuse
 600  * the well-know SCSI majors 8, 65--71, 136--143.
 601  */
 602 static int sd_major(int major_idx)
 603 {
 604         switch (major_idx) {
 605         case 0:
 606                 return SCSI_DISK0_MAJOR;
 607         case 1 ... 7:
 608                 return SCSI_DISK1_MAJOR + major_idx - 1;
 609         case 8 ... 15:
 610                 return SCSI_DISK8_MAJOR + major_idx - 8;
 611         default:
 612                 BUG();
 613                 return 0;       /* shut up gcc */
 614         }
 615 }
 616
 617 static struct scsi_disk *scsi_disk_get(struct gendisk *disk)
 618 {
 619         struct scsi_disk *sdkp = NULL;
 620
 621         mutex_lock(&sd_ref_mutex);
 622
 623         if (disk->private_data) {
 624                 sdkp = scsi_disk(disk);
 625                 if (scsi_device_get(sdkp->device) == 0)
 626                         get_device(&sdkp->dev);
 627                 else
 628                         sdkp = NULL;
 629         }
 630         mutex_unlock(&sd_ref_mutex);
 631         return sdkp;
 632 }
 633
 634 static void scsi_disk_put(struct scsi_disk *sdkp)
 635 {
 636         struct scsi_device *sdev = sdkp->device;
 637
 638         mutex_lock(&sd_ref_mutex);
 639         put_device(&sdkp->dev);
 640         scsi_device_put(sdev);
 641         mutex_unlock(&sd_ref_mutex);
 642 }
 643
 644 static unsigned char sd_setup_protect_cmnd(struct scsi_cmnd *scmd,
 645                                            unsigned int dix, unsigned int dif)
 646 {
 647         struct bio *bio = scmd->request->bio;
 648         unsigned int prot_op = sd_prot_op(rq_data_dir(scmd->request), dix, dif);
 649         unsigned int protect = 0;
 650
 651         if (dix) {                              /* DIX Type 0, 1, 2, 3 */
 652                 if (bio_integrity_flagged(bio, BIP_IP_CHECKSUM))
 653                         scmd->prot_flags |= SCSI_PROT_IP_CHECKSUM;
 654
 655                 if (bio_integrity_flagged(bio, BIP_CTRL_NOCHECK) == false)
 656                         scmd->prot_flags |= SCSI_PROT_GUARD_CHECK;
 657         }
 658
 659         if (dif != T10_PI_TYPE3_PROTECTION) {   /* DIX/DIF Type 0, 1, 2 */
 660                 scmd->prot_flags |= SCSI_PROT_REF_INCREMENT;
 661
 662                 if (bio_integrity_flagged(bio, BIP_CTRL_NOCHECK) == false)
 663                         scmd->prot_flags |= SCSI_PROT_REF_CHECK;
 664         }
 665
 666         if (dif) {                              /* DIX/DIF Type 1, 2, 3 */
 667                 scmd->prot_flags |= SCSI_PROT_TRANSFER_PI;
 668
 669                 if (bio_integrity_flagged(bio, BIP_DISK_NOCHECK))
 670                         protect = 3 << 5;       /* Disable target PI checking */
 671                 else
 672                         protect = 1 << 5;       /* Enable target PI checking */
 673         }
 674
 675         scsi_set_prot_op(scmd, prot_op);
 676         scsi_set_prot_type(scmd, dif);
 677         scmd->prot_flags &= sd_prot_flag_mask(prot_op);
 678
 679         return protect;
 680 }
 681
 682 static void sd_config_discard(struct scsi_disk *sdkp, unsigned int mode)
 683 {
 684         struct request_queue *q = sdkp->disk->queue;
 685         unsigned int logical_block_size = sdkp->device->sector_size;
 686         unsigned int max_blocks = 0;
 687
 688         q->limits.discard_alignment =
 689                 sdkp->unmap_alignment * logical_block_size;
 690         q->limits.discard_granularity =
 691                 max(sdkp->physical_block_size,
 692                     sdkp->unmap_granularity * logical_block_size);
 693         sdkp->provisioning_mode = mode;
 694
 695         switch (mode) {
 696
 697         case SD_LBP_DISABLE:
 698                 blk_queue_max_discard_sectors(q, 0);
 699                 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD, q);
 700                 return;
 701
 702         case SD_LBP_UNMAP:
 703                 max_blocks = min_not_zero(sdkp->max_unmap_blocks,
 704                                           (u32)SD_MAX_WS16_BLOCKS);
 705                 break;
 706
 707         case SD_LBP_WS16:
 708                 max_blocks = min_not_zero(sdkp->max_ws_blocks,
 709                                           (u32)SD_MAX_WS16_BLOCKS);
 710                 break;
 711
 712         case SD_LBP_WS10:
 713                 max_blocks = min_not_zero(sdkp->max_ws_blocks,
 714                                           (u32)SD_MAX_WS10_BLOCKS);
 715                 break;
 716
 717         case SD_LBP_ZERO:
 718                 max_blocks = min_not_zero(sdkp->max_ws_blocks,
 719                                           (u32)SD_MAX_WS10_BLOCKS);
 720                 break;
 721         }
 722
 723         blk_queue_max_discard_sectors(q, max_blocks * (logical_block_size >> 9));
 724         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, q);
 725 }
 726
 727 static int sd_setup_unmap_cmnd(struct scsi_cmnd *cmd)
 728 {
 729         struct scsi_device *sdp = cmd->device;
 730         struct request *rq = cmd->request;
 731         u64 sector = blk_rq_pos(rq) >> (ilog2(sdp->sector_size) - 9);
 732         u32 nr_sectors = blk_rq_sectors(rq) >> (ilog2(sdp->sector_size) - 9);
 733         unsigned int data_len = 24;
 734         char *buf;
 735
 736         rq->special_vec.bv_page = alloc_page(GFP_ATOMIC | __GFP_ZERO);
 737         if (!rq->special_vec.bv_page)
 738                 return BLKPREP_DEFER;
 739         rq->special_vec.bv_offset = 0;
 740         rq->special_vec.bv_len = data_len;
 741         rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
 742
 743         cmd->cmd_len = 10;
 744         cmd->cmnd[0] = UNMAP;
 745         cmd->cmnd[8] = 24;
 746
 747         buf = page_address(rq->special_vec.bv_page);
 748         put_unaligned_be16(6 + 16, &buf[0]);
 749         put_unaligned_be16(16, &buf[2]);
 750         put_unaligned_be64(sector, &buf[8]);
 751         put_unaligned_be32(nr_sectors, &buf[16]);
 752
 753         cmd->allowed = SD_MAX_RETRIES;
 754         cmd->transfersize = data_len;
 755         rq->timeout = SD_TIMEOUT;
 756         scsi_req(rq)->resid_len = data_len;
 757
 758         return scsi_init_io(cmd);
 759 }
 760
 761 static int sd_setup_write_same16_cmnd(struct scsi_cmnd *cmd, bool unmap)
 762 {
 763         struct scsi_device *sdp = cmd->device;
 764         struct request *rq = cmd->request;
 765         u64 sector = blk_rq_pos(rq) >> (ilog2(sdp->sector_size) - 9);
 766         u32 nr_sectors = blk_rq_sectors(rq) >> (ilog2(sdp->sector_size) - 9);
 767         u32 data_len = sdp->sector_size;
 768
 769         rq->special_vec.bv_page = alloc_page(GFP_ATOMIC | __GFP_ZERO);
 770         if (!rq->special_vec.bv_page)
 771                 return BLKPREP_DEFER;
 772         rq->special_vec.bv_offset = 0;
 773         rq->special_vec.bv_len = data_len;
 774         rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
 775
 776         cmd->cmd_len = 16;
 777         cmd->cmnd[0] = WRITE_SAME_16;
 778         if (unmap)
 779                 cmd->cmnd[1] = 0x8; /* UNMAP */
 780         put_unaligned_be64(sector, &cmd->cmnd[2]);
 781         put_unaligned_be32(nr_sectors, &cmd->cmnd[10]);
 782
 783         cmd->allowed = SD_MAX_RETRIES;
 784         cmd->transfersize = data_len;
 785         rq->timeout = unmap ? SD_TIMEOUT : SD_WRITE_SAME_TIMEOUT;
 786         scsi_req(rq)->resid_len = data_len;
 787
 788         return scsi_init_io(cmd);
 789 }
 790
 791 static int sd_setup_write_same10_cmnd(struct scsi_cmnd *cmd, bool unmap)
 792 {
 793         struct scsi_device *sdp = cmd->device;
 794         struct request *rq = cmd->request;
 795         u64 sector = blk_rq_pos(rq) >> (ilog2(sdp->sector_size) - 9);
 796         u32 nr_sectors = blk_rq_sectors(rq) >> (ilog2(sdp->sector_size) - 9);
 797         u32 data_len = sdp->sector_size;
 798
 799         rq->special_vec.bv_page = alloc_page(GFP_ATOMIC | __GFP_ZERO);
 800         if (!rq->special_vec.bv_page)
 801                 return BLKPREP_DEFER;
 802         rq->special_vec.bv_offset = 0;
 803         rq->special_vec.bv_len = data_len;
 804         rq->rq_flags |= RQF_SPECIAL_PAYLOAD;
 805
 806         cmd->cmd_len = 10;
 807         cmd->cmnd[0] = WRITE_SAME;
 808         if (unmap)
 809                 cmd->cmnd[1] = 0x8; /* UNMAP */
 810         put_unaligned_be32(sector, &cmd->cmnd[2]);
 811         put_unaligned_be16(nr_sectors, &cmd->cmnd[7]);
 812
 813         cmd->allowed = SD_MAX_RETRIES;
 814         cmd->transfersize = data_len;
 815         rq->timeout = unmap ? SD_TIMEOUT : SD_WRITE_SAME_TIMEOUT;
 816         scsi_req(rq)->resid_len = data_len;
 817
 818         return scsi_init_io(cmd);
 819 }
 820
 821 static int sd_setup_write_zeroes_cmnd(struct scsi_cmnd *cmd)
 822 {
 823         struct request *rq = cmd->request;
 824         struct scsi_device *sdp = cmd->device;
 825         struct scsi_disk *sdkp = scsi_disk(rq->rq_disk);
 826         u64 sector = blk_rq_pos(rq) >> (ilog2(sdp->sector_size) - 9);
 827         u32 nr_sectors = blk_rq_sectors(rq) >> (ilog2(sdp->sector_size) - 9);
 828
 829         if (!(rq->cmd_flags & REQ_NOUNMAP)) {
 830                 switch (sdkp->zeroing_mode) {
 831                 case SD_ZERO_WS16_UNMAP:
 832                         return sd_setup_write_same16_cmnd(cmd, true);
 833                 case SD_ZERO_WS10_UNMAP:
 834                         return sd_setup_write_same10_cmnd(cmd, true);
 835                 }
 836         }
 837
 838         if (sdp->no_write_same)
 839                 return BLKPREP_INVALID;
 840         if (sdkp->ws16 || sector > 0xffffffff || nr_sectors > 0xffff)
 841                 return sd_setup_write_same16_cmnd(cmd, false);
 842         return sd_setup_write_same10_cmnd(cmd, false);
 843 }
 844
 845 static void sd_config_write_same(struct scsi_disk *sdkp)
 846 {
 847         struct request_queue *q = sdkp->disk->queue;
 848         unsigned int logical_block_size = sdkp->device->sector_size;
 849
 850         if (sdkp->device->no_write_same) {
 851                 sdkp->max_ws_blocks = 0;
 852                 goto out;
 853         }
 854
 855         /* Some devices can not handle block counts above 0xffff despite
 856          * supporting WRITE SAME(16). Consequently we default to 64k
 857          * blocks per I/O unless the device explicitly advertises a
 858          * bigger limit.
 859          */
 860         if (sdkp->max_ws_blocks > SD_MAX_WS10_BLOCKS)
 861                 sdkp->max_ws_blocks = min_not_zero(sdkp->max_ws_blocks,
 862                                                    (u32)SD_MAX_WS16_BLOCKS);
 863         else if (sdkp->ws16 || sdkp->ws10 || sdkp->device->no_report_opcodes)
 864                 sdkp->max_ws_blocks = min_not_zero(sdkp->max_ws_blocks,
 865                                                    (u32)SD_MAX_WS10_BLOCKS);
 866         else {
 867                 sdkp->device->no_write_same = 1;
 868                 sdkp->max_ws_blocks = 0;
 869         }
 870
 871         if (sdkp->lbprz && sdkp->lbpws)
 872                 sdkp->zeroing_mode = SD_ZERO_WS16_UNMAP;
 873         else if (sdkp->lbprz && sdkp->lbpws10)
 874                 sdkp->zeroing_mode = SD_ZERO_WS10_UNMAP;
 875         else if (sdkp->max_ws_blocks)
 876                 sdkp->zeroing_mode = SD_ZERO_WS;
 877         else
 878                 sdkp->zeroing_mode = SD_ZERO_WRITE;
 879
 880 out:
 881         blk_queue_max_write_same_sectors(q, sdkp->max_ws_blocks *
 882                                          (logical_block_size >> 9));
 883         blk_queue_max_write_zeroes_sectors(q, sdkp->max_ws_blocks *
 884                                          (logical_block_size >> 9));
 885 }
 886
 887 /**
 888  * sd_setup_write_same_cmnd - write the same data to multiple blocks
 889  * @cmd: command to prepare
 890  *
 891  * Will issue either WRITE SAME(10) or WRITE SAME(16) depending on
 892  * preference indicated by target device.
 893  **/
 894 static int sd_setup_write_same_cmnd(struct scsi_cmnd *cmd)
 895 {
 896         struct request *rq = cmd->request;
 897         struct scsi_device *sdp = cmd->device;
 898         struct scsi_disk *sdkp = scsi_disk(rq->rq_disk);
 899         struct bio *bio = rq->bio;
 900         sector_t sector = blk_rq_pos(rq);
 901         unsigned int nr_sectors = blk_rq_sectors(rq);
 902         unsigned int nr_bytes = blk_rq_bytes(rq);
 903         int ret;
 904
 905         if (sdkp->device->no_write_same)
 906                 return BLKPREP_INVALID;
 907
 908         BUG_ON(bio_offset(bio) || bio_iovec(bio).bv_len != sdp->sector_size);
 909
 910         if (sd_is_zoned(sdkp)) {
 911                 ret = sd_zbc_setup_write_cmnd(cmd);
 912                 if (ret != BLKPREP_OK)
 913                         return ret;
 914         }
 915
 916         sector >>= ilog2(sdp->sector_size) - 9;
 917         nr_sectors >>= ilog2(sdp->sector_size) - 9;
 918
 919         rq->timeout = SD_WRITE_SAME_TIMEOUT;
 920
 921         if (sdkp->ws16 || sector > 0xffffffff || nr_sectors > 0xffff) {
 922                 cmd->cmd_len = 16;
 923                 cmd->cmnd[0] = WRITE_SAME_16;
 924                 put_unaligned_be64(sector, &cmd->cmnd[2]);
 925                 put_unaligned_be32(nr_sectors, &cmd->cmnd[10]);
 926         } else {
 927                 cmd->cmd_len = 10;
 928                 cmd->cmnd[0] = WRITE_SAME;
 929                 put_unaligned_be32(sector, &cmd->cmnd[2]);
 930                 put_unaligned_be16(nr_sectors, &cmd->cmnd[7]);
 931         }
 932
 933         cmd->transfersize = sdp->sector_size;
 934         cmd->allowed = SD_MAX_RETRIES;
 935
 936         /*
 937          * For WRITE SAME the data transferred via the DATA OUT buffer is
 938          * different from the amount of data actually written to the target.
 939          *
 940          * We set up __data_len to the amount of data transferred via the
 941          * DATA OUT buffer so that blk_rq_map_sg sets up the proper S/G list
 942          * to transfer a single sector of data first, but then reset it to
 943          * the amount of data to be written right after so that the I/O path
 944          * knows how much to actually write.
 945          */
 946         rq->__data_len = sdp->sector_size;
 947         ret = scsi_init_io(cmd);
 948         rq->__data_len = nr_bytes;
 949         return ret;
 950 }
 951
 952 static int sd_setup_flush_cmnd(struct scsi_cmnd *cmd)
 953 {
 954         struct request *rq = cmd->request;
 955
 956         /* flush requests don't perform I/O, zero the S/G table */
 957         memset(&cmd->sdb, 0, sizeof(cmd->sdb));
 958
 959         cmd->cmnd[0] = SYNCHRONIZE_CACHE;
 960         cmd->cmd_len = 10;
 961         cmd->transfersize = 0;
 962         cmd->allowed = SD_MAX_RETRIES;
 963
 964         rq->timeout = rq->q->rq_timeout * SD_FLUSH_TIMEOUT_MULTIPLIER;
 965         return BLKPREP_OK;
 966 }
 967
 968 static int sd_setup_read_write_cmnd(struct scsi_cmnd *SCpnt)
 969 {
 970         struct request *rq = SCpnt->request;
 971         struct scsi_device *sdp = SCpnt->device;
 972         struct gendisk *disk = rq->rq_disk;
 973         struct scsi_disk *sdkp = scsi_disk(disk);
 974         sector_t block = blk_rq_pos(rq);
 975         sector_t threshold;
 976         unsigned int this_count = blk_rq_sectors(rq);
 977         unsigned int dif, dix;
 978         bool zoned_write = sd_is_zoned(sdkp) && rq_data_dir(rq) == WRITE;
 979         int ret;
 980         unsigned char protect;
 981
 982         if (zoned_write) {
 983                 ret = sd_zbc_setup_write_cmnd(SCpnt);
 984                 if (ret != BLKPREP_OK)
 985                         return ret;
 986         }
 987
 988         ret = scsi_init_io(SCpnt);
 989         if (ret != BLKPREP_OK)
 990                 goto out;
 991         SCpnt = rq->special;
 992
 993         /* from here on until we're complete, any goto out
 994          * is used for a killable error condition */
 995         ret = BLKPREP_KILL;
 996
 997         SCSI_LOG_HLQUEUE(1,
 998                 scmd_printk(KERN_INFO, SCpnt,
 999                         "%s: block=%llu, count=%d\n",
1000                         __func__, (unsigned long long)block, this_count));
1001
1002         if (!sdp || !scsi_device_online(sdp) ||
1003             block + blk_rq_sectors(rq) > get_capacity(disk)) {
1004                 SCSI_LOG_HLQUEUE(2, scmd_printk(KERN_INFO, SCpnt,
1005                                                 "Finishing %u sectors\n",
1006                                                 blk_rq_sectors(rq)));
1007                 SCSI_LOG_HLQUEUE(2, scmd_printk(KERN_INFO, SCpnt,
1008                                                 "Retry with 0x%p\n", SCpnt));
1009                 goto out;
1010         }
1011
1012         if (sdp->changed) {
1013                 /*
1014                  * quietly refuse to do anything to a changed disc until
1015                  * the changed bit has been reset
1016                  */
1017                 /* printk("SCSI disk has been changed or is not present. Prohibiting further I/O.\n"); */
1018                 goto out;
1019         }
1020
1021         /*
1022          * Some SD card readers can't handle multi-sector accesses which touch
1023          * the last one or two hardware sectors.  Split accesses as needed.
1024          */
1025         threshold = get_capacity(disk) - SD_LAST_BUGGY_SECTORS *
1026                 (sdp->sector_size / 512);
1027
1028         if (unlikely(sdp->last_sector_bug && block + this_count > threshold)) {
1029                 if (block < threshold) {
1030                         /* Access up to the threshold but not beyond */
1031                         this_count = threshold - block;
1032                 } else {
1033                         /* Access only a single hardware sector */
1034                         this_count = sdp->sector_size / 512;
1035                 }
1036         }
1037
1038         SCSI_LOG_HLQUEUE(2, scmd_printk(KERN_INFO, SCpnt, "block=%llu\n",
1039                                         (unsigned long long)block));
1040
1041         /*
1042          * If we have a 1K hardware sectorsize, prevent access to single
1043          * 512 byte sectors.  In theory we could handle this - in fact
1044          * the scsi cdrom driver must be able to handle this because
1045          * we typically use 1K blocksizes, and cdroms typically have
1046          * 2K hardware sectorsizes.  Of course, things are simpler
1047          * with the cdrom, since it is read-only.  For performance
1048          * reasons, the filesystems should be able to handle this
1049          * and not force the scsi disk driver to use bounce buffers
1050          * for this.
1051          */
1052         if (sdp->sector_size == 1024) {
1053                 if ((block & 1) || (blk_rq_sectors(rq) & 1)) {
1054                         scmd_printk(KERN_ERR, SCpnt,
1055                                     "Bad block number requested\n");
1056                         goto out;
1057                 } else {
1058                         block = block >> 1;
1059                         this_count = this_count >> 1;
1060                 }
1061         }
1062         if (sdp->sector_size == 2048) {
1063                 if ((block & 3) || (blk_rq_sectors(rq) & 3)) {
1064                         scmd_printk(KERN_ERR, SCpnt,
1065                                     "Bad block number requested\n");
1066                         goto out;
1067                 } else {
1068                         block = block >> 2;
1069                         this_count = this_count >> 2;
1070                 }
1071         }
1072         if (sdp->sector_size == 4096) {
1073                 if ((block & 7) || (blk_rq_sectors(rq) & 7)) {
1074                         scmd_printk(KERN_ERR, SCpnt,
1075                                     "Bad block number requested\n");
1076                         goto out;
1077                 } else {
1078                         block = block >> 3;
1079                         this_count = this_count >> 3;
1080                 }
1081         }
1082         if (rq_data_dir(rq) == WRITE) {
1083                 SCpnt->cmnd[0] = WRITE_6;
1084
1085                 if (blk_integrity_rq(rq))
1086                         sd_dif_prepare(SCpnt);
1087
1088         } else if (rq_data_dir(rq) == READ) {
1089                 SCpnt->cmnd[0] = READ_6;
1090         } else {
1091                 scmd_printk(KERN_ERR, SCpnt, "Unknown command %d\n", req_op(rq));
1092                 goto out;
1093         }
1094
1095         SCSI_LOG_HLQUEUE(2, scmd_printk(KERN_INFO, SCpnt,
1096                                         "%s %d/%u 512 byte blocks.\n",
1097                                         (rq_data_dir(rq) == WRITE) ?
1098                                         "writing" : "reading", this_count,
1099                                         blk_rq_sectors(rq)));
1100
1101         dix = scsi_prot_sg_count(SCpnt);
1102         dif = scsi_host_dif_capable(SCpnt->device->host, sdkp->protection_type);
1103
1104         if (dif || dix)
1105                 protect = sd_setup_protect_cmnd(SCpnt, dix, dif);
1106         else
1107                 protect = 0;
1108
1109         if (protect && sdkp->protection_type == T10_PI_TYPE2_PROTECTION) {
1110                 SCpnt->cmnd = mempool_alloc(sd_cdb_pool, GFP_ATOMIC);
1111
1112                 if (unlikely(SCpnt->cmnd == NULL)) {
1113                         ret = BLKPREP_DEFER;
1114                         goto out;
1115                 }
1116
1117                 SCpnt->cmd_len = SD_EXT_CDB_SIZE;
1118                 memset(SCpnt->cmnd, 0, SCpnt->cmd_len);
1119                 SCpnt->cmnd[0] = VARIABLE_LENGTH_CMD;
1120                 SCpnt->cmnd[7] = 0x18;
1121                 SCpnt->cmnd[9] = (rq_data_dir(rq) == READ) ? READ_32 : WRITE_32;
1122                 SCpnt->cmnd[10] = protect | ((rq->cmd_flags & REQ_FUA) ? 0x8 : 0);
1123
1124                 /* LBA */
1125                 SCpnt->cmnd[12] = sizeof(block) > 4 ? (unsigned char) (block >> 56) & 0xff : 0;
1126                 SCpnt->cmnd[13] = sizeof(block) > 4 ? (unsigned char) (block >> 48) & 0xff : 0;
1127                 SCpnt->cmnd[14] = sizeof(block) > 4 ? (unsigned char) (block >> 40) & 0xff : 0;
1128                 SCpnt->cmnd[15] = sizeof(block) > 4 ? (unsigned char) (block >> 32) & 0xff : 0;
1129                 SCpnt->cmnd[16] = (unsigned char) (block >> 24) & 0xff;
1130                 SCpnt->cmnd[17] = (unsigned char) (block >> 16) & 0xff;
1131                 SCpnt->cmnd[18] = (unsigned char) (block >> 8) & 0xff;
1132                 SCpnt->cmnd[19] = (unsigned char) block & 0xff;
1133
1134                 /* Expected Indirect LBA */
1135                 SCpnt->cmnd[20] = (unsigned char) (block >> 24) & 0xff;
1136                 SCpnt->cmnd[21] = (unsigned char) (block >> 16) & 0xff;
1137                 SCpnt->cmnd[22] = (unsigned char) (block >> 8) & 0xff;
1138                 SCpnt->cmnd[23] = (unsigned char) block & 0xff;
1139
1140                 /* Transfer length */
1141                 SCpnt->cmnd[28] = (unsigned char) (this_count >> 24) & 0xff;
1142                 SCpnt->cmnd[29] = (unsigned char) (this_count >> 16) & 0xff;
1143                 SCpnt->cmnd[30] = (unsigned char) (this_count >> 8) & 0xff;
1144                 SCpnt->cmnd[31] = (unsigned char) this_count & 0xff;
1145         } else if (sdp->use_16_for_rw || (this_count > 0xffff)) {
1146                 SCpnt->cmnd[0] += READ_16 - READ_6;
1147                 SCpnt->cmnd[1] = protect | ((rq->cmd_flags & REQ_FUA) ? 0x8 : 0);
1148                 SCpnt->cmnd[2] = sizeof(block) > 4 ? (unsigned char) (block >> 56) & 0xff : 0;
1149                 SCpnt->cmnd[3] = sizeof(block) > 4 ? (unsigned char) (block >> 48) & 0xff : 0;
1150                 SCpnt->cmnd[4] = sizeof(block) > 4 ? (unsigned char) (block >> 40) & 0xff : 0;
1151                 SCpnt->cmnd[5] = sizeof(block) > 4 ? (unsigned char) (block >> 32) & 0xff : 0;
1152                 SCpnt->cmnd[6] = (unsigned char) (block >> 24) & 0xff;
1153                 SCpnt->cmnd[7] = (unsigned char) (block >> 16) & 0xff;
1154                 SCpnt->cmnd[8] = (unsigned char) (block >> 8) & 0xff;
1155                 SCpnt->cmnd[9] = (unsigned char) block & 0xff;
1156                 SCpnt->cmnd[10] = (unsigned char) (this_count >> 24) & 0xff;
1157                 SCpnt->cmnd[11] = (unsigned char) (this_count >> 16) & 0xff;
1158                 SCpnt->cmnd[12] = (unsigned char) (this_count >> 8) & 0xff;
1159                 SCpnt->cmnd[13] = (unsigned char) this_count & 0xff;
1160                 SCpnt->cmnd[14] = SCpnt->cmnd[15] = 0;
1161         } else if ((this_count > 0xff) || (block > 0x1fffff) ||
1162                    scsi_device_protection(SCpnt->device) ||
1163                    SCpnt->device->use_10_for_rw) {
1164                 SCpnt->cmnd[0] += READ_10 - READ_6;
1165                 SCpnt->cmnd[1] = protect | ((rq->cmd_flags & REQ_FUA) ? 0x8 : 0);
1166                 SCpnt->cmnd[2] = (unsigned char) (block >> 24) & 0xff;
1167                 SCpnt->cmnd[3] = (unsigned char) (block >> 16) & 0xff;
1168                 SCpnt->cmnd[4] = (unsigned char) (block >> 8) & 0xff;
1169                 SCpnt->cmnd[5] = (unsigned char) block & 0xff;
1170                 SCpnt->cmnd[6] = SCpnt->cmnd[9] = 0;
1171                 SCpnt->cmnd[7] = (unsigned char) (this_count >> 8) & 0xff;
1172                 SCpnt->cmnd[8] = (unsigned char) this_count & 0xff;
1173         } else {
1174                 if (unlikely(rq->cmd_flags & REQ_FUA)) {
1175                         /*
1176                          * This happens only if this drive failed
1177                          * 10byte rw command with ILLEGAL_REQUEST
1178                          * during operation and thus turned off
1179                          * use_10_for_rw.
1180                          */
1181                         scmd_printk(KERN_ERR, SCpnt,
1182                                     "FUA write on READ/WRITE(6) drive\n");
1183                         goto out;
1184                 }
1185
1186                 SCpnt->cmnd[1] |= (unsigned char) ((block >> 16) & 0x1f);
1187                 SCpnt->cmnd[2] = (unsigned char) ((block >> 8) & 0xff);
1188                 SCpnt->cmnd[3] = (unsigned char) block & 0xff;
1189                 SCpnt->cmnd[4] = (unsigned char) this_count;
1190                 SCpnt->cmnd[5] = 0;
1191         }
1192         SCpnt->sdb.length = this_count * sdp->sector_size;
1193
1194         /*
1195          * We shouldn't disconnect in the middle of a sector, so with a dumb
1196          * host adapter, it's safe to assume that we can at least transfer
1197          * this many bytes between each connect / disconnect.
1198          */
1199         SCpnt->transfersize = sdp->sector_size;
1200         SCpnt->underflow = this_count << 9;
1201         SCpnt->allowed = SD_MAX_RETRIES;
1202
1203         /*
1204          * This indicates that the command is ready from our end to be
1205          * queued.
1206          */
1207         ret = BLKPREP_OK;
1208  out:
1209         if (zoned_write && ret != BLKPREP_OK)
1210                 sd_zbc_cancel_write_cmnd(SCpnt);
1211
1212         return ret;
1213 }
1214
1215 static int sd_init_command(struct scsi_cmnd *cmd)
1216 {
1217         struct request *rq = cmd->request;
1218
1219         switch (req_op(rq)) {
1220         case REQ_OP_DISCARD:
1221                 switch (scsi_disk(rq->rq_disk)->provisioning_mode) {
1222                 case SD_LBP_UNMAP:
1223                         return sd_setup_unmap_cmnd(cmd);
1224                 case SD_LBP_WS16:
1225                         return sd_setup_write_same16_cmnd(cmd, true);
1226                 case SD_LBP_WS10:
1227                         return sd_setup_write_same10_cmnd(cmd, true);
1228                 case SD_LBP_ZERO:
1229                         return sd_setup_write_same10_cmnd(cmd, false);
1230                 default:
1231                         return BLKPREP_INVALID;
1232                 }
1233         case REQ_OP_WRITE_ZEROES:
1234                 return sd_setup_write_zeroes_cmnd(cmd);
1235         case REQ_OP_WRITE_SAME:
1236                 return sd_setup_write_same_cmnd(cmd);
1237         case REQ_OP_FLUSH:
1238                 return sd_setup_flush_cmnd(cmd);
1239         case REQ_OP_READ:
1240         case REQ_OP_WRITE:
1241                 return sd_setup_read_write_cmnd(cmd);
1242         case REQ_OP_ZONE_REPORT:
1243                 return sd_zbc_setup_report_cmnd(cmd);
1244         case REQ_OP_ZONE_RESET:
1245                 return sd_zbc_setup_reset_cmnd(cmd);
1246         default:
1247                 BUG();
1248         }
1249 }
1250
1251 static void sd_uninit_command(struct scsi_cmnd *SCpnt)
1252 {
1253         struct request *rq = SCpnt->request;
1254
1255         if (rq->rq_flags & RQF_SPECIAL_PAYLOAD)
1256                 __free_page(rq->special_vec.bv_page);
1257
1258         if (SCpnt->cmnd != scsi_req(rq)->cmd) {
1259                 mempool_free(SCpnt->cmnd, sd_cdb_pool);
1260                 SCpnt->cmnd = NULL;
1261                 SCpnt->cmd_len = 0;
1262         }
1263 }
1264
1265 /**
1266  *      sd_open - open a scsi disk device
1267  *      @inode: only i_rdev member may be used
1268  *      @filp: only f_mode and f_flags may be used
1269  *
1270  *      Returns 0 if successful. Returns a negated errno value in case
1271  *      of error.
1272  *
1273  *      Note: This can be called from a user context (e.g. fsck(1) )
1274  *      or from within the kernel (e.g. as a result of a mount(1) ).
1275  *      In the latter case @inode and @filp carry an abridged amount
1276  *      of information as noted above.
1277  *
1278  *      Locking: called with bdev->bd_mutex held.
1279  **/
1280 static int sd_open(struct block_device *bdev, fmode_t mode)
1281 {
1282         struct scsi_disk *sdkp = scsi_disk_get(bdev->bd_disk);
1283         struct scsi_device *sdev;
1284         int retval;
1285
1286         if (!sdkp)
1287                 return -ENXIO;
1288
1289         SCSI_LOG_HLQUEUE(3, sd_printk(KERN_INFO, sdkp, "sd_open\n"));
1290
1291         sdev = sdkp->device;
1292
1293         /*
1294          * If the device is in error recovery, wait until it is done.
1295          * If the device is offline, then disallow any access to it.
1296          */
1297         retval = -ENXIO;
1298         if (!scsi_block_when_processing_errors(sdev))
1299                 goto error_out;
1300
1301         if (sdev->removable || sdkp->write_prot)
1302                 check_disk_change(bdev);
1303
1304         /*
1305          * If the drive is empty, just let the open fail.
1306          */
1307         retval = -ENOMEDIUM;
1308         if (sdev->removable && !sdkp->media_present && !(mode & FMODE_NDELAY))
1309                 goto error_out;
1310
1311         /*
1312          * If the device has the write protect tab set, have the open fail
1313          * if the user expects to be able to write to the thing.
1314          */
1315         retval = -EROFS;
1316         if (sdkp->write_prot && (mode & FMODE_WRITE))
1317                 goto error_out;
1318
1319         /*
1320          * It is possible that the disk changing stuff resulted in
1321          * the device being taken offline.  If this is the case,
1322          * report this to the user, and don't pretend that the
1323          * open actually succeeded.
1324          */
1325         retval = -ENXIO;
1326         if (!scsi_device_online(sdev))
1327                 goto error_out;
1328
1329         if ((atomic_inc_return(&sdkp->openers) == 1) && sdev->removable) {
1330                 if (scsi_block_when_processing_errors(sdev))
1331                         scsi_set_medium_removal(sdev, SCSI_REMOVAL_PREVENT);
1332         }
1333
1334         return 0;
1335
1336 error_out:
1337         scsi_disk_put(sdkp);
1338         return retval;
1339 }
1340
1341 /**
1342  *      sd_release - invoked when the (last) close(2) is called on this
1343  *      scsi disk.
1344  *      @inode: only i_rdev member may be used
1345  *      @filp: only f_mode and f_flags may be used
1346  *
1347  *      Returns 0.
1348  *
1349  *      Note: may block (uninterruptible) if error recovery is underway
1350  *      on this disk.
1351  *
1352  *      Locking: called with bdev->bd_mutex held.
1353  **/
1354 static void sd_release(struct gendisk *disk, fmode_t mode)
1355 {
1356         struct scsi_disk *sdkp = scsi_disk(disk);
1357         struct scsi_device *sdev = sdkp->device;
1358
1359         SCSI_LOG_HLQUEUE(3, sd_printk(KERN_INFO, sdkp, "sd_release\n"));
1360
1361         if (atomic_dec_return(&sdkp->openers) == 0 && sdev->removable) {
1362                 if (scsi_block_when_processing_errors(sdev))
1363                         scsi_set_medium_removal(sdev, SCSI_REMOVAL_ALLOW);
1364         }
1365
1366         /*
1367          * XXX and what if there are packets in flight and this close()
1368          * XXX is followed by a "rmmod sd_mod"?
1369          */
1370
1371         scsi_disk_put(sdkp);
1372 }
1373
1374 static int sd_getgeo(struct block_device *bdev, struct hd_geometry *geo)
1375 {
1376         struct scsi_disk *sdkp = scsi_disk(bdev->bd_disk);
1377         struct scsi_device *sdp = sdkp->device;
1378         struct Scsi_Host *host = sdp->host;
1379         sector_t capacity = logical_to_sectors(sdp, sdkp->capacity);
1380         int diskinfo[4];
1381
1382         /* default to most commonly used values */
1383         diskinfo[0] = 0x40;     /* 1 << 6 */
1384         diskinfo[1] = 0x20;     /* 1 << 5 */
1385         diskinfo[2] = capacity >> 11;
1386
1387         /* override with calculated, extended default, or driver values */
1388         if (host->hostt->bios_param)
1389                 host->hostt->bios_param(sdp, bdev, capacity, diskinfo);
1390         else
1391                 scsicam_bios_param(bdev, capacity, diskinfo);
1392
1393         geo->heads = diskinfo[0];
1394         geo->sectors = diskinfo[1];
1395         geo->cylinders = diskinfo[2];
1396         return 0;
1397 }
1398
1399 /**
1400  *      sd_ioctl - process an ioctl
1401  *      @inode: only i_rdev/i_bdev members may be used
1402  *      @filp: only f_mode and f_flags may be used
1403  *      @cmd: ioctl command number
1404  *      @arg: this is third argument given to ioctl(2) system call.
1405  *      Often contains a pointer.
1406  *
1407  *      Returns 0 if successful (some ioctls return positive numbers on
1408  *      success as well). Returns a negated errno value in case of error.
1409  *
1410  *      Note: most ioctls are forward onto the block subsystem or further
1411  *      down in the scsi subsystem.
1412  **/
1413 static int sd_ioctl(struct block_device *bdev, fmode_t mode,
1414                     unsigned int cmd, unsigned long arg)
1415 {
1416         struct gendisk *disk = bdev->bd_disk;
1417         struct scsi_disk *sdkp = scsi_disk(disk);
1418         struct scsi_device *sdp = sdkp->device;
1419         void __user *p = (void __user *)arg;
1420         int error;
1421
1422         SCSI_LOG_IOCTL(1, sd_printk(KERN_INFO, sdkp, "sd_ioctl: disk=%s, "
1423                                     "cmd=0x%x\n", disk->disk_name, cmd));
1424
1425         error = scsi_verify_blk_ioctl(bdev, cmd);
1426         if (error < 0)
1427                 return error;
1428
1429         /*
1430          * If we are in the middle of error recovery, don't let anyone
1431          * else try and use this device.  Also, if error recovery fails, it
1432          * may try and take the device offline, in which case all further
1433          * access to the device is prohibited.
1434          */
1435         error = scsi_ioctl_block_when_processing_errors(sdp, cmd,
1436                         (mode & FMODE_NDELAY) != 0);
1437         if (error)
1438                 goto out;
1439
1440         /*
1441          * Send SCSI addressing ioctls directly to mid level, send other
1442          * ioctls to block level and then onto mid level if they can't be
1443          * resolved.
1444          */
1445         switch (cmd) {
1446                 case SCSI_IOCTL_GET_IDLUN:
1447                 case SCSI_IOCTL_GET_BUS_NUMBER:
1448                         error = scsi_ioctl(sdp, cmd, p);
1449                         break;
1450                 default:
1451                         error = scsi_cmd_blk_ioctl(bdev, mode, cmd, p);
1452                         if (error != -ENOTTY)
1453                                 break;
1454                         error = scsi_ioctl(sdp, cmd, p);
1455                         break;
1456         }
1457 out:
1458         return error;
1459 }
1460
1461 static void set_media_not_present(struct scsi_disk *sdkp)
1462 {
1463         if (sdkp->media_present)
1464                 sdkp->device->changed = 1;
1465
1466         if (sdkp->device->removable) {
1467                 sdkp->media_present = 0;
1468                 sdkp->capacity = 0;
1469         }
1470 }
1471
1472 static int media_not_present(struct scsi_disk *sdkp,
1473                              struct scsi_sense_hdr *sshdr)
1474 {
1475         if (!scsi_sense_valid(sshdr))
1476                 return 0;
1477
1478         /* not invoked for commands that could return deferred errors */
1479         switch (sshdr->sense_key) {
1480         case UNIT_ATTENTION:
1481         case NOT_READY:
1482                 /* medium not present */
1483                 if (sshdr->asc == 0x3A) {
1484                         set_media_not_present(sdkp);
1485                         return 1;
1486                 }
1487         }
1488         return 0;
1489 }
1490
1491 /**
1492  *      sd_check_events - check media events
1493  *      @disk: kernel device descriptor
1494  *      @clearing: disk events currently being cleared
1495  *
1496  *      Returns mask of DISK_EVENT_*.
1497  *
1498  *      Note: this function is invoked from the block subsystem.
1499  **/
1500 static unsigned int sd_check_events(struct gendisk *disk, unsigned int clearing)
1501 {
1502         struct scsi_disk *sdkp = scsi_disk_get(disk);
1503         struct scsi_device *sdp;
1504         int retval;
1505
1506         if (!sdkp)
1507                 return 0;
1508
1509         sdp = sdkp->device;
1510         SCSI_LOG_HLQUEUE(3, sd_printk(KERN_INFO, sdkp, "sd_check_events\n"));
1511
1512         /*
1513          * If the device is offline, don't send any commands - just pretend as
1514          * if the command failed.  If the device ever comes back online, we
1515          * can deal with it then.  It is only because of unrecoverable errors
1516          * that we would ever take a device offline in the first place.
1517          */
1518         if (!scsi_device_online(sdp)) {
1519                 set_media_not_present(sdkp);
1520                 goto out;
1521         }
1522
1523         /*
1524          * Using TEST_UNIT_READY enables differentiation between drive with
1525          * no cartridge loaded - NOT READY, drive with changed cartridge -
1526          * UNIT ATTENTION, or with same cartridge - GOOD STATUS.
1527          *
1528          * Drives that auto spin down. eg iomega jaz 1G, will be started
1529          * by sd_spinup_disk() from sd_revalidate_disk(), which happens whenever
1530          * sd_revalidate() is called.
1531          */
1532         if (scsi_block_when_processing_errors(sdp)) {
1533                 struct scsi_sense_hdr sshdr = { 0, };
1534
1535                 retval = scsi_test_unit_ready(sdp, SD_TIMEOUT, SD_MAX_RETRIES,
1536                                               &sshdr);
1537
1538                 /* failed to execute TUR, assume media not present */
1539                 if (host_byte(retval)) {
1540                         set_media_not_present(sdkp);
1541                         goto out;
1542                 }
1543
1544                 if (media_not_present(sdkp, &sshdr))
1545                         goto out;
1546         }
1547
1548         /*
1549          * For removable scsi disk we have to recognise the presence
1550          * of a disk in the drive.
1551          */
1552         if (!sdkp->media_present)
1553                 sdp->changed = 1;
1554         sdkp->media_present = 1;
1555 out:
1556         /*
1557          * sdp->changed is set under the following conditions:
1558          *
1559          *      Medium present state has changed in either direction.
1560          *      Device has indicated UNIT_ATTENTION.
1561          */
1562         retval = sdp->changed ? DISK_EVENT_MEDIA_CHANGE : 0;
1563         sdp->changed = 0;
1564         scsi_disk_put(sdkp);
1565         return retval;
1566 }
1567
1568 static int sd_sync_cache(struct scsi_disk *sdkp)
1569 {
1570         int retries, res;
1571         struct scsi_device *sdp = sdkp->device;
1572         const int timeout = sdp->request_queue->rq_timeout
1573                 * SD_FLUSH_TIMEOUT_MULTIPLIER;
1574         struct scsi_sense_hdr sshdr;
1575
1576         if (!scsi_device_online(sdp))
1577                 return -ENODEV;
1578
1579         for (retries = 3; retries > 0; --retries) {
1580                 unsigned char cmd[10] = { 0 };
1581
1582                 cmd[0] = SYNCHRONIZE_CACHE;
1583                 /*
1584                  * Leave the rest of the command zero to indicate
1585                  * flush everything.
1586                  */
1587                 res = scsi_execute(sdp, cmd, DMA_NONE, NULL, 0, NULL, &sshdr,
1588                                 timeout, SD_MAX_RETRIES, 0, RQF_PM, NULL);
1589                 if (res == 0)
1590                         break;
1591         }
1592
1593         if (res) {
1594                 sd_print_result(sdkp, "Synchronize Cache(10) failed", res);
1595
1596                 if (driver_byte(res) & DRIVER_SENSE)
1597                         sd_print_sense_hdr(sdkp, &sshdr);
1598                 /* we need to evaluate the error return  */
1599                 if (scsi_sense_valid(&sshdr) &&
1600                         (sshdr.asc == 0x3a ||   /* medium not present */
1601                          sshdr.asc == 0x20))    /* invalid command */
1602                                 /* this is no error here */
1603                                 return 0;
1604
1605                 switch (host_byte(res)) {
1606                 /* ignore errors due to racing a disconnection */
1607                 case DID_BAD_TARGET:
1608                 case DID_NO_CONNECT:
1609                         return 0;
1610                 /* signal the upper layer it might try again */
1611                 case DID_BUS_BUSY:
1612                 case DID_IMM_RETRY:
1613                 case DID_REQUEUE:
1614                 case DID_SOFT_ERROR:
1615                         return -EBUSY;
1616                 default:
1617                         return -EIO;
1618                 }
1619         }
1620         return 0;
1621 }
1622
1623 static void sd_rescan(struct device *dev)
1624 {
1625         struct scsi_disk *sdkp = dev_get_drvdata(dev);
1626
1627         revalidate_disk(sdkp->disk);
1628 }
1629
1630
1631 #ifdef CONFIG_COMPAT
1632 /*
1633  * This gets directly called from VFS. When the ioctl
1634  * is not recognized we go back to the other translation paths.
1635  */
1636 static int sd_compat_ioctl(struct block_device *bdev, fmode_t mode,
1637                            unsigned int cmd, unsigned long arg)
1638 {
1639         struct scsi_device *sdev = scsi_disk(bdev->bd_disk)->device;
1640         int error;
1641
1642         error = scsi_ioctl_block_when_processing_errors(sdev, cmd,
1643                         (mode & FMODE_NDELAY) != 0);
1644         if (error)
1645                 return error;
1646
1647         /*
1648          * Let the static ioctl translation table take care of it.
1649          */
1650         if (!sdev->host->hostt->compat_ioctl)
1651                 return -ENOIOCTLCMD;
1652         return sdev->host->hostt->compat_ioctl(sdev, cmd, (void __user *)arg);
1653 }
1654 #endif
1655
1656 static char sd_pr_type(enum pr_type type)
1657 {
1658         switch (type) {
1659         case PR_WRITE_EXCLUSIVE:
1660                 return 0x01;
1661         case PR_EXCLUSIVE_ACCESS:
1662                 return 0x03;
1663         case PR_WRITE_EXCLUSIVE_REG_ONLY:
1664                 return 0x05;
1665         case PR_EXCLUSIVE_ACCESS_REG_ONLY:
1666                 return 0x06;
1667         case PR_WRITE_EXCLUSIVE_ALL_REGS:
1668                 return 0x07;
1669         case PR_EXCLUSIVE_ACCESS_ALL_REGS:
1670                 return 0x08;
1671         default:
1672                 return 0;
1673         }
1674 };
1675
1676 static int sd_pr_command(struct block_device *bdev, u8 sa,
1677                 u64 key, u64 sa_key, u8 type, u8 flags)
1678 {
1679         struct scsi_device *sdev = scsi_disk(bdev->bd_disk)->device;
1680         struct scsi_sense_hdr sshdr;
1681         int result;
1682         u8 cmd[16] = { 0, };
1683         u8 data[24] = { 0, };
1684
1685         cmd[0] = PERSISTENT_RESERVE_OUT;
1686         cmd[1] = sa;
1687         cmd[2] = type;
1688         put_unaligned_be32(sizeof(data), &cmd[5]);
1689
1690         put_unaligned_be64(key, &data[0]);
1691         put_unaligned_be64(sa_key, &data[8]);
1692         data[20] = flags;
1693
1694         result = scsi_execute_req(sdev, cmd, DMA_TO_DEVICE, &data, sizeof(data),
1695                         &sshdr, SD_TIMEOUT, SD_MAX_RETRIES, NULL);
1696
1697         if ((driver_byte(result) & DRIVER_SENSE) &&
1698             (scsi_sense_valid(&sshdr))) {
1699                 sdev_printk(KERN_INFO, sdev, "PR command failed: %d\n", result);
1700                 scsi_print_sense_hdr(sdev, NULL, &sshdr);
1701         }
1702
1703         return result;
1704 }
1705
1706 static int sd_pr_register(struct block_device *bdev, u64 old_key, u64 new_key,
1707                 u32 flags)
1708 {
1709         if (flags & ~PR_FL_IGNORE_KEY)
1710                 return -EOPNOTSUPP;
1711         return sd_pr_command(bdev, (flags & PR_FL_IGNORE_KEY) ? 0x06 : 0x00,
1712                         old_key, new_key, 0,
1713                         (1 << 0) /* APTPL */);
1714 }
1715
1716 static int sd_pr_reserve(struct block_device *bdev, u64 key, enum pr_type type,
1717                 u32 flags)
1718 {
1719         if (flags)
1720                 return -EOPNOTSUPP;
1721         return sd_pr_command(bdev, 0x01, key, 0, sd_pr_type(type), 0);
1722 }
1723
1724 static int sd_pr_release(struct block_device *bdev, u64 key, enum pr_type type)
1725 {
1726         return sd_pr_command(bdev, 0x02, key, 0, sd_pr_type(type), 0);
1727 }
1728
1729 static int sd_pr_preempt(struct block_device *bdev, u64 old_key, u64 new_key,
1730                 enum pr_type type, bool abort)
1731 {
1732         return sd_pr_command(bdev, abort ? 0x05 : 0x04, old_key, new_key,
1733                              sd_pr_type(type), 0);
1734 }
1735
1736 static int sd_pr_clear(struct block_device *bdev, u64 key)
1737 {
1738         return sd_pr_command(bdev, 0x03, key, 0, 0, 0);
1739 }
1740
1741 static const struct pr_ops sd_pr_ops = {
1742         .pr_register    = sd_pr_register,
1743         .pr_reserve     = sd_pr_reserve,
1744         .pr_release     = sd_pr_release,
1745         .pr_preempt     = sd_pr_preempt,
1746         .pr_clear       = sd_pr_clear,
1747 };
1748
1749 static const struct block_device_operations sd_fops = {
1750         .owner                  = THIS_MODULE,
1751         .open                   = sd_open,
1752         .release                = sd_release,
1753         .ioctl                  = sd_ioctl,
1754         .getgeo                 = sd_getgeo,
1755 #ifdef CONFIG_COMPAT
1756         .compat_ioctl           = sd_compat_ioctl,
1757 #endif
1758         .check_events           = sd_check_events,
1759         .revalidate_disk        = sd_revalidate_disk,
1760         .unlock_native_capacity = sd_unlock_native_capacity,
1761         .pr_ops                 = &sd_pr_ops,
1762 };
1763
1764 /**
1765  *      sd_eh_action - error handling callback
1766  *      @scmd:          sd-issued command that has failed
1767  *      @eh_disp:       The recovery disposition suggested by the midlayer
1768  *
1769  *      This function is called by the SCSI midlayer upon completion of an
1770  *      error test command (currently TEST UNIT READY). The result of sending
1771  *      the eh command is passed in eh_disp.  We're looking for devices that
1772  *      fail medium access commands but are OK with non access commands like
1773  *      test unit ready (so wrongly see the device as having a successful
1774  *      recovery)
1775  **/
1776 static int sd_eh_action(struct scsi_cmnd *scmd, int eh_disp)
1777 {
1778         struct scsi_disk *sdkp = scsi_disk(scmd->request->rq_disk);
1779
1780         if (!scsi_device_online(scmd->device) ||
1781             !scsi_medium_access_command(scmd) ||
1782             host_byte(scmd->result) != DID_TIME_OUT ||
1783             eh_disp != SUCCESS)
1784                 return eh_disp;
1785
1786         /*
1787          * The device has timed out executing a medium access command.
1788          * However, the TEST UNIT READY command sent during error
1789          * handling completed successfully. Either the device is in the
1790          * process of recovering or has it suffered an internal failure
1791          * that prevents access to the storage medium.
1792          */
1793         sdkp->medium_access_timed_out++;
1794
1795         /*
1796          * If the device keeps failing read/write commands but TEST UNIT
1797          * READY always completes successfully we assume that medium
1798          * access is no longer possible and take the device offline.
1799          */
1800         if (sdkp->medium_access_timed_out >= sdkp->max_medium_access_timeouts) {
1801                 scmd_printk(KERN_ERR, scmd,
1802                             "Medium access timeout failure. Offlining disk!\n");
1803                 scsi_device_set_state(scmd->device, SDEV_OFFLINE);
1804
1805                 return FAILED;
1806         }
1807
1808         return eh_disp;
1809 }
1810
1811 static unsigned int sd_completed_bytes(struct scsi_cmnd *scmd)
1812 {
1813         u64 start_lba = blk_rq_pos(scmd->request);
1814         u64 end_lba = blk_rq_pos(scmd->request) + (scsi_bufflen(scmd) / 512);
1815         u64 factor = scmd->device->sector_size / 512;
1816         u64 bad_lba;
1817         int info_valid;
1818         /*
1819          * resid is optional but mostly filled in.  When it's unused,
1820          * its value is zero, so we assume the whole buffer transferred
1821          */
1822         unsigned int transferred = scsi_bufflen(scmd) - scsi_get_resid(scmd);
1823         unsigned int good_bytes;
1824
1825         info_valid = scsi_get_sense_info_fld(scmd->sense_buffer,
1826                                              SCSI_SENSE_BUFFERSIZE,
1827                                              &bad_lba);
1828         if (!info_valid)
1829                 return 0;
1830
1831         if (scsi_bufflen(scmd) <= scmd->device->sector_size)
1832                 return 0;
1833
1834         /* be careful ... don't want any overflows */
1835         do_div(start_lba, factor);
1836         do_div(end_lba, factor);
1837
1838         /* The bad lba was reported incorrectly, we have no idea where
1839          * the error is.
1840          */
1841         if (bad_lba < start_lba  || bad_lba >= end_lba)
1842                 return 0;
1843
1844         /* This computation should always be done in terms of
1845          * the resolution of the device's medium.
1846          */
1847         good_bytes = (bad_lba - start_lba) * scmd->device->sector_size;
1848         return min(good_bytes, transferred);
1849 }
1850
1851 /**
1852  *      sd_done - bottom half handler: called when the lower level
1853  *      driver has completed (successfully or otherwise) a scsi command.
1854  *      @SCpnt: mid-level's per command structure.
1855  *
1856  *      Note: potentially run from within an ISR. Must not block.
1857  **/
1858 static int sd_done(struct scsi_cmnd *SCpnt)
1859 {
1860         int result = SCpnt->result;
1861         unsigned int good_bytes = result ? 0 : scsi_bufflen(SCpnt);
1862         unsigned int sector_size = SCpnt->device->sector_size;
1863         unsigned int resid;
1864         struct scsi_sense_hdr sshdr;
1865         struct scsi_disk *sdkp = scsi_disk(SCpnt->request->rq_disk);
1866         struct request *req = SCpnt->request;
1867         int sense_valid = 0;
1868         int sense_deferred = 0;
1869         unsigned char op = SCpnt->cmnd[0];
1870         unsigned char unmap = SCpnt->cmnd[1] & 8;
1871
1872         switch (req_op(req)) {
1873         case REQ_OP_DISCARD:
1874         case REQ_OP_WRITE_ZEROES:
1875         case REQ_OP_WRITE_SAME:
1876         case REQ_OP_ZONE_RESET:
1877                 if (!result) {
1878                         good_bytes = blk_rq_bytes(req);
1879                         scsi_set_resid(SCpnt, 0);
1880                 } else {
1881                         good_bytes = 0;
1882                         scsi_set_resid(SCpnt, blk_rq_bytes(req));
1883                 }
1884                 break;
1885         case REQ_OP_ZONE_REPORT:
1886                 if (!result) {
1887                         good_bytes = scsi_bufflen(SCpnt)
1888                                 - scsi_get_resid(SCpnt);
1889                         scsi_set_resid(SCpnt, 0);
1890                 } else {
1891                         good_bytes = 0;
1892                         scsi_set_resid(SCpnt, blk_rq_bytes(req));
1893                 }
1894                 break;
1895         default:
1896                 /*
1897                  * In case of bogus fw or device, we could end up having
1898                  * an unaligned partial completion. Check this here and force
1899                  * alignment.
1900                  */
1901                 resid = scsi_get_resid(SCpnt);
1902                 if (resid & (sector_size - 1)) {
1903                         sd_printk(KERN_INFO, sdkp,
1904                                 "Unaligned partial completion (resid=%u, sector_sz=%u)\n",
1905                                 resid, sector_size);
1906                         resid = min(scsi_bufflen(SCpnt),
1907                                     round_up(resid, sector_size));
1908                         scsi_set_resid(SCpnt, resid);
1909                 }
1910         }
1911
1912         if (result) {
1913                 sense_valid = scsi_command_normalize_sense(SCpnt, &sshdr);
1914                 if (sense_valid)
1915                         sense_deferred = scsi_sense_is_deferred(&sshdr);
1916         }
1917         sdkp->medium_access_timed_out = 0;
1918
1919         if (driver_byte(result) != DRIVER_SENSE &&
1920             (!sense_valid || sense_deferred))
1921                 goto out;
1922
1923         switch (sshdr.sense_key) {
1924         case HARDWARE_ERROR:
1925         case MEDIUM_ERROR:
1926                 good_bytes = sd_completed_bytes(SCpnt);
1927                 break;
1928         case RECOVERED_ERROR:
1929                 good_bytes = scsi_bufflen(SCpnt);
1930                 break;
1931         case NO_SENSE:
1932                 /* This indicates a false check condition, so ignore it.  An
1933                  * unknown amount of data was transferred so treat it as an
1934                  * error.
1935                  */
1936                 SCpnt->result = 0;
1937                 memset(SCpnt->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
1938                 break;
1939         case ABORTED_COMMAND:
1940                 if (sshdr.asc == 0x10)  /* DIF: Target detected corruption */
1941                         good_bytes = sd_completed_bytes(SCpnt);
1942                 break;
1943         case ILLEGAL_REQUEST:
1944                 if (sshdr.asc == 0x10)  /* DIX: Host detected corruption */
1945                         good_bytes = sd_completed_bytes(SCpnt);
1946                 /* INVALID COMMAND OPCODE or INVALID FIELD IN CDB */
1947                 if (sshdr.asc == 0x20 || sshdr.asc == 0x24) {
1948                         switch (op) {
1949                         case UNMAP:
1950                                 sd_config_discard(sdkp, SD_LBP_DISABLE);
1951                                 break;
1952                         case WRITE_SAME_16:
1953                         case WRITE_SAME:
1954                                 if (unmap)
1955                                         sd_config_discard(sdkp, SD_LBP_DISABLE);
1956                                 else {
1957                                         sdkp->device->no_write_same = 1;
1958                                         sd_config_write_same(sdkp);
1959
1960                                         good_bytes = 0;
1961                                         req->__data_len = blk_rq_bytes(req);
1962                                         req->rq_flags |= RQF_QUIET;
1963                                 }
1964                         }
1965                 }
1966                 break;
1967         default:
1968                 break;
1969         }
1970
1971  out:
1972         if (sd_is_zoned(sdkp))
1973                 sd_zbc_complete(SCpnt, good_bytes, &sshdr);
1974
1975         SCSI_LOG_HLCOMPLETE(1, scmd_printk(KERN_INFO, SCpnt,
1976                                            "sd_done: completed %d of %d bytes\n",
1977                                            good_bytes, scsi_bufflen(SCpnt)));
1978
1979         if (rq_data_dir(SCpnt->request) == READ && scsi_prot_sg_count(SCpnt))
1980                 sd_dif_complete(SCpnt, good_bytes);
1981
1982         return good_bytes;
1983 }
1984
1985 /*
1986  * spinup disk - called only in sd_revalidate_disk()
1987  */
1988 static void
1989 sd_spinup_disk(struct scsi_disk *sdkp)
1990 {
1991         unsigned char cmd[10];
1992         unsigned long spintime_expire = 0;
1993         int retries, spintime;
1994         unsigned int the_result;
1995         struct scsi_sense_hdr sshdr;
1996         int sense_valid = 0;
1997
1998         spintime = 0;
1999
2000         /* Spin up drives, as required.  Only do this at boot time */
2001         /* Spinup needs to be done for module loads too. */
2002         do {
2003                 retries = 0;
2004
2005                 do {
2006                         cmd[0] = TEST_UNIT_READY;
2007                         memset((void *) &cmd[1], 0, 9);
2008
2009                         the_result = scsi_execute_req(sdkp->device, cmd,
2010                                                       DMA_NONE, NULL, 0,
2011                                                       &sshdr, SD_TIMEOUT,
2012                                                       SD_MAX_RETRIES, NULL);
2013
2014                         /*
2015                          * If the drive has indicated to us that it
2016                          * doesn't have any media in it, don't bother
2017                          * with any more polling.
2018                          */
2019                         if (media_not_present(sdkp, &sshdr))
2020                                 return;
2021
2022                         if (the_result)
2023                                 sense_valid = scsi_sense_valid(&sshdr);
2024                         retries++;
2025                 } while (retries < 3 &&
2026                          (!scsi_status_is_good(the_result) ||
2027                           ((driver_byte(the_result) & DRIVER_SENSE) &&
2028                           sense_valid && sshdr.sense_key == UNIT_ATTENTION)));
2029
2030                 if ((driver_byte(the_result) & DRIVER_SENSE) == 0) {
2031                         /* no sense, TUR either succeeded or failed
2032                          * with a status error */
2033                         if(!spintime && !scsi_status_is_good(the_result)) {
2034                                 sd_print_result(sdkp, "Test Unit Ready failed",
2035                                                 the_result);
2036                         }
2037                         break;
2038                 }
2039
2040                 /*
2041                  * The device does not want the automatic start to be issued.
2042                  */
2043                 if (sdkp->device->no_start_on_add)
2044                         break;
2045
2046                 if (sense_valid && sshdr.sense_key == NOT_READY) {
2047                         if (sshdr.asc == 4 && sshdr.ascq == 3)
2048                                 break;  /* manual intervention required */
2049                         if (sshdr.asc == 4 && sshdr.ascq == 0xb)
2050                                 break;  /* standby */
2051                         if (sshdr.asc == 4 && sshdr.ascq == 0xc)
2052                                 break;  /* unavailable */
2053                         /*
2054                          * Issue command to spin up drive when not ready
2055                          */
2056                         if (!spintime) {
2057                                 sd_printk(KERN_NOTICE, sdkp, "Spinning up disk...");
2058                                 cmd[0] = START_STOP;
2059                                 cmd[1] = 1;     /* Return immediately */
2060                                 memset((void *) &cmd[2], 0, 8);
2061                                 cmd[4] = 1;     /* Start spin cycle */
2062                                 if (sdkp->device->start_stop_pwr_cond)
2063                                         cmd[4] |= 1 << 4;
2064                                 scsi_execute_req(sdkp->device, cmd, DMA_NONE,
2065                                                  NULL, 0, &sshdr,
2066                                                  SD_TIMEOUT, SD_MAX_RETRIES,
2067                                                  NULL);
2068                                 spintime_expire = jiffies + 100 * HZ;
2069                                 spintime = 1;
2070                         }
2071                         /* Wait 1 second for next try */
2072                         msleep(1000);
2073                         printk(".");
2074
2075                 /*
2076                  * Wait for USB flash devices with slow firmware.
2077                  * Yes, this sense key/ASC combination shouldn't
2078                  * occur here.  It's characteristic of these devices.
2079                  */
2080                 } else if (sense_valid &&
2081                                 sshdr.sense_key == UNIT_ATTENTION &&
2082                                 sshdr.asc == 0x28) {
2083                         if (!spintime) {
2084                                 spintime_expire = jiffies + 5 * HZ;
2085                                 spintime = 1;
2086                         }
2087                         /* Wait 1 second for next try */
2088                         msleep(1000);
2089                 } else {
2090                         /* we don't understand the sense code, so it's
2091                          * probably pointless to loop */
2092                         if(!spintime) {
2093                                 sd_printk(KERN_NOTICE, sdkp, "Unit Not Ready\n");
2094                                 sd_print_sense_hdr(sdkp, &sshdr);
2095                         }
2096                         break;
2097                 }
2098
2099         } while (spintime && time_before_eq(jiffies, spintime_expire));
2100
2101         if (spintime) {
2102                 if (scsi_status_is_good(the_result))
2103                         printk("ready\n");
2104                 else
2105                         printk("not responding...\n");
2106         }
2107 }
2108
2109 /*
2110  * Determine whether disk supports Data Integrity Field.
2111  */
2112 static int sd_read_protection_type(struct scsi_disk *sdkp, unsigned char *buffer)
2113 {
2114         struct scsi_device *sdp = sdkp->device;
2115         u8 type;
2116         int ret = 0;
2117
2118         if (scsi_device_protection(sdp) == 0 || (buffer[12] & 1) == 0)
2119                 return ret;
2120
2121         type = ((buffer[12] >> 1) & 7) + 1; /* P_TYPE 0 = Type 1 */
2122
2123         if (type > T10_PI_TYPE3_PROTECTION)
2124                 ret = -ENODEV;
2125         else if (scsi_host_dif_capable(sdp->host, type))
2126                 ret = 1;
2127
2128         if (sdkp->first_scan || type != sdkp->protection_type)
2129                 switch (ret) {
2130                 case -ENODEV:
2131                         sd_printk(KERN_ERR, sdkp, "formatted with unsupported" \
2132                                   " protection type %u. Disabling disk!\n",
2133                                   type);
2134                         break;
2135                 case 1:
2136                         sd_printk(KERN_NOTICE, sdkp,
2137                                   "Enabling DIF Type %u protection\n", type);
2138                         break;
2139                 case 0:
2140                         sd_printk(KERN_NOTICE, sdkp,
2141                                   "Disabling DIF Type %u protection\n", type);
2142                         break;
2143                 }
2144
2145         sdkp->protection_type = type;
2146
2147         return ret;
2148 }
2149
2150 static void read_capacity_error(struct scsi_disk *sdkp, struct scsi_device *sdp,
2151                         struct scsi_sense_hdr *sshdr, int sense_valid,
2152                         int the_result)
2153 {
2154         if (driver_byte(the_result) & DRIVER_SENSE)
2155                 sd_print_sense_hdr(sdkp, sshdr);
2156         else
2157                 sd_printk(KERN_NOTICE, sdkp, "Sense not available.\n");
2158
2159         /*
2160          * Set dirty bit for removable devices if not ready -
2161          * sometimes drives will not report this properly.
2162          */
2163         if (sdp->removable &&
2164             sense_valid && sshdr->sense_key == NOT_READY)
2165                 set_media_not_present(sdkp);
2166
2167         /*
2168          * We used to set media_present to 0 here to indicate no media
2169          * in the drive, but some drives fail read capacity even with
2170          * media present, so we can't do that.
2171          */
2172         sdkp->capacity = 0; /* unknown mapped to zero - as usual */
2173 }
2174
2175 #define RC16_LEN 32
2176 #if RC16_LEN > SD_BUF_SIZE
2177 #error RC16_LEN must not be more than SD_BUF_SIZE
2178 #endif
2179
2180 #define READ_CAPACITY_RETRIES_ON_RESET  10
2181
2182 static int read_capacity_16(struct scsi_disk *sdkp, struct scsi_device *sdp,
2183                                                 unsigned char *buffer)
2184 {
2185         unsigned char cmd[16];
2186         struct scsi_sense_hdr sshdr;
2187         int sense_valid = 0;
2188         int the_result;
2189         int retries = 3, reset_retries = READ_CAPACITY_RETRIES_ON_RESET;
2190         unsigned int alignment;
2191         unsigned long long lba;
2192         unsigned sector_size;
2193
2194         if (sdp->no_read_capacity_16)
2195                 return -EINVAL;
2196
2197         do {
2198                 memset(cmd, 0, 16);
2199                 cmd[0] = SERVICE_ACTION_IN_16;
2200                 cmd[1] = SAI_READ_CAPACITY_16;
2201                 cmd[13] = RC16_LEN;
2202                 memset(buffer, 0, RC16_LEN);
2203
2204                 the_result = scsi_execute_req(sdp, cmd, DMA_FROM_DEVICE,
2205                                         buffer, RC16_LEN, &sshdr,
2206                                         SD_TIMEOUT, SD_MAX_RETRIES, NULL);
2207
2208                 if (media_not_present(sdkp, &sshdr))
2209                         return -ENODEV;
2210
2211                 if (the_result) {
2212                         sense_valid = scsi_sense_valid(&sshdr);
2213                         if (sense_valid &&
2214                             sshdr.sense_key == ILLEGAL_REQUEST &&
2215                             (sshdr.asc == 0x20 || sshdr.asc == 0x24) &&
2216                             sshdr.ascq == 0x00)
2217                                 /* Invalid Command Operation Code or
2218                                  * Invalid Field in CDB, just retry
2219                                  * silently with RC10 */
2220                                 return -EINVAL;
2221                         if (sense_valid &&
2222                             sshdr.sense_key == UNIT_ATTENTION &&
2223                             sshdr.asc == 0x29 && sshdr.ascq == 0x00)
2224                                 /* Device reset might occur several times,
2225                                  * give it one more chance */
2226                                 if (--reset_retries > 0)
2227                                         continue;
2228                 }
2229                 retries--;
2230
2231         } while (the_result && retries);
2232
2233         if (the_result) {
2234                 sd_print_result(sdkp, "Read Capacity(16) failed", the_result);
2235                 read_capacity_error(sdkp, sdp, &sshdr, sense_valid, the_result);
2236                 return -EINVAL;
2237         }
2238
2239         sector_size = get_unaligned_be32(&buffer[8]);
2240         lba = get_unaligned_be64(&buffer[0]);
2241
2242         if (sd_read_protection_type(sdkp, buffer) < 0) {
2243                 sdkp->capacity = 0;
2244                 return -ENODEV;
2245         }
2246
2247         if ((sizeof(sdkp->capacity) == 4) && (lba >= 0xffffffffULL)) {
2248                 sd_printk(KERN_ERR, sdkp, "Too big for this kernel. Use a "
2249                         "kernel compiled with support for large block "
2250                         "devices.\n");
2251                 sdkp->capacity = 0;
2252                 return -EOVERFLOW;
2253         }
2254
2255         /* Logical blocks per physical block exponent */
2256         sdkp->physical_block_size = (1 << (buffer[13] & 0xf)) * sector_size;
2257
2258         /* RC basis */
2259         sdkp->rc_basis = (buffer[12] >> 4) & 0x3;
2260
2261         /* Lowest aligned logical block */
2262         alignment = ((buffer[14] & 0x3f) << 8 | buffer[15]) * sector_size;
2263         blk_queue_alignment_offset(sdp->request_queue, alignment);
2264         if (alignment && sdkp->first_scan)
2265                 sd_printk(KERN_NOTICE, sdkp,
2266                           "physical block alignment offset: %u\n", alignment);
2267
2268         if (buffer[14] & 0x80) { /* LBPME */
2269                 sdkp->lbpme = 1;
2270
2271                 if (buffer[14] & 0x40) /* LBPRZ */
2272                         sdkp->lbprz = 1;
2273
2274                 sd_config_discard(sdkp, SD_LBP_WS16);
2275         }
2276
2277         sdkp->capacity = lba + 1;
2278         return sector_size;
2279 }
2280
2281 static int read_capacity_10(struct scsi_disk *sdkp, struct scsi_device *sdp,
2282                                                 unsigned char *buffer)
2283 {
2284         unsigned char cmd[16];
2285         struct scsi_sense_hdr sshdr;
2286         int sense_valid = 0;
2287         int the_result;
2288         int retries = 3, reset_retries = READ_CAPACITY_RETRIES_ON_RESET;
2289         sector_t lba;
2290         unsigned sector_size;
2291
2292         do {
2293                 cmd[0] = READ_CAPACITY;
2294                 memset(&cmd[1], 0, 9);
2295                 memset(buffer, 0, 8);
2296
2297                 the_result = scsi_execute_req(sdp, cmd, DMA_FROM_DEVICE,
2298                                         buffer, 8, &sshdr,
2299                                         SD_TIMEOUT, SD_MAX_RETRIES, NULL);
2300
2301                 if (media_not_present(sdkp, &sshdr))
2302                         return -ENODEV;
2303
2304                 if (the_result) {
2305                         sense_valid = scsi_sense_valid(&sshdr);
2306                         if (sense_valid &&
2307                             sshdr.sense_key == UNIT_ATTENTION &&
2308                             sshdr.asc == 0x29 && sshdr.ascq == 0x00)
2309                                 /* Device reset might occur several times,
2310                                  * give it one more chance */
2311                                 if (--reset_retries > 0)
2312                                         continue;
2313                 }
2314                 retries--;
2315
2316         } while (the_result && retries);
2317
2318         if (the_result) {
2319                 sd_print_result(sdkp, "Read Capacity(10) failed", the_result);
2320                 read_capacity_error(sdkp, sdp, &sshdr, sense_valid, the_result);
2321                 return -EINVAL;
2322         }
2323
2324         sector_size = get_unaligned_be32(&buffer[4]);
2325         lba = get_unaligned_be32(&buffer[0]);
2326
2327         if (sdp->no_read_capacity_16 && (lba == 0xffffffff)) {
2328                 /* Some buggy (usb cardreader) devices return an lba of
2329                    0xffffffff when the want to report a size of 0 (with
2330                    which they really mean no media is present) */
2331                 sdkp->capacity = 0;
2332                 sdkp->physical_block_size = sector_size;
2333                 return sector_size;
2334         }
2335
2336         if ((sizeof(sdkp->capacity) == 4) && (lba == 0xffffffff)) {
2337                 sd_printk(KERN_ERR, sdkp, "Too big for this kernel. Use a "
2338                         "kernel compiled with support for large block "
2339                         "devices.\n");
2340                 sdkp->capacity = 0;
2341                 return -EOVERFLOW;
2342         }
2343
2344         sdkp->capacity = lba + 1;
2345         sdkp->physical_block_size = sector_size;
2346         return sector_size;
2347 }
2348
2349 static int sd_try_rc16_first(struct scsi_device *sdp)
2350 {
2351         if (sdp->host->max_cmd_len < 16)
2352                 return 0;
2353         if (sdp->try_rc_10_first)
2354                 return 0;
2355         if (sdp->scsi_level > SCSI_SPC_2)
2356                 return 1;
2357         if (scsi_device_protection(sdp))
2358                 return 1;
2359         return 0;
2360 }
2361
2362 /*
2363  * read disk capacity
2364  */
2365 static void
2366 sd_read_capacity(struct scsi_disk *sdkp, unsigned char *buffer)
2367 {
2368         int sector_size;
2369         struct scsi_device *sdp = sdkp->device;
2370
2371         if (sd_try_rc16_first(sdp)) {
2372                 sector_size = read_capacity_16(sdkp, sdp, buffer);
2373                 if (sector_size == -EOVERFLOW)
2374                         goto got_data;
2375                 if (sector_size == -ENODEV)
2376                         return;
2377                 if (sector_size < 0)
2378                         sector_size = read_capacity_10(sdkp, sdp, buffer);
2379                 if (sector_size < 0)
2380                         return;
2381         } else {
2382                 sector_size = read_capacity_10(sdkp, sdp, buffer);
2383                 if (sector_size == -EOVERFLOW)
2384                         goto got_data;
2385                 if (sector_size < 0)
2386                         return;
2387                 if ((sizeof(sdkp->capacity) > 4) &&
2388                     (sdkp->capacity > 0xffffffffULL)) {
2389                         int old_sector_size = sector_size;
2390                         sd_printk(KERN_NOTICE, sdkp, "Very big device. "
2391                                         "Trying to use READ CAPACITY(16).\n");
2392                         sector_size = read_capacity_16(sdkp, sdp, buffer);
2393                         if (sector_size < 0) {
2394                                 sd_printk(KERN_NOTICE, sdkp,
2395                                         "Using 0xffffffff as device size\n");
2396                                 sdkp->capacity = 1 + (sector_t) 0xffffffff;
2397                                 sector_size = old_sector_size;
2398                                 goto got_data;
2399                         }
2400                 }
2401         }
2402
2403         /* Some devices are known to return the total number of blocks,
2404          * not the highest block number.  Some devices have versions
2405          * which do this and others which do not.  Some devices we might
2406          * suspect of doing this but we don't know for certain.
2407          *
2408          * If we know the reported capacity is wrong, decrement it.  If
2409          * we can only guess, then assume the number of blocks is even
2410          * (usually true but not always) and err on the side of lowering
2411          * the capacity.
2412          */
2413         if (sdp->fix_capacity ||
2414             (sdp->guess_capacity && (sdkp->capacity & 0x01))) {
2415                 sd_printk(KERN_INFO, sdkp, "Adjusting the sector count "
2416                                 "from its reported value: %llu\n",
2417                                 (unsigned long long) sdkp->capacity);
2418                 --sdkp->capacity;
2419         }
2420
2421 got_data:
2422         if (sector_size == 0) {
2423                 sector_size = 512;
2424                 sd_printk(KERN_NOTICE, sdkp, "Sector size 0 reported, "
2425                           "assuming 512.\n");
2426         }
2427
2428         if (sector_size != 512 &&
2429             sector_size != 1024 &&
2430             sector_size != 2048 &&
2431             sector_size != 4096) {
2432                 sd_printk(KERN_NOTICE, sdkp, "Unsupported sector size %d.\n",
2433                           sector_size);
2434                 /*
2435                  * The user might want to re-format the drive with
2436                  * a supported sectorsize.  Once this happens, it
2437                  * would be relatively trivial to set the thing up.
2438                  * For this reason, we leave the thing in the table.
2439                  */
2440                 sdkp->capacity = 0;
2441                 /*
2442                  * set a bogus sector size so the normal read/write
2443                  * logic in the block layer will eventually refuse any
2444                  * request on this device without tripping over power
2445                  * of two sector size assumptions
2446                  */
2447                 sector_size = 512;
2448         }
2449         blk_queue_logical_block_size(sdp->request_queue, sector_size);
2450         blk_queue_physical_block_size(sdp->request_queue,
2451                                       sdkp->physical_block_size);
2452         sdkp->device->sector_size = sector_size;
2453
2454         if (sdkp->capacity > 0xffffffff)
2455                 sdp->use_16_for_rw = 1;
2456
2457 }
2458
2459 /*
2460  * Print disk capacity
2461  */
2462 static void
2463 sd_print_capacity(struct scsi_disk *sdkp,
2464                   sector_t old_capacity)
2465 {
2466         int sector_size = sdkp->device->sector_size;
2467         char cap_str_2[10], cap_str_10[10];
2468
2469         string_get_size(sdkp->capacity, sector_size,
2470                         STRING_UNITS_2, cap_str_2, sizeof(cap_str_2));
2471         string_get_size(sdkp->capacity, sector_size,
2472                         STRING_UNITS_10, cap_str_10,
2473                         sizeof(cap_str_10));
2474
2475         if (sdkp->first_scan || old_capacity != sdkp->capacity) {
2476                 sd_printk(KERN_NOTICE, sdkp,
2477                           "%llu %d-byte logical blocks: (%s/%s)\n",
2478                           (unsigned long long)sdkp->capacity,
2479                           sector_size, cap_str_10, cap_str_2);
2480
2481                 if (sdkp->physical_block_size != sector_size)
2482                         sd_printk(KERN_NOTICE, sdkp,
2483                                   "%u-byte physical blocks\n",
2484                                   sdkp->physical_block_size);
2485
2486                 sd_zbc_print_zones(sdkp);
2487         }
2488 }
2489
2490 /* called with buffer of length 512 */
2491 static inline int
2492 sd_do_mode_sense(struct scsi_device *sdp, int dbd, int modepage,
2493                  unsigned char *buffer, int len, struct scsi_mode_data *data,
2494                  struct scsi_sense_hdr *sshdr)
2495 {
2496         return scsi_mode_sense(sdp, dbd, modepage, buffer, len,
2497                                SD_TIMEOUT, SD_MAX_RETRIES, data,
2498                                sshdr);
2499 }
2500
2501 /*
2502  * read write protect setting, if possible - called only in sd_revalidate_disk()
2503  * called with buffer of length SD_BUF_SIZE
2504  */
2505 static void
2506 sd_read_write_protect_flag(struct scsi_disk *sdkp, unsigned char *buffer)
2507 {
2508         int res;
2509         struct scsi_device *sdp = sdkp->device;
2510         struct scsi_mode_data data;
2511         int old_wp = sdkp->write_prot;
2512
2513         set_disk_ro(sdkp->disk, 0);
2514         if (sdp->skip_ms_page_3f) {
2515                 sd_first_printk(KERN_NOTICE, sdkp, "Assuming Write Enabled\n");
2516                 return;
2517         }
2518
2519         if (sdp->use_192_bytes_for_3f) {
2520                 res = sd_do_mode_sense(sdp, 0, 0x3F, buffer, 192, &data, NULL);
2521         } else {
2522                 /*
2523                  * First attempt: ask for all pages (0x3F), but only 4 bytes.
2524                  * We have to start carefully: some devices hang if we ask
2525                  * for more than is available.
2526                  */
2527                 res = sd_do_mode_sense(sdp, 0, 0x3F, buffer, 4, &data, NULL);
2528
2529                 /*
2530                  * Second attempt: ask for page 0 When only page 0 is
2531                  * implemented, a request for page 3F may return Sense Key
2532                  * 5: Illegal Request, Sense Code 24: Invalid field in
2533                  * CDB.
2534                  */
2535                 if (!scsi_status_is_good(res))
2536                         res = sd_do_mode_sense(sdp, 0, 0, buffer, 4, &data, NULL);
2537
2538                 /*
2539                  * Third attempt: ask 255 bytes, as we did earlier.
2540                  */
2541                 if (!scsi_status_is_good(res))
2542                         res = sd_do_mode_sense(sdp, 0, 0x3F, buffer, 255,
2543                                                &data, NULL);
2544         }
2545
2546         if (!scsi_status_is_good(res)) {
2547                 sd_first_printk(KERN_WARNING, sdkp,
2548                           "Test WP failed, assume Write Enabled\n");
2549         } else {
2550                 sdkp->write_prot = ((data.device_specific & 0x80) != 0);
2551                 set_disk_ro(sdkp->disk, sdkp->write_prot);
2552                 if (sdkp->first_scan || old_wp != sdkp->write_prot) {
2553                         sd_printk(KERN_NOTICE, sdkp, "Write Protect is %s\n",
2554                                   sdkp->write_prot ? "on" : "off");
2555                         sd_printk(KERN_DEBUG, sdkp, "Mode Sense: %4ph\n", buffer);
2556                 }
2557         }
2558 }
2559
2560 /*
2561  * sd_read_cache_type - called only from sd_revalidate_disk()
2562  * called with buffer of length SD_BUF_SIZE
2563  */
2564 static void
2565 sd_read_cache_type(struct scsi_disk *sdkp, unsigned char *buffer)
2566 {
2567         int len = 0, res;
2568         struct scsi_device *sdp = sdkp->device;
2569
2570         int dbd;
2571         int modepage;
2572         int first_len;
2573         struct scsi_mode_data data;
2574         struct scsi_sense_hdr sshdr;
2575         int old_wce = sdkp->WCE;
2576         int old_rcd = sdkp->RCD;
2577         int old_dpofua = sdkp->DPOFUA;
2578
2579
2580         if (sdkp->cache_override)
2581                 return;
2582
2583         first_len = 4;
2584         if (sdp->skip_ms_page_8) {
2585                 if (sdp->type == TYPE_RBC)
2586                         goto defaults;
2587                 else {
2588                         if (sdp->skip_ms_page_3f)
2589                                 goto defaults;
2590                         modepage = 0x3F;
2591                         if (sdp->use_192_bytes_for_3f)
2592                                 first_len = 192;
2593                         dbd = 0;
2594                 }
2595         } else if (sdp->type == TYPE_RBC) {
2596                 modepage = 6;
2597                 dbd = 8;
2598         } else {
2599                 modepage = 8;
2600                 dbd = 0;
2601         }
2602
2603         /* cautiously ask */
2604         res = sd_do_mode_sense(sdp, dbd, modepage, buffer, first_len,
2605                         &data, &sshdr);
2606
2607         if (!scsi_status_is_good(res))
2608                 goto bad_sense;
2609
2610         if (!data.header_length) {
2611                 modepage = 6;
2612                 first_len = 0;
2613                 sd_first_printk(KERN_ERR, sdkp,
2614                                 "Missing header in MODE_SENSE response\n");
2615         }
2616
2617         /* that went OK, now ask for the proper length */
2618         len = data.length;
2619
2620         /*
2621          * We're only interested in the first three bytes, actually.
2622          * But the data cache page is defined for the first 20.
2623          */
2624         if (len < 3)
2625                 goto bad_sense;
2626         else if (len > SD_BUF_SIZE) {
2627                 sd_first_printk(KERN_NOTICE, sdkp, "Truncating mode parameter "
2628                           "data from %d to %d bytes\n", len, SD_BUF_SIZE);
2629                 len = SD_BUF_SIZE;
2630         }
2631         if (modepage == 0x3F && sdp->use_192_bytes_for_3f)
2632                 len = 192;
2633
2634         /* Get the data */
2635         if (len > first_len)
2636                 res = sd_do_mode_sense(sdp, dbd, modepage, buffer, len,
2637                                 &data, &sshdr);
2638
2639         if (scsi_status_is_good(res)) {
2640                 int offset = data.header_length + data.block_descriptor_length;
2641
2642                 while (offset < len) {
2643                         u8 page_code = buffer[offset] & 0x3F;
2644                         u8 spf       = buffer[offset] & 0x40;
2645
2646                         if (page_code == 8 || page_code == 6) {
2647                                 /* We're interested only in the first 3 bytes.
2648                                  */
2649                                 if (len - offset <= 2) {
2650                                         sd_first_printk(KERN_ERR, sdkp,
2651                                                 "Incomplete mode parameter "
2652                                                         "data\n");
2653                                         goto defaults;
2654                                 } else {
2655                                         modepage = page_code;
2656                                         goto Page_found;
2657                                 }
2658                         } else {
2659                                 /* Go to the next page */
2660                                 if (spf && len - offset > 3)
2661                                         offset += 4 + (buffer[offset+2] << 8) +
2662                                                 buffer[offset+3];
2663                                 else if (!spf && len - offset > 1)
2664                                         offset += 2 + buffer[offset+1];
2665                                 else {
2666                                         sd_first_printk(KERN_ERR, sdkp,
2667                                                         "Incomplete mode "
2668                                                         "parameter data\n");
2669                                         goto defaults;
2670                                 }
2671                         }
2672                 }
2673
2674                 sd_first_printk(KERN_ERR, sdkp, "No Caching mode page found\n");
2675                 goto defaults;
2676
2677         Page_found:
2678                 if (modepage == 8) {
2679                         sdkp->WCE = ((buffer[offset + 2] & 0x04) != 0);
2680                         sdkp->RCD = ((buffer[offset + 2] & 0x01) != 0);
2681                 } else {
2682                         sdkp->WCE = ((buffer[offset + 2] & 0x01) == 0);
2683                         sdkp->RCD = 0;
2684                 }
2685
2686                 sdkp->DPOFUA = (data.device_specific & 0x10) != 0;
2687                 if (sdp->broken_fua) {
2688                         sd_first_printk(KERN_NOTICE, sdkp, "Disabling FUA\n");
2689                         sdkp->DPOFUA = 0;
2690                 } else if (sdkp->DPOFUA && !sdkp->device->use_10_for_rw &&
2691                            !sdkp->device->use_16_for_rw) {
2692                         sd_first_printk(KERN_NOTICE, sdkp,
2693                                   "Uses READ/WRITE(6), disabling FUA\n");
2694                         sdkp->DPOFUA = 0;
2695                 }
2696
2697                 /* No cache flush allowed for write protected devices */
2698                 if (sdkp->WCE && sdkp->write_prot)
2699                         sdkp->WCE = 0;
2700
2701                 if (sdkp->first_scan || old_wce != sdkp->WCE ||
2702                     old_rcd != sdkp->RCD || old_dpofua != sdkp->DPOFUA)
2703                         sd_printk(KERN_NOTICE, sdkp,
2704                                   "Write cache: %s, read cache: %s, %s\n",
2705                                   sdkp->WCE ? "enabled" : "disabled",
2706                                   sdkp->RCD ? "disabled" : "enabled",
2707                                   sdkp->DPOFUA ? "supports DPO and FUA"
2708                                   : "doesn't support DPO or FUA");
2709
2710                 return;
2711         }
2712
2713 bad_sense:
2714         if (scsi_sense_valid(&sshdr) &&
2715             sshdr.sense_key == ILLEGAL_REQUEST &&
2716             sshdr.asc == 0x24 && sshdr.ascq == 0x0)
2717                 /* Invalid field in CDB */
2718                 sd_first_printk(KERN_NOTICE, sdkp, "Cache data unavailable\n");
2719         else
2720                 sd_first_printk(KERN_ERR, sdkp,
2721                                 "Asking for cache data failed\n");
2722
2723 defaults:
2724         if (sdp->wce_default_on) {
2725                 sd_first_printk(KERN_NOTICE, sdkp,
2726                                 "Assuming drive cache: write back\n");
2727                 sdkp->WCE = 1;
2728         } else {
2729                 sd_first_printk(KERN_ERR, sdkp,
2730                                 "Assuming drive cache: write through\n");
2731                 sdkp->WCE = 0;
2732         }
2733         sdkp->RCD = 0;
2734         sdkp->DPOFUA = 0;
2735 }
2736
2737 /*
2738  * The ATO bit indicates whether the DIF application tag is available
2739  * for use by the operating system.
2740  */
2741 static void sd_read_app_tag_own(struct scsi_disk *sdkp, unsigned char *buffer)
2742 {
2743         int res, offset;
2744         struct scsi_device *sdp = sdkp->device;
2745         struct scsi_mode_data data;
2746         struct scsi_sense_hdr sshdr;
2747
2748         if (sdp->type != TYPE_DISK && sdp->type != TYPE_ZBC)
2749                 return;
2750
2751         if (sdkp->protection_type == 0)
2752                 return;
2753
2754         res = scsi_mode_sense(sdp, 1, 0x0a, buffer, 36, SD_TIMEOUT,
2755                               SD_MAX_RETRIES, &data, &sshdr);
2756
2757         if (!scsi_status_is_good(res) || !data.header_length ||
2758             data.length < 6) {
2759                 sd_first_printk(KERN_WARNING, sdkp,
2760                           "getting Control mode page failed, assume no ATO\n");
2761
2762                 if (scsi_sense_valid(&sshdr))
2763                         sd_print_sense_hdr(sdkp, &sshdr);
2764
2765                 return;
2766         }
2767
2768         offset = data.header_length + data.block_descriptor_length;
2769
2770         if ((buffer[offset] & 0x3f) != 0x0a) {
2771                 sd_first_printk(KERN_ERR, sdkp, "ATO Got wrong page\n");
2772                 return;
2773         }
2774
2775         if ((buffer[offset + 5] & 0x80) == 0)
2776                 return;
2777
2778         sdkp->ATO = 1;
2779
2780         return;
2781 }
2782
2783 /**
2784  * sd_read_block_limits - Query disk device for preferred I/O sizes.
2785  * @disk: disk to query
2786  */
2787 static void sd_read_block_limits(struct scsi_disk *sdkp)
2788 {
2789         unsigned int sector_sz = sdkp->device->sector_size;
2790         const int vpd_len = 64;
2791         unsigned char *buffer = kmalloc(vpd_len, GFP_KERNEL);
2792
2793         if (!buffer ||
2794             /* Block Limits VPD */
2795             scsi_get_vpd_page(sdkp->device, 0xb0, buffer, vpd_len))
2796                 goto out;
2797
2798         blk_queue_io_min(sdkp->disk->queue,
2799                          get_unaligned_be16(&buffer[6]) * sector_sz);
2800
2801         sdkp->max_xfer_blocks = get_unaligned_be32(&buffer[8]);
2802         sdkp->opt_xfer_blocks = get_unaligned_be32(&buffer[12]);
2803
2804         if (buffer[3] == 0x3c) {
2805                 unsigned int lba_count, desc_count;
2806
2807                 sdkp->max_ws_blocks = (u32)get_unaligned_be64(&buffer[36]);
2808
2809                 if (!sdkp->lbpme)
2810                         goto out;
2811
2812                 lba_count = get_unaligned_be32(&buffer[20]);
2813                 desc_count = get_unaligned_be32(&buffer[24]);
2814
2815                 if (lba_count && desc_count)
2816                         sdkp->max_unmap_blocks = lba_count;
2817
2818                 sdkp->unmap_granularity = get_unaligned_be32(&buffer[28]);
2819
2820                 if (buffer[32] & 0x80)
2821                         sdkp->unmap_alignment =
2822                                 get_unaligned_be32(&buffer[32]) & ~(1 << 31);
2823
2824                 if (!sdkp->lbpvpd) { /* LBP VPD page not provided */
2825
2826                         if (sdkp->max_unmap_blocks)
2827                                 sd_config_discard(sdkp, SD_LBP_UNMAP);
2828                         else
2829                                 sd_config_discard(sdkp, SD_LBP_WS16);
2830
2831                 } else {        /* LBP VPD page tells us what to use */
2832                         if (sdkp->lbpu && sdkp->max_unmap_blocks)
2833                                 sd_config_discard(sdkp, SD_LBP_UNMAP);
2834                         else if (sdkp->lbpws)
2835                                 sd_config_discard(sdkp, SD_LBP_WS16);
2836                         else if (sdkp->lbpws10)
2837                                 sd_config_discard(sdkp, SD_LBP_WS10);
2838                         else if (sdkp->lbpu && sdkp->max_unmap_blocks)
2839                                 sd_config_discard(sdkp, SD_LBP_UNMAP);
2840                         else
2841                                 sd_config_discard(sdkp, SD_LBP_DISABLE);
2842                 }
2843         }
2844
2845  out:
2846         kfree(buffer);
2847 }
2848
2849 /**
2850  * sd_read_block_characteristics - Query block dev. characteristics
2851  * @disk: disk to query
2852  */
2853 static void sd_read_block_characteristics(struct scsi_disk *sdkp)
2854 {
2855         struct request_queue *q = sdkp->disk->queue;
2856         unsigned char *buffer;
2857         u16 rot;
2858         const int vpd_len = 64;
2859
2860         buffer = kmalloc(vpd_len, GFP_KERNEL);
2861
2862         if (!buffer ||
2863             /* Block Device Characteristics VPD */
2864             scsi_get_vpd_page(sdkp->device, 0xb1, buffer, vpd_len))
2865                 goto out;
2866
2867         rot = get_unaligned_be16(&buffer[4]);
2868
2869         if (rot == 1) {
2870                 queue_flag_set_unlocked(QUEUE_FLAG_NONROT, q);
2871                 queue_flag_clear_unlocked(QUEUE_FLAG_ADD_RANDOM, q);
2872         }
2873
2874         if (sdkp->device->type == TYPE_ZBC) {
2875                 /* Host-managed */
2876                 q->limits.zoned = BLK_ZONED_HM;
2877         } else {
2878                 sdkp->zoned = (buffer[8] >> 4) & 3;
2879                 if (sdkp->zoned == 1)
2880                         /* Host-aware */
2881                         q->limits.zoned = BLK_ZONED_HA;
2882                 else
2883                         /*
2884                          * Treat drive-managed devices as
2885                          * regular block devices.
2886                          */
2887                         q->limits.zoned = BLK_ZONED_NONE;
2888         }
2889         if (blk_queue_is_zoned(q) && sdkp->first_scan)
2890                 sd_printk(KERN_NOTICE, sdkp, "Host-%s zoned block device\n",
2891                       q->limits.zoned == BLK_ZONED_HM ? "managed" : "aware");
2892
2893  out:
2894         kfree(buffer);
2895 }
2896
2897 /**
2898  * sd_read_block_provisioning - Query provisioning VPD page
2899  * @disk: disk to query
2900  */
2901 static void sd_read_block_provisioning(struct scsi_disk *sdkp)
2902 {
2903         unsigned char *buffer;
2904         const int vpd_len = 8;
2905
2906         if (sdkp->lbpme == 0)
2907                 return;
2908
2909         buffer = kmalloc(vpd_len, GFP_KERNEL);
2910
2911         if (!buffer || scsi_get_vpd_page(sdkp->device, 0xb2, buffer, vpd_len))
2912                 goto out;
2913
2914         sdkp->lbpvpd    = 1;
2915         sdkp->lbpu      = (buffer[5] >> 7) & 1; /* UNMAP */
2916         sdkp->lbpws     = (buffer[5] >> 6) & 1; /* WRITE SAME(16) with UNMAP */
2917         sdkp->lbpws10   = (buffer[5] >> 5) & 1; /* WRITE SAME(10) with UNMAP */
2918
2919  out:
2920         kfree(buffer);
2921 }
2922
2923 static void sd_read_write_same(struct scsi_disk *sdkp, unsigned char *buffer)
2924 {
2925         struct scsi_device *sdev = sdkp->device;
2926
2927         if (sdev->host->no_write_same) {
2928                 sdev->no_write_same = 1;
2929
2930                 return;
2931         }
2932
2933         if (scsi_report_opcode(sdev, buffer, SD_BUF_SIZE, INQUIRY) < 0) {
2934                 /* too large values might cause issues with arcmsr */
2935                 int vpd_buf_len = 64;
2936
2937                 sdev->no_report_opcodes = 1;
2938
2939                 /* Disable WRITE SAME if REPORT SUPPORTED OPERATION
2940                  * CODES is unsupported and the device has an ATA
2941                  * Information VPD page (SAT).
2942                  */
2943                 if (!scsi_get_vpd_page(sdev, 0x89, buffer, vpd_buf_len))
2944                         sdev->no_write_same = 1;
2945         }
2946
2947         if (scsi_report_opcode(sdev, buffer, SD_BUF_SIZE, WRITE_SAME_16) == 1)
2948                 sdkp->ws16 = 1;
2949
2950         if (scsi_report_opcode(sdev, buffer, SD_BUF_SIZE, WRITE_SAME) == 1)
2951                 sdkp->ws10 = 1;
2952 }
2953
2954 /**
2955  *      sd_revalidate_disk - called the first time a new disk is seen,
2956  *      performs disk spin up, read_capacity, etc.
2957  *      @disk: struct gendisk we care about
2958  **/
2959 static int sd_revalidate_disk(struct gendisk *disk)
2960 {
2961         struct scsi_disk *sdkp = scsi_disk(disk);
2962         struct scsi_device *sdp = sdkp->device;
2963         struct request_queue *q = sdkp->disk->queue;
2964         sector_t old_capacity = sdkp->capacity;
2965         unsigned char *buffer;
2966         unsigned int dev_max, rw_max;
2967
2968         SCSI_LOG_HLQUEUE(3, sd_printk(KERN_INFO, sdkp,
2969                                       "sd_revalidate_disk\n"));
2970
2971         /*
2972          * If the device is offline, don't try and read capacity or any
2973          * of the other niceties.
2974          */
2975         if (!scsi_device_online(sdp))
2976                 goto out;
2977
2978         buffer = kmalloc(SD_BUF_SIZE, GFP_KERNEL);
2979         if (!buffer) {
2980                 sd_printk(KERN_WARNING, sdkp, "sd_revalidate_disk: Memory "
2981                           "allocation failure.\n");
2982                 goto out;
2983         }
2984
2985         sd_spinup_disk(sdkp);
2986
2987         /*
2988          * Without media there is no reason to ask; moreover, some devices
2989          * react badly if we do.
2990          */
2991         if (sdkp->media_present) {
2992                 sd_read_capacity(sdkp, buffer);
2993
2994                 if (scsi_device_supports_vpd(sdp)) {
2995                         sd_read_block_provisioning(sdkp);
2996                         sd_read_block_limits(sdkp);
2997                         sd_read_block_characteristics(sdkp);
2998                         sd_zbc_read_zones(sdkp, buffer);
2999                 }
3000
3001                 sd_print_capacity(sdkp, old_capacity);
3002
3003                 sd_read_write_protect_flag(sdkp, buffer);
3004                 sd_read_cache_type(sdkp, buffer);
3005                 sd_read_app_tag_own(sdkp, buffer);
3006                 sd_read_write_same(sdkp, buffer);
3007         }
3008
3009         sdkp->first_scan = 0;
3010
3011         /*
3012          * We now have all cache related info, determine how we deal
3013          * with flush requests.
3014          */
3015         sd_set_flush_flag(sdkp);
3016
3017         /* Initial block count limit based on CDB TRANSFER LENGTH field size. */
3018         dev_max = sdp->use_16_for_rw ? SD_MAX_XFER_BLOCKS : SD_DEF_XFER_BLOCKS;
3019
3020         /* Some devices report a maximum block count for READ/WRITE requests. */
3021         dev_max = min_not_zero(dev_max, sdkp->max_xfer_blocks);
3022         q->limits.max_dev_sectors = logical_to_sectors(sdp, dev_max);
3023
3024         /*
3025          * Use the device's preferred I/O size for reads and writes
3026          * unless the reported value is unreasonably small, large, or
3027          * garbage.
3028          */
3029         if (sdkp->opt_xfer_blocks &&
3030             sdkp->opt_xfer_blocks <= dev_max &&
3031             sdkp->opt_xfer_blocks <= SD_DEF_XFER_BLOCKS &&
3032             logical_to_bytes(sdp, sdkp->opt_xfer_blocks) >= PAGE_SIZE) {
3033                 q->limits.io_opt = logical_to_bytes(sdp, sdkp->opt_xfer_blocks);
3034                 rw_max = logical_to_sectors(sdp, sdkp->opt_xfer_blocks);
3035         } else
3036                 rw_max = BLK_DEF_MAX_SECTORS;
3037
3038         /* Combine with controller limits */
3039         q->limits.max_sectors = min(rw_max, queue_max_hw_sectors(q));
3040
3041         set_capacity(disk, logical_to_sectors(sdp, sdkp->capacity));
3042         sd_config_write_same(sdkp);
3043         kfree(buffer);
3044
3045  out:
3046         return 0;
3047 }
3048
3049 /**
3050  *      sd_unlock_native_capacity - unlock native capacity
3051  *      @disk: struct gendisk to set capacity for
3052  *
3053  *      Block layer calls this function if it detects that partitions
3054  *      on @disk reach beyond the end of the device.  If the SCSI host
3055  *      implements ->unlock_native_capacity() method, it's invoked to
3056  *      give it a chance to adjust the device capacity.
3057  *
3058  *      CONTEXT:
3059  *      Defined by block layer.  Might sleep.
3060  */
3061 static void sd_unlock_native_capacity(struct gendisk *disk)
3062 {
3063         struct scsi_device *sdev = scsi_disk(disk)->device;
3064
3065         if (sdev->host->hostt->unlock_native_capacity)
3066                 sdev->host->hostt->unlock_native_capacity(sdev);
3067 }
3068
3069 /**
3070  *      sd_format_disk_name - format disk name
3071  *      @prefix: name prefix - ie. "sd" for SCSI disks
3072  *      @index: index of the disk to format name for
3073  *      @buf: output buffer
3074  *      @buflen: length of the output buffer
3075  *
3076  *      SCSI disk names starts at sda.  The 26th device is sdz and the
3077  *      27th is sdaa.  The last one for two lettered suffix is sdzz
3078  *      which is followed by sdaaa.
3079  *
3080  *      This is basically 26 base counting with one extra 'nil' entry
3081  *      at the beginning from the second digit on and can be
3082  *      determined using similar method as 26 base conversion with the
3083  *      index shifted -1 after each digit is computed.
3084  *
3085  *      CONTEXT:
3086  *      Don't care.
3087  *
3088  *      RETURNS:
3089  *      0 on success, -errno on failure.
3090  */
3091 static int sd_format_disk_name(char *prefix, int index, char *buf, int buflen)
3092 {
3093         const int base = 'z' - 'a' + 1;
3094         char *begin = buf + strlen(prefix);
3095         char *end = buf + buflen;
3096         char *p;
3097         int unit;
3098
3099         p = end - 1;
3100         *p = '\0';
3101         unit = base;
3102         do {
3103                 if (p == begin)
3104                         return -EINVAL;
3105                 *--p = 'a' + (index % unit);
3106                 index = (index / unit) - 1;
3107         } while (index >= 0);
3108
3109         memmove(begin, p, end - p);
3110         memcpy(buf, prefix, strlen(prefix));
3111
3112         return 0;
3113 }
3114
3115 /*
3116  * The asynchronous part of sd_probe
3117  */
3118 static void sd_probe_async(void *data, async_cookie_t cookie)
3119 {
3120         struct scsi_disk *sdkp = data;
3121         struct scsi_device *sdp;
3122         struct gendisk *gd;
3123         u32 index;
3124         struct device *dev;
3125
3126         sdp = sdkp->device;
3127         gd = sdkp->disk;
3128         index = sdkp->index;
3129         dev = &sdp->sdev_gendev;
3130
3131         gd->major = sd_major((index & 0xf0) >> 4);
3132         gd->first_minor = ((index & 0xf) << 4) | (index & 0xfff00);
3133         gd->minors = SD_MINORS;
3134
3135         gd->fops = &sd_fops;
3136         gd->private_data = &sdkp->driver;
3137         gd->queue = sdkp->device->request_queue;
3138
3139         /* defaults, until the device tells us otherwise */
3140         sdp->sector_size = 512;
3141         sdkp->capacity = 0;
3142         sdkp->media_present = 1;
3143         sdkp->write_prot = 0;
3144         sdkp->cache_override = 0;
3145         sdkp->WCE = 0;
3146         sdkp->RCD = 0;
3147         sdkp->ATO = 0;
3148         sdkp->first_scan = 1;
3149         sdkp->max_medium_access_timeouts = SD_MAX_MEDIUM_TIMEOUTS;
3150
3151         sd_revalidate_disk(gd);
3152
3153         gd->flags = GENHD_FL_EXT_DEVT;
3154         if (sdp->removable) {
3155                 gd->flags |= GENHD_FL_REMOVABLE;
3156                 gd->events |= DISK_EVENT_MEDIA_CHANGE;
3157         }
3158
3159         blk_pm_runtime_init(sdp->request_queue, dev);
3160         device_add_disk(dev, gd);
3161         if (sdkp->capacity)
3162                 sd_dif_config_host(sdkp);
3163
3164         sd_revalidate_disk(gd);
3165
3166         sd_printk(KERN_NOTICE, sdkp, "Attached SCSI %sdisk\n",
3167                   sdp->removable ? "removable " : "");
3168         scsi_autopm_put_device(sdp);
3169         put_device(&sdkp->dev);
3170 }
3171
3172 /**
3173  *      sd_probe - called during driver initialization and whenever a
3174  *      new scsi device is attached to the system. It is called once
3175  *      for each scsi device (not just disks) present.
3176  *      @dev: pointer to device object
3177  *
3178  *      Returns 0 if successful (or not interested in this scsi device
3179  *      (e.g. scanner)); 1 when there is an error.
3180  *
3181  *      Note: this function is invoked from the scsi mid-level.
3182  *      This function sets up the mapping between a given
3183  *      <host,channel,id,lun> (found in sdp) and new device name
3184  *      (e.g. /dev/sda). More precisely it is the block device major
3185  *      and minor number that is chosen here.
3186  *
3187  *      Assume sd_probe is not re-entrant (for time being)
3188  *      Also think about sd_probe() and sd_remove() running coincidentally.
3189  **/
3190 static int sd_probe(struct device *dev)
3191 {
3192         struct scsi_device *sdp = to_scsi_device(dev);
3193         struct scsi_disk *sdkp;
3194         struct gendisk *gd;
3195         int index;
3196         int error;
3197
3198         scsi_autopm_get_device(sdp);
3199         error = -ENODEV;
3200         if (sdp->type != TYPE_DISK &&
3201             sdp->type != TYPE_ZBC &&
3202             sdp->type != TYPE_MOD &&
3203             sdp->type != TYPE_RBC)
3204                 goto out;
3205
3206 #ifndef CONFIG_BLK_DEV_ZONED
3207         if (sdp->type == TYPE_ZBC)
3208                 goto out;
3209 #endif
3210         SCSI_LOG_HLQUEUE(3, sdev_printk(KERN_INFO, sdp,
3211                                         "sd_probe\n"));
3212
3213         error = -ENOMEM;
3214         sdkp = kzalloc(sizeof(*sdkp), GFP_KERNEL);
3215         if (!sdkp)
3216                 goto out;
3217
3218         gd = alloc_disk(SD_MINORS);
3219         if (!gd)
3220                 goto out_free;
3221
3222         do {
3223                 if (!ida_pre_get(&sd_index_ida, GFP_KERNEL))
3224                         goto out_put;
3225
3226                 spin_lock(&sd_index_lock);
3227                 error = ida_get_new(&sd_index_ida, &index);
3228                 spin_unlock(&sd_index_lock);
3229         } while (error == -EAGAIN);
3230
3231         if (error) {
3232                 sdev_printk(KERN_WARNING, sdp, "sd_probe: memory exhausted.\n");
3233                 goto out_put;
3234         }
3235
3236         error = sd_format_disk_name("sd", index, gd->disk_name, DISK_NAME_LEN);
3237         if (error) {
3238                 sdev_printk(KERN_WARNING, sdp, "SCSI disk (sd) name length exceeded.\n");
3239                 goto out_free_index;
3240         }
3241
3242         sdkp->device = sdp;
3243         sdkp->driver = &sd_template;
3244         sdkp->disk = gd;
3245         sdkp->index = index;
3246         atomic_set(&sdkp->openers, 0);
3247         atomic_set(&sdkp->device->ioerr_cnt, 0);
3248
3249         if (!sdp->request_queue->rq_timeout) {
3250                 if (sdp->type != TYPE_MOD)
3251                         blk_queue_rq_timeout(sdp->request_queue, SD_TIMEOUT);
3252                 else
3253                         blk_queue_rq_timeout(sdp->request_queue,
3254                                              SD_MOD_TIMEOUT);
3255         }
3256
3257         device_initialize(&sdkp->dev);
3258         sdkp->dev.parent = dev;
3259         sdkp->dev.class = &sd_disk_class;
3260         dev_set_name(&sdkp->dev, "%s", dev_name(dev));
3261
3262         error = device_add(&sdkp->dev);
3263         if (error)
3264                 goto out_free_index;
3265
3266         get_device(dev);
3267         dev_set_drvdata(dev, sdkp);
3268
3269         get_device(&sdkp->dev); /* prevent release before async_schedule */
3270         async_schedule_domain(sd_probe_async, sdkp, &scsi_sd_probe_domain);
3271
3272         return 0;
3273
3274  out_free_index:
3275         spin_lock(&sd_index_lock);
3276         ida_remove(&sd_index_ida, index);
3277         spin_unlock(&sd_index_lock);
3278  out_put:
3279         put_disk(gd);
3280  out_free:
3281         kfree(sdkp);
3282  out:
3283         scsi_autopm_put_device(sdp);
3284         return error;
3285 }
3286
3287 /**
3288  *      sd_remove - called whenever a scsi disk (previously recognized by
3289  *      sd_probe) is detached from the system. It is called (potentially
3290  *      multiple times) during sd module unload.
3291  *      @dev: pointer to device object
3292  *
3293  *      Note: this function is invoked from the scsi mid-level.
3294  *      This function potentially frees up a device name (e.g. /dev/sdc)
3295  *      that could be re-used by a subsequent sd_probe().
3296  *      This function is not called when the built-in sd driver is "exit-ed".
3297  **/
3298 static int sd_remove(struct device *dev)
3299 {
3300         struct scsi_disk *sdkp;
3301         dev_t devt;
3302
3303         sdkp = dev_get_drvdata(dev);
3304         devt = disk_devt(sdkp->disk);
3305         scsi_autopm_get_device(sdkp->device);
3306
3307         async_synchronize_full_domain(&scsi_sd_pm_domain);
3308         async_synchronize_full_domain(&scsi_sd_probe_domain);
3309         device_del(&sdkp->dev);
3310         del_gendisk(sdkp->disk);
3311         sd_shutdown(dev);
3312
3313         sd_zbc_remove(sdkp);
3314
3315         blk_register_region(devt, SD_MINORS, NULL,
3316                             sd_default_probe, NULL, NULL);
3317
3318         mutex_lock(&sd_ref_mutex);
3319         dev_set_drvdata(dev, NULL);
3320         put_device(&sdkp->dev);
3321         mutex_unlock(&sd_ref_mutex);
3322
3323         return 0;
3324 }
3325
3326 /**
3327  *      scsi_disk_release - Called to free the scsi_disk structure
3328  *      @dev: pointer to embedded class device
3329  *
3330  *      sd_ref_mutex must be held entering this routine.  Because it is
3331  *      called on last put, you should always use the scsi_disk_get()
3332  *      scsi_disk_put() helpers which manipulate the semaphore directly
3333  *      and never do a direct put_device.
3334  **/
3335 static void scsi_disk_release(struct device *dev)
3336 {
3337         struct scsi_disk *sdkp = to_scsi_disk(dev);
3338         struct gendisk *disk = sdkp->disk;
3339
3340         spin_lock(&sd_index_lock);
3341         ida_remove(&sd_index_ida, sdkp->index);
3342         spin_unlock(&sd_index_lock);
3343
3344         disk->private_data = NULL;
3345         put_disk(disk);
3346         put_device(&sdkp->device->sdev_gendev);
3347
3348         kfree(sdkp);
3349 }
3350
3351 static int sd_start_stop_device(struct scsi_disk *sdkp, int start)
3352 {
3353         unsigned char cmd[6] = { START_STOP };  /* START_VALID */
3354         struct scsi_sense_hdr sshdr;
3355         struct scsi_device *sdp = sdkp->device;
3356         int res;
3357
3358         if (start)
3359                 cmd[4] |= 1;    /* START */
3360
3361         if (sdp->start_stop_pwr_cond)
3362                 cmd[4] |= start ? 1 << 4 : 3 << 4;      /* Active or Standby */
3363
3364         if (!scsi_device_online(sdp))
3365                 return -ENODEV;
3366
3367         res = scsi_execute(sdp, cmd, DMA_NONE, NULL, 0, NULL, &sshdr,
3368                         SD_TIMEOUT, SD_MAX_RETRIES, 0, RQF_PM, NULL);
3369         if (res) {
3370                 sd_print_result(sdkp, "Start/Stop Unit failed", res);
3371                 if (driver_byte(res) & DRIVER_SENSE)
3372                         sd_print_sense_hdr(sdkp, &sshdr);
3373                 if (scsi_sense_valid(&sshdr) &&
3374                         /* 0x3a is medium not present */
3375                         sshdr.asc == 0x3a)
3376                         res = 0;
3377         }
3378
3379         /* SCSI error codes must not go to the generic layer */
3380         if (res)
3381                 return -EIO;
3382
3383         return 0;
3384 }
3385
3386 /*
3387  * Send a SYNCHRONIZE CACHE instruction down to the device through
3388  * the normal SCSI command structure.  Wait for the command to
3389  * complete.
3390  */
3391 static void sd_shutdown(struct device *dev)
3392 {
3393         struct scsi_disk *sdkp = dev_get_drvdata(dev);
3394
3395         if (!sdkp)
3396                 return;         /* this can happen */
3397
3398         if (pm_runtime_suspended(dev))
3399                 return;
3400
3401         if (sdkp->WCE && sdkp->media_present) {
3402                 sd_printk(KERN_NOTICE, sdkp, "Synchronizing SCSI cache\n");
3403                 sd_sync_cache(sdkp);
3404         }
3405
3406         if (system_state != SYSTEM_RESTART && sdkp->device->manage_start_stop) {
3407                 sd_printk(KERN_NOTICE, sdkp, "Stopping disk\n");
3408                 sd_start_stop_device(sdkp, 0);
3409         }
3410 }
3411
3412 static int sd_suspend_common(struct device *dev, bool ignore_stop_errors)
3413 {
3414         struct scsi_disk *sdkp = dev_get_drvdata(dev);
3415         int ret = 0;
3416
3417         if (!sdkp)      /* E.g.: runtime suspend following sd_remove() */
3418                 return 0;
3419
3420         if (sdkp->WCE && sdkp->media_present) {
3421                 sd_printk(KERN_NOTICE, sdkp, "Synchronizing SCSI cache\n");
3422                 ret = sd_sync_cache(sdkp);
3423                 if (ret) {
3424                         /* ignore OFFLINE device */
3425                         if (ret == -ENODEV)
3426                                 ret = 0;
3427                         goto done;
3428                 }
3429         }
3430
3431         if (sdkp->device->manage_start_stop) {
3432                 sd_printk(KERN_NOTICE, sdkp, "Stopping disk\n");
3433                 /* an error is not worth aborting a system sleep */
3434                 ret = sd_start_stop_device(sdkp, 0);
3435                 if (ignore_stop_errors)
3436                         ret = 0;
3437         }
3438
3439 done:
3440         return ret;
3441 }
3442
3443 static int sd_suspend_system(struct device *dev)
3444 {
3445         return sd_suspend_common(dev, true);
3446 }
3447
3448 static int sd_suspend_runtime(struct device *dev)
3449 {
3450         return sd_suspend_common(dev, false);
3451 }
3452
3453 static int sd_resume(struct device *dev)
3454 {
3455         struct scsi_disk *sdkp = dev_get_drvdata(dev);
3456
3457         if (!sdkp)      /* E.g.: runtime resume at the start of sd_probe() */
3458                 return 0;
3459
3460         if (!sdkp->device->manage_start_stop)
3461                 return 0;
3462
3463         sd_printk(KERN_NOTICE, sdkp, "Starting disk\n");
3464         return sd_start_stop_device(sdkp, 1);
3465 }
3466
3467 /**
3468  *      init_sd - entry point for this driver (both when built in or when
3469  *      a module).
3470  *
3471  *      Note: this function registers this driver with the scsi mid-level.
3472  **/
3473 static int __init init_sd(void)
3474 {
3475         int majors = 0, i, err;
3476
3477         SCSI_LOG_HLQUEUE(3, printk("init_sd: sd driver entry point\n"));
3478
3479         for (i = 0; i < SD_MAJORS; i++) {
3480                 if (register_blkdev(sd_major(i), "sd") != 0)
3481                         continue;
3482                 majors++;
3483                 blk_register_region(sd_major(i), SD_MINORS, NULL,
3484                                     sd_default_probe, NULL, NULL);
3485         }
3486
3487         if (!majors)
3488                 return -ENODEV;
3489
3490         err = class_register(&sd_disk_class);
3491         if (err)
3492                 goto err_out;
3493
3494         sd_cdb_cache = kmem_cache_create("sd_ext_cdb", SD_EXT_CDB_SIZE,
3495                                          0, 0, NULL);
3496         if (!sd_cdb_cache) {
3497                 printk(KERN_ERR "sd: can't init extended cdb cache\n");
3498                 err = -ENOMEM;
3499                 goto err_out_class;
3500         }
3501
3502         sd_cdb_pool = mempool_create_slab_pool(SD_MEMPOOL_SIZE, sd_cdb_cache);
3503         if (!sd_cdb_pool) {
3504                 printk(KERN_ERR "sd: can't init extended cdb pool\n");
3505                 err = -ENOMEM;
3506                 goto err_out_cache;
3507         }
3508
3509         err = scsi_register_driver(&sd_template.gendrv);
3510         if (err)
3511                 goto err_out_driver;
3512
3513         return 0;
3514
3515 err_out_driver:
3516         mempool_destroy(sd_cdb_pool);
3517
3518 err_out_cache:
3519         kmem_cache_destroy(sd_cdb_cache);
3520
3521 err_out_class:
3522         class_unregister(&sd_disk_class);
3523 err_out:
3524         for (i = 0; i < SD_MAJORS; i++)
3525                 unregister_blkdev(sd_major(i), "sd");
3526         return err;
3527 }
3528
3529 /**
3530  *      exit_sd - exit point for this driver (when it is a module).
3531  *
3532  *      Note: this function unregisters this driver from the scsi mid-level.
3533  **/
3534 static void __exit exit_sd(void)
3535 {
3536         int i;
3537
3538         SCSI_LOG_HLQUEUE(3, printk("exit_sd: exiting sd driver\n"));
3539
3540         scsi_unregister_driver(&sd_template.gendrv);
3541         mempool_destroy(sd_cdb_pool);
3542         kmem_cache_destroy(sd_cdb_cache);
3543
3544         class_unregister(&sd_disk_class);
3545
3546         for (i = 0; i < SD_MAJORS; i++) {
3547                 blk_unregister_region(sd_major(i), SD_MINORS);
3548                 unregister_blkdev(sd_major(i), "sd");
3549         }
3550 }
3551
3552 module_init(init_sd);
3553 module_exit(exit_sd);
3554
3555 static void sd_print_sense_hdr(struct scsi_disk *sdkp,
3556                                struct scsi_sense_hdr *sshdr)
3557 {
3558         scsi_print_sense_hdr(sdkp->device,
3559                              sdkp->disk ? sdkp->disk->disk_name : NULL, sshdr);
3560 }
3561
3562 static void sd_print_result(const struct scsi_disk *sdkp, const char *msg,
3563                             int result)
3564 {
3565         const char *hb_string = scsi_hostbyte_string(result);
3566         const char *db_string = scsi_driverbyte_string(result);
3567
3568         if (hb_string || db_string)
3569                 sd_printk(KERN_INFO, sdkp,
3570                           "%s: Result: hostbyte=%s driverbyte=%s\n", msg,
3571                           hb_string ? hb_string : "invalid",
3572                           db_string ? db_string : "invalid");
3573         else
3574                 sd_printk(KERN_INFO, sdkp,
3575                           "%s: Result: hostbyte=0x%02x driverbyte=0x%02x\n",
3576                           msg, host_byte(result), driver_byte(result));
3577 }
3578