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