block/blk-settings.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Functions related to setting various queue properties from drivers
   4  */
   5 #include <linux/kernel.h>
   6 #include <linux/module.h>
   7 #include <linux/init.h>
   8 #include <linux/bio.h>
   9 #include <linux/blk-integrity.h>
  10 #include <linux/pagemap.h>
  11 #include <linux/backing-dev-defs.h>
  12 #include <linux/gcd.h>
  13 #include <linux/lcm.h>
  14 #include <linux/jiffies.h>
  15 #include <linux/gfp.h>
  16 #include <linux/dma-mapping.h>
  17
  18 #include "blk.h"
  19 #include "blk-rq-qos.h"
  20 #include "blk-wbt.h"
  21
  22 void blk_queue_rq_timeout(struct request_queue *q, unsigned int timeout)
  23 {
  24         q->rq_timeout = timeout;
  25 }
  26 EXPORT_SYMBOL_GPL(blk_queue_rq_timeout);
  27
  28 /**
  29  * blk_set_stacking_limits - set default limits for stacking devices
  30  * @lim:  the queue_limits structure to reset
  31  *
  32  * Prepare queue limits for applying limits from underlying devices using
  33  * blk_stack_limits().
  34  */
  35 void blk_set_stacking_limits(struct queue_limits *lim)
  36 {
  37         memset(lim, 0, sizeof(*lim));
  38         lim->logical_block_size = SECTOR_SIZE;
  39         lim->physical_block_size = SECTOR_SIZE;
  40         lim->io_min = SECTOR_SIZE;
  41         lim->discard_granularity = SECTOR_SIZE;
  42         lim->dma_alignment = SECTOR_SIZE - 1;
  43         lim->seg_boundary_mask = BLK_SEG_BOUNDARY_MASK;
  44
  45         /* Inherit limits from component devices */
  46         lim->max_segments = USHRT_MAX;
  47         lim->max_discard_segments = USHRT_MAX;
  48         lim->max_hw_sectors = UINT_MAX;
  49         lim->max_segment_size = UINT_MAX;
  50         lim->max_sectors = UINT_MAX;
  51         lim->max_dev_sectors = UINT_MAX;
  52         lim->max_write_zeroes_sectors = UINT_MAX;
  53         lim->max_hw_zone_append_sectors = UINT_MAX;
  54         lim->max_user_discard_sectors = UINT_MAX;
  55 }
  56 EXPORT_SYMBOL(blk_set_stacking_limits);
  57
  58 void blk_apply_bdi_limits(struct backing_dev_info *bdi,
  59                 struct queue_limits *lim)
  60 {
  61         /*
  62          * For read-ahead of large files to be effective, we need to read ahead
  63          * at least twice the optimal I/O size.
  64          */
  65         bdi->ra_pages = max(lim->io_opt * 2 / PAGE_SIZE, VM_READAHEAD_PAGES);
  66         bdi->io_pages = lim->max_sectors >> PAGE_SECTORS_SHIFT;
  67 }
  68
  69 static int blk_validate_zoned_limits(struct queue_limits *lim)
  70 {
  71         if (!(lim->features & BLK_FEAT_ZONED)) {
  72                 if (WARN_ON_ONCE(lim->max_open_zones) ||
  73                     WARN_ON_ONCE(lim->max_active_zones) ||
  74                     WARN_ON_ONCE(lim->zone_write_granularity) ||
  75                     WARN_ON_ONCE(lim->max_zone_append_sectors))
  76                         return -EINVAL;
  77                 return 0;
  78         }
  79
  80         if (WARN_ON_ONCE(!IS_ENABLED(CONFIG_BLK_DEV_ZONED)))
  81                 return -EINVAL;
  82
  83         /*
  84          * Given that active zones include open zones, the maximum number of
  85          * open zones cannot be larger than the maximum number of active zones.
  86          */
  87         if (lim->max_active_zones &&
  88             lim->max_open_zones > lim->max_active_zones)
  89                 return -EINVAL;
  90
  91         if (lim->zone_write_granularity < lim->logical_block_size)
  92                 lim->zone_write_granularity = lim->logical_block_size;
  93
  94         /*
  95          * The Zone Append size is limited by the maximum I/O size and the zone
  96          * size given that it can't span zones.
  97          *
  98          * If no max_hw_zone_append_sectors limit is provided, the block layer
  99          * will emulated it, else we're also bound by the hardware limit.
 100          */
 101         lim->max_zone_append_sectors =
 102                 min_not_zero(lim->max_hw_zone_append_sectors,
 103                         min(lim->chunk_sectors, lim->max_hw_sectors));
 104         return 0;
 105 }
 106
 107 static int blk_validate_integrity_limits(struct queue_limits *lim)
 108 {
 109         struct blk_integrity *bi = &lim->integrity;
 110
 111         if (!bi->tuple_size) {
 112                 if (bi->csum_type != BLK_INTEGRITY_CSUM_NONE ||
 113                     bi->tag_size || ((bi->flags & BLK_INTEGRITY_REF_TAG))) {
 114                         pr_warn("invalid PI settings.\n");
 115                         return -EINVAL;
 116                 }
 117                 return 0;
 118         }
 119
 120         if (!IS_ENABLED(CONFIG_BLK_DEV_INTEGRITY)) {
 121                 pr_warn("integrity support disabled.\n");
 122                 return -EINVAL;
 123         }
 124
 125         if (bi->csum_type == BLK_INTEGRITY_CSUM_NONE &&
 126             (bi->flags & BLK_INTEGRITY_REF_TAG)) {
 127                 pr_warn("ref tag not support without checksum.\n");
 128                 return -EINVAL;
 129         }
 130
 131         if (!bi->interval_exp)
 132                 bi->interval_exp = ilog2(lim->logical_block_size);
 133
 134         return 0;
 135 }
 136
 137 /*
 138  * Returns max guaranteed bytes which we can fit in a bio.
 139  *
 140  * We request that an atomic_write is ITER_UBUF iov_iter (so a single vector),
 141  * so we assume that we can fit in at least PAGE_SIZE in a segment, apart from
 142  * the first and last segments.
 143  */
 144 static unsigned int blk_queue_max_guaranteed_bio(struct queue_limits *lim)
 145 {
 146         unsigned int max_segments = min(BIO_MAX_VECS, lim->max_segments);
 147         unsigned int length;
 148
 149         length = min(max_segments, 2) * lim->logical_block_size;
 150         if (max_segments > 2)
 151                 length += (max_segments - 2) * PAGE_SIZE;
 152
 153         return length;
 154 }
 155
 156 static void blk_atomic_writes_update_limits(struct queue_limits *lim)
 157 {
 158         unsigned int unit_limit = min(lim->max_hw_sectors << SECTOR_SHIFT,
 159                                         blk_queue_max_guaranteed_bio(lim));
 160
 161         unit_limit = rounddown_pow_of_two(unit_limit);
 162
 163         lim->atomic_write_max_sectors =
 164                 min(lim->atomic_write_hw_max >> SECTOR_SHIFT,
 165                         lim->max_hw_sectors);
 166         lim->atomic_write_unit_min =
 167                 min(lim->atomic_write_hw_unit_min, unit_limit);
 168         lim->atomic_write_unit_max =
 169                 min(lim->atomic_write_hw_unit_max, unit_limit);
 170         lim->atomic_write_boundary_sectors =
 171                 lim->atomic_write_hw_boundary >> SECTOR_SHIFT;
 172 }
 173
 174 static void blk_validate_atomic_write_limits(struct queue_limits *lim)
 175 {
 176         unsigned int boundary_sectors;
 177
 178         if (!lim->atomic_write_hw_max)
 179                 goto unsupported;
 180
 181         if (WARN_ON_ONCE(!is_power_of_2(lim->atomic_write_hw_unit_min)))
 182                 goto unsupported;
 183
 184         if (WARN_ON_ONCE(!is_power_of_2(lim->atomic_write_hw_unit_max)))
 185                 goto unsupported;
 186
 187         if (WARN_ON_ONCE(lim->atomic_write_hw_unit_min >
 188                          lim->atomic_write_hw_unit_max))
 189                 goto unsupported;
 190
 191         if (WARN_ON_ONCE(lim->atomic_write_hw_unit_max >
 192                          lim->atomic_write_hw_max))
 193                 goto unsupported;
 194
 195         boundary_sectors = lim->atomic_write_hw_boundary >> SECTOR_SHIFT;
 196
 197         if (boundary_sectors) {
 198                 if (WARN_ON_ONCE(lim->atomic_write_hw_max >
 199                                  lim->atomic_write_hw_boundary))
 200                         goto unsupported;
 201                 /*
 202                  * A feature of boundary support is that it disallows bios to
 203                  * be merged which would result in a merged request which
 204                  * crosses either a chunk sector or atomic write HW boundary,
 205                  * even though chunk sectors may be just set for performance.
 206                  * For simplicity, disallow atomic writes for a chunk sector
 207                  * which is non-zero and smaller than atomic write HW boundary.
 208                  * Furthermore, chunk sectors must be a multiple of atomic
 209                  * write HW boundary. Otherwise boundary support becomes
 210                  * complicated.
 211                  * Devices which do not conform to these rules can be dealt
 212                  * with if and when they show up.
 213                  */
 214                 if (WARN_ON_ONCE(lim->chunk_sectors % boundary_sectors))
 215                         goto unsupported;
 216
 217                 /*
 218                  * The boundary size just needs to be a multiple of unit_max
 219                  * (and not necessarily a power-of-2), so this following check
 220                  * could be relaxed in future.
 221                  * Furthermore, if needed, unit_max could even be reduced so
 222                  * that it is compliant with a !power-of-2 boundary.
 223                  */
 224                 if (!is_power_of_2(boundary_sectors))
 225                         goto unsupported;
 226         }
 227
 228         blk_atomic_writes_update_limits(lim);
 229         return;
 230
 231 unsupported:
 232         lim->atomic_write_max_sectors = 0;
 233         lim->atomic_write_boundary_sectors = 0;
 234         lim->atomic_write_unit_min = 0;
 235         lim->atomic_write_unit_max = 0;
 236 }
 237
 238 /*
 239  * Check that the limits in lim are valid, initialize defaults for unset
 240  * values, and cap values based on others where needed.
 241  */
 242 int blk_validate_limits(struct queue_limits *lim)
 243 {
 244         unsigned int max_hw_sectors;
 245         unsigned int logical_block_sectors;
 246         int err;
 247
 248         /*
 249          * Unless otherwise specified, default to 512 byte logical blocks and a
 250          * physical block size equal to the logical block size.
 251          */
 252         if (!lim->logical_block_size)
 253                 lim->logical_block_size = SECTOR_SIZE;
 254         else if (blk_validate_block_size(lim->logical_block_size)) {
 255                 pr_warn("Invalid logical block size (%d)\n", lim->logical_block_size);
 256                 return -EINVAL;
 257         }
 258         if (lim->physical_block_size < lim->logical_block_size)
 259                 lim->physical_block_size = lim->logical_block_size;
 260
 261         /*
 262          * The minimum I/O size defaults to the physical block size unless
 263          * explicitly overridden.
 264          */
 265         if (lim->io_min < lim->physical_block_size)
 266                 lim->io_min = lim->physical_block_size;
 267
 268         /*
 269          * The optimal I/O size may not be aligned to physical block size
 270          * (because it may be limited by dma engines which have no clue about
 271          * block size of the disks attached to them), so we round it down here.
 272          */
 273         lim->io_opt = round_down(lim->io_opt, lim->physical_block_size);
 274
 275         /*
 276          * max_hw_sectors has a somewhat weird default for historical reason,
 277          * but driver really should set their own instead of relying on this
 278          * value.
 279          *
 280          * The block layer relies on the fact that every driver can
 281          * handle at lest a page worth of data per I/O, and needs the value
 282          * aligned to the logical block size.
 283          */
 284         if (!lim->max_hw_sectors)
 285                 lim->max_hw_sectors = BLK_SAFE_MAX_SECTORS;
 286         if (WARN_ON_ONCE(lim->max_hw_sectors < PAGE_SECTORS))
 287                 return -EINVAL;
 288         logical_block_sectors = lim->logical_block_size >> SECTOR_SHIFT;
 289         if (WARN_ON_ONCE(logical_block_sectors > lim->max_hw_sectors))
 290                 return -EINVAL;
 291         lim->max_hw_sectors = round_down(lim->max_hw_sectors,
 292                         logical_block_sectors);
 293
 294         /*
 295          * The actual max_sectors value is a complex beast and also takes the
 296          * max_dev_sectors value (set by SCSI ULPs) and a user configurable
 297          * value into account.  The ->max_sectors value is always calculated
 298          * from these, so directly setting it won't have any effect.
 299          */
 300         max_hw_sectors = min_not_zero(lim->max_hw_sectors,
 301                                 lim->max_dev_sectors);
 302         if (lim->max_user_sectors) {
 303                 if (lim->max_user_sectors < PAGE_SIZE / SECTOR_SIZE)
 304                         return -EINVAL;
 305                 lim->max_sectors = min(max_hw_sectors, lim->max_user_sectors);
 306         } else if (lim->io_opt > (BLK_DEF_MAX_SECTORS_CAP << SECTOR_SHIFT)) {
 307                 lim->max_sectors =
 308                         min(max_hw_sectors, lim->io_opt >> SECTOR_SHIFT);
 309         } else if (lim->io_min > (BLK_DEF_MAX_SECTORS_CAP << SECTOR_SHIFT)) {
 310                 lim->max_sectors =
 311                         min(max_hw_sectors, lim->io_min >> SECTOR_SHIFT);
 312         } else {
 313                 lim->max_sectors = min(max_hw_sectors, BLK_DEF_MAX_SECTORS_CAP);
 314         }
 315         lim->max_sectors = round_down(lim->max_sectors,
 316                         logical_block_sectors);
 317
 318         /*
 319          * Random default for the maximum number of segments.  Driver should not
 320          * rely on this and set their own.
 321          */
 322         if (!lim->max_segments)
 323                 lim->max_segments = BLK_MAX_SEGMENTS;
 324
 325         lim->max_discard_sectors =
 326                 min(lim->max_hw_discard_sectors, lim->max_user_discard_sectors);
 327
 328         if (!lim->max_discard_segments)
 329                 lim->max_discard_segments = 1;
 330
 331         if (lim->discard_granularity < lim->physical_block_size)
 332                 lim->discard_granularity = lim->physical_block_size;
 333
 334         /*
 335          * By default there is no limit on the segment boundary alignment,
 336          * but if there is one it can't be smaller than the page size as
 337          * that would break all the normal I/O patterns.
 338          */
 339         if (!lim->seg_boundary_mask)
 340                 lim->seg_boundary_mask = BLK_SEG_BOUNDARY_MASK;
 341         if (WARN_ON_ONCE(lim->seg_boundary_mask < PAGE_SIZE - 1))
 342                 return -EINVAL;
 343
 344         /*
 345          * Stacking device may have both virtual boundary and max segment
 346          * size limit, so allow this setting now, and long-term the two
 347          * might need to move out of stacking limits since we have immutable
 348          * bvec and lower layer bio splitting is supposed to handle the two
 349          * correctly.
 350          */
 351         if (lim->virt_boundary_mask) {
 352                 if (!lim->max_segment_size)
 353                         lim->max_segment_size = UINT_MAX;
 354         } else {
 355                 /*
 356                  * The maximum segment size has an odd historic 64k default that
 357                  * drivers probably should override.  Just like the I/O size we
 358                  * require drivers to at least handle a full page per segment.
 359                  */
 360                 if (!lim->max_segment_size)
 361                         lim->max_segment_size = BLK_MAX_SEGMENT_SIZE;
 362                 if (WARN_ON_ONCE(lim->max_segment_size < PAGE_SIZE))
 363                         return -EINVAL;
 364         }
 365
 366         /*
 367          * We require drivers to at least do logical block aligned I/O, but
 368          * historically could not check for that due to the separate calls
 369          * to set the limits.  Once the transition is finished the check
 370          * below should be narrowed down to check the logical block size.
 371          */
 372         if (!lim->dma_alignment)
 373                 lim->dma_alignment = SECTOR_SIZE - 1;
 374         if (WARN_ON_ONCE(lim->dma_alignment > PAGE_SIZE))
 375                 return -EINVAL;
 376
 377         if (lim->alignment_offset) {
 378                 lim->alignment_offset &= (lim->physical_block_size - 1);
 379                 lim->flags &= ~BLK_FLAG_MISALIGNED;
 380         }
 381
 382         if (!(lim->features & BLK_FEAT_WRITE_CACHE))
 383                 lim->features &= ~BLK_FEAT_FUA;
 384
 385         blk_validate_atomic_write_limits(lim);
 386
 387         err = blk_validate_integrity_limits(lim);
 388         if (err)
 389                 return err;
 390         return blk_validate_zoned_limits(lim);
 391 }
 392 EXPORT_SYMBOL_GPL(blk_validate_limits);
 393
 394 /*
 395  * Set the default limits for a newly allocated queue.  @lim contains the
 396  * initial limits set by the driver, which could be no limit in which case
 397  * all fields are cleared to zero.
 398  */
 399 int blk_set_default_limits(struct queue_limits *lim)
 400 {
 401         /*
 402          * Most defaults are set by capping the bounds in blk_validate_limits,
 403          * but max_user_discard_sectors is special and needs an explicit
 404          * initialization to the max value here.
 405          */
 406         lim->max_user_discard_sectors = UINT_MAX;
 407         return blk_validate_limits(lim);
 408 }
 409
 410 /**
 411  * queue_limits_commit_update - commit an atomic update of queue limits
 412  * @q:          queue to update
 413  * @lim:        limits to apply
 414  *
 415  * Apply the limits in @lim that were obtained from queue_limits_start_update()
 416  * and updated by the caller to @q.
 417  *
 418  * Returns 0 if successful, else a negative error code.
 419  */
 420 int queue_limits_commit_update(struct request_queue *q,
 421                 struct queue_limits *lim)
 422 {
 423         int error;
 424
 425         error = blk_validate_limits(lim);
 426         if (error)
 427                 goto out_unlock;
 428
 429 #ifdef CONFIG_BLK_INLINE_ENCRYPTION
 430         if (q->crypto_profile && lim->integrity.tag_size) {
 431                 pr_warn("blk-integrity: Integrity and hardware inline encryption are not supported together.\n");
 432                 error = -EINVAL;
 433                 goto out_unlock;
 434         }
 435 #endif
 436
 437         q->limits = *lim;
 438         if (q->disk)
 439                 blk_apply_bdi_limits(q->disk->bdi, lim);
 440 out_unlock:
 441         mutex_unlock(&q->limits_lock);
 442         return error;
 443 }
 444 EXPORT_SYMBOL_GPL(queue_limits_commit_update);
 445
 446 /**
 447  * queue_limits_set - apply queue limits to queue
 448  * @q:          queue to update
 449  * @lim:        limits to apply
 450  *
 451  * Apply the limits in @lim that were freshly initialized to @q.
 452  * To update existing limits use queue_limits_start_update() and
 453  * queue_limits_commit_update() instead.
 454  *
 455  * Returns 0 if successful, else a negative error code.
 456  */
 457 int queue_limits_set(struct request_queue *q, struct queue_limits *lim)
 458 {
 459         mutex_lock(&q->limits_lock);
 460         return queue_limits_commit_update(q, lim);
 461 }
 462 EXPORT_SYMBOL_GPL(queue_limits_set);
 463
 464 static int queue_limit_alignment_offset(const struct queue_limits *lim,
 465                 sector_t sector)
 466 {
 467         unsigned int granularity = max(lim->physical_block_size, lim->io_min);
 468         unsigned int alignment = sector_div(sector, granularity >> SECTOR_SHIFT)
 469                 << SECTOR_SHIFT;
 470
 471         return (granularity + lim->alignment_offset - alignment) % granularity;
 472 }
 473
 474 static unsigned int queue_limit_discard_alignment(
 475                 const struct queue_limits *lim, sector_t sector)
 476 {
 477         unsigned int alignment, granularity, offset;
 478
 479         if (!lim->max_discard_sectors)
 480                 return 0;
 481
 482         /* Why are these in bytes, not sectors? */
 483         alignment = lim->discard_alignment >> SECTOR_SHIFT;
 484         granularity = lim->discard_granularity >> SECTOR_SHIFT;
 485
 486         /* Offset of the partition start in 'granularity' sectors */
 487         offset = sector_div(sector, granularity);
 488
 489         /* And why do we do this modulus *again* in blkdev_issue_discard()? */
 490         offset = (granularity + alignment - offset) % granularity;
 491
 492         /* Turn it back into bytes, gaah */
 493         return offset << SECTOR_SHIFT;
 494 }
 495
 496 static unsigned int blk_round_down_sectors(unsigned int sectors, unsigned int lbs)
 497 {
 498         sectors = round_down(sectors, lbs >> SECTOR_SHIFT);
 499         if (sectors < PAGE_SIZE >> SECTOR_SHIFT)
 500                 sectors = PAGE_SIZE >> SECTOR_SHIFT;
 501         return sectors;
 502 }
 503
 504 /* Check if second and later bottom devices are compliant */
 505 static bool blk_stack_atomic_writes_tail(struct queue_limits *t,
 506                                 struct queue_limits *b)
 507 {
 508         /* We're not going to support different boundary sizes.. yet */
 509         if (t->atomic_write_hw_boundary != b->atomic_write_hw_boundary)
 510                 return false;
 511
 512         /* Can't support this */
 513         if (t->atomic_write_hw_unit_min > b->atomic_write_hw_unit_max)
 514                 return false;
 515
 516         /* Or this */
 517         if (t->atomic_write_hw_unit_max < b->atomic_write_hw_unit_min)
 518                 return false;
 519
 520         t->atomic_write_hw_max = min(t->atomic_write_hw_max,
 521                                 b->atomic_write_hw_max);
 522         t->atomic_write_hw_unit_min = max(t->atomic_write_hw_unit_min,
 523                                 b->atomic_write_hw_unit_min);
 524         t->atomic_write_hw_unit_max = min(t->atomic_write_hw_unit_max,
 525                                 b->atomic_write_hw_unit_max);
 526         return true;
 527 }
 528
 529 /* Check for valid boundary of first bottom device */
 530 static bool blk_stack_atomic_writes_boundary_head(struct queue_limits *t,
 531                                 struct queue_limits *b)
 532 {
 533         /*
 534          * Ensure atomic write boundary is aligned with chunk sectors. Stacked
 535          * devices store chunk sectors in t->io_min.
 536          */
 537         if (b->atomic_write_hw_boundary > t->io_min &&
 538             b->atomic_write_hw_boundary % t->io_min)
 539                 return false;
 540         if (t->io_min > b->atomic_write_hw_boundary &&
 541             t->io_min % b->atomic_write_hw_boundary)
 542                 return false;
 543
 544         t->atomic_write_hw_boundary = b->atomic_write_hw_boundary;
 545         return true;
 546 }
 547
 548
 549 /* Check stacking of first bottom device */
 550 static bool blk_stack_atomic_writes_head(struct queue_limits *t,
 551                                 struct queue_limits *b)
 552 {
 553         if (b->atomic_write_hw_boundary &&
 554             !blk_stack_atomic_writes_boundary_head(t, b))
 555                 return false;
 556
 557         if (t->io_min <= SECTOR_SIZE) {
 558                 /* No chunk sectors, so use bottom device values directly */
 559                 t->atomic_write_hw_unit_max = b->atomic_write_hw_unit_max;
 560                 t->atomic_write_hw_unit_min = b->atomic_write_hw_unit_min;
 561                 t->atomic_write_hw_max = b->atomic_write_hw_max;
 562                 return true;
 563         }
 564
 565         /*
 566          * Find values for limits which work for chunk size.
 567          * b->atomic_write_hw_unit_{min, max} may not be aligned with chunk
 568          * size (t->io_min), as chunk size is not restricted to a power-of-2.
 569          * So we need to find highest power-of-2 which works for the chunk
 570          * size.
 571          * As an example scenario, we could have b->unit_max = 16K and
 572          * t->io_min = 24K. For this case, reduce t->unit_max to a value
 573          * aligned with both limits, i.e. 8K in this example.
 574          */
 575         t->atomic_write_hw_unit_max = b->atomic_write_hw_unit_max;
 576         while (t->io_min % t->atomic_write_hw_unit_max)
 577                 t->atomic_write_hw_unit_max /= 2;
 578
 579         t->atomic_write_hw_unit_min = min(b->atomic_write_hw_unit_min,
 580                                           t->atomic_write_hw_unit_max);
 581         t->atomic_write_hw_max = min(b->atomic_write_hw_max, t->io_min);
 582
 583         return true;
 584 }
 585
 586 static void blk_stack_atomic_writes_limits(struct queue_limits *t,
 587                                 struct queue_limits *b)
 588 {
 589         if (!(t->features & BLK_FEAT_ATOMIC_WRITES_STACKED))
 590                 goto unsupported;
 591
 592         if (!b->atomic_write_unit_min)
 593                 goto unsupported;
 594
 595         /*
 596          * If atomic_write_hw_max is set, we have already stacked 1x bottom
 597          * device, so check for compliance.
 598          */
 599         if (t->atomic_write_hw_max) {
 600                 if (!blk_stack_atomic_writes_tail(t, b))
 601                         goto unsupported;
 602                 return;
 603         }
 604
 605         if (!blk_stack_atomic_writes_head(t, b))
 606                 goto unsupported;
 607         return;
 608
 609 unsupported:
 610         t->atomic_write_hw_max = 0;
 611         t->atomic_write_hw_unit_max = 0;
 612         t->atomic_write_hw_unit_min = 0;
 613         t->atomic_write_hw_boundary = 0;
 614         t->features &= ~BLK_FEAT_ATOMIC_WRITES_STACKED;
 615 }
 616
 617 /**
 618  * blk_stack_limits - adjust queue_limits for stacked devices
 619  * @t:  the stacking driver limits (top device)
 620  * @b:  the underlying queue limits (bottom, component device)
 621  * @start:  first data sector within component device
 622  *
 623  * Description:
 624  *    This function is used by stacking drivers like MD and DM to ensure
 625  *    that all component devices have compatible block sizes and
 626  *    alignments.  The stacking driver must provide a queue_limits
 627  *    struct (top) and then iteratively call the stacking function for
 628  *    all component (bottom) devices.  The stacking function will
 629  *    attempt to combine the values and ensure proper alignment.
 630  *
 631  *    Returns 0 if the top and bottom queue_limits are compatible.  The
 632  *    top device's block sizes and alignment offsets may be adjusted to
 633  *    ensure alignment with the bottom device. If no compatible sizes
 634  *    and alignments exist, -1 is returned and the resulting top
 635  *    queue_limits will have the misaligned flag set to indicate that
 636  *    the alignment_offset is undefined.
 637  */
 638 int blk_stack_limits(struct queue_limits *t, struct queue_limits *b,
 639                      sector_t start)
 640 {
 641         unsigned int top, bottom, alignment, ret = 0;
 642
 643         t->features |= (b->features & BLK_FEAT_INHERIT_MASK);
 644
 645         /*
 646          * Some feaures need to be supported both by the stacking driver and all
 647          * underlying devices.  The stacking driver sets these flags before
 648          * stacking the limits, and this will clear the flags if any of the
 649          * underlying devices does not support it.
 650          */
 651         if (!(b->features & BLK_FEAT_NOWAIT))
 652                 t->features &= ~BLK_FEAT_NOWAIT;
 653         if (!(b->features & BLK_FEAT_POLL))
 654                 t->features &= ~BLK_FEAT_POLL;
 655
 656         t->flags |= (b->flags & BLK_FLAG_MISALIGNED);
 657
 658         t->max_sectors = min_not_zero(t->max_sectors, b->max_sectors);
 659         t->max_user_sectors = min_not_zero(t->max_user_sectors,
 660                         b->max_user_sectors);
 661         t->max_hw_sectors = min_not_zero(t->max_hw_sectors, b->max_hw_sectors);
 662         t->max_dev_sectors = min_not_zero(t->max_dev_sectors, b->max_dev_sectors);
 663         t->max_write_zeroes_sectors = min(t->max_write_zeroes_sectors,
 664                                         b->max_write_zeroes_sectors);
 665         t->max_hw_zone_append_sectors = min(t->max_hw_zone_append_sectors,
 666                                         b->max_hw_zone_append_sectors);
 667
 668         t->seg_boundary_mask = min_not_zero(t->seg_boundary_mask,
 669                                             b->seg_boundary_mask);
 670         t->virt_boundary_mask = min_not_zero(t->virt_boundary_mask,
 671                                             b->virt_boundary_mask);
 672
 673         t->max_segments = min_not_zero(t->max_segments, b->max_segments);
 674         t->max_discard_segments = min_not_zero(t->max_discard_segments,
 675                                                b->max_discard_segments);
 676         t->max_integrity_segments = min_not_zero(t->max_integrity_segments,
 677                                                  b->max_integrity_segments);
 678
 679         t->max_segment_size = min_not_zero(t->max_segment_size,
 680                                            b->max_segment_size);
 681
 682         alignment = queue_limit_alignment_offset(b, start);
 683
 684         /* Bottom device has different alignment.  Check that it is
 685          * compatible with the current top alignment.
 686          */
 687         if (t->alignment_offset != alignment) {
 688
 689                 top = max(t->physical_block_size, t->io_min)
 690                         + t->alignment_offset;
 691                 bottom = max(b->physical_block_size, b->io_min) + alignment;
 692
 693                 /* Verify that top and bottom intervals line up */
 694                 if (max(top, bottom) % min(top, bottom)) {
 695                         t->flags |= BLK_FLAG_MISALIGNED;
 696                         ret = -1;
 697                 }
 698         }
 699
 700         t->logical_block_size = max(t->logical_block_size,
 701                                     b->logical_block_size);
 702
 703         t->physical_block_size = max(t->physical_block_size,
 704                                      b->physical_block_size);
 705
 706         t->io_min = max(t->io_min, b->io_min);
 707         t->io_opt = lcm_not_zero(t->io_opt, b->io_opt);
 708         t->dma_alignment = max(t->dma_alignment, b->dma_alignment);
 709
 710         /* Set non-power-of-2 compatible chunk_sectors boundary */
 711         if (b->chunk_sectors)
 712                 t->chunk_sectors = gcd(t->chunk_sectors, b->chunk_sectors);
 713
 714         /* Physical block size a multiple of the logical block size? */
 715         if (t->physical_block_size & (t->logical_block_size - 1)) {
 716                 t->physical_block_size = t->logical_block_size;
 717                 t->flags |= BLK_FLAG_MISALIGNED;
 718                 ret = -1;
 719         }
 720
 721         /* Minimum I/O a multiple of the physical block size? */
 722         if (t->io_min & (t->physical_block_size - 1)) {
 723                 t->io_min = t->physical_block_size;
 724                 t->flags |= BLK_FLAG_MISALIGNED;
 725                 ret = -1;
 726         }
 727
 728         /* Optimal I/O a multiple of the physical block size? */
 729         if (t->io_opt & (t->physical_block_size - 1)) {
 730                 t->io_opt = 0;
 731                 t->flags |= BLK_FLAG_MISALIGNED;
 732                 ret = -1;
 733         }
 734
 735         /* chunk_sectors a multiple of the physical block size? */
 736         if ((t->chunk_sectors << 9) & (t->physical_block_size - 1)) {
 737                 t->chunk_sectors = 0;
 738                 t->flags |= BLK_FLAG_MISALIGNED;
 739                 ret = -1;
 740         }
 741
 742         /* Find lowest common alignment_offset */
 743         t->alignment_offset = lcm_not_zero(t->alignment_offset, alignment)
 744                 % max(t->physical_block_size, t->io_min);
 745
 746         /* Verify that new alignment_offset is on a logical block boundary */
 747         if (t->alignment_offset & (t->logical_block_size - 1)) {
 748                 t->flags |= BLK_FLAG_MISALIGNED;
 749                 ret = -1;
 750         }
 751
 752         t->max_sectors = blk_round_down_sectors(t->max_sectors, t->logical_block_size);
 753         t->max_hw_sectors = blk_round_down_sectors(t->max_hw_sectors, t->logical_block_size);
 754         t->max_dev_sectors = blk_round_down_sectors(t->max_dev_sectors, t->logical_block_size);
 755
 756         /* Discard alignment and granularity */
 757         if (b->discard_granularity) {
 758                 alignment = queue_limit_discard_alignment(b, start);
 759
 760                 t->max_discard_sectors = min_not_zero(t->max_discard_sectors,
 761                                                       b->max_discard_sectors);
 762                 t->max_hw_discard_sectors = min_not_zero(t->max_hw_discard_sectors,
 763                                                          b->max_hw_discard_sectors);
 764                 t->discard_granularity = max(t->discard_granularity,
 765                                              b->discard_granularity);
 766                 t->discard_alignment = lcm_not_zero(t->discard_alignment, alignment) %
 767                         t->discard_granularity;
 768         }
 769         t->max_secure_erase_sectors = min_not_zero(t->max_secure_erase_sectors,
 770                                                    b->max_secure_erase_sectors);
 771         t->zone_write_granularity = max(t->zone_write_granularity,
 772                                         b->zone_write_granularity);
 773         if (!(t->features & BLK_FEAT_ZONED)) {
 774                 t->zone_write_granularity = 0;
 775                 t->max_zone_append_sectors = 0;
 776         }
 777         blk_stack_atomic_writes_limits(t, b);
 778
 779         return ret;
 780 }
 781 EXPORT_SYMBOL(blk_stack_limits);
 782
 783 /**
 784  * queue_limits_stack_bdev - adjust queue_limits for stacked devices
 785  * @t:  the stacking driver limits (top device)
 786  * @bdev:  the underlying block device (bottom)
 787  * @offset:  offset to beginning of data within component device
 788  * @pfx: prefix to use for warnings logged
 789  *
 790  * Description:
 791  *    This function is used by stacking drivers like MD and DM to ensure
 792  *    that all component devices have compatible block sizes and
 793  *    alignments.  The stacking driver must provide a queue_limits
 794  *    struct (top) and then iteratively call the stacking function for
 795  *    all component (bottom) devices.  The stacking function will
 796  *    attempt to combine the values and ensure proper alignment.
 797  */
 798 void queue_limits_stack_bdev(struct queue_limits *t, struct block_device *bdev,
 799                 sector_t offset, const char *pfx)
 800 {
 801         if (blk_stack_limits(t, bdev_limits(bdev),
 802                         get_start_sect(bdev) + offset))
 803                 pr_notice("%s: Warning: Device %pg is misaligned\n",
 804                         pfx, bdev);
 805 }
 806 EXPORT_SYMBOL_GPL(queue_limits_stack_bdev);
 807
 808 /**
 809  * queue_limits_stack_integrity - stack integrity profile
 810  * @t: target queue limits
 811  * @b: base queue limits
 812  *
 813  * Check if the integrity profile in the @b can be stacked into the
 814  * target @t.  Stacking is possible if either:
 815  *
 816  *   a) does not have any integrity information stacked into it yet
 817  *   b) the integrity profile in @b is identical to the one in @t
 818  *
 819  * If @b can be stacked into @t, return %true.  Else return %false and clear the
 820  * integrity information in @t.
 821  */
 822 bool queue_limits_stack_integrity(struct queue_limits *t,
 823                 struct queue_limits *b)
 824 {
 825         struct blk_integrity *ti = &t->integrity;
 826         struct blk_integrity *bi = &b->integrity;
 827
 828         if (!IS_ENABLED(CONFIG_BLK_DEV_INTEGRITY))
 829                 return true;
 830
 831         if (!ti->tuple_size) {
 832                 /* inherit the settings from the first underlying device */
 833                 if (!(ti->flags & BLK_INTEGRITY_STACKED)) {
 834                         ti->flags = BLK_INTEGRITY_DEVICE_CAPABLE |
 835                                 (bi->flags & BLK_INTEGRITY_REF_TAG);
 836                         ti->csum_type = bi->csum_type;
 837                         ti->tuple_size = bi->tuple_size;
 838                         ti->pi_offset = bi->pi_offset;
 839                         ti->interval_exp = bi->interval_exp;
 840                         ti->tag_size = bi->tag_size;
 841                         goto done;
 842                 }
 843                 if (!bi->tuple_size)
 844                         goto done;
 845         }
 846
 847         if (ti->tuple_size != bi->tuple_size)
 848                 goto incompatible;
 849         if (ti->interval_exp != bi->interval_exp)
 850                 goto incompatible;
 851         if (ti->tag_size != bi->tag_size)
 852                 goto incompatible;
 853         if (ti->csum_type != bi->csum_type)
 854                 goto incompatible;
 855         if ((ti->flags & BLK_INTEGRITY_REF_TAG) !=
 856             (bi->flags & BLK_INTEGRITY_REF_TAG))
 857                 goto incompatible;
 858
 859 done:
 860         ti->flags |= BLK_INTEGRITY_STACKED;
 861         return true;
 862
 863 incompatible:
 864         memset(ti, 0, sizeof(*ti));
 865         return false;
 866 }
 867 EXPORT_SYMBOL_GPL(queue_limits_stack_integrity);
 868
 869 /**
 870  * blk_set_queue_depth - tell the block layer about the device queue depth
 871  * @q:          the request queue for the device
 872  * @depth:              queue depth
 873  *
 874  */
 875 void blk_set_queue_depth(struct request_queue *q, unsigned int depth)
 876 {
 877         q->queue_depth = depth;
 878         rq_qos_queue_depth_changed(q);
 879 }
 880 EXPORT_SYMBOL(blk_set_queue_depth);
 881
 882 int bdev_alignment_offset(struct block_device *bdev)
 883 {
 884         struct request_queue *q = bdev_get_queue(bdev);
 885
 886         if (q->limits.flags & BLK_FLAG_MISALIGNED)
 887                 return -1;
 888         if (bdev_is_partition(bdev))
 889                 return queue_limit_alignment_offset(&q->limits,
 890                                 bdev->bd_start_sect);
 891         return q->limits.alignment_offset;
 892 }
 893 EXPORT_SYMBOL_GPL(bdev_alignment_offset);
 894
 895 unsigned int bdev_discard_alignment(struct block_device *bdev)
 896 {
 897         struct request_queue *q = bdev_get_queue(bdev);
 898
 899         if (bdev_is_partition(bdev))
 900                 return queue_limit_discard_alignment(&q->limits,
 901                                 bdev->bd_start_sect);
 902         return q->limits.discard_alignment;
 903 }
 904 EXPORT_SYMBOL_GPL(bdev_discard_alignment);