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 <asm/uaccess.h>
  55 #include <asm/unaligned.h>
  56
  57 #include <scsi/scsi.h>
  58 #include <scsi/scsi_cmnd.h>
  59 #include <scsi/scsi_dbg.h>
  60 #include <scsi/scsi_device.h>
  61 #include <scsi/scsi_driver.h>
  62 #include <scsi/scsi_eh.h>
  63 #include <scsi/scsi_host.h>
  64 #include <scsi/scsi_ioctl.h>
  65 #include <scsi/scsicam.h>
  66
  67 #include "sd.h"
  68 #include "scsi_logging.h"
  69
  70 MODULE_AUTHOR("Eric Youngdale");
  71 MODULE_DESCRIPTION("SCSI disk (sd) driver");
  72 MODULE_LICENSE("GPL");
  73
  74 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK0_MAJOR);
  75 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK1_MAJOR);
  76 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK2_MAJOR);
  77 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK3_MAJOR);
  78 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK4_MAJOR);
  79 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK5_MAJOR);
  80 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK6_MAJOR);
  81 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK7_MAJOR);
  82 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK8_MAJOR);
  83 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK9_MAJOR);
  84 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK10_MAJOR);
  85 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK11_MAJOR);
  86 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK12_MAJOR);
  87 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK13_MAJOR);
  88 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK14_MAJOR);
  89 MODULE_ALIAS_BLOCKDEV_MAJOR(SCSI_DISK15_MAJOR);
  90 MODULE_ALIAS_SCSI_DEVICE(TYPE_DISK);
  91 MODULE_ALIAS_SCSI_DEVICE(TYPE_MOD);
  92 MODULE_ALIAS_SCSI_DEVICE(TYPE_RBC);
  93
  94 #if !defined(CONFIG_DEBUG_BLOCK_EXT_DEVT)
  95 #define SD_MINORS       16
  96 #else
  97 #define SD_MINORS       0
  98 #endif
  99
 100 static void sd_config_discard(struct scsi_disk *, unsigned int);
 101 static int  sd_revalidate_disk(struct gendisk *);
 102 static void sd_unlock_native_capacity(struct gendisk *disk);
 103 static int  sd_probe(struct device *);
 104 static int  sd_remove(struct device *);
 105 static void sd_shutdown(struct device *);
 106 static int sd_suspend(struct device *, pm_message_t state);
 107 static int sd_resume(struct device *);
 108 static void sd_rescan(struct device *);
 109 static int sd_done(struct scsi_cmnd *);
 110 static int sd_eh_action(struct scsi_cmnd *, unsigned char *, int, int);
 111 static void sd_read_capacity(struct scsi_disk *sdkp, unsigned char *buffer);
 112 static void scsi_disk_release(struct device *cdev);
 113 static void sd_print_sense_hdr(struct scsi_disk *, struct scsi_sense_hdr *);
 114 static void sd_print_result(struct scsi_disk *, int);
 115
 116 static DEFINE_SPINLOCK(sd_index_lock);
 117 static DEFINE_IDA(sd_index_ida);
 118
 119 /* This semaphore is used to mediate the 0->1 reference get in the
 120  * face of object destruction (i.e. we can't allow a get on an
 121  * object after last put) */
 122 static DEFINE_MUTEX(sd_ref_mutex);
 123
 124 static struct kmem_cache *sd_cdb_cache;
 125 static mempool_t *sd_cdb_pool;
 126
 127 static const char *sd_cache_types[] = {
 128         "write through", "none", "write back",
 129         "write back, no read (daft)"
 130 };
 131
 132 static ssize_t
 133 sd_store_cache_type(struct device *dev, struct device_attribute *attr,
 134                     const char *buf, size_t count)
 135 {
 136         int i, ct = -1, rcd, wce, sp;
 137         struct scsi_disk *sdkp = to_scsi_disk(dev);
 138         struct scsi_device *sdp = sdkp->device;
 139         char buffer[64];
 140         char *buffer_data;
 141         struct scsi_mode_data data;
 142         struct scsi_sense_hdr sshdr;
 143         int len;
 144
 145         if (sdp->type != TYPE_DISK)
 146                 /* no cache control on RBC devices; theoretically they
 147                  * can do it, but there's probably so many exceptions
 148                  * it's not worth the risk */
 149                 return -EINVAL;
 150
 151         for (i = 0; i < ARRAY_SIZE(sd_cache_types); i++) {
 152                 len = strlen(sd_cache_types[i]);
 153                 if (strncmp(sd_cache_types[i], buf, len) == 0 &&
 154                     buf[len] == '\n') {
 155                         ct = i;
 156                         break;
 157                 }
 158         }
 159         if (ct < 0)
 160                 return -EINVAL;
 161         rcd = ct & 0x01 ? 1 : 0;
 162         wce = ct & 0x02 ? 1 : 0;
 163         if (scsi_mode_sense(sdp, 0x08, 8, buffer, sizeof(buffer), SD_TIMEOUT,
 164                             SD_MAX_RETRIES, &data, NULL))
 165                 return -EINVAL;
 166         len = min_t(size_t, sizeof(buffer), data.length - data.header_length -
 167                   data.block_descriptor_length);
 168         buffer_data = buffer + data.header_length +
 169                 data.block_descriptor_length;
 170         buffer_data[2] &= ~0x05;
 171         buffer_data[2] |= wce << 2 | rcd;
 172         sp = buffer_data[0] & 0x80 ? 1 : 0;
 173
 174         if (scsi_mode_select(sdp, 1, sp, 8, buffer_data, len, SD_TIMEOUT,
 175                              SD_MAX_RETRIES, &data, &sshdr)) {
 176                 if (scsi_sense_valid(&sshdr))
 177                         sd_print_sense_hdr(sdkp, &sshdr);
 178                 return -EINVAL;
 179         }
 180         revalidate_disk(sdkp->disk);
 181         return count;
 182 }
 183
 184 static ssize_t
 185 sd_store_manage_start_stop(struct device *dev, struct device_attribute *attr,
 186                            const char *buf, size_t count)
 187 {
 188         struct scsi_disk *sdkp = to_scsi_disk(dev);
 189         struct scsi_device *sdp = sdkp->device;
 190
 191         if (!capable(CAP_SYS_ADMIN))
 192                 return -EACCES;
 193
 194         sdp->manage_start_stop = simple_strtoul(buf, NULL, 10);
 195
 196         return count;
 197 }
 198
 199 static ssize_t
 200 sd_store_allow_restart(struct device *dev, struct device_attribute *attr,
 201                        const char *buf, size_t count)
 202 {
 203         struct scsi_disk *sdkp = to_scsi_disk(dev);
 204         struct scsi_device *sdp = sdkp->device;
 205
 206         if (!capable(CAP_SYS_ADMIN))
 207                 return -EACCES;
 208
 209         if (sdp->type != TYPE_DISK)
 210                 return -EINVAL;
 211
 212         sdp->allow_restart = simple_strtoul(buf, NULL, 10);
 213
 214         return count;
 215 }
 216
 217 static ssize_t
 218 sd_show_cache_type(struct device *dev, struct device_attribute *attr,
 219                    char *buf)
 220 {
 221         struct scsi_disk *sdkp = to_scsi_disk(dev);
 222         int ct = sdkp->RCD + 2*sdkp->WCE;
 223
 224         return snprintf(buf, 40, "%s\n", sd_cache_types[ct]);
 225 }
 226
 227 static ssize_t
 228 sd_show_fua(struct device *dev, struct device_attribute *attr, char *buf)
 229 {
 230         struct scsi_disk *sdkp = to_scsi_disk(dev);
 231
 232         return snprintf(buf, 20, "%u\n", sdkp->DPOFUA);
 233 }
 234
 235 static ssize_t
 236 sd_show_manage_start_stop(struct device *dev, struct device_attribute *attr,
 237                           char *buf)
 238 {
 239         struct scsi_disk *sdkp = to_scsi_disk(dev);
 240         struct scsi_device *sdp = sdkp->device;
 241
 242         return snprintf(buf, 20, "%u\n", sdp->manage_start_stop);
 243 }
 244
 245 static ssize_t
 246 sd_show_allow_restart(struct device *dev, struct device_attribute *attr,
 247                       char *buf)
 248 {
 249         struct scsi_disk *sdkp = to_scsi_disk(dev);
 250
 251         return snprintf(buf, 40, "%d\n", sdkp->device->allow_restart);
 252 }
 253
 254 static ssize_t
 255 sd_show_protection_type(struct device *dev, struct device_attribute *attr,
 256                         char *buf)
 257 {
 258         struct scsi_disk *sdkp = to_scsi_disk(dev);
 259
 260         return snprintf(buf, 20, "%u\n", sdkp->protection_type);
 261 }
 262
 263 static ssize_t
 264 sd_show_protection_mode(struct device *dev, struct device_attribute *attr,
 265                         char *buf)
 266 {
 267         struct scsi_disk *sdkp = to_scsi_disk(dev);
 268         struct scsi_device *sdp = sdkp->device;
 269         unsigned int dif, dix;
 270
 271         dif = scsi_host_dif_capable(sdp->host, sdkp->protection_type);
 272         dix = scsi_host_dix_capable(sdp->host, sdkp->protection_type);
 273
 274         if (!dix && scsi_host_dix_capable(sdp->host, SD_DIF_TYPE0_PROTECTION)) {
 275                 dif = 0;
 276                 dix = 1;
 277         }
 278
 279         if (!dif && !dix)
 280                 return snprintf(buf, 20, "none\n");
 281
 282         return snprintf(buf, 20, "%s%u\n", dix ? "dix" : "dif", dif);
 283 }
 284
 285 static ssize_t
 286 sd_show_app_tag_own(struct device *dev, struct device_attribute *attr,
 287                     char *buf)
 288 {
 289         struct scsi_disk *sdkp = to_scsi_disk(dev);
 290
 291         return snprintf(buf, 20, "%u\n", sdkp->ATO);
 292 }
 293
 294 static ssize_t
 295 sd_show_thin_provisioning(struct device *dev, struct device_attribute *attr,
 296                           char *buf)
 297 {
 298         struct scsi_disk *sdkp = to_scsi_disk(dev);
 299
 300         return snprintf(buf, 20, "%u\n", sdkp->lbpme);
 301 }
 302
 303 static const char *lbp_mode[] = {
 304         [SD_LBP_FULL]           = "full",
 305         [SD_LBP_UNMAP]          = "unmap",
 306         [SD_LBP_WS16]           = "writesame_16",
 307         [SD_LBP_WS10]           = "writesame_10",
 308         [SD_LBP_ZERO]           = "writesame_zero",
 309         [SD_LBP_DISABLE]        = "disabled",
 310 };
 311
 312 static ssize_t
 313 sd_show_provisioning_mode(struct device *dev, struct device_attribute *attr,
 314                           char *buf)
 315 {
 316         struct scsi_disk *sdkp = to_scsi_disk(dev);
 317
 318         return snprintf(buf, 20, "%s\n", lbp_mode[sdkp->provisioning_mode]);
 319 }
 320
 321 static ssize_t
 322 sd_store_provisioning_mode(struct device *dev, struct device_attribute *attr,
 323                            const char *buf, size_t count)
 324 {
 325         struct scsi_disk *sdkp = to_scsi_disk(dev);
 326         struct scsi_device *sdp = sdkp->device;
 327
 328         if (!capable(CAP_SYS_ADMIN))
 329                 return -EACCES;
 330
 331         if (sdp->type != TYPE_DISK)
 332                 return -EINVAL;
 333
 334         if (!strncmp(buf, lbp_mode[SD_LBP_UNMAP], 20))
 335                 sd_config_discard(sdkp, SD_LBP_UNMAP);
 336         else if (!strncmp(buf, lbp_mode[SD_LBP_WS16], 20))
 337                 sd_config_discard(sdkp, SD_LBP_WS16);
 338         else if (!strncmp(buf, lbp_mode[SD_LBP_WS10], 20))
 339                 sd_config_discard(sdkp, SD_LBP_WS10);
 340         else if (!strncmp(buf, lbp_mode[SD_LBP_ZERO], 20))
 341                 sd_config_discard(sdkp, SD_LBP_ZERO);
 342         else if (!strncmp(buf, lbp_mode[SD_LBP_DISABLE], 20))
 343                 sd_config_discard(sdkp, SD_LBP_DISABLE);
 344         else
 345                 return -EINVAL;
 346
 347         return count;
 348 }
 349
 350 static ssize_t
 351 sd_show_max_medium_access_timeouts(struct device *dev,
 352                                    struct device_attribute *attr, char *buf)
 353 {
 354         struct scsi_disk *sdkp = to_scsi_disk(dev);
 355
 356         return snprintf(buf, 20, "%u\n", sdkp->max_medium_access_timeouts);
 357 }
 358
 359 static ssize_t
 360 sd_store_max_medium_access_timeouts(struct device *dev,
 361                                     struct device_attribute *attr,
 362                                     const char *buf, size_t count)
 363 {
 364         struct scsi_disk *sdkp = to_scsi_disk(dev);
 365         int err;
 366
 367         if (!capable(CAP_SYS_ADMIN))
 368                 return -EACCES;
 369
 370         err = kstrtouint(buf, 10, &sdkp->max_medium_access_timeouts);
 371
 372         return err ? err : count;
 373 }
 374
 375 static struct device_attribute sd_disk_attrs[] = {
 376         __ATTR(cache_type, S_IRUGO|S_IWUSR, sd_show_cache_type,
 377                sd_store_cache_type),
 378         __ATTR(FUA, S_IRUGO, sd_show_fua, NULL),
 379         __ATTR(allow_restart, S_IRUGO|S_IWUSR, sd_show_allow_restart,
 380                sd_store_allow_restart),
 381         __ATTR(manage_start_stop, S_IRUGO|S_IWUSR, sd_show_manage_start_stop,
 382                sd_store_manage_start_stop),
 383         __ATTR(protection_type, S_IRUGO, sd_show_protection_type, NULL),
 384         __ATTR(protection_mode, S_IRUGO, sd_show_protection_mode, NULL),
 385         __ATTR(app_tag_own, S_IRUGO, sd_show_app_tag_own, NULL),
 386         __ATTR(thin_provisioning, S_IRUGO, sd_show_thin_provisioning, NULL),
 387         __ATTR(provisioning_mode, S_IRUGO|S_IWUSR, sd_show_provisioning_mode,
 388                sd_store_provisioning_mode),
 389         __ATTR(max_medium_access_timeouts, S_IRUGO|S_IWUSR,
 390                sd_show_max_medium_access_timeouts,
 391                sd_store_max_medium_access_timeouts),
 392         __ATTR_NULL,
 393 };
 394
 395 static struct class sd_disk_class = {
 396         .name           = "scsi_disk",
 397         .owner          = THIS_MODULE,
 398         .dev_release    = scsi_disk_release,
 399         .dev_attrs      = sd_disk_attrs,
 400 };
 401
 402 static struct scsi_driver sd_template = {
 403         .owner                  = THIS_MODULE,
 404         .gendrv = {
 405                 .name           = "sd",
 406                 .probe          = sd_probe,
 407                 .remove         = sd_remove,
 408                 .suspend        = sd_suspend,
 409                 .resume         = sd_resume,
 410                 .shutdown       = sd_shutdown,
 411         },
 412         .rescan                 = sd_rescan,
 413         .done                   = sd_done,
 414         .eh_action              = sd_eh_action,
 415 };
 416
 417 /*
 418  * Device no to disk mapping:
 419  *
 420  *       major         disc2     disc  p1
 421  *   |............|.............|....|....| <- dev_t
 422  *    31        20 19          8 7  4 3  0
 423  *
 424  * Inside a major, we have 16k disks, however mapped non-
 425  * contiguously. The first 16 disks are for major0, the next
 426  * ones with major1, ... Disk 256 is for major0 again, disk 272
 427  * for major1, ...
 428  * As we stay compatible with our numbering scheme, we can reuse
 429  * the well-know SCSI majors 8, 65--71, 136--143.
 430  */
 431 static int sd_major(int major_idx)
 432 {
 433         switch (major_idx) {
 434         case 0:
 435                 return SCSI_DISK0_MAJOR;
 436         case 1 ... 7:
 437                 return SCSI_DISK1_MAJOR + major_idx - 1;
 438         case 8 ... 15:
 439                 return SCSI_DISK8_MAJOR + major_idx - 8;
 440         default:
 441                 BUG();
 442                 return 0;       /* shut up gcc */
 443         }
 444 }
 445
 446 static struct scsi_disk *__scsi_disk_get(struct gendisk *disk)
 447 {
 448         struct scsi_disk *sdkp = NULL;
 449
 450         if (disk->private_data) {
 451                 sdkp = scsi_disk(disk);
 452                 if (scsi_device_get(sdkp->device) == 0)
 453                         get_device(&sdkp->dev);
 454                 else
 455                         sdkp = NULL;
 456         }
 457         return sdkp;
 458 }
 459
 460 static struct scsi_disk *scsi_disk_get(struct gendisk *disk)
 461 {
 462         struct scsi_disk *sdkp;
 463
 464         mutex_lock(&sd_ref_mutex);
 465         sdkp = __scsi_disk_get(disk);
 466         mutex_unlock(&sd_ref_mutex);
 467         return sdkp;
 468 }
 469
 470 static struct scsi_disk *scsi_disk_get_from_dev(struct device *dev)
 471 {
 472         struct scsi_disk *sdkp;
 473
 474         mutex_lock(&sd_ref_mutex);
 475         sdkp = dev_get_drvdata(dev);
 476         if (sdkp)
 477                 sdkp = __scsi_disk_get(sdkp->disk);
 478         mutex_unlock(&sd_ref_mutex);
 479         return sdkp;
 480 }
 481
 482 static void scsi_disk_put(struct scsi_disk *sdkp)
 483 {
 484         struct scsi_device *sdev = sdkp->device;
 485
 486         mutex_lock(&sd_ref_mutex);
 487         put_device(&sdkp->dev);
 488         scsi_device_put(sdev);
 489         mutex_unlock(&sd_ref_mutex);
 490 }
 491
 492 static void sd_prot_op(struct scsi_cmnd *scmd, unsigned int dif)
 493 {
 494         unsigned int prot_op = SCSI_PROT_NORMAL;
 495         unsigned int dix = scsi_prot_sg_count(scmd);
 496
 497         if (scmd->sc_data_direction == DMA_FROM_DEVICE) {
 498                 if (dif && dix)
 499                         prot_op = SCSI_PROT_READ_PASS;
 500                 else if (dif && !dix)
 501                         prot_op = SCSI_PROT_READ_STRIP;
 502                 else if (!dif && dix)
 503                         prot_op = SCSI_PROT_READ_INSERT;
 504         } else {
 505                 if (dif && dix)
 506                         prot_op = SCSI_PROT_WRITE_PASS;
 507                 else if (dif && !dix)
 508                         prot_op = SCSI_PROT_WRITE_INSERT;
 509                 else if (!dif && dix)
 510                         prot_op = SCSI_PROT_WRITE_STRIP;
 511         }
 512
 513         scsi_set_prot_op(scmd, prot_op);
 514         scsi_set_prot_type(scmd, dif);
 515 }
 516
 517 static void sd_config_discard(struct scsi_disk *sdkp, unsigned int mode)
 518 {
 519         struct request_queue *q = sdkp->disk->queue;
 520         unsigned int logical_block_size = sdkp->device->sector_size;
 521         unsigned int max_blocks = 0;
 522
 523         q->limits.discard_zeroes_data = sdkp->lbprz;
 524         q->limits.discard_alignment = sdkp->unmap_alignment *
 525                 logical_block_size;
 526         q->limits.discard_granularity =
 527                 max(sdkp->physical_block_size,
 528                     sdkp->unmap_granularity * logical_block_size);
 529
 530         sdkp->provisioning_mode = mode;
 531
 532         switch (mode) {
 533
 534         case SD_LBP_DISABLE:
 535                 q->limits.max_discard_sectors = 0;
 536                 queue_flag_clear_unlocked(QUEUE_FLAG_DISCARD, q);
 537                 return;
 538
 539         case SD_LBP_UNMAP:
 540                 max_blocks = min_not_zero(sdkp->max_unmap_blocks, 0xffffffff);
 541                 break;
 542
 543         case SD_LBP_WS16:
 544                 max_blocks = min_not_zero(sdkp->max_ws_blocks, 0xffffffff);
 545                 break;
 546
 547         case SD_LBP_WS10:
 548                 max_blocks = min_not_zero(sdkp->max_ws_blocks, (u32)0xffff);
 549                 break;
 550
 551         case SD_LBP_ZERO:
 552                 max_blocks = min_not_zero(sdkp->max_ws_blocks, (u32)0xffff);
 553                 q->limits.discard_zeroes_data = 1;
 554                 break;
 555         }
 556
 557         q->limits.max_discard_sectors = max_blocks * (logical_block_size >> 9);
 558         queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, q);
 559 }
 560
 561 /**
 562  * scsi_setup_discard_cmnd - unmap blocks on thinly provisioned device
 563  * @sdp: scsi device to operate one
 564  * @rq: Request to prepare
 565  *
 566  * Will issue either UNMAP or WRITE SAME(16) depending on preference
 567  * indicated by target device.
 568  **/
 569 static int scsi_setup_discard_cmnd(struct scsi_device *sdp, struct request *rq)
 570 {
 571         struct scsi_disk *sdkp = scsi_disk(rq->rq_disk);
 572         struct bio *bio = rq->bio;
 573         sector_t sector = bio->bi_sector;
 574         unsigned int nr_sectors = bio_sectors(bio);
 575         unsigned int len;
 576         int ret;
 577         char *buf;
 578         struct page *page;
 579
 580         if (sdkp->device->sector_size == 4096) {
 581                 sector >>= 3;
 582                 nr_sectors >>= 3;
 583         }
 584
 585         rq->timeout = SD_TIMEOUT;
 586
 587         memset(rq->cmd, 0, rq->cmd_len);
 588
 589         page = alloc_page(GFP_ATOMIC | __GFP_ZERO);
 590         if (!page)
 591                 return BLKPREP_DEFER;
 592
 593         switch (sdkp->provisioning_mode) {
 594         case SD_LBP_UNMAP:
 595                 buf = page_address(page);
 596
 597                 rq->cmd_len = 10;
 598                 rq->cmd[0] = UNMAP;
 599                 rq->cmd[8] = 24;
 600
 601                 put_unaligned_be16(6 + 16, &buf[0]);
 602                 put_unaligned_be16(16, &buf[2]);
 603                 put_unaligned_be64(sector, &buf[8]);
 604                 put_unaligned_be32(nr_sectors, &buf[16]);
 605
 606                 len = 24;
 607                 break;
 608
 609         case SD_LBP_WS16:
 610                 rq->cmd_len = 16;
 611                 rq->cmd[0] = WRITE_SAME_16;
 612                 rq->cmd[1] = 0x8; /* UNMAP */
 613                 put_unaligned_be64(sector, &rq->cmd[2]);
 614                 put_unaligned_be32(nr_sectors, &rq->cmd[10]);
 615
 616                 len = sdkp->device->sector_size;
 617                 break;
 618
 619         case SD_LBP_WS10:
 620         case SD_LBP_ZERO:
 621                 rq->cmd_len = 10;
 622                 rq->cmd[0] = WRITE_SAME;
 623                 if (sdkp->provisioning_mode == SD_LBP_WS10)
 624                         rq->cmd[1] = 0x8; /* UNMAP */
 625                 put_unaligned_be32(sector, &rq->cmd[2]);
 626                 put_unaligned_be16(nr_sectors, &rq->cmd[7]);
 627
 628                 len = sdkp->device->sector_size;
 629                 break;
 630
 631         default:
 632                 ret = BLKPREP_KILL;
 633                 goto out;
 634         }
 635
 636         blk_add_request_payload(rq, page, len);
 637         ret = scsi_setup_blk_pc_cmnd(sdp, rq);
 638         rq->buffer = page_address(page);
 639
 640 out:
 641         if (ret != BLKPREP_OK) {
 642                 __free_page(page);
 643                 rq->buffer = NULL;
 644         }
 645         return ret;
 646 }
 647
 648 static int scsi_setup_flush_cmnd(struct scsi_device *sdp, struct request *rq)
 649 {
 650         rq->timeout = SD_FLUSH_TIMEOUT;
 651         rq->retries = SD_MAX_RETRIES;
 652         rq->cmd[0] = SYNCHRONIZE_CACHE;
 653         rq->cmd_len = 10;
 654
 655         return scsi_setup_blk_pc_cmnd(sdp, rq);
 656 }
 657
 658 static void sd_unprep_fn(struct request_queue *q, struct request *rq)
 659 {
 660         if (rq->cmd_flags & REQ_DISCARD) {
 661                 free_page((unsigned long)rq->buffer);
 662                 rq->buffer = NULL;
 663         }
 664 }
 665
 666 /**
 667  *      sd_prep_fn - build a scsi (read or write) command from
 668  *      information in the request structure.
 669  *      @SCpnt: pointer to mid-level's per scsi command structure that
 670  *      contains request and into which the scsi command is written
 671  *
 672  *      Returns 1 if successful and 0 if error (or cannot be done now).
 673  **/
 674 static int sd_prep_fn(struct request_queue *q, struct request *rq)
 675 {
 676         struct scsi_cmnd *SCpnt;
 677         struct scsi_device *sdp = q->queuedata;
 678         struct gendisk *disk = rq->rq_disk;
 679         struct scsi_disk *sdkp;
 680         sector_t block = blk_rq_pos(rq);
 681         sector_t threshold;
 682         unsigned int this_count = blk_rq_sectors(rq);
 683         int ret, host_dif;
 684         unsigned char protect;
 685
 686         /*
 687          * Discard request come in as REQ_TYPE_FS but we turn them into
 688          * block PC requests to make life easier.
 689          */
 690         if (rq->cmd_flags & REQ_DISCARD) {
 691                 ret = scsi_setup_discard_cmnd(sdp, rq);
 692                 goto out;
 693         } else if (rq->cmd_flags & REQ_FLUSH) {
 694                 ret = scsi_setup_flush_cmnd(sdp, rq);
 695                 goto out;
 696         } else if (rq->cmd_type == REQ_TYPE_BLOCK_PC) {
 697                 ret = scsi_setup_blk_pc_cmnd(sdp, rq);
 698                 goto out;
 699         } else if (rq->cmd_type != REQ_TYPE_FS) {
 700                 ret = BLKPREP_KILL;
 701                 goto out;
 702         }
 703         ret = scsi_setup_fs_cmnd(sdp, rq);
 704         if (ret != BLKPREP_OK)
 705                 goto out;
 706         SCpnt = rq->special;
 707         sdkp = scsi_disk(disk);
 708
 709         /* from here on until we're complete, any goto out
 710          * is used for a killable error condition */
 711         ret = BLKPREP_KILL;
 712
 713         SCSI_LOG_HLQUEUE(1, scmd_printk(KERN_INFO, SCpnt,
 714                                         "sd_prep_fn: block=%llu, "
 715                                         "count=%d\n",
 716                                         (unsigned long long)block,
 717                                         this_count));
 718
 719         if (!sdp || !scsi_device_online(sdp) ||
 720             block + blk_rq_sectors(rq) > get_capacity(disk)) {
 721                 SCSI_LOG_HLQUEUE(2, scmd_printk(KERN_INFO, SCpnt,
 722                                                 "Finishing %u sectors\n",
 723                                                 blk_rq_sectors(rq)));
 724                 SCSI_LOG_HLQUEUE(2, scmd_printk(KERN_INFO, SCpnt,
 725                                                 "Retry with 0x%p\n", SCpnt));
 726                 goto out;
 727         }
 728
 729         if (sdp->changed) {
 730                 /*
 731                  * quietly refuse to do anything to a changed disc until
 732                  * the changed bit has been reset
 733                  */
 734                 /* printk("SCSI disk has been changed or is not present. Prohibiting further I/O.\n"); */
 735                 goto out;
 736         }
 737
 738         /*
 739          * Some SD card readers can't handle multi-sector accesses which touch
 740          * the last one or two hardware sectors.  Split accesses as needed.
 741          */
 742         threshold = get_capacity(disk) - SD_LAST_BUGGY_SECTORS *
 743                 (sdp->sector_size / 512);
 744
 745         if (unlikely(sdp->last_sector_bug && block + this_count > threshold)) {
 746                 if (block < threshold) {
 747                         /* Access up to the threshold but not beyond */
 748                         this_count = threshold - block;
 749                 } else {
 750                         /* Access only a single hardware sector */
 751                         this_count = sdp->sector_size / 512;
 752                 }
 753         }
 754
 755         SCSI_LOG_HLQUEUE(2, scmd_printk(KERN_INFO, SCpnt, "block=%llu\n",
 756                                         (unsigned long long)block));
 757
 758         /*
 759          * If we have a 1K hardware sectorsize, prevent access to single
 760          * 512 byte sectors.  In theory we could handle this - in fact
 761          * the scsi cdrom driver must be able to handle this because
 762          * we typically use 1K blocksizes, and cdroms typically have
 763          * 2K hardware sectorsizes.  Of course, things are simpler
 764          * with the cdrom, since it is read-only.  For performance
 765          * reasons, the filesystems should be able to handle this
 766          * and not force the scsi disk driver to use bounce buffers
 767          * for this.
 768          */
 769         if (sdp->sector_size == 1024) {
 770                 if ((block & 1) || (blk_rq_sectors(rq) & 1)) {
 771                         scmd_printk(KERN_ERR, SCpnt,
 772                                     "Bad block number requested\n");
 773                         goto out;
 774                 } else {
 775                         block = block >> 1;
 776                         this_count = this_count >> 1;
 777                 }
 778         }
 779         if (sdp->sector_size == 2048) {
 780                 if ((block & 3) || (blk_rq_sectors(rq) & 3)) {
 781                         scmd_printk(KERN_ERR, SCpnt,
 782                                     "Bad block number requested\n");
 783                         goto out;
 784                 } else {
 785                         block = block >> 2;
 786                         this_count = this_count >> 2;
 787                 }
 788         }
 789         if (sdp->sector_size == 4096) {
 790                 if ((block & 7) || (blk_rq_sectors(rq) & 7)) {
 791                         scmd_printk(KERN_ERR, SCpnt,
 792                                     "Bad block number requested\n");
 793                         goto out;
 794                 } else {
 795                         block = block >> 3;
 796                         this_count = this_count >> 3;
 797                 }
 798         }
 799         if (rq_data_dir(rq) == WRITE) {
 800                 if (!sdp->writeable) {
 801                         goto out;
 802                 }
 803                 SCpnt->cmnd[0] = WRITE_6;
 804                 SCpnt->sc_data_direction = DMA_TO_DEVICE;
 805
 806                 if (blk_integrity_rq(rq) &&
 807                     sd_dif_prepare(rq, block, sdp->sector_size) == -EIO)
 808                         goto out;
 809
 810         } else if (rq_data_dir(rq) == READ) {
 811                 SCpnt->cmnd[0] = READ_6;
 812                 SCpnt->sc_data_direction = DMA_FROM_DEVICE;
 813         } else {
 814                 scmd_printk(KERN_ERR, SCpnt, "Unknown command %x\n", rq->cmd_flags);
 815                 goto out;
 816         }
 817
 818         SCSI_LOG_HLQUEUE(2, scmd_printk(KERN_INFO, SCpnt,
 819                                         "%s %d/%u 512 byte blocks.\n",
 820                                         (rq_data_dir(rq) == WRITE) ?
 821                                         "writing" : "reading", this_count,
 822                                         blk_rq_sectors(rq)));
 823
 824         /* Set RDPROTECT/WRPROTECT if disk is formatted with DIF */
 825         host_dif = scsi_host_dif_capable(sdp->host, sdkp->protection_type);
 826         if (host_dif)
 827                 protect = 1 << 5;
 828         else
 829                 protect = 0;
 830
 831         if (host_dif == SD_DIF_TYPE2_PROTECTION) {
 832                 SCpnt->cmnd = mempool_alloc(sd_cdb_pool, GFP_ATOMIC);
 833
 834                 if (unlikely(SCpnt->cmnd == NULL)) {
 835                         ret = BLKPREP_DEFER;
 836                         goto out;
 837                 }
 838
 839                 SCpnt->cmd_len = SD_EXT_CDB_SIZE;
 840                 memset(SCpnt->cmnd, 0, SCpnt->cmd_len);
 841                 SCpnt->cmnd[0] = VARIABLE_LENGTH_CMD;
 842                 SCpnt->cmnd[7] = 0x18;
 843                 SCpnt->cmnd[9] = (rq_data_dir(rq) == READ) ? READ_32 : WRITE_32;
 844                 SCpnt->cmnd[10] = protect | ((rq->cmd_flags & REQ_FUA) ? 0x8 : 0);
 845
 846                 /* LBA */
 847                 SCpnt->cmnd[12] = sizeof(block) > 4 ? (unsigned char) (block >> 56) & 0xff : 0;
 848                 SCpnt->cmnd[13] = sizeof(block) > 4 ? (unsigned char) (block >> 48) & 0xff : 0;
 849                 SCpnt->cmnd[14] = sizeof(block) > 4 ? (unsigned char) (block >> 40) & 0xff : 0;
 850                 SCpnt->cmnd[15] = sizeof(block) > 4 ? (unsigned char) (block >> 32) & 0xff : 0;
 851                 SCpnt->cmnd[16] = (unsigned char) (block >> 24) & 0xff;
 852                 SCpnt->cmnd[17] = (unsigned char) (block >> 16) & 0xff;
 853                 SCpnt->cmnd[18] = (unsigned char) (block >> 8) & 0xff;
 854                 SCpnt->cmnd[19] = (unsigned char) block & 0xff;
 855
 856                 /* Expected Indirect LBA */
 857                 SCpnt->cmnd[20] = (unsigned char) (block >> 24) & 0xff;
 858                 SCpnt->cmnd[21] = (unsigned char) (block >> 16) & 0xff;
 859                 SCpnt->cmnd[22] = (unsigned char) (block >> 8) & 0xff;
 860                 SCpnt->cmnd[23] = (unsigned char) block & 0xff;
 861
 862                 /* Transfer length */
 863                 SCpnt->cmnd[28] = (unsigned char) (this_count >> 24) & 0xff;
 864                 SCpnt->cmnd[29] = (unsigned char) (this_count >> 16) & 0xff;
 865                 SCpnt->cmnd[30] = (unsigned char) (this_count >> 8) & 0xff;
 866                 SCpnt->cmnd[31] = (unsigned char) this_count & 0xff;
 867         } else if (block > 0xffffffff) {
 868                 SCpnt->cmnd[0] += READ_16 - READ_6;
 869                 SCpnt->cmnd[1] = protect | ((rq->cmd_flags & REQ_FUA) ? 0x8 : 0);
 870                 SCpnt->cmnd[2] = sizeof(block) > 4 ? (unsigned char) (block >> 56) & 0xff : 0;
 871                 SCpnt->cmnd[3] = sizeof(block) > 4 ? (unsigned char) (block >> 48) & 0xff : 0;
 872                 SCpnt->cmnd[4] = sizeof(block) > 4 ? (unsigned char) (block >> 40) & 0xff : 0;
 873                 SCpnt->cmnd[5] = sizeof(block) > 4 ? (unsigned char) (block >> 32) & 0xff : 0;
 874                 SCpnt->cmnd[6] = (unsigned char) (block >> 24) & 0xff;
 875                 SCpnt->cmnd[7] = (unsigned char) (block >> 16) & 0xff;
 876                 SCpnt->cmnd[8] = (unsigned char) (block >> 8) & 0xff;
 877                 SCpnt->cmnd[9] = (unsigned char) block & 0xff;
 878                 SCpnt->cmnd[10] = (unsigned char) (this_count >> 24) & 0xff;
 879                 SCpnt->cmnd[11] = (unsigned char) (this_count >> 16) & 0xff;
 880                 SCpnt->cmnd[12] = (unsigned char) (this_count >> 8) & 0xff;
 881                 SCpnt->cmnd[13] = (unsigned char) this_count & 0xff;
 882                 SCpnt->cmnd[14] = SCpnt->cmnd[15] = 0;
 883         } else if ((this_count > 0xff) || (block > 0x1fffff) ||
 884                    scsi_device_protection(SCpnt->device) ||
 885                    SCpnt->device->use_10_for_rw) {
 886                 if (this_count > 0xffff)
 887                         this_count = 0xffff;
 888
 889                 SCpnt->cmnd[0] += READ_10 - READ_6;
 890                 SCpnt->cmnd[1] = protect | ((rq->cmd_flags & REQ_FUA) ? 0x8 : 0);
 891                 SCpnt->cmnd[2] = (unsigned char) (block >> 24) & 0xff;
 892                 SCpnt->cmnd[3] = (unsigned char) (block >> 16) & 0xff;
 893                 SCpnt->cmnd[4] = (unsigned char) (block >> 8) & 0xff;
 894                 SCpnt->cmnd[5] = (unsigned char) block & 0xff;
 895                 SCpnt->cmnd[6] = SCpnt->cmnd[9] = 0;
 896                 SCpnt->cmnd[7] = (unsigned char) (this_count >> 8) & 0xff;
 897                 SCpnt->cmnd[8] = (unsigned char) this_count & 0xff;
 898         } else {
 899                 if (unlikely(rq->cmd_flags & REQ_FUA)) {
 900                         /*
 901                          * This happens only if this drive failed
 902                          * 10byte rw command with ILLEGAL_REQUEST
 903                          * during operation and thus turned off
 904                          * use_10_for_rw.
 905                          */
 906                         scmd_printk(KERN_ERR, SCpnt,
 907                                     "FUA write on READ/WRITE(6) drive\n");
 908                         goto out;
 909                 }
 910
 911                 SCpnt->cmnd[1] |= (unsigned char) ((block >> 16) & 0x1f);
 912                 SCpnt->cmnd[2] = (unsigned char) ((block >> 8) & 0xff);
 913                 SCpnt->cmnd[3] = (unsigned char) block & 0xff;
 914                 SCpnt->cmnd[4] = (unsigned char) this_count;
 915                 SCpnt->cmnd[5] = 0;
 916         }
 917         SCpnt->sdb.length = this_count * sdp->sector_size;
 918
 919         /* If DIF or DIX is enabled, tell HBA how to handle request */
 920         if (host_dif || scsi_prot_sg_count(SCpnt))
 921                 sd_prot_op(SCpnt, host_dif);
 922
 923         /*
 924          * We shouldn't disconnect in the middle of a sector, so with a dumb
 925          * host adapter, it's safe to assume that we can at least transfer
 926          * this many bytes between each connect / disconnect.
 927          */
 928         SCpnt->transfersize = sdp->sector_size;
 929         SCpnt->underflow = this_count << 9;
 930         SCpnt->allowed = SD_MAX_RETRIES;
 931
 932         /*
 933          * This indicates that the command is ready from our end to be
 934          * queued.
 935          */
 936         ret = BLKPREP_OK;
 937  out:
 938         return scsi_prep_return(q, rq, ret);
 939 }
 940
 941 /**
 942  *      sd_open - open a scsi disk device
 943  *      @inode: only i_rdev member may be used
 944  *      @filp: only f_mode and f_flags may be used
 945  *
 946  *      Returns 0 if successful. Returns a negated errno value in case
 947  *      of error.
 948  *
 949  *      Note: This can be called from a user context (e.g. fsck(1) )
 950  *      or from within the kernel (e.g. as a result of a mount(1) ).
 951  *      In the latter case @inode and @filp carry an abridged amount
 952  *      of information as noted above.
 953  *
 954  *      Locking: called with bdev->bd_mutex held.
 955  **/
 956 static int sd_open(struct block_device *bdev, fmode_t mode)
 957 {
 958         struct scsi_disk *sdkp = scsi_disk_get(bdev->bd_disk);
 959         struct scsi_device *sdev;
 960         int retval;
 961
 962         if (!sdkp)
 963                 return -ENXIO;
 964
 965         SCSI_LOG_HLQUEUE(3, sd_printk(KERN_INFO, sdkp, "sd_open\n"));
 966
 967         sdev = sdkp->device;
 968
 969         retval = scsi_autopm_get_device(sdev);
 970         if (retval)
 971                 goto error_autopm;
 972
 973         /*
 974          * If the device is in error recovery, wait until it is done.
 975          * If the device is offline, then disallow any access to it.
 976          */
 977         retval = -ENXIO;
 978         if (!scsi_block_when_processing_errors(sdev))
 979                 goto error_out;
 980
 981         if (sdev->removable || sdkp->write_prot)
 982                 check_disk_change(bdev);
 983
 984         /*
 985          * If the drive is empty, just let the open fail.
 986          */
 987         retval = -ENOMEDIUM;
 988         if (sdev->removable && !sdkp->media_present && !(mode & FMODE_NDELAY))
 989                 goto error_out;
 990
 991         /*
 992          * If the device has the write protect tab set, have the open fail
 993          * if the user expects to be able to write to the thing.
 994          */
 995         retval = -EROFS;
 996         if (sdkp->write_prot && (mode & FMODE_WRITE))
 997                 goto error_out;
 998
 999         /*
1000          * It is possible that the disk changing stuff resulted in
1001          * the device being taken offline.  If this is the case,
1002          * report this to the user, and don't pretend that the
1003          * open actually succeeded.
1004          */
1005         retval = -ENXIO;
1006         if (!scsi_device_online(sdev))
1007                 goto error_out;
1008
1009         if ((atomic_inc_return(&sdkp->openers) == 1) && sdev->removable) {
1010                 if (scsi_block_when_processing_errors(sdev))
1011                         scsi_set_medium_removal(sdev, SCSI_REMOVAL_PREVENT);
1012         }
1013
1014         return 0;
1015
1016 error_out:
1017         scsi_autopm_put_device(sdev);
1018 error_autopm:
1019         scsi_disk_put(sdkp);
1020         return retval;
1021 }
1022
1023 /**
1024  *      sd_release - invoked when the (last) close(2) is called on this
1025  *      scsi disk.
1026  *      @inode: only i_rdev member may be used
1027  *      @filp: only f_mode and f_flags may be used
1028  *
1029  *      Returns 0.
1030  *
1031  *      Note: may block (uninterruptible) if error recovery is underway
1032  *      on this disk.
1033  *
1034  *      Locking: called with bdev->bd_mutex held.
1035  **/
1036 static int sd_release(struct gendisk *disk, fmode_t mode)
1037 {
1038         struct scsi_disk *sdkp = scsi_disk(disk);
1039         struct scsi_device *sdev = sdkp->device;
1040
1041         SCSI_LOG_HLQUEUE(3, sd_printk(KERN_INFO, sdkp, "sd_release\n"));
1042
1043         if (atomic_dec_return(&sdkp->openers) == 0 && sdev->removable) {
1044                 if (scsi_block_when_processing_errors(sdev))
1045                         scsi_set_medium_removal(sdev, SCSI_REMOVAL_ALLOW);
1046         }
1047
1048         /*
1049          * XXX and what if there are packets in flight and this close()
1050          * XXX is followed by a "rmmod sd_mod"?
1051          */
1052
1053         scsi_autopm_put_device(sdev);
1054         scsi_disk_put(sdkp);
1055         return 0;
1056 }
1057
1058 static int sd_getgeo(struct block_device *bdev, struct hd_geometry *geo)
1059 {
1060         struct scsi_disk *sdkp = scsi_disk(bdev->bd_disk);
1061         struct scsi_device *sdp = sdkp->device;
1062         struct Scsi_Host *host = sdp->host;
1063         int diskinfo[4];
1064
1065         /* default to most commonly used values */
1066         diskinfo[0] = 0x40;     /* 1 << 6 */
1067         diskinfo[1] = 0x20;     /* 1 << 5 */
1068         diskinfo[2] = sdkp->capacity >> 11;
1069
1070         /* override with calculated, extended default, or driver values */
1071         if (host->hostt->bios_param)
1072                 host->hostt->bios_param(sdp, bdev, sdkp->capacity, diskinfo);
1073         else
1074                 scsicam_bios_param(bdev, sdkp->capacity, diskinfo);
1075
1076         geo->heads = diskinfo[0];
1077         geo->sectors = diskinfo[1];
1078         geo->cylinders = diskinfo[2];
1079         return 0;
1080 }
1081
1082 /**
1083  *      sd_ioctl - process an ioctl
1084  *      @inode: only i_rdev/i_bdev members may be used
1085  *      @filp: only f_mode and f_flags may be used
1086  *      @cmd: ioctl command number
1087  *      @arg: this is third argument given to ioctl(2) system call.
1088  *      Often contains a pointer.
1089  *
1090  *      Returns 0 if successful (some ioctls return positive numbers on
1091  *      success as well). Returns a negated errno value in case of error.
1092  *
1093  *      Note: most ioctls are forward onto the block subsystem or further
1094  *      down in the scsi subsystem.
1095  **/
1096 static int sd_ioctl(struct block_device *bdev, fmode_t mode,
1097                     unsigned int cmd, unsigned long arg)
1098 {
1099         struct gendisk *disk = bdev->bd_disk;
1100         struct scsi_disk *sdkp = scsi_disk(disk);
1101         struct scsi_device *sdp = sdkp->device;
1102         void __user *p = (void __user *)arg;
1103         int error;
1104
1105         SCSI_LOG_IOCTL(1, sd_printk(KERN_INFO, sdkp, "sd_ioctl: disk=%s, "
1106                                     "cmd=0x%x\n", disk->disk_name, cmd));
1107
1108         error = scsi_verify_blk_ioctl(bdev, cmd);
1109         if (error < 0)
1110                 return error;
1111
1112         /*
1113          * If we are in the middle of error recovery, don't let anyone
1114          * else try and use this device.  Also, if error recovery fails, it
1115          * may try and take the device offline, in which case all further
1116          * access to the device is prohibited.
1117          */
1118         error = scsi_nonblockable_ioctl(sdp, cmd, p,
1119                                         (mode & FMODE_NDELAY) != 0);
1120         if (!scsi_block_when_processing_errors(sdp) || !error)
1121                 goto out;
1122
1123         /*
1124          * Send SCSI addressing ioctls directly to mid level, send other
1125          * ioctls to block level and then onto mid level if they can't be
1126          * resolved.
1127          */
1128         switch (cmd) {
1129                 case SCSI_IOCTL_GET_IDLUN:
1130                 case SCSI_IOCTL_GET_BUS_NUMBER:
1131                         error = scsi_ioctl(sdp, cmd, p);
1132                         break;
1133                 default:
1134                         error = scsi_cmd_blk_ioctl(bdev, mode, cmd, p);
1135                         if (error != -ENOTTY)
1136                                 break;
1137                         error = scsi_ioctl(sdp, cmd, p);
1138                         break;
1139         }
1140 out:
1141         return error;
1142 }
1143
1144 static void set_media_not_present(struct scsi_disk *sdkp)
1145 {
1146         if (sdkp->media_present)
1147                 sdkp->device->changed = 1;
1148
1149         if (sdkp->device->removable) {
1150                 sdkp->media_present = 0;
1151                 sdkp->capacity = 0;
1152         }
1153 }
1154
1155 static int media_not_present(struct scsi_disk *sdkp,
1156                              struct scsi_sense_hdr *sshdr)
1157 {
1158         if (!scsi_sense_valid(sshdr))
1159                 return 0;
1160
1161         /* not invoked for commands that could return deferred errors */
1162         switch (sshdr->sense_key) {
1163         case UNIT_ATTENTION:
1164         case NOT_READY:
1165                 /* medium not present */
1166                 if (sshdr->asc == 0x3A) {
1167                         set_media_not_present(sdkp);
1168                         return 1;
1169                 }
1170         }
1171         return 0;
1172 }
1173
1174 /**
1175  *      sd_check_events - check media events
1176  *      @disk: kernel device descriptor
1177  *      @clearing: disk events currently being cleared
1178  *
1179  *      Returns mask of DISK_EVENT_*.
1180  *
1181  *      Note: this function is invoked from the block subsystem.
1182  **/
1183 static unsigned int sd_check_events(struct gendisk *disk, unsigned int clearing)
1184 {
1185         struct scsi_disk *sdkp = scsi_disk(disk);
1186         struct scsi_device *sdp = sdkp->device;
1187         struct scsi_sense_hdr *sshdr = NULL;
1188         int retval;
1189
1190         SCSI_LOG_HLQUEUE(3, sd_printk(KERN_INFO, sdkp, "sd_check_events\n"));
1191
1192         /*
1193          * If the device is offline, don't send any commands - just pretend as
1194          * if the command failed.  If the device ever comes back online, we
1195          * can deal with it then.  It is only because of unrecoverable errors
1196          * that we would ever take a device offline in the first place.
1197          */
1198         if (!scsi_device_online(sdp)) {
1199                 set_media_not_present(sdkp);
1200                 goto out;
1201         }
1202
1203         /*
1204          * Using TEST_UNIT_READY enables differentiation between drive with
1205          * no cartridge loaded - NOT READY, drive with changed cartridge -
1206          * UNIT ATTENTION, or with same cartridge - GOOD STATUS.
1207          *
1208          * Drives that auto spin down. eg iomega jaz 1G, will be started
1209          * by sd_spinup_disk() from sd_revalidate_disk(), which happens whenever
1210          * sd_revalidate() is called.
1211          */
1212         retval = -ENODEV;
1213
1214         if (scsi_block_when_processing_errors(sdp)) {
1215                 retval = scsi_autopm_get_device(sdp);
1216                 if (retval)
1217                         goto out;
1218
1219                 sshdr  = kzalloc(sizeof(*sshdr), GFP_KERNEL);
1220                 retval = scsi_test_unit_ready(sdp, SD_TIMEOUT, SD_MAX_RETRIES,
1221                                               sshdr);
1222                 scsi_autopm_put_device(sdp);
1223         }
1224
1225         /* failed to execute TUR, assume media not present */
1226         if (host_byte(retval)) {
1227                 set_media_not_present(sdkp);
1228                 goto out;
1229         }
1230
1231         if (media_not_present(sdkp, sshdr))
1232                 goto out;
1233
1234         /*
1235          * For removable scsi disk we have to recognise the presence
1236          * of a disk in the drive.
1237          */
1238         if (!sdkp->media_present)
1239                 sdp->changed = 1;
1240         sdkp->media_present = 1;
1241 out:
1242         /*
1243          * sdp->changed is set under the following conditions:
1244          *
1245          *      Medium present state has changed in either direction.
1246          *      Device has indicated UNIT_ATTENTION.
1247          */
1248         kfree(sshdr);
1249         retval = sdp->changed ? DISK_EVENT_MEDIA_CHANGE : 0;
1250         sdp->changed = 0;
1251         return retval;
1252 }
1253
1254 static int sd_sync_cache(struct scsi_disk *sdkp)
1255 {
1256         int retries, res;
1257         struct scsi_device *sdp = sdkp->device;
1258         struct scsi_sense_hdr sshdr;
1259
1260         if (!scsi_device_online(sdp))
1261                 return -ENODEV;
1262
1263
1264         for (retries = 3; retries > 0; --retries) {
1265                 unsigned char cmd[10] = { 0 };
1266
1267                 cmd[0] = SYNCHRONIZE_CACHE;
1268                 /*
1269                  * Leave the rest of the command zero to indicate
1270                  * flush everything.
1271                  */
1272                 res = scsi_execute_req(sdp, cmd, DMA_NONE, NULL, 0, &sshdr,
1273                                        SD_FLUSH_TIMEOUT, SD_MAX_RETRIES, NULL);
1274                 if (res == 0)
1275                         break;
1276         }
1277
1278         if (res) {
1279                 sd_print_result(sdkp, res);
1280                 if (driver_byte(res) & DRIVER_SENSE)
1281                         sd_print_sense_hdr(sdkp, &sshdr);
1282         }
1283
1284         if (res)
1285                 return -EIO;
1286         return 0;
1287 }
1288
1289 static void sd_rescan(struct device *dev)
1290 {
1291         struct scsi_disk *sdkp = scsi_disk_get_from_dev(dev);
1292
1293         if (sdkp) {
1294                 revalidate_disk(sdkp->disk);
1295                 scsi_disk_put(sdkp);
1296         }
1297 }
1298
1299
1300 #ifdef CONFIG_COMPAT
1301 /*
1302  * This gets directly called from VFS. When the ioctl
1303  * is not recognized we go back to the other translation paths.
1304  */
1305 static int sd_compat_ioctl(struct block_device *bdev, fmode_t mode,
1306                            unsigned int cmd, unsigned long arg)
1307 {
1308         struct scsi_device *sdev = scsi_disk(bdev->bd_disk)->device;
1309         int ret;
1310
1311         ret = scsi_verify_blk_ioctl(bdev, cmd);
1312         if (ret < 0)
1313                 return ret;
1314
1315         /*
1316          * If we are in the middle of error recovery, don't let anyone
1317          * else try and use this device.  Also, if error recovery fails, it
1318          * may try and take the device offline, in which case all further
1319          * access to the device is prohibited.
1320          */
1321         if (!scsi_block_when_processing_errors(sdev))
1322                 return -ENODEV;
1323
1324         if (sdev->host->hostt->compat_ioctl) {
1325                 ret = sdev->host->hostt->compat_ioctl(sdev, cmd, (void __user *)arg);
1326
1327                 return ret;
1328         }
1329
1330         /*
1331          * Let the static ioctl translation table take care of it.
1332          */
1333         return -ENOIOCTLCMD;
1334 }
1335 #endif
1336
1337 static const struct block_device_operations sd_fops = {
1338         .owner                  = THIS_MODULE,
1339         .open                   = sd_open,
1340         .release                = sd_release,
1341         .ioctl                  = sd_ioctl,
1342         .getgeo                 = sd_getgeo,
1343 #ifdef CONFIG_COMPAT
1344         .compat_ioctl           = sd_compat_ioctl,
1345 #endif
1346         .check_events           = sd_check_events,
1347         .revalidate_disk        = sd_revalidate_disk,
1348         .unlock_native_capacity = sd_unlock_native_capacity,
1349 };
1350
1351 /**
1352  *      sd_eh_action - error handling callback
1353  *      @scmd:          sd-issued command that has failed
1354  *      @eh_cmnd:       The command that was sent during error handling
1355  *      @eh_cmnd_len:   Length of eh_cmnd in bytes
1356  *      @eh_disp:       The recovery disposition suggested by the midlayer
1357  *
1358  *      This function is called by the SCSI midlayer upon completion of
1359  *      an error handling command (TEST UNIT READY, START STOP UNIT,
1360  *      etc.) The command sent to the device by the error handler is
1361  *      stored in eh_cmnd. The result of sending the eh command is
1362  *      passed in eh_disp.
1363  **/
1364 static int sd_eh_action(struct scsi_cmnd *scmd, unsigned char *eh_cmnd,
1365                         int eh_cmnd_len, int eh_disp)
1366 {
1367         struct scsi_disk *sdkp = scsi_disk(scmd->request->rq_disk);
1368
1369         if (!scsi_device_online(scmd->device) ||
1370             !scsi_medium_access_command(scmd))
1371                 return eh_disp;
1372
1373         /*
1374          * The device has timed out executing a medium access command.
1375          * However, the TEST UNIT READY command sent during error
1376          * handling completed successfully. Either the device is in the
1377          * process of recovering or has it suffered an internal failure
1378          * that prevents access to the storage medium.
1379          */
1380         if (host_byte(scmd->result) == DID_TIME_OUT && eh_disp == SUCCESS &&
1381             eh_cmnd_len && eh_cmnd[0] == TEST_UNIT_READY)
1382                 sdkp->medium_access_timed_out++;
1383
1384         /*
1385          * If the device keeps failing read/write commands but TEST UNIT
1386          * READY always completes successfully we assume that medium
1387          * access is no longer possible and take the device offline.
1388          */
1389         if (sdkp->medium_access_timed_out >= sdkp->max_medium_access_timeouts) {
1390                 scmd_printk(KERN_ERR, scmd,
1391                             "Medium access timeout failure. Offlining disk!\n");
1392                 scsi_device_set_state(scmd->device, SDEV_OFFLINE);
1393
1394                 return FAILED;
1395         }
1396
1397         return eh_disp;
1398 }
1399
1400 static unsigned int sd_completed_bytes(struct scsi_cmnd *scmd)
1401 {
1402         u64 start_lba = blk_rq_pos(scmd->request);
1403         u64 end_lba = blk_rq_pos(scmd->request) + (scsi_bufflen(scmd) / 512);
1404         u64 bad_lba;
1405         int info_valid;
1406         /*
1407          * resid is optional but mostly filled in.  When it's unused,
1408          * its value is zero, so we assume the whole buffer transferred
1409          */
1410         unsigned int transferred = scsi_bufflen(scmd) - scsi_get_resid(scmd);
1411         unsigned int good_bytes;
1412
1413         if (scmd->request->cmd_type != REQ_TYPE_FS)
1414                 return 0;
1415
1416         info_valid = scsi_get_sense_info_fld(scmd->sense_buffer,
1417                                              SCSI_SENSE_BUFFERSIZE,
1418                                              &bad_lba);
1419         if (!info_valid)
1420                 return 0;
1421
1422         if (scsi_bufflen(scmd) <= scmd->device->sector_size)
1423                 return 0;
1424
1425         if (scmd->device->sector_size < 512) {
1426                 /* only legitimate sector_size here is 256 */
1427                 start_lba <<= 1;
1428                 end_lba <<= 1;
1429         } else {
1430                 /* be careful ... don't want any overflows */
1431                 u64 factor = scmd->device->sector_size / 512;
1432                 do_div(start_lba, factor);
1433                 do_div(end_lba, factor);
1434         }
1435
1436         /* The bad lba was reported incorrectly, we have no idea where
1437          * the error is.
1438          */
1439         if (bad_lba < start_lba  || bad_lba >= end_lba)
1440                 return 0;
1441
1442         /* This computation should always be done in terms of
1443          * the resolution of the device's medium.
1444          */
1445         good_bytes = (bad_lba - start_lba) * scmd->device->sector_size;
1446         return min(good_bytes, transferred);
1447 }
1448
1449 /**
1450  *      sd_done - bottom half handler: called when the lower level
1451  *      driver has completed (successfully or otherwise) a scsi command.
1452  *      @SCpnt: mid-level's per command structure.
1453  *
1454  *      Note: potentially run from within an ISR. Must not block.
1455  **/
1456 static int sd_done(struct scsi_cmnd *SCpnt)
1457 {
1458         int result = SCpnt->result;
1459         unsigned int good_bytes = result ? 0 : scsi_bufflen(SCpnt);
1460         struct scsi_sense_hdr sshdr;
1461         struct scsi_disk *sdkp = scsi_disk(SCpnt->request->rq_disk);
1462         int sense_valid = 0;
1463         int sense_deferred = 0;
1464         unsigned char op = SCpnt->cmnd[0];
1465
1466         if ((SCpnt->request->cmd_flags & REQ_DISCARD) && !result)
1467                 scsi_set_resid(SCpnt, 0);
1468
1469         if (result) {
1470                 sense_valid = scsi_command_normalize_sense(SCpnt, &sshdr);
1471                 if (sense_valid)
1472                         sense_deferred = scsi_sense_is_deferred(&sshdr);
1473         }
1474 #ifdef CONFIG_SCSI_LOGGING
1475         SCSI_LOG_HLCOMPLETE(1, scsi_print_result(SCpnt));
1476         if (sense_valid) {
1477                 SCSI_LOG_HLCOMPLETE(1, scmd_printk(KERN_INFO, SCpnt,
1478                                                    "sd_done: sb[respc,sk,asc,"
1479                                                    "ascq]=%x,%x,%x,%x\n",
1480                                                    sshdr.response_code,
1481                                                    sshdr.sense_key, sshdr.asc,
1482                                                    sshdr.ascq));
1483         }
1484 #endif
1485         if (driver_byte(result) != DRIVER_SENSE &&
1486             (!sense_valid || sense_deferred))
1487                 goto out;
1488
1489         sdkp->medium_access_timed_out = 0;
1490
1491         switch (sshdr.sense_key) {
1492         case HARDWARE_ERROR:
1493         case MEDIUM_ERROR:
1494                 good_bytes = sd_completed_bytes(SCpnt);
1495                 break;
1496         case RECOVERED_ERROR:
1497                 good_bytes = scsi_bufflen(SCpnt);
1498                 break;
1499         case NO_SENSE:
1500                 /* This indicates a false check condition, so ignore it.  An
1501                  * unknown amount of data was transferred so treat it as an
1502                  * error.
1503                  */
1504                 scsi_print_sense("sd", SCpnt);
1505                 SCpnt->result = 0;
1506                 memset(SCpnt->sense_buffer, 0, SCSI_SENSE_BUFFERSIZE);
1507                 break;
1508         case ABORTED_COMMAND:
1509                 if (sshdr.asc == 0x10)  /* DIF: Target detected corruption */
1510                         good_bytes = sd_completed_bytes(SCpnt);
1511                 break;
1512         case ILLEGAL_REQUEST:
1513                 if (sshdr.asc == 0x10)  /* DIX: Host detected corruption */
1514                         good_bytes = sd_completed_bytes(SCpnt);
1515                 /* INVALID COMMAND OPCODE or INVALID FIELD IN CDB */
1516                 if ((sshdr.asc == 0x20 || sshdr.asc == 0x24) &&
1517                     (op == UNMAP || op == WRITE_SAME_16 || op == WRITE_SAME))
1518                         sd_config_discard(sdkp, SD_LBP_DISABLE);
1519                 break;
1520         default:
1521                 break;
1522         }
1523  out:
1524         if (rq_data_dir(SCpnt->request) == READ && scsi_prot_sg_count(SCpnt))
1525                 sd_dif_complete(SCpnt, good_bytes);
1526
1527         if (scsi_host_dif_capable(sdkp->device->host, sdkp->protection_type)
1528             == SD_DIF_TYPE2_PROTECTION && SCpnt->cmnd != SCpnt->request->cmd) {
1529
1530                 /* We have to print a failed command here as the
1531                  * extended CDB gets freed before scsi_io_completion()
1532                  * is called.
1533                  */
1534                 if (result)
1535                         scsi_print_command(SCpnt);
1536
1537                 mempool_free(SCpnt->cmnd, sd_cdb_pool);
1538                 SCpnt->cmnd = NULL;
1539                 SCpnt->cmd_len = 0;
1540         }
1541
1542         return good_bytes;
1543 }
1544
1545 /*
1546  * spinup disk - called only in sd_revalidate_disk()
1547  */
1548 static void
1549 sd_spinup_disk(struct scsi_disk *sdkp)
1550 {
1551         unsigned char cmd[10];
1552         unsigned long spintime_expire = 0;
1553         int retries, spintime;
1554         unsigned int the_result;
1555         struct scsi_sense_hdr sshdr;
1556         int sense_valid = 0;
1557
1558         spintime = 0;
1559
1560         /* Spin up drives, as required.  Only do this at boot time */
1561         /* Spinup needs to be done for module loads too. */
1562         do {
1563                 retries = 0;
1564
1565                 do {
1566                         cmd[0] = TEST_UNIT_READY;
1567                         memset((void *) &cmd[1], 0, 9);
1568
1569                         the_result = scsi_execute_req(sdkp->device, cmd,
1570                                                       DMA_NONE, NULL, 0,
1571                                                       &sshdr, SD_TIMEOUT,
1572                                                       SD_MAX_RETRIES, NULL);
1573
1574                         /*
1575                          * If the drive has indicated to us that it
1576                          * doesn't have any media in it, don't bother
1577                          * with any more polling.
1578                          */
1579                         if (media_not_present(sdkp, &sshdr))
1580                                 return;
1581
1582                         if (the_result)
1583                                 sense_valid = scsi_sense_valid(&sshdr);
1584                         retries++;
1585                 } while (retries < 3 &&
1586                          (!scsi_status_is_good(the_result) ||
1587                           ((driver_byte(the_result) & DRIVER_SENSE) &&
1588                           sense_valid && sshdr.sense_key == UNIT_ATTENTION)));
1589
1590                 if ((driver_byte(the_result) & DRIVER_SENSE) == 0) {
1591                         /* no sense, TUR either succeeded or failed
1592                          * with a status error */
1593                         if(!spintime && !scsi_status_is_good(the_result)) {
1594                                 sd_printk(KERN_NOTICE, sdkp, "Unit Not Ready\n");
1595                                 sd_print_result(sdkp, the_result);
1596                         }
1597                         break;
1598                 }
1599
1600                 /*
1601                  * The device does not want the automatic start to be issued.
1602                  */
1603                 if (sdkp->device->no_start_on_add)
1604                         break;
1605
1606                 if (sense_valid && sshdr.sense_key == NOT_READY) {
1607                         if (sshdr.asc == 4 && sshdr.ascq == 3)
1608                                 break;  /* manual intervention required */
1609                         if (sshdr.asc == 4 && sshdr.ascq == 0xb)
1610                                 break;  /* standby */
1611                         if (sshdr.asc == 4 && sshdr.ascq == 0xc)
1612                                 break;  /* unavailable */
1613                         /*
1614                          * Issue command to spin up drive when not ready
1615                          */
1616                         if (!spintime) {
1617                                 sd_printk(KERN_NOTICE, sdkp, "Spinning up disk...");
1618                                 cmd[0] = START_STOP;
1619                                 cmd[1] = 1;     /* Return immediately */
1620                                 memset((void *) &cmd[2], 0, 8);
1621                                 cmd[4] = 1;     /* Start spin cycle */
1622                                 if (sdkp->device->start_stop_pwr_cond)
1623                                         cmd[4] |= 1 << 4;
1624                                 scsi_execute_req(sdkp->device, cmd, DMA_NONE,
1625                                                  NULL, 0, &sshdr,
1626                                                  SD_TIMEOUT, SD_MAX_RETRIES,
1627                                                  NULL);
1628                                 spintime_expire = jiffies + 100 * HZ;
1629                                 spintime = 1;
1630                         }
1631                         /* Wait 1 second for next try */
1632                         msleep(1000);
1633                         printk(".");
1634
1635                 /*
1636                  * Wait for USB flash devices with slow firmware.
1637                  * Yes, this sense key/ASC combination shouldn't
1638                  * occur here.  It's characteristic of these devices.
1639                  */
1640                 } else if (sshdr.sense_key == UNIT_ATTENTION &&
1641                                 sshdr.asc == 0x28) {
1642                         if (!spintime) {
1643                                 spintime_expire = jiffies + 5 * HZ;
1644                                 spintime = 1;
1645                         }
1646                         /* Wait 1 second for next try */
1647                         msleep(1000);
1648                 } else {
1649                         /* we don't understand the sense code, so it's
1650                          * probably pointless to loop */
1651                         if(!spintime) {
1652                                 sd_printk(KERN_NOTICE, sdkp, "Unit Not Ready\n");
1653                                 sd_print_sense_hdr(sdkp, &sshdr);
1654                         }
1655                         break;
1656                 }
1657
1658         } while (spintime && time_before_eq(jiffies, spintime_expire));
1659
1660         if (spintime) {
1661                 if (scsi_status_is_good(the_result))
1662                         printk("ready\n");
1663                 else
1664                         printk("not responding...\n");
1665         }
1666 }
1667
1668
1669 /*
1670  * Determine whether disk supports Data Integrity Field.
1671  */
1672 static void sd_read_protection_type(struct scsi_disk *sdkp, unsigned char *buffer)
1673 {
1674         struct scsi_device *sdp = sdkp->device;
1675         u8 type;
1676
1677         if (scsi_device_protection(sdp) == 0 || (buffer[12] & 1) == 0)
1678                 return;
1679
1680         type = ((buffer[12] >> 1) & 7) + 1; /* P_TYPE 0 = Type 1 */
1681
1682         if (type == sdkp->protection_type || !sdkp->first_scan)
1683                 return;
1684
1685         sdkp->protection_type = type;
1686
1687         if (type > SD_DIF_TYPE3_PROTECTION) {
1688                 sd_printk(KERN_ERR, sdkp, "formatted with unsupported " \
1689                           "protection type %u. Disabling disk!\n", type);
1690                 sdkp->capacity = 0;
1691                 return;
1692         }
1693
1694         if (scsi_host_dif_capable(sdp->host, type))
1695                 sd_printk(KERN_NOTICE, sdkp,
1696                           "Enabling DIF Type %u protection\n", type);
1697         else
1698                 sd_printk(KERN_NOTICE, sdkp,
1699                           "Disabling DIF Type %u protection\n", type);
1700 }
1701
1702 static void read_capacity_error(struct scsi_disk *sdkp, struct scsi_device *sdp,
1703                         struct scsi_sense_hdr *sshdr, int sense_valid,
1704                         int the_result)
1705 {
1706         sd_print_result(sdkp, the_result);
1707         if (driver_byte(the_result) & DRIVER_SENSE)
1708                 sd_print_sense_hdr(sdkp, sshdr);
1709         else
1710                 sd_printk(KERN_NOTICE, sdkp, "Sense not available.\n");
1711
1712         /*
1713          * Set dirty bit for removable devices if not ready -
1714          * sometimes drives will not report this properly.
1715          */
1716         if (sdp->removable &&
1717             sense_valid && sshdr->sense_key == NOT_READY)
1718                 set_media_not_present(sdkp);
1719
1720         /*
1721          * We used to set media_present to 0 here to indicate no media
1722          * in the drive, but some drives fail read capacity even with
1723          * media present, so we can't do that.
1724          */
1725         sdkp->capacity = 0; /* unknown mapped to zero - as usual */
1726 }
1727
1728 #define RC16_LEN 32
1729 #if RC16_LEN > SD_BUF_SIZE
1730 #error RC16_LEN must not be more than SD_BUF_SIZE
1731 #endif
1732
1733 #define READ_CAPACITY_RETRIES_ON_RESET  10
1734
1735 static int read_capacity_16(struct scsi_disk *sdkp, struct scsi_device *sdp,
1736                                                 unsigned char *buffer)
1737 {
1738         unsigned char cmd[16];
1739         struct scsi_sense_hdr sshdr;
1740         int sense_valid = 0;
1741         int the_result;
1742         int retries = 3, reset_retries = READ_CAPACITY_RETRIES_ON_RESET;
1743         unsigned int alignment;
1744         unsigned long long lba;
1745         unsigned sector_size;
1746
1747         if (sdp->no_read_capacity_16)
1748                 return -EINVAL;
1749
1750         do {
1751                 memset(cmd, 0, 16);
1752                 cmd[0] = SERVICE_ACTION_IN;
1753                 cmd[1] = SAI_READ_CAPACITY_16;
1754                 cmd[13] = RC16_LEN;
1755                 memset(buffer, 0, RC16_LEN);
1756
1757                 the_result = scsi_execute_req(sdp, cmd, DMA_FROM_DEVICE,
1758                                         buffer, RC16_LEN, &sshdr,
1759                                         SD_TIMEOUT, SD_MAX_RETRIES, NULL);
1760
1761                 if (media_not_present(sdkp, &sshdr))
1762                         return -ENODEV;
1763
1764                 if (the_result) {
1765                         sense_valid = scsi_sense_valid(&sshdr);
1766                         if (sense_valid &&
1767                             sshdr.sense_key == ILLEGAL_REQUEST &&
1768                             (sshdr.asc == 0x20 || sshdr.asc == 0x24) &&
1769                             sshdr.ascq == 0x00)
1770                                 /* Invalid Command Operation Code or
1771                                  * Invalid Field in CDB, just retry
1772                                  * silently with RC10 */
1773                                 return -EINVAL;
1774                         if (sense_valid &&
1775                             sshdr.sense_key == UNIT_ATTENTION &&
1776                             sshdr.asc == 0x29 && sshdr.ascq == 0x00)
1777                                 /* Device reset might occur several times,
1778                                  * give it one more chance */
1779                                 if (--reset_retries > 0)
1780                                         continue;
1781                 }
1782                 retries--;
1783
1784         } while (the_result && retries);
1785
1786         if (the_result) {
1787                 sd_printk(KERN_NOTICE, sdkp, "READ CAPACITY(16) failed\n");
1788                 read_capacity_error(sdkp, sdp, &sshdr, sense_valid, the_result);
1789                 return -EINVAL;
1790         }
1791
1792         sector_size = get_unaligned_be32(&buffer[8]);
1793         lba = get_unaligned_be64(&buffer[0]);
1794
1795         sd_read_protection_type(sdkp, buffer);
1796
1797         if ((sizeof(sdkp->capacity) == 4) && (lba >= 0xffffffffULL)) {
1798                 sd_printk(KERN_ERR, sdkp, "Too big for this kernel. Use a "
1799                         "kernel compiled with support for large block "
1800                         "devices.\n");
1801                 sdkp->capacity = 0;
1802                 return -EOVERFLOW;
1803         }
1804
1805         /* Logical blocks per physical block exponent */
1806         sdkp->physical_block_size = (1 << (buffer[13] & 0xf)) * sector_size;
1807
1808         /* Lowest aligned logical block */
1809         alignment = ((buffer[14] & 0x3f) << 8 | buffer[15]) * sector_size;
1810         blk_queue_alignment_offset(sdp->request_queue, alignment);
1811         if (alignment && sdkp->first_scan)
1812                 sd_printk(KERN_NOTICE, sdkp,
1813                           "physical block alignment offset: %u\n", alignment);
1814
1815         if (buffer[14] & 0x80) { /* LBPME */
1816                 sdkp->lbpme = 1;
1817
1818                 if (buffer[14] & 0x40) /* LBPRZ */
1819                         sdkp->lbprz = 1;
1820
1821                 sd_config_discard(sdkp, SD_LBP_WS16);
1822         }
1823
1824         sdkp->capacity = lba + 1;
1825         return sector_size;
1826 }
1827
1828 static int read_capacity_10(struct scsi_disk *sdkp, struct scsi_device *sdp,
1829                                                 unsigned char *buffer)
1830 {
1831         unsigned char cmd[16];
1832         struct scsi_sense_hdr sshdr;
1833         int sense_valid = 0;
1834         int the_result;
1835         int retries = 3, reset_retries = READ_CAPACITY_RETRIES_ON_RESET;
1836         sector_t lba;
1837         unsigned sector_size;
1838
1839         do {
1840                 cmd[0] = READ_CAPACITY;
1841                 memset(&cmd[1], 0, 9);
1842                 memset(buffer, 0, 8);
1843
1844                 the_result = scsi_execute_req(sdp, cmd, DMA_FROM_DEVICE,
1845                                         buffer, 8, &sshdr,
1846                                         SD_TIMEOUT, SD_MAX_RETRIES, NULL);
1847
1848                 if (media_not_present(sdkp, &sshdr))
1849                         return -ENODEV;
1850
1851                 if (the_result) {
1852                         sense_valid = scsi_sense_valid(&sshdr);
1853                         if (sense_valid &&
1854                             sshdr.sense_key == UNIT_ATTENTION &&
1855                             sshdr.asc == 0x29 && sshdr.ascq == 0x00)
1856                                 /* Device reset might occur several times,
1857                                  * give it one more chance */
1858                                 if (--reset_retries > 0)
1859                                         continue;
1860                 }
1861                 retries--;
1862
1863         } while (the_result && retries);
1864
1865         if (the_result) {
1866                 sd_printk(KERN_NOTICE, sdkp, "READ CAPACITY failed\n");
1867                 read_capacity_error(sdkp, sdp, &sshdr, sense_valid, the_result);
1868                 return -EINVAL;
1869         }
1870
1871         sector_size = get_unaligned_be32(&buffer[4]);
1872         lba = get_unaligned_be32(&buffer[0]);
1873
1874         if (sdp->no_read_capacity_16 && (lba == 0xffffffff)) {
1875                 /* Some buggy (usb cardreader) devices return an lba of
1876                    0xffffffff when the want to report a size of 0 (with
1877                    which they really mean no media is present) */
1878                 sdkp->capacity = 0;
1879                 sdkp->physical_block_size = sector_size;
1880                 return sector_size;
1881         }
1882
1883         if ((sizeof(sdkp->capacity) == 4) && (lba == 0xffffffff)) {
1884                 sd_printk(KERN_ERR, sdkp, "Too big for this kernel. Use a "
1885                         "kernel compiled with support for large block "
1886                         "devices.\n");
1887                 sdkp->capacity = 0;
1888                 return -EOVERFLOW;
1889         }
1890
1891         sdkp->capacity = lba + 1;
1892         sdkp->physical_block_size = sector_size;
1893         return sector_size;
1894 }
1895
1896 static int sd_try_rc16_first(struct scsi_device *sdp)
1897 {
1898         if (sdp->host->max_cmd_len < 16)
1899                 return 0;
1900         if (sdp->scsi_level > SCSI_SPC_2)
1901                 return 1;
1902         if (scsi_device_protection(sdp))
1903                 return 1;
1904         return 0;
1905 }
1906
1907 /*
1908  * read disk capacity
1909  */
1910 static void
1911 sd_read_capacity(struct scsi_disk *sdkp, unsigned char *buffer)
1912 {
1913         int sector_size;
1914         struct scsi_device *sdp = sdkp->device;
1915         sector_t old_capacity = sdkp->capacity;
1916
1917         if (sd_try_rc16_first(sdp)) {
1918                 sector_size = read_capacity_16(sdkp, sdp, buffer);
1919                 if (sector_size == -EOVERFLOW)
1920                         goto got_data;
1921                 if (sector_size == -ENODEV)
1922                         return;
1923                 if (sector_size < 0)
1924                         sector_size = read_capacity_10(sdkp, sdp, buffer);
1925                 if (sector_size < 0)
1926                         return;
1927         } else {
1928                 sector_size = read_capacity_10(sdkp, sdp, buffer);
1929                 if (sector_size == -EOVERFLOW)
1930                         goto got_data;
1931                 if (sector_size < 0)
1932                         return;
1933                 if ((sizeof(sdkp->capacity) > 4) &&
1934                     (sdkp->capacity > 0xffffffffULL)) {
1935                         int old_sector_size = sector_size;
1936                         sd_printk(KERN_NOTICE, sdkp, "Very big device. "
1937                                         "Trying to use READ CAPACITY(16).\n");
1938                         sector_size = read_capacity_16(sdkp, sdp, buffer);
1939                         if (sector_size < 0) {
1940                                 sd_printk(KERN_NOTICE, sdkp,
1941                                         "Using 0xffffffff as device size\n");
1942                                 sdkp->capacity = 1 + (sector_t) 0xffffffff;
1943                                 sector_size = old_sector_size;
1944                                 goto got_data;
1945                         }
1946                 }
1947         }
1948
1949         /* Some devices are known to return the total number of blocks,
1950          * not the highest block number.  Some devices have versions
1951          * which do this and others which do not.  Some devices we might
1952          * suspect of doing this but we don't know for certain.
1953          *
1954          * If we know the reported capacity is wrong, decrement it.  If
1955          * we can only guess, then assume the number of blocks is even
1956          * (usually true but not always) and err on the side of lowering
1957          * the capacity.
1958          */
1959         if (sdp->fix_capacity ||
1960             (sdp->guess_capacity && (sdkp->capacity & 0x01))) {
1961                 sd_printk(KERN_INFO, sdkp, "Adjusting the sector count "
1962                                 "from its reported value: %llu\n",
1963                                 (unsigned long long) sdkp->capacity);
1964                 --sdkp->capacity;
1965         }
1966
1967 got_data:
1968         if (sector_size == 0) {
1969                 sector_size = 512;
1970                 sd_printk(KERN_NOTICE, sdkp, "Sector size 0 reported, "
1971                           "assuming 512.\n");
1972         }
1973
1974         if (sector_size != 512 &&
1975             sector_size != 1024 &&
1976             sector_size != 2048 &&
1977             sector_size != 4096 &&
1978             sector_size != 256) {
1979                 sd_printk(KERN_NOTICE, sdkp, "Unsupported sector size %d.\n",
1980                           sector_size);
1981                 /*
1982                  * The user might want to re-format the drive with
1983                  * a supported sectorsize.  Once this happens, it
1984                  * would be relatively trivial to set the thing up.
1985                  * For this reason, we leave the thing in the table.
1986                  */
1987                 sdkp->capacity = 0;
1988                 /*
1989                  * set a bogus sector size so the normal read/write
1990                  * logic in the block layer will eventually refuse any
1991                  * request on this device without tripping over power
1992                  * of two sector size assumptions
1993                  */
1994                 sector_size = 512;
1995         }
1996         blk_queue_logical_block_size(sdp->request_queue, sector_size);
1997
1998         {
1999                 char cap_str_2[10], cap_str_10[10];
2000                 u64 sz = (u64)sdkp->capacity << ilog2(sector_size);
2001
2002                 string_get_size(sz, STRING_UNITS_2, cap_str_2,
2003                                 sizeof(cap_str_2));
2004                 string_get_size(sz, STRING_UNITS_10, cap_str_10,
2005                                 sizeof(cap_str_10));
2006
2007                 if (sdkp->first_scan || old_capacity != sdkp->capacity) {
2008                         sd_printk(KERN_NOTICE, sdkp,
2009                                   "%llu %d-byte logical blocks: (%s/%s)\n",
2010                                   (unsigned long long)sdkp->capacity,
2011                                   sector_size, cap_str_10, cap_str_2);
2012
2013                         if (sdkp->physical_block_size != sector_size)
2014                                 sd_printk(KERN_NOTICE, sdkp,
2015                                           "%u-byte physical blocks\n",
2016                                           sdkp->physical_block_size);
2017                 }
2018         }
2019
2020         /* Rescale capacity to 512-byte units */
2021         if (sector_size == 4096)
2022                 sdkp->capacity <<= 3;
2023         else if (sector_size == 2048)
2024                 sdkp->capacity <<= 2;
2025         else if (sector_size == 1024)
2026                 sdkp->capacity <<= 1;
2027         else if (sector_size == 256)
2028                 sdkp->capacity >>= 1;
2029
2030         blk_queue_physical_block_size(sdp->request_queue,
2031                                       sdkp->physical_block_size);
2032         sdkp->device->sector_size = sector_size;
2033 }
2034
2035 /* called with buffer of length 512 */
2036 static inline int
2037 sd_do_mode_sense(struct scsi_device *sdp, int dbd, int modepage,
2038                  unsigned char *buffer, int len, struct scsi_mode_data *data,
2039                  struct scsi_sense_hdr *sshdr)
2040 {
2041         return scsi_mode_sense(sdp, dbd, modepage, buffer, len,
2042                                SD_TIMEOUT, SD_MAX_RETRIES, data,
2043                                sshdr);
2044 }
2045
2046 /*
2047  * read write protect setting, if possible - called only in sd_revalidate_disk()
2048  * called with buffer of length SD_BUF_SIZE
2049  */
2050 static void
2051 sd_read_write_protect_flag(struct scsi_disk *sdkp, unsigned char *buffer)
2052 {
2053         int res;
2054         struct scsi_device *sdp = sdkp->device;
2055         struct scsi_mode_data data;
2056         int old_wp = sdkp->write_prot;
2057
2058         set_disk_ro(sdkp->disk, 0);
2059         if (sdp->skip_ms_page_3f) {
2060                 sd_printk(KERN_NOTICE, sdkp, "Assuming Write Enabled\n");
2061                 return;
2062         }
2063
2064         if (sdp->use_192_bytes_for_3f) {
2065                 res = sd_do_mode_sense(sdp, 0, 0x3F, buffer, 192, &data, NULL);
2066         } else {
2067                 /*
2068                  * First attempt: ask for all pages (0x3F), but only 4 bytes.
2069                  * We have to start carefully: some devices hang if we ask
2070                  * for more than is available.
2071                  */
2072                 res = sd_do_mode_sense(sdp, 0, 0x3F, buffer, 4, &data, NULL);
2073
2074                 /*
2075                  * Second attempt: ask for page 0 When only page 0 is
2076                  * implemented, a request for page 3F may return Sense Key
2077                  * 5: Illegal Request, Sense Code 24: Invalid field in
2078                  * CDB.
2079                  */
2080                 if (!scsi_status_is_good(res))
2081                         res = sd_do_mode_sense(sdp, 0, 0, buffer, 4, &data, NULL);
2082
2083                 /*
2084                  * Third attempt: ask 255 bytes, as we did earlier.
2085                  */
2086                 if (!scsi_status_is_good(res))
2087                         res = sd_do_mode_sense(sdp, 0, 0x3F, buffer, 255,
2088                                                &data, NULL);
2089         }
2090
2091         if (!scsi_status_is_good(res)) {
2092                 sd_printk(KERN_WARNING, sdkp,
2093                           "Test WP failed, assume Write Enabled\n");
2094         } else {
2095                 sdkp->write_prot = ((data.device_specific & 0x80) != 0);
2096                 set_disk_ro(sdkp->disk, sdkp->write_prot);
2097                 if (sdkp->first_scan || old_wp != sdkp->write_prot) {
2098                         sd_printk(KERN_NOTICE, sdkp, "Write Protect is %s\n",
2099                                   sdkp->write_prot ? "on" : "off");
2100                         sd_printk(KERN_DEBUG, sdkp,
2101                                   "Mode Sense: %02x %02x %02x %02x\n",
2102                                   buffer[0], buffer[1], buffer[2], buffer[3]);
2103                 }
2104         }
2105 }
2106
2107 /*
2108  * sd_read_cache_type - called only from sd_revalidate_disk()
2109  * called with buffer of length SD_BUF_SIZE
2110  */
2111 static void
2112 sd_read_cache_type(struct scsi_disk *sdkp, unsigned char *buffer)
2113 {
2114         int len = 0, res;
2115         struct scsi_device *sdp = sdkp->device;
2116
2117         int dbd;
2118         int modepage;
2119         int first_len;
2120         struct scsi_mode_data data;
2121         struct scsi_sense_hdr sshdr;
2122         int old_wce = sdkp->WCE;
2123         int old_rcd = sdkp->RCD;
2124         int old_dpofua = sdkp->DPOFUA;
2125
2126         first_len = 4;
2127         if (sdp->skip_ms_page_8) {
2128                 if (sdp->type == TYPE_RBC)
2129                         goto defaults;
2130                 else {
2131                         if (sdp->skip_ms_page_3f)
2132                                 goto defaults;
2133                         modepage = 0x3F;
2134                         if (sdp->use_192_bytes_for_3f)
2135                                 first_len = 192;
2136                         dbd = 0;
2137                 }
2138         } else if (sdp->type == TYPE_RBC) {
2139                 modepage = 6;
2140                 dbd = 8;
2141         } else {
2142                 modepage = 8;
2143                 dbd = 0;
2144         }
2145
2146         /* cautiously ask */
2147         res = sd_do_mode_sense(sdp, dbd, modepage, buffer, first_len,
2148                         &data, &sshdr);
2149
2150         if (!scsi_status_is_good(res))
2151                 goto bad_sense;
2152
2153         if (!data.header_length) {
2154                 modepage = 6;
2155                 first_len = 0;
2156                 sd_printk(KERN_ERR, sdkp, "Missing header in MODE_SENSE response\n");
2157         }
2158
2159         /* that went OK, now ask for the proper length */
2160         len = data.length;
2161
2162         /*
2163          * We're only interested in the first three bytes, actually.
2164          * But the data cache page is defined for the first 20.
2165          */
2166         if (len < 3)
2167                 goto bad_sense;
2168         else if (len > SD_BUF_SIZE) {
2169                 sd_printk(KERN_NOTICE, sdkp, "Truncating mode parameter "
2170                           "data from %d to %d bytes\n", len, SD_BUF_SIZE);
2171                 len = SD_BUF_SIZE;
2172         }
2173         if (modepage == 0x3F && sdp->use_192_bytes_for_3f)
2174                 len = 192;
2175
2176         /* Get the data */
2177         if (len > first_len)
2178                 res = sd_do_mode_sense(sdp, dbd, modepage, buffer, len,
2179                                 &data, &sshdr);
2180
2181         if (scsi_status_is_good(res)) {
2182                 int offset = data.header_length + data.block_descriptor_length;
2183
2184                 while (offset < len) {
2185                         u8 page_code = buffer[offset] & 0x3F;
2186                         u8 spf       = buffer[offset] & 0x40;
2187
2188                         if (page_code == 8 || page_code == 6) {
2189                                 /* We're interested only in the first 3 bytes.
2190                                  */
2191                                 if (len - offset <= 2) {
2192                                         sd_printk(KERN_ERR, sdkp, "Incomplete "
2193                                                   "mode parameter data\n");
2194                                         goto defaults;
2195                                 } else {
2196                                         modepage = page_code;
2197                                         goto Page_found;
2198                                 }
2199                         } else {
2200                                 /* Go to the next page */
2201                                 if (spf && len - offset > 3)
2202                                         offset += 4 + (buffer[offset+2] << 8) +
2203                                                 buffer[offset+3];
2204                                 else if (!spf && len - offset > 1)
2205                                         offset += 2 + buffer[offset+1];
2206                                 else {
2207                                         sd_printk(KERN_ERR, sdkp, "Incomplete "
2208                                                   "mode parameter data\n");
2209                                         goto defaults;
2210                                 }
2211                         }
2212                 }
2213
2214                 if (modepage == 0x3F) {
2215                         sd_printk(KERN_ERR, sdkp, "No Caching mode page "
2216                                   "present\n");
2217                         goto defaults;
2218                 } else if ((buffer[offset] & 0x3f) != modepage) {
2219                         sd_printk(KERN_ERR, sdkp, "Got wrong page\n");
2220                         goto defaults;
2221                 }
2222         Page_found:
2223                 if (modepage == 8) {
2224                         sdkp->WCE = ((buffer[offset + 2] & 0x04) != 0);
2225                         sdkp->RCD = ((buffer[offset + 2] & 0x01) != 0);
2226                 } else {
2227                         sdkp->WCE = ((buffer[offset + 2] & 0x01) == 0);
2228                         sdkp->RCD = 0;
2229                 }
2230
2231                 sdkp->DPOFUA = (data.device_specific & 0x10) != 0;
2232                 if (sdkp->DPOFUA && !sdkp->device->use_10_for_rw) {
2233                         sd_printk(KERN_NOTICE, sdkp,
2234                                   "Uses READ/WRITE(6), disabling FUA\n");
2235                         sdkp->DPOFUA = 0;
2236                 }
2237
2238                 if (sdkp->first_scan || old_wce != sdkp->WCE ||
2239                     old_rcd != sdkp->RCD || old_dpofua != sdkp->DPOFUA)
2240                         sd_printk(KERN_NOTICE, sdkp,
2241                                   "Write cache: %s, read cache: %s, %s\n",
2242                                   sdkp->WCE ? "enabled" : "disabled",
2243                                   sdkp->RCD ? "disabled" : "enabled",
2244                                   sdkp->DPOFUA ? "supports DPO and FUA"
2245                                   : "doesn't support DPO or FUA");
2246
2247                 return;
2248         }
2249
2250 bad_sense:
2251         if (scsi_sense_valid(&sshdr) &&
2252             sshdr.sense_key == ILLEGAL_REQUEST &&
2253             sshdr.asc == 0x24 && sshdr.ascq == 0x0)
2254                 /* Invalid field in CDB */
2255                 sd_printk(KERN_NOTICE, sdkp, "Cache data unavailable\n");
2256         else
2257                 sd_printk(KERN_ERR, sdkp, "Asking for cache data failed\n");
2258
2259 defaults:
2260         sd_printk(KERN_ERR, sdkp, "Assuming drive cache: write through\n");
2261         sdkp->WCE = 0;
2262         sdkp->RCD = 0;
2263         sdkp->DPOFUA = 0;
2264 }
2265
2266 /*
2267  * The ATO bit indicates whether the DIF application tag is available
2268  * for use by the operating system.
2269  */
2270 static void sd_read_app_tag_own(struct scsi_disk *sdkp, unsigned char *buffer)
2271 {
2272         int res, offset;
2273         struct scsi_device *sdp = sdkp->device;
2274         struct scsi_mode_data data;
2275         struct scsi_sense_hdr sshdr;
2276
2277         if (sdp->type != TYPE_DISK)
2278                 return;
2279
2280         if (sdkp->protection_type == 0)
2281                 return;
2282
2283         res = scsi_mode_sense(sdp, 1, 0x0a, buffer, 36, SD_TIMEOUT,
2284                               SD_MAX_RETRIES, &data, &sshdr);
2285
2286         if (!scsi_status_is_good(res) || !data.header_length ||
2287             data.length < 6) {
2288                 sd_printk(KERN_WARNING, sdkp,
2289                           "getting Control mode page failed, assume no ATO\n");
2290
2291                 if (scsi_sense_valid(&sshdr))
2292                         sd_print_sense_hdr(sdkp, &sshdr);
2293
2294                 return;
2295         }
2296
2297         offset = data.header_length + data.block_descriptor_length;
2298
2299         if ((buffer[offset] & 0x3f) != 0x0a) {
2300                 sd_printk(KERN_ERR, sdkp, "ATO Got wrong page\n");
2301                 return;
2302         }
2303
2304         if ((buffer[offset + 5] & 0x80) == 0)
2305                 return;
2306
2307         sdkp->ATO = 1;
2308
2309         return;
2310 }
2311
2312 /**
2313  * sd_read_block_limits - Query disk device for preferred I/O sizes.
2314  * @disk: disk to query
2315  */
2316 static void sd_read_block_limits(struct scsi_disk *sdkp)
2317 {
2318         unsigned int sector_sz = sdkp->device->sector_size;
2319         const int vpd_len = 64;
2320         unsigned char *buffer = kmalloc(vpd_len, GFP_KERNEL);
2321
2322         if (!buffer ||
2323             /* Block Limits VPD */
2324             scsi_get_vpd_page(sdkp->device, 0xb0, buffer, vpd_len))
2325                 goto out;
2326
2327         blk_queue_io_min(sdkp->disk->queue,
2328                          get_unaligned_be16(&buffer[6]) * sector_sz);
2329         blk_queue_io_opt(sdkp->disk->queue,
2330                          get_unaligned_be32(&buffer[12]) * sector_sz);
2331
2332         if (buffer[3] == 0x3c) {
2333                 unsigned int lba_count, desc_count;
2334
2335                 sdkp->max_ws_blocks =
2336                         (u32) min_not_zero(get_unaligned_be64(&buffer[36]),
2337                                            (u64)0xffffffff);
2338
2339                 if (!sdkp->lbpme)
2340                         goto out;
2341
2342                 lba_count = get_unaligned_be32(&buffer[20]);
2343                 desc_count = get_unaligned_be32(&buffer[24]);
2344
2345                 if (lba_count && desc_count)
2346                         sdkp->max_unmap_blocks = lba_count;
2347
2348                 sdkp->unmap_granularity = get_unaligned_be32(&buffer[28]);
2349
2350                 if (buffer[32] & 0x80)
2351                         sdkp->unmap_alignment =
2352                                 get_unaligned_be32(&buffer[32]) & ~(1 << 31);
2353
2354                 if (!sdkp->lbpvpd) { /* LBP VPD page not provided */
2355
2356                         if (sdkp->max_unmap_blocks)
2357                                 sd_config_discard(sdkp, SD_LBP_UNMAP);
2358                         else
2359                                 sd_config_discard(sdkp, SD_LBP_WS16);
2360
2361                 } else {        /* LBP VPD page tells us what to use */
2362
2363                         if (sdkp->lbpu && sdkp->max_unmap_blocks)
2364                                 sd_config_discard(sdkp, SD_LBP_UNMAP);
2365                         else if (sdkp->lbpws)
2366                                 sd_config_discard(sdkp, SD_LBP_WS16);
2367                         else if (sdkp->lbpws10)
2368                                 sd_config_discard(sdkp, SD_LBP_WS10);
2369                         else
2370                                 sd_config_discard(sdkp, SD_LBP_DISABLE);
2371                 }
2372         }
2373
2374  out:
2375         kfree(buffer);
2376 }
2377
2378 /**
2379  * sd_read_block_characteristics - Query block dev. characteristics
2380  * @disk: disk to query
2381  */
2382 static void sd_read_block_characteristics(struct scsi_disk *sdkp)
2383 {
2384         unsigned char *buffer;
2385         u16 rot;
2386         const int vpd_len = 64;
2387
2388         buffer = kmalloc(vpd_len, GFP_KERNEL);
2389
2390         if (!buffer ||
2391             /* Block Device Characteristics VPD */
2392             scsi_get_vpd_page(sdkp->device, 0xb1, buffer, vpd_len))
2393                 goto out;
2394
2395         rot = get_unaligned_be16(&buffer[4]);
2396
2397         if (rot == 1)
2398                 queue_flag_set_unlocked(QUEUE_FLAG_NONROT, sdkp->disk->queue);
2399
2400  out:
2401         kfree(buffer);
2402 }
2403
2404 /**
2405  * sd_read_block_provisioning - Query provisioning VPD page
2406  * @disk: disk to query
2407  */
2408 static void sd_read_block_provisioning(struct scsi_disk *sdkp)
2409 {
2410         unsigned char *buffer;
2411         const int vpd_len = 8;
2412
2413         if (sdkp->lbpme == 0)
2414                 return;
2415
2416         buffer = kmalloc(vpd_len, GFP_KERNEL);
2417
2418         if (!buffer || scsi_get_vpd_page(sdkp->device, 0xb2, buffer, vpd_len))
2419                 goto out;
2420
2421         sdkp->lbpvpd    = 1;
2422         sdkp->lbpu      = (buffer[5] >> 7) & 1; /* UNMAP */
2423         sdkp->lbpws     = (buffer[5] >> 6) & 1; /* WRITE SAME(16) with UNMAP */
2424         sdkp->lbpws10   = (buffer[5] >> 5) & 1; /* WRITE SAME(10) with UNMAP */
2425
2426  out:
2427         kfree(buffer);
2428 }
2429
2430 static int sd_try_extended_inquiry(struct scsi_device *sdp)
2431 {
2432         /*
2433          * Although VPD inquiries can go to SCSI-2 type devices,
2434          * some USB ones crash on receiving them, and the pages
2435          * we currently ask for are for SPC-3 and beyond
2436          */
2437         if (sdp->scsi_level > SCSI_SPC_2 && !sdp->skip_vpd_pages)
2438                 return 1;
2439         return 0;
2440 }
2441
2442 /**
2443  *      sd_revalidate_disk - called the first time a new disk is seen,
2444  *      performs disk spin up, read_capacity, etc.
2445  *      @disk: struct gendisk we care about
2446  **/
2447 static int sd_revalidate_disk(struct gendisk *disk)
2448 {
2449         struct scsi_disk *sdkp = scsi_disk(disk);
2450         struct scsi_device *sdp = sdkp->device;
2451         unsigned char *buffer;
2452         unsigned flush = 0;
2453
2454         SCSI_LOG_HLQUEUE(3, sd_printk(KERN_INFO, sdkp,
2455                                       "sd_revalidate_disk\n"));
2456
2457         /*
2458          * If the device is offline, don't try and read capacity or any
2459          * of the other niceties.
2460          */
2461         if (!scsi_device_online(sdp))
2462                 goto out;
2463
2464         buffer = kmalloc(SD_BUF_SIZE, GFP_KERNEL);
2465         if (!buffer) {
2466                 sd_printk(KERN_WARNING, sdkp, "sd_revalidate_disk: Memory "
2467                           "allocation failure.\n");
2468                 goto out;
2469         }
2470
2471         sd_spinup_disk(sdkp);
2472
2473         /*
2474          * Without media there is no reason to ask; moreover, some devices
2475          * react badly if we do.
2476          */
2477         if (sdkp->media_present) {
2478                 sd_read_capacity(sdkp, buffer);
2479
2480                 if (sd_try_extended_inquiry(sdp)) {
2481                         sd_read_block_provisioning(sdkp);
2482                         sd_read_block_limits(sdkp);
2483                         sd_read_block_characteristics(sdkp);
2484                 }
2485
2486                 sd_read_write_protect_flag(sdkp, buffer);
2487                 sd_read_cache_type(sdkp, buffer);
2488                 sd_read_app_tag_own(sdkp, buffer);
2489         }
2490
2491         sdkp->first_scan = 0;
2492
2493         /*
2494          * We now have all cache related info, determine how we deal
2495          * with flush requests.
2496          */
2497         if (sdkp->WCE) {
2498                 flush |= REQ_FLUSH;
2499                 if (sdkp->DPOFUA)
2500                         flush |= REQ_FUA;
2501         }
2502
2503         blk_queue_flush(sdkp->disk->queue, flush);
2504
2505         set_capacity(disk, sdkp->capacity);
2506         kfree(buffer);
2507
2508  out:
2509         return 0;
2510 }
2511
2512 /**
2513  *      sd_unlock_native_capacity - unlock native capacity
2514  *      @disk: struct gendisk to set capacity for
2515  *
2516  *      Block layer calls this function if it detects that partitions
2517  *      on @disk reach beyond the end of the device.  If the SCSI host
2518  *      implements ->unlock_native_capacity() method, it's invoked to
2519  *      give it a chance to adjust the device capacity.
2520  *
2521  *      CONTEXT:
2522  *      Defined by block layer.  Might sleep.
2523  */
2524 static void sd_unlock_native_capacity(struct gendisk *disk)
2525 {
2526         struct scsi_device *sdev = scsi_disk(disk)->device;
2527
2528         if (sdev->host->hostt->unlock_native_capacity)
2529                 sdev->host->hostt->unlock_native_capacity(sdev);
2530 }
2531
2532 /**
2533  *      sd_format_disk_name - format disk name
2534  *      @prefix: name prefix - ie. "sd" for SCSI disks
2535  *      @index: index of the disk to format name for
2536  *      @buf: output buffer
2537  *      @buflen: length of the output buffer
2538  *
2539  *      SCSI disk names starts at sda.  The 26th device is sdz and the
2540  *      27th is sdaa.  The last one for two lettered suffix is sdzz
2541  *      which is followed by sdaaa.
2542  *
2543  *      This is basically 26 base counting with one extra 'nil' entry
2544  *      at the beginning from the second digit on and can be
2545  *      determined using similar method as 26 base conversion with the
2546  *      index shifted -1 after each digit is computed.
2547  *
2548  *      CONTEXT:
2549  *      Don't care.
2550  *
2551  *      RETURNS:
2552  *      0 on success, -errno on failure.
2553  */
2554 static int sd_format_disk_name(char *prefix, int index, char *buf, int buflen)
2555 {
2556         const int base = 'z' - 'a' + 1;
2557         char *begin = buf + strlen(prefix);
2558         char *end = buf + buflen;
2559         char *p;
2560         int unit;
2561
2562         p = end - 1;
2563         *p = '\0';
2564         unit = base;
2565         do {
2566                 if (p == begin)
2567                         return -EINVAL;
2568                 *--p = 'a' + (index % unit);
2569                 index = (index / unit) - 1;
2570         } while (index >= 0);
2571
2572         memmove(begin, p, end - p);
2573         memcpy(buf, prefix, strlen(prefix));
2574
2575         return 0;
2576 }
2577
2578 /*
2579  * The asynchronous part of sd_probe
2580  */
2581 static void sd_probe_async(void *data, async_cookie_t cookie)
2582 {
2583         struct scsi_disk *sdkp = data;
2584         struct scsi_device *sdp;
2585         struct gendisk *gd;
2586         u32 index;
2587         struct device *dev;
2588
2589         sdp = sdkp->device;
2590         gd = sdkp->disk;
2591         index = sdkp->index;
2592         dev = &sdp->sdev_gendev;
2593
2594         gd->major = sd_major((index & 0xf0) >> 4);
2595         gd->first_minor = ((index & 0xf) << 4) | (index & 0xfff00);
2596         gd->minors = SD_MINORS;
2597
2598         gd->fops = &sd_fops;
2599         gd->private_data = &sdkp->driver;
2600         gd->queue = sdkp->device->request_queue;
2601
2602         /* defaults, until the device tells us otherwise */
2603         sdp->sector_size = 512;
2604         sdkp->capacity = 0;
2605         sdkp->media_present = 1;
2606         sdkp->write_prot = 0;
2607         sdkp->WCE = 0;
2608         sdkp->RCD = 0;
2609         sdkp->ATO = 0;
2610         sdkp->first_scan = 1;
2611         sdkp->max_medium_access_timeouts = SD_MAX_MEDIUM_TIMEOUTS;
2612
2613         sd_revalidate_disk(gd);
2614
2615         blk_queue_prep_rq(sdp->request_queue, sd_prep_fn);
2616         blk_queue_unprep_rq(sdp->request_queue, sd_unprep_fn);
2617
2618         gd->driverfs_dev = &sdp->sdev_gendev;
2619         gd->flags = GENHD_FL_EXT_DEVT;
2620         if (sdp->removable) {
2621                 gd->flags |= GENHD_FL_REMOVABLE;
2622                 gd->events |= DISK_EVENT_MEDIA_CHANGE;
2623         }
2624
2625         add_disk(gd);
2626         sd_dif_config_host(sdkp);
2627
2628         sd_revalidate_disk(gd);
2629
2630         sd_printk(KERN_NOTICE, sdkp, "Attached SCSI %sdisk\n",
2631                   sdp->removable ? "removable " : "");
2632         scsi_autopm_put_device(sdp);
2633         put_device(&sdkp->dev);
2634 }
2635
2636 static int sd_get_index(int *index)
2637 {
2638         int error = -ENOMEM;
2639         do {
2640                 if (!ida_pre_get(&sd_index_ida, GFP_KERNEL))
2641                         break;
2642
2643                 spin_lock(&sd_index_lock);
2644                 error = ida_get_new(&sd_index_ida, index);
2645                 spin_unlock(&sd_index_lock);
2646         } while (error == -EAGAIN);
2647
2648         return error;
2649 }
2650 /**
2651  *      sd_probe - called during driver initialization and whenever a
2652  *      new scsi device is attached to the system. It is called once
2653  *      for each scsi device (not just disks) present.
2654  *      @dev: pointer to device object
2655  *
2656  *      Returns 0 if successful (or not interested in this scsi device
2657  *      (e.g. scanner)); 1 when there is an error.
2658  *
2659  *      Note: this function is invoked from the scsi mid-level.
2660  *      This function sets up the mapping between a given
2661  *      <host,channel,id,lun> (found in sdp) and new device name
2662  *      (e.g. /dev/sda). More precisely it is the block device major
2663  *      and minor number that is chosen here.
2664  *
2665  *      Assume sd_probe is not re-entrant (for time being)
2666  *      Also think about sd_probe() and sd_remove() running coincidentally.
2667  **/
2668 static int sd_probe(struct device *dev)
2669 {
2670         struct scsi_device *sdp = to_scsi_device(dev);
2671         struct scsi_disk *sdkp;
2672         struct gendisk *gd;
2673         int index;
2674         int error;
2675
2676         error = -ENODEV;
2677         if (sdp->type != TYPE_DISK && sdp->type != TYPE_MOD && sdp->type != TYPE_RBC)
2678                 goto out;
2679
2680         SCSI_LOG_HLQUEUE(3, sdev_printk(KERN_INFO, sdp,
2681                                         "sd_probe\n"));
2682
2683         error = -ENOMEM;
2684         sdkp = kzalloc(sizeof(*sdkp), GFP_KERNEL);
2685         if (!sdkp)
2686                 goto out;
2687
2688         gd = alloc_disk(SD_MINORS);
2689         if (!gd)
2690                 goto out_free;
2691
2692         error = sd_get_index(&index);
2693         if (error) {
2694                 sdev_printk(KERN_WARNING, sdp, "sd_probe: memory exhausted.\n");
2695                 goto out_put;
2696         }
2697
2698         error = sd_format_disk_name("sd", index, gd->disk_name, DISK_NAME_LEN);
2699         if (error) {
2700                 sdev_printk(KERN_WARNING, sdp, "SCSI disk (sd) name length exceeded.\n");
2701                 goto out_free_index;
2702         }
2703
2704         sdkp->device = sdp;
2705         sdkp->driver = &sd_template;
2706         sdkp->disk = gd;
2707         sdkp->index = index;
2708         atomic_set(&sdkp->openers, 0);
2709
2710         if (!sdp->request_queue->rq_timeout) {
2711                 if (sdp->type != TYPE_MOD)
2712                         blk_queue_rq_timeout(sdp->request_queue, SD_TIMEOUT);
2713                 else
2714                         blk_queue_rq_timeout(sdp->request_queue,
2715                                              SD_MOD_TIMEOUT);
2716         }
2717
2718         device_initialize(&sdkp->dev);
2719         sdkp->dev.parent = dev;
2720         sdkp->dev.class = &sd_disk_class;
2721         dev_set_name(&sdkp->dev, dev_name(dev));
2722
2723         if (device_add(&sdkp->dev))
2724                 goto out_free_index;
2725
2726         get_device(dev);
2727         dev_set_drvdata(dev, sdkp);
2728
2729         get_device(&sdkp->dev); /* prevent release before async_schedule */
2730         async_schedule(sd_probe_async, sdkp);
2731
2732         return 0;
2733
2734  out_free_index:
2735         spin_lock(&sd_index_lock);
2736         ida_remove(&sd_index_ida, index);
2737         spin_unlock(&sd_index_lock);
2738  out_put:
2739         put_disk(gd);
2740  out_free:
2741         kfree(sdkp);
2742  out:
2743         return error;
2744 }
2745
2746 /**
2747  *      sd_remove - called whenever a scsi disk (previously recognized by
2748  *      sd_probe) is detached from the system. It is called (potentially
2749  *      multiple times) during sd module unload.
2750  *      @sdp: pointer to mid level scsi device object
2751  *
2752  *      Note: this function is invoked from the scsi mid-level.
2753  *      This function potentially frees up a device name (e.g. /dev/sdc)
2754  *      that could be re-used by a subsequent sd_probe().
2755  *      This function is not called when the built-in sd driver is "exit-ed".
2756  **/
2757 static int sd_remove(struct device *dev)
2758 {
2759         struct scsi_disk *sdkp;
2760
2761         sdkp = dev_get_drvdata(dev);
2762         scsi_autopm_get_device(sdkp->device);
2763
2764         async_synchronize_full();
2765         blk_queue_prep_rq(sdkp->device->request_queue, scsi_prep_fn);
2766         blk_queue_unprep_rq(sdkp->device->request_queue, NULL);
2767         device_del(&sdkp->dev);
2768         del_gendisk(sdkp->disk);
2769         sd_shutdown(dev);
2770
2771         mutex_lock(&sd_ref_mutex);
2772         dev_set_drvdata(dev, NULL);
2773         put_device(&sdkp->dev);
2774         mutex_unlock(&sd_ref_mutex);
2775
2776         return 0;
2777 }
2778
2779 /**
2780  *      scsi_disk_release - Called to free the scsi_disk structure
2781  *      @dev: pointer to embedded class device
2782  *
2783  *      sd_ref_mutex must be held entering this routine.  Because it is
2784  *      called on last put, you should always use the scsi_disk_get()
2785  *      scsi_disk_put() helpers which manipulate the semaphore directly
2786  *      and never do a direct put_device.
2787  **/
2788 static void scsi_disk_release(struct device *dev)
2789 {
2790         struct scsi_disk *sdkp = to_scsi_disk(dev);
2791         struct gendisk *disk = sdkp->disk;
2792
2793         spin_lock(&sd_index_lock);
2794         ida_remove(&sd_index_ida, sdkp->index);
2795         spin_unlock(&sd_index_lock);
2796
2797         disk->private_data = NULL;
2798         put_disk(disk);
2799         put_device(&sdkp->device->sdev_gendev);
2800
2801         kfree(sdkp);
2802 }
2803
2804 static int sd_start_stop_device(struct scsi_disk *sdkp, int start)
2805 {
2806         unsigned char cmd[6] = { START_STOP };  /* START_VALID */
2807         struct scsi_sense_hdr sshdr;
2808         struct scsi_device *sdp = sdkp->device;
2809         int res;
2810
2811         if (start)
2812                 cmd[4] |= 1;    /* START */
2813
2814         if (sdp->start_stop_pwr_cond)
2815                 cmd[4] |= start ? 1 << 4 : 3 << 4;      /* Active or Standby */
2816
2817         if (!scsi_device_online(sdp))
2818                 return -ENODEV;
2819
2820         res = scsi_execute_req(sdp, cmd, DMA_NONE, NULL, 0, &sshdr,
2821                                SD_TIMEOUT, SD_MAX_RETRIES, NULL);
2822         if (res) {
2823                 sd_printk(KERN_WARNING, sdkp, "START_STOP FAILED\n");
2824                 sd_print_result(sdkp, res);
2825                 if (driver_byte(res) & DRIVER_SENSE)
2826                         sd_print_sense_hdr(sdkp, &sshdr);
2827         }
2828
2829         return res;
2830 }
2831
2832 /*
2833  * Send a SYNCHRONIZE CACHE instruction down to the device through
2834  * the normal SCSI command structure.  Wait for the command to
2835  * complete.
2836  */
2837 static void sd_shutdown(struct device *dev)
2838 {
2839         struct scsi_disk *sdkp = scsi_disk_get_from_dev(dev);
2840
2841         if (!sdkp)
2842                 return;         /* this can happen */
2843
2844         if (pm_runtime_suspended(dev))
2845                 goto exit;
2846
2847         if (sdkp->WCE) {
2848                 sd_printk(KERN_NOTICE, sdkp, "Synchronizing SCSI cache\n");
2849                 sd_sync_cache(sdkp);
2850         }
2851
2852         if (system_state != SYSTEM_RESTART && sdkp->device->manage_start_stop) {
2853                 sd_printk(KERN_NOTICE, sdkp, "Stopping disk\n");
2854                 sd_start_stop_device(sdkp, 0);
2855         }
2856
2857 exit:
2858         scsi_disk_put(sdkp);
2859 }
2860
2861 static int sd_suspend(struct device *dev, pm_message_t mesg)
2862 {
2863         struct scsi_disk *sdkp = scsi_disk_get_from_dev(dev);
2864         int ret = 0;
2865
2866         if (!sdkp)
2867                 return 0;       /* this can happen */
2868
2869         if (sdkp->WCE) {
2870                 sd_printk(KERN_NOTICE, sdkp, "Synchronizing SCSI cache\n");
2871                 ret = sd_sync_cache(sdkp);
2872                 if (ret)
2873                         goto done;
2874         }
2875
2876         if ((mesg.event & PM_EVENT_SLEEP) && sdkp->device->manage_start_stop) {
2877                 sd_printk(KERN_NOTICE, sdkp, "Stopping disk\n");
2878                 ret = sd_start_stop_device(sdkp, 0);
2879         }
2880
2881 done:
2882         scsi_disk_put(sdkp);
2883         return ret;
2884 }
2885
2886 static int sd_resume(struct device *dev)
2887 {
2888         struct scsi_disk *sdkp = scsi_disk_get_from_dev(dev);
2889         int ret = 0;
2890
2891         if (!sdkp->device->manage_start_stop)
2892                 goto done;
2893
2894         sd_printk(KERN_NOTICE, sdkp, "Starting disk\n");
2895         ret = sd_start_stop_device(sdkp, 1);
2896
2897 done:
2898         scsi_disk_put(sdkp);
2899         return ret;
2900 }
2901
2902 /*
2903 * Each major represents 16 disks. A minor is used for the disk itself and 15
2904 * partitions. Mark each disk busy so that sd_probe can not reclaim this major.
2905 */
2906 static int __init init_sd_ida(int *error)
2907 {
2908         int *index, i, j, err;
2909
2910         index = kmalloc(SD_MAJORS * (256 / SD_MINORS) * sizeof(int), GFP_KERNEL);
2911         if (!index)
2912                 return -ENOMEM;
2913
2914         /* Mark minors for all majors as busy */
2915         for (i = 0; i < SD_MAJORS; i++)
2916         {
2917                 for (j = 0; j < (256 / SD_MINORS); j++) {
2918                         err = sd_get_index(&index[i * (256 / SD_MINORS) + j]);
2919                         if (err) {
2920                                 kfree(index);
2921                                 return err;
2922                         }
2923                 }
2924         }
2925
2926         /* Mark minors for claimed majors as free */
2927         for (i = 0; i < SD_MAJORS; i++)
2928         {
2929                 if (error[i])
2930                         continue;
2931                 for (j = 0; j < (256 / SD_MINORS); j++)
2932                         ida_remove(&sd_index_ida, index[i * (256 / SD_MINORS) + j]);
2933         }
2934         kfree(index);
2935         return 0;
2936 }
2937
2938 /**
2939  *      init_sd - entry point for this driver (both when built in or when
2940  *      a module).
2941  *
2942  *      Note: this function registers this driver with the scsi mid-level.
2943  **/
2944 static int __init init_sd(void)
2945 {
2946         int majors = 0, i, err;
2947         int error[SD_MAJORS];
2948
2949         SCSI_LOG_HLQUEUE(3, printk("init_sd: sd driver entry point\n"));
2950
2951         for (i = 0; i < SD_MAJORS; i++)
2952         {
2953                 error[i] = register_blkdev(sd_major(i), "sd");
2954                 if (error[i] == 0)
2955                         majors++;
2956         }
2957
2958         if (!majors)
2959                 return -ENODEV;
2960
2961         if (majors < SD_MAJORS) {
2962                 err = init_sd_ida(error);
2963                 if (err)
2964                         return err;
2965         }
2966
2967         err = class_register(&sd_disk_class);
2968         if (err)
2969                 goto err_out;
2970
2971         err = scsi_register_driver(&sd_template.gendrv);
2972         if (err)
2973                 goto err_out_class;
2974
2975         sd_cdb_cache = kmem_cache_create("sd_ext_cdb", SD_EXT_CDB_SIZE,
2976                                          0, 0, NULL);
2977         if (!sd_cdb_cache) {
2978                 printk(KERN_ERR "sd: can't init extended cdb cache\n");
2979                 goto err_out_class;
2980         }
2981
2982         sd_cdb_pool = mempool_create_slab_pool(SD_MEMPOOL_SIZE, sd_cdb_cache);
2983         if (!sd_cdb_pool) {
2984                 printk(KERN_ERR "sd: can't init extended cdb pool\n");
2985                 goto err_out_cache;
2986         }
2987
2988         return 0;
2989
2990 err_out_cache:
2991         kmem_cache_destroy(sd_cdb_cache);
2992
2993 err_out_class:
2994         class_unregister(&sd_disk_class);
2995 err_out:
2996         for (i = 0; i < SD_MAJORS; i++)
2997                 unregister_blkdev(sd_major(i), "sd");
2998         return err;
2999 }
3000
3001 /**
3002  *      exit_sd - exit point for this driver (when it is a module).
3003  *
3004  *      Note: this function unregisters this driver from the scsi mid-level.
3005  **/
3006 static void __exit exit_sd(void)
3007 {
3008         int i;
3009
3010         SCSI_LOG_HLQUEUE(3, printk("exit_sd: exiting sd driver\n"));
3011
3012         mempool_destroy(sd_cdb_pool);
3013         kmem_cache_destroy(sd_cdb_cache);
3014
3015         scsi_unregister_driver(&sd_template.gendrv);
3016         class_unregister(&sd_disk_class);
3017
3018         for (i = 0; i < SD_MAJORS; i++)
3019                 unregister_blkdev(sd_major(i), "sd");
3020 }
3021
3022 module_init(init_sd);
3023 module_exit(exit_sd);
3024
3025 static void sd_print_sense_hdr(struct scsi_disk *sdkp,
3026                                struct scsi_sense_hdr *sshdr)
3027 {
3028         sd_printk(KERN_INFO, sdkp, " ");
3029         scsi_show_sense_hdr(sshdr);
3030         sd_printk(KERN_INFO, sdkp, " ");
3031         scsi_show_extd_sense(sshdr->asc, sshdr->ascq);
3032 }
3033
3034 static void sd_print_result(struct scsi_disk *sdkp, int result)
3035 {
3036         sd_printk(KERN_INFO, sdkp, " ");
3037         scsi_show_result(result);
3038 }
3039