drivers/md/raid10.c

   1 /*
   2  * raid10.c : Multiple Devices driver for Linux
   3  *
   4  * Copyright (C) 2000-2004 Neil Brown
   5  *
   6  * RAID-10 support for md.
   7  *
   8  * Base on code in raid1.c.  See raid1.c for futher copyright information.
   9  *
  10  *
  11  * This program is free software; you can redistribute it and/or modify
  12  * it under the terms of the GNU General Public License as published by
  13  * the Free Software Foundation; either version 2, or (at your option)
  14  * any later version.
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * (for example /usr/src/linux/COPYING); if not, write to the Free
  18  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  19  */
  20
  21 #include <linux/delay.h>
  22 #include <linux/blkdev.h>
  23 #include <linux/seq_file.h>
  24 #include "md.h"
  25 #include "raid10.h"
  26 #include "bitmap.h"
  27
  28 /*
  29  * RAID10 provides a combination of RAID0 and RAID1 functionality.
  30  * The layout of data is defined by
  31  *    chunk_size
  32  *    raid_disks
  33  *    near_copies (stored in low byte of layout)
  34  *    far_copies (stored in second byte of layout)
  35  *    far_offset (stored in bit 16 of layout )
  36  *
  37  * The data to be stored is divided into chunks using chunksize.
  38  * Each device is divided into far_copies sections.
  39  * In each section, chunks are laid out in a style similar to raid0, but
  40  * near_copies copies of each chunk is stored (each on a different drive).
  41  * The starting device for each section is offset near_copies from the starting
  42  * device of the previous section.
  43  * Thus they are (near_copies*far_copies) of each chunk, and each is on a different
  44  * drive.
  45  * near_copies and far_copies must be at least one, and their product is at most
  46  * raid_disks.
  47  *
  48  * If far_offset is true, then the far_copies are handled a bit differently.
  49  * The copies are still in different stripes, but instead of be very far apart
  50  * on disk, there are adjacent stripes.
  51  */
  52
  53 /*
  54  * Number of guaranteed r10bios in case of extreme VM load:
  55  */
  56 #define NR_RAID10_BIOS 256
  57
  58 static void unplug_slaves(mddev_t *mddev);
  59
  60 static void allow_barrier(conf_t *conf);
  61 static void lower_barrier(conf_t *conf);
  62
  63 static void * r10bio_pool_alloc(gfp_t gfp_flags, void *data)
  64 {
  65         conf_t *conf = data;
  66         r10bio_t *r10_bio;
  67         int size = offsetof(struct r10bio_s, devs[conf->copies]);
  68
  69         /* allocate a r10bio with room for raid_disks entries in the bios array */
  70         r10_bio = kzalloc(size, gfp_flags);
  71         if (!r10_bio)
  72                 unplug_slaves(conf->mddev);
  73
  74         return r10_bio;
  75 }
  76
  77 static void r10bio_pool_free(void *r10_bio, void *data)
  78 {
  79         kfree(r10_bio);
  80 }
  81
  82 /* Maximum size of each resync request */
  83 #define RESYNC_BLOCK_SIZE (64*1024)
  84 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  85 /* amount of memory to reserve for resync requests */
  86 #define RESYNC_WINDOW (1024*1024)
  87 /* maximum number of concurrent requests, memory permitting */
  88 #define RESYNC_DEPTH (32*1024*1024/RESYNC_BLOCK_SIZE)
  89
  90 /*
  91  * When performing a resync, we need to read and compare, so
  92  * we need as many pages are there are copies.
  93  * When performing a recovery, we need 2 bios, one for read,
  94  * one for write (we recover only one drive per r10buf)
  95  *
  96  */
  97 static void * r10buf_pool_alloc(gfp_t gfp_flags, void *data)
  98 {
  99         conf_t *conf = data;
 100         struct page *page;
 101         r10bio_t *r10_bio;
 102         struct bio *bio;
 103         int i, j;
 104         int nalloc;
 105
 106         r10_bio = r10bio_pool_alloc(gfp_flags, conf);
 107         if (!r10_bio) {
 108                 unplug_slaves(conf->mddev);
 109                 return NULL;
 110         }
 111
 112         if (test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
 113                 nalloc = conf->copies; /* resync */
 114         else
 115                 nalloc = 2; /* recovery */
 116
 117         /*
 118          * Allocate bios.
 119          */
 120         for (j = nalloc ; j-- ; ) {
 121                 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
 122                 if (!bio)
 123                         goto out_free_bio;
 124                 r10_bio->devs[j].bio = bio;
 125         }
 126         /*
 127          * Allocate RESYNC_PAGES data pages and attach them
 128          * where needed.
 129          */
 130         for (j = 0 ; j < nalloc; j++) {
 131                 bio = r10_bio->devs[j].bio;
 132                 for (i = 0; i < RESYNC_PAGES; i++) {
 133                         page = alloc_page(gfp_flags);
 134                         if (unlikely(!page))
 135                                 goto out_free_pages;
 136
 137                         bio->bi_io_vec[i].bv_page = page;
 138                 }
 139         }
 140
 141         return r10_bio;
 142
 143 out_free_pages:
 144         for ( ; i > 0 ; i--)
 145                 safe_put_page(bio->bi_io_vec[i-1].bv_page);
 146         while (j--)
 147                 for (i = 0; i < RESYNC_PAGES ; i++)
 148                         safe_put_page(r10_bio->devs[j].bio->bi_io_vec[i].bv_page);
 149         j = -1;
 150 out_free_bio:
 151         while ( ++j < nalloc )
 152                 bio_put(r10_bio->devs[j].bio);
 153         r10bio_pool_free(r10_bio, conf);
 154         return NULL;
 155 }
 156
 157 static void r10buf_pool_free(void *__r10_bio, void *data)
 158 {
 159         int i;
 160         conf_t *conf = data;
 161         r10bio_t *r10bio = __r10_bio;
 162         int j;
 163
 164         for (j=0; j < conf->copies; j++) {
 165                 struct bio *bio = r10bio->devs[j].bio;
 166                 if (bio) {
 167                         for (i = 0; i < RESYNC_PAGES; i++) {
 168                                 safe_put_page(bio->bi_io_vec[i].bv_page);
 169                                 bio->bi_io_vec[i].bv_page = NULL;
 170                         }
 171                         bio_put(bio);
 172                 }
 173         }
 174         r10bio_pool_free(r10bio, conf);
 175 }
 176
 177 static void put_all_bios(conf_t *conf, r10bio_t *r10_bio)
 178 {
 179         int i;
 180
 181         for (i = 0; i < conf->copies; i++) {
 182                 struct bio **bio = & r10_bio->devs[i].bio;
 183                 if (*bio && *bio != IO_BLOCKED)
 184                         bio_put(*bio);
 185                 *bio = NULL;
 186         }
 187 }
 188
 189 static void free_r10bio(r10bio_t *r10_bio)
 190 {
 191         conf_t *conf = r10_bio->mddev->private;
 192
 193         /*
 194          * Wake up any possible resync thread that waits for the device
 195          * to go idle.
 196          */
 197         allow_barrier(conf);
 198
 199         put_all_bios(conf, r10_bio);
 200         mempool_free(r10_bio, conf->r10bio_pool);
 201 }
 202
 203 static void put_buf(r10bio_t *r10_bio)
 204 {
 205         conf_t *conf = r10_bio->mddev->private;
 206
 207         mempool_free(r10_bio, conf->r10buf_pool);
 208
 209         lower_barrier(conf);
 210 }
 211
 212 static void reschedule_retry(r10bio_t *r10_bio)
 213 {
 214         unsigned long flags;
 215         mddev_t *mddev = r10_bio->mddev;
 216         conf_t *conf = mddev->private;
 217
 218         spin_lock_irqsave(&conf->device_lock, flags);
 219         list_add(&r10_bio->retry_list, &conf->retry_list);
 220         conf->nr_queued ++;
 221         spin_unlock_irqrestore(&conf->device_lock, flags);
 222
 223         /* wake up frozen array... */
 224         wake_up(&conf->wait_barrier);
 225
 226         md_wakeup_thread(mddev->thread);
 227 }
 228
 229 /*
 230  * raid_end_bio_io() is called when we have finished servicing a mirrored
 231  * operation and are ready to return a success/failure code to the buffer
 232  * cache layer.
 233  */
 234 static void raid_end_bio_io(r10bio_t *r10_bio)
 235 {
 236         struct bio *bio = r10_bio->master_bio;
 237
 238         bio_endio(bio,
 239                 test_bit(R10BIO_Uptodate, &r10_bio->state) ? 0 : -EIO);
 240         free_r10bio(r10_bio);
 241 }
 242
 243 /*
 244  * Update disk head position estimator based on IRQ completion info.
 245  */
 246 static inline void update_head_pos(int slot, r10bio_t *r10_bio)
 247 {
 248         conf_t *conf = r10_bio->mddev->private;
 249
 250         conf->mirrors[r10_bio->devs[slot].devnum].head_position =
 251                 r10_bio->devs[slot].addr + (r10_bio->sectors);
 252 }
 253
 254 static void raid10_end_read_request(struct bio *bio, int error)
 255 {
 256         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 257         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
 258         int slot, dev;
 259         conf_t *conf = r10_bio->mddev->private;
 260
 261
 262         slot = r10_bio->read_slot;
 263         dev = r10_bio->devs[slot].devnum;
 264         /*
 265          * this branch is our 'one mirror IO has finished' event handler:
 266          */
 267         update_head_pos(slot, r10_bio);
 268
 269         if (uptodate) {
 270                 /*
 271                  * Set R10BIO_Uptodate in our master bio, so that
 272                  * we will return a good error code to the higher
 273                  * levels even if IO on some other mirrored buffer fails.
 274                  *
 275                  * The 'master' represents the composite IO operation to
 276                  * user-side. So if something waits for IO, then it will
 277                  * wait for the 'master' bio.
 278                  */
 279                 set_bit(R10BIO_Uptodate, &r10_bio->state);
 280                 raid_end_bio_io(r10_bio);
 281         } else {
 282                 /*
 283                  * oops, read error:
 284                  */
 285                 char b[BDEVNAME_SIZE];
 286                 if (printk_ratelimit())
 287                         printk(KERN_ERR "raid10: %s: rescheduling sector %llu\n",
 288                                bdevname(conf->mirrors[dev].rdev->bdev,b), (unsigned long long)r10_bio->sector);
 289                 reschedule_retry(r10_bio);
 290         }
 291
 292         rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
 293 }
 294
 295 static void raid10_end_write_request(struct bio *bio, int error)
 296 {
 297         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 298         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
 299         int slot, dev;
 300         conf_t *conf = r10_bio->mddev->private;
 301
 302         for (slot = 0; slot < conf->copies; slot++)
 303                 if (r10_bio->devs[slot].bio == bio)
 304                         break;
 305         dev = r10_bio->devs[slot].devnum;
 306
 307         /*
 308          * this branch is our 'one mirror IO has finished' event handler:
 309          */
 310         if (!uptodate) {
 311                 md_error(r10_bio->mddev, conf->mirrors[dev].rdev);
 312                 /* an I/O failed, we can't clear the bitmap */
 313                 set_bit(R10BIO_Degraded, &r10_bio->state);
 314         } else
 315                 /*
 316                  * Set R10BIO_Uptodate in our master bio, so that
 317                  * we will return a good error code for to the higher
 318                  * levels even if IO on some other mirrored buffer fails.
 319                  *
 320                  * The 'master' represents the composite IO operation to
 321                  * user-side. So if something waits for IO, then it will
 322                  * wait for the 'master' bio.
 323                  */
 324                 set_bit(R10BIO_Uptodate, &r10_bio->state);
 325
 326         update_head_pos(slot, r10_bio);
 327
 328         /*
 329          *
 330          * Let's see if all mirrored write operations have finished
 331          * already.
 332          */
 333         if (atomic_dec_and_test(&r10_bio->remaining)) {
 334                 /* clear the bitmap if all writes complete successfully */
 335                 bitmap_endwrite(r10_bio->mddev->bitmap, r10_bio->sector,
 336                                 r10_bio->sectors,
 337                                 !test_bit(R10BIO_Degraded, &r10_bio->state),
 338                                 0);
 339                 md_write_end(r10_bio->mddev);
 340                 raid_end_bio_io(r10_bio);
 341         }
 342
 343         rdev_dec_pending(conf->mirrors[dev].rdev, conf->mddev);
 344 }
 345
 346
 347 /*
 348  * RAID10 layout manager
 349  * Aswell as the chunksize and raid_disks count, there are two
 350  * parameters: near_copies and far_copies.
 351  * near_copies * far_copies must be <= raid_disks.
 352  * Normally one of these will be 1.
 353  * If both are 1, we get raid0.
 354  * If near_copies == raid_disks, we get raid1.
 355  *
 356  * Chunks are layed out in raid0 style with near_copies copies of the
 357  * first chunk, followed by near_copies copies of the next chunk and
 358  * so on.
 359  * If far_copies > 1, then after 1/far_copies of the array has been assigned
 360  * as described above, we start again with a device offset of near_copies.
 361  * So we effectively have another copy of the whole array further down all
 362  * the drives, but with blocks on different drives.
 363  * With this layout, and block is never stored twice on the one device.
 364  *
 365  * raid10_find_phys finds the sector offset of a given virtual sector
 366  * on each device that it is on.
 367  *
 368  * raid10_find_virt does the reverse mapping, from a device and a
 369  * sector offset to a virtual address
 370  */
 371
 372 static void raid10_find_phys(conf_t *conf, r10bio_t *r10bio)
 373 {
 374         int n,f;
 375         sector_t sector;
 376         sector_t chunk;
 377         sector_t stripe;
 378         int dev;
 379
 380         int slot = 0;
 381
 382         /* now calculate first sector/dev */
 383         chunk = r10bio->sector >> conf->chunk_shift;
 384         sector = r10bio->sector & conf->chunk_mask;
 385
 386         chunk *= conf->near_copies;
 387         stripe = chunk;
 388         dev = sector_div(stripe, conf->raid_disks);
 389         if (conf->far_offset)
 390                 stripe *= conf->far_copies;
 391
 392         sector += stripe << conf->chunk_shift;
 393
 394         /* and calculate all the others */
 395         for (n=0; n < conf->near_copies; n++) {
 396                 int d = dev;
 397                 sector_t s = sector;
 398                 r10bio->devs[slot].addr = sector;
 399                 r10bio->devs[slot].devnum = d;
 400                 slot++;
 401
 402                 for (f = 1; f < conf->far_copies; f++) {
 403                         d += conf->near_copies;
 404                         if (d >= conf->raid_disks)
 405                                 d -= conf->raid_disks;
 406                         s += conf->stride;
 407                         r10bio->devs[slot].devnum = d;
 408                         r10bio->devs[slot].addr = s;
 409                         slot++;
 410                 }
 411                 dev++;
 412                 if (dev >= conf->raid_disks) {
 413                         dev = 0;
 414                         sector += (conf->chunk_mask + 1);
 415                 }
 416         }
 417         BUG_ON(slot != conf->copies);
 418 }
 419
 420 static sector_t raid10_find_virt(conf_t *conf, sector_t sector, int dev)
 421 {
 422         sector_t offset, chunk, vchunk;
 423
 424         offset = sector & conf->chunk_mask;
 425         if (conf->far_offset) {
 426                 int fc;
 427                 chunk = sector >> conf->chunk_shift;
 428                 fc = sector_div(chunk, conf->far_copies);
 429                 dev -= fc * conf->near_copies;
 430                 if (dev < 0)
 431                         dev += conf->raid_disks;
 432         } else {
 433                 while (sector >= conf->stride) {
 434                         sector -= conf->stride;
 435                         if (dev < conf->near_copies)
 436                                 dev += conf->raid_disks - conf->near_copies;
 437                         else
 438                                 dev -= conf->near_copies;
 439                 }
 440                 chunk = sector >> conf->chunk_shift;
 441         }
 442         vchunk = chunk * conf->raid_disks + dev;
 443         sector_div(vchunk, conf->near_copies);
 444         return (vchunk << conf->chunk_shift) + offset;
 445 }
 446
 447 /**
 448  *      raid10_mergeable_bvec -- tell bio layer if a two requests can be merged
 449  *      @q: request queue
 450  *      @bvm: properties of new bio
 451  *      @biovec: the request that could be merged to it.
 452  *
 453  *      Return amount of bytes we can accept at this offset
 454  *      If near_copies == raid_disk, there are no striping issues,
 455  *      but in that case, the function isn't called at all.
 456  */
 457 static int raid10_mergeable_bvec(struct request_queue *q,
 458                                  struct bvec_merge_data *bvm,
 459                                  struct bio_vec *biovec)
 460 {
 461         mddev_t *mddev = q->queuedata;
 462         sector_t sector = bvm->bi_sector + get_start_sect(bvm->bi_bdev);
 463         int max;
 464         unsigned int chunk_sectors = mddev->chunk_sectors;
 465         unsigned int bio_sectors = bvm->bi_size >> 9;
 466
 467         max =  (chunk_sectors - ((sector & (chunk_sectors - 1)) + bio_sectors)) << 9;
 468         if (max < 0) max = 0; /* bio_add cannot handle a negative return */
 469         if (max <= biovec->bv_len && bio_sectors == 0)
 470                 return biovec->bv_len;
 471         else
 472                 return max;
 473 }
 474
 475 /*
 476  * This routine returns the disk from which the requested read should
 477  * be done. There is a per-array 'next expected sequential IO' sector
 478  * number - if this matches on the next IO then we use the last disk.
 479  * There is also a per-disk 'last know head position' sector that is
 480  * maintained from IRQ contexts, both the normal and the resync IO
 481  * completion handlers update this position correctly. If there is no
 482  * perfect sequential match then we pick the disk whose head is closest.
 483  *
 484  * If there are 2 mirrors in the same 2 devices, performance degrades
 485  * because position is mirror, not device based.
 486  *
 487  * The rdev for the device selected will have nr_pending incremented.
 488  */
 489
 490 /*
 491  * FIXME: possibly should rethink readbalancing and do it differently
 492  * depending on near_copies / far_copies geometry.
 493  */
 494 static int read_balance(conf_t *conf, r10bio_t *r10_bio)
 495 {
 496         const unsigned long this_sector = r10_bio->sector;
 497         int disk, slot, nslot;
 498         const int sectors = r10_bio->sectors;
 499         sector_t new_distance, current_distance;
 500         mdk_rdev_t *rdev;
 501
 502         raid10_find_phys(conf, r10_bio);
 503         rcu_read_lock();
 504         /*
 505          * Check if we can balance. We can balance on the whole
 506          * device if no resync is going on (recovery is ok), or below
 507          * the resync window. We take the first readable disk when
 508          * above the resync window.
 509          */
 510         if (conf->mddev->recovery_cp < MaxSector
 511             && (this_sector + sectors >= conf->next_resync)) {
 512                 /* make sure that disk is operational */
 513                 slot = 0;
 514                 disk = r10_bio->devs[slot].devnum;
 515
 516                 while ((rdev = rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
 517                        r10_bio->devs[slot].bio == IO_BLOCKED ||
 518                        !test_bit(In_sync, &rdev->flags)) {
 519                         slot++;
 520                         if (slot == conf->copies) {
 521                                 slot = 0;
 522                                 disk = -1;
 523                                 break;
 524                         }
 525                         disk = r10_bio->devs[slot].devnum;
 526                 }
 527                 goto rb_out;
 528         }
 529
 530
 531         /* make sure the disk is operational */
 532         slot = 0;
 533         disk = r10_bio->devs[slot].devnum;
 534         while ((rdev=rcu_dereference(conf->mirrors[disk].rdev)) == NULL ||
 535                r10_bio->devs[slot].bio == IO_BLOCKED ||
 536                !test_bit(In_sync, &rdev->flags)) {
 537                 slot ++;
 538                 if (slot == conf->copies) {
 539                         disk = -1;
 540                         goto rb_out;
 541                 }
 542                 disk = r10_bio->devs[slot].devnum;
 543         }
 544
 545
 546         current_distance = abs(r10_bio->devs[slot].addr -
 547                                conf->mirrors[disk].head_position);
 548
 549         /* Find the disk whose head is closest,
 550          * or - for far > 1 - find the closest to partition beginning */
 551
 552         for (nslot = slot; nslot < conf->copies; nslot++) {
 553                 int ndisk = r10_bio->devs[nslot].devnum;
 554
 555
 556                 if ((rdev=rcu_dereference(conf->mirrors[ndisk].rdev)) == NULL ||
 557                     r10_bio->devs[nslot].bio == IO_BLOCKED ||
 558                     !test_bit(In_sync, &rdev->flags))
 559                         continue;
 560
 561                 /* This optimisation is debatable, and completely destroys
 562                  * sequential read speed for 'far copies' arrays.  So only
 563                  * keep it for 'near' arrays, and review those later.
 564                  */
 565                 if (conf->near_copies > 1 && !atomic_read(&rdev->nr_pending)) {
 566                         disk = ndisk;
 567                         slot = nslot;
 568                         break;
 569                 }
 570
 571                 /* for far > 1 always use the lowest address */
 572                 if (conf->far_copies > 1)
 573                         new_distance = r10_bio->devs[nslot].addr;
 574                 else
 575                         new_distance = abs(r10_bio->devs[nslot].addr -
 576                                            conf->mirrors[ndisk].head_position);
 577                 if (new_distance < current_distance) {
 578                         current_distance = new_distance;
 579                         disk = ndisk;
 580                         slot = nslot;
 581                 }
 582         }
 583
 584 rb_out:
 585         r10_bio->read_slot = slot;
 586 /*      conf->next_seq_sect = this_sector + sectors;*/
 587
 588         if (disk >= 0 && (rdev=rcu_dereference(conf->mirrors[disk].rdev))!= NULL)
 589                 atomic_inc(&conf->mirrors[disk].rdev->nr_pending);
 590         else
 591                 disk = -1;
 592         rcu_read_unlock();
 593
 594         return disk;
 595 }
 596
 597 static void unplug_slaves(mddev_t *mddev)
 598 {
 599         conf_t *conf = mddev->private;
 600         int i;
 601
 602         rcu_read_lock();
 603         for (i=0; i<mddev->raid_disks; i++) {
 604                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 605                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
 606                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
 607
 608                         atomic_inc(&rdev->nr_pending);
 609                         rcu_read_unlock();
 610
 611                         blk_unplug(r_queue);
 612
 613                         rdev_dec_pending(rdev, mddev);
 614                         rcu_read_lock();
 615                 }
 616         }
 617         rcu_read_unlock();
 618 }
 619
 620 static void raid10_unplug(struct request_queue *q)
 621 {
 622         mddev_t *mddev = q->queuedata;
 623
 624         unplug_slaves(q->queuedata);
 625         md_wakeup_thread(mddev->thread);
 626 }
 627
 628 static int raid10_congested(void *data, int bits)
 629 {
 630         mddev_t *mddev = data;
 631         conf_t *conf = mddev->private;
 632         int i, ret = 0;
 633
 634         rcu_read_lock();
 635         for (i = 0; i < mddev->raid_disks && ret == 0; i++) {
 636                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 637                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 638                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 639
 640                         ret |= bdi_congested(&q->backing_dev_info, bits);
 641                 }
 642         }
 643         rcu_read_unlock();
 644         return ret;
 645 }
 646
 647 static int flush_pending_writes(conf_t *conf)
 648 {
 649         /* Any writes that have been queued but are awaiting
 650          * bitmap updates get flushed here.
 651          * We return 1 if any requests were actually submitted.
 652          */
 653         int rv = 0;
 654
 655         spin_lock_irq(&conf->device_lock);
 656
 657         if (conf->pending_bio_list.head) {
 658                 struct bio *bio;
 659                 bio = bio_list_get(&conf->pending_bio_list);
 660                 blk_remove_plug(conf->mddev->queue);
 661                 spin_unlock_irq(&conf->device_lock);
 662                 /* flush any pending bitmap writes to disk
 663                  * before proceeding w/ I/O */
 664                 bitmap_unplug(conf->mddev->bitmap);
 665
 666                 while (bio) { /* submit pending writes */
 667                         struct bio *next = bio->bi_next;
 668                         bio->bi_next = NULL;
 669                         generic_make_request(bio);
 670                         bio = next;
 671                 }
 672                 rv = 1;
 673         } else
 674                 spin_unlock_irq(&conf->device_lock);
 675         return rv;
 676 }
 677 /* Barriers....
 678  * Sometimes we need to suspend IO while we do something else,
 679  * either some resync/recovery, or reconfigure the array.
 680  * To do this we raise a 'barrier'.
 681  * The 'barrier' is a counter that can be raised multiple times
 682  * to count how many activities are happening which preclude
 683  * normal IO.
 684  * We can only raise the barrier if there is no pending IO.
 685  * i.e. if nr_pending == 0.
 686  * We choose only to raise the barrier if no-one is waiting for the
 687  * barrier to go down.  This means that as soon as an IO request
 688  * is ready, no other operations which require a barrier will start
 689  * until the IO request has had a chance.
 690  *
 691  * So: regular IO calls 'wait_barrier'.  When that returns there
 692  *    is no backgroup IO happening,  It must arrange to call
 693  *    allow_barrier when it has finished its IO.
 694  * backgroup IO calls must call raise_barrier.  Once that returns
 695  *    there is no normal IO happeing.  It must arrange to call
 696  *    lower_barrier when the particular background IO completes.
 697  */
 698
 699 static void raise_barrier(conf_t *conf, int force)
 700 {
 701         BUG_ON(force && !conf->barrier);
 702         spin_lock_irq(&conf->resync_lock);
 703
 704         /* Wait until no block IO is waiting (unless 'force') */
 705         wait_event_lock_irq(conf->wait_barrier, force || !conf->nr_waiting,
 706                             conf->resync_lock,
 707                             raid10_unplug(conf->mddev->queue));
 708
 709         /* block any new IO from starting */
 710         conf->barrier++;
 711
 712         /* No wait for all pending IO to complete */
 713         wait_event_lock_irq(conf->wait_barrier,
 714                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 715                             conf->resync_lock,
 716                             raid10_unplug(conf->mddev->queue));
 717
 718         spin_unlock_irq(&conf->resync_lock);
 719 }
 720
 721 static void lower_barrier(conf_t *conf)
 722 {
 723         unsigned long flags;
 724         spin_lock_irqsave(&conf->resync_lock, flags);
 725         conf->barrier--;
 726         spin_unlock_irqrestore(&conf->resync_lock, flags);
 727         wake_up(&conf->wait_barrier);
 728 }
 729
 730 static void wait_barrier(conf_t *conf)
 731 {
 732         spin_lock_irq(&conf->resync_lock);
 733         if (conf->barrier) {
 734                 conf->nr_waiting++;
 735                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
 736                                     conf->resync_lock,
 737                                     raid10_unplug(conf->mddev->queue));
 738                 conf->nr_waiting--;
 739         }
 740         conf->nr_pending++;
 741         spin_unlock_irq(&conf->resync_lock);
 742 }
 743
 744 static void allow_barrier(conf_t *conf)
 745 {
 746         unsigned long flags;
 747         spin_lock_irqsave(&conf->resync_lock, flags);
 748         conf->nr_pending--;
 749         spin_unlock_irqrestore(&conf->resync_lock, flags);
 750         wake_up(&conf->wait_barrier);
 751 }
 752
 753 static void freeze_array(conf_t *conf)
 754 {
 755         /* stop syncio and normal IO and wait for everything to
 756          * go quiet.
 757          * We increment barrier and nr_waiting, and then
 758          * wait until nr_pending match nr_queued+1
 759          * This is called in the context of one normal IO request
 760          * that has failed. Thus any sync request that might be pending
 761          * will be blocked by nr_pending, and we need to wait for
 762          * pending IO requests to complete or be queued for re-try.
 763          * Thus the number queued (nr_queued) plus this request (1)
 764          * must match the number of pending IOs (nr_pending) before
 765          * we continue.
 766          */
 767         spin_lock_irq(&conf->resync_lock);
 768         conf->barrier++;
 769         conf->nr_waiting++;
 770         wait_event_lock_irq(conf->wait_barrier,
 771                             conf->nr_pending == conf->nr_queued+1,
 772                             conf->resync_lock,
 773                             ({ flush_pending_writes(conf);
 774                                raid10_unplug(conf->mddev->queue); }));
 775         spin_unlock_irq(&conf->resync_lock);
 776 }
 777
 778 static void unfreeze_array(conf_t *conf)
 779 {
 780         /* reverse the effect of the freeze */
 781         spin_lock_irq(&conf->resync_lock);
 782         conf->barrier--;
 783         conf->nr_waiting--;
 784         wake_up(&conf->wait_barrier);
 785         spin_unlock_irq(&conf->resync_lock);
 786 }
 787
 788 static int make_request(struct request_queue *q, struct bio * bio)
 789 {
 790         mddev_t *mddev = q->queuedata;
 791         conf_t *conf = mddev->private;
 792         mirror_info_t *mirror;
 793         r10bio_t *r10_bio;
 794         struct bio *read_bio;
 795         int cpu;
 796         int i;
 797         int chunk_sects = conf->chunk_mask + 1;
 798         const int rw = bio_data_dir(bio);
 799         const int do_sync = bio_sync(bio);
 800         struct bio_list bl;
 801         unsigned long flags;
 802         mdk_rdev_t *blocked_rdev;
 803
 804         if (unlikely(bio_barrier(bio))) {
 805                 bio_endio(bio, -EOPNOTSUPP);
 806                 return 0;
 807         }
 808
 809         /* If this request crosses a chunk boundary, we need to
 810          * split it.  This will only happen for 1 PAGE (or less) requests.
 811          */
 812         if (unlikely( (bio->bi_sector & conf->chunk_mask) + (bio->bi_size >> 9)
 813                       > chunk_sects &&
 814                     conf->near_copies < conf->raid_disks)) {
 815                 struct bio_pair *bp;
 816                 /* Sanity check -- queue functions should prevent this happening */
 817                 if (bio->bi_vcnt != 1 ||
 818                     bio->bi_idx != 0)
 819                         goto bad_map;
 820                 /* This is a one page bio that upper layers
 821                  * refuse to split for us, so we need to split it.
 822                  */
 823                 bp = bio_split(bio,
 824                                chunk_sects - (bio->bi_sector & (chunk_sects - 1)) );
 825                 if (make_request(q, &bp->bio1))
 826                         generic_make_request(&bp->bio1);
 827                 if (make_request(q, &bp->bio2))
 828                         generic_make_request(&bp->bio2);
 829
 830                 bio_pair_release(bp);
 831                 return 0;
 832         bad_map:
 833                 printk("raid10_make_request bug: can't convert block across chunks"
 834                        " or bigger than %dk %llu %d\n", chunk_sects/2,
 835                        (unsigned long long)bio->bi_sector, bio->bi_size >> 10);
 836
 837                 bio_io_error(bio);
 838                 return 0;
 839         }
 840
 841         md_write_start(mddev, bio);
 842
 843         /*
 844          * Register the new request and wait if the reconstruction
 845          * thread has put up a bar for new requests.
 846          * Continue immediately if no resync is active currently.
 847          */
 848         wait_barrier(conf);
 849
 850         cpu = part_stat_lock();
 851         part_stat_inc(cpu, &mddev->gendisk->part0, ios[rw]);
 852         part_stat_add(cpu, &mddev->gendisk->part0, sectors[rw],
 853                       bio_sectors(bio));
 854         part_stat_unlock();
 855
 856         r10_bio = mempool_alloc(conf->r10bio_pool, GFP_NOIO);
 857
 858         r10_bio->master_bio = bio;
 859         r10_bio->sectors = bio->bi_size >> 9;
 860
 861         r10_bio->mddev = mddev;
 862         r10_bio->sector = bio->bi_sector;
 863         r10_bio->state = 0;
 864
 865         if (rw == READ) {
 866                 /*
 867                  * read balancing logic:
 868                  */
 869                 int disk = read_balance(conf, r10_bio);
 870                 int slot = r10_bio->read_slot;
 871                 if (disk < 0) {
 872                         raid_end_bio_io(r10_bio);
 873                         return 0;
 874                 }
 875                 mirror = conf->mirrors + disk;
 876
 877                 read_bio = bio_clone(bio, GFP_NOIO);
 878
 879                 r10_bio->devs[slot].bio = read_bio;
 880
 881                 read_bio->bi_sector = r10_bio->devs[slot].addr +
 882                         mirror->rdev->data_offset;
 883                 read_bio->bi_bdev = mirror->rdev->bdev;
 884                 read_bio->bi_end_io = raid10_end_read_request;
 885                 read_bio->bi_rw = READ | do_sync;
 886                 read_bio->bi_private = r10_bio;
 887
 888                 generic_make_request(read_bio);
 889                 return 0;
 890         }
 891
 892         /*
 893          * WRITE:
 894          */
 895         /* first select target devices under rcu_lock and
 896          * inc refcount on their rdev.  Record them by setting
 897          * bios[x] to bio
 898          */
 899         raid10_find_phys(conf, r10_bio);
 900  retry_write:
 901         blocked_rdev = NULL;
 902         rcu_read_lock();
 903         for (i = 0;  i < conf->copies; i++) {
 904                 int d = r10_bio->devs[i].devnum;
 905                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[d].rdev);
 906                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
 907                         atomic_inc(&rdev->nr_pending);
 908                         blocked_rdev = rdev;
 909                         break;
 910                 }
 911                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 912                         atomic_inc(&rdev->nr_pending);
 913                         r10_bio->devs[i].bio = bio;
 914                 } else {
 915                         r10_bio->devs[i].bio = NULL;
 916                         set_bit(R10BIO_Degraded, &r10_bio->state);
 917                 }
 918         }
 919         rcu_read_unlock();
 920
 921         if (unlikely(blocked_rdev)) {
 922                 /* Have to wait for this device to get unblocked, then retry */
 923                 int j;
 924                 int d;
 925
 926                 for (j = 0; j < i; j++)
 927                         if (r10_bio->devs[j].bio) {
 928                                 d = r10_bio->devs[j].devnum;
 929                                 rdev_dec_pending(conf->mirrors[d].rdev, mddev);
 930                         }
 931                 allow_barrier(conf);
 932                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
 933                 wait_barrier(conf);
 934                 goto retry_write;
 935         }
 936
 937         atomic_set(&r10_bio->remaining, 0);
 938
 939         bio_list_init(&bl);
 940         for (i = 0; i < conf->copies; i++) {
 941                 struct bio *mbio;
 942                 int d = r10_bio->devs[i].devnum;
 943                 if (!r10_bio->devs[i].bio)
 944                         continue;
 945
 946                 mbio = bio_clone(bio, GFP_NOIO);
 947                 r10_bio->devs[i].bio = mbio;
 948
 949                 mbio->bi_sector = r10_bio->devs[i].addr+
 950                         conf->mirrors[d].rdev->data_offset;
 951                 mbio->bi_bdev = conf->mirrors[d].rdev->bdev;
 952                 mbio->bi_end_io = raid10_end_write_request;
 953                 mbio->bi_rw = WRITE | do_sync;
 954                 mbio->bi_private = r10_bio;
 955
 956                 atomic_inc(&r10_bio->remaining);
 957                 bio_list_add(&bl, mbio);
 958         }
 959
 960         if (unlikely(!atomic_read(&r10_bio->remaining))) {
 961                 /* the array is dead */
 962                 md_write_end(mddev);
 963                 raid_end_bio_io(r10_bio);
 964                 return 0;
 965         }
 966
 967         bitmap_startwrite(mddev->bitmap, bio->bi_sector, r10_bio->sectors, 0);
 968         spin_lock_irqsave(&conf->device_lock, flags);
 969         bio_list_merge(&conf->pending_bio_list, &bl);
 970         blk_plug_device(mddev->queue);
 971         spin_unlock_irqrestore(&conf->device_lock, flags);
 972
 973         /* In case raid10d snuck in to freeze_array */
 974         wake_up(&conf->wait_barrier);
 975
 976         if (do_sync)
 977                 md_wakeup_thread(mddev->thread);
 978
 979         return 0;
 980 }
 981
 982 static void status(struct seq_file *seq, mddev_t *mddev)
 983 {
 984         conf_t *conf = mddev->private;
 985         int i;
 986
 987         if (conf->near_copies < conf->raid_disks)
 988                 seq_printf(seq, " %dK chunks", mddev->chunk_sectors / 2);
 989         if (conf->near_copies > 1)
 990                 seq_printf(seq, " %d near-copies", conf->near_copies);
 991         if (conf->far_copies > 1) {
 992                 if (conf->far_offset)
 993                         seq_printf(seq, " %d offset-copies", conf->far_copies);
 994                 else
 995                         seq_printf(seq, " %d far-copies", conf->far_copies);
 996         }
 997         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
 998                                         conf->raid_disks - mddev->degraded);
 999         for (i = 0; i < conf->raid_disks; i++)
1000                 seq_printf(seq, "%s",
1001                               conf->mirrors[i].rdev &&
1002                               test_bit(In_sync, &conf->mirrors[i].rdev->flags) ? "U" : "_");
1003         seq_printf(seq, "]");
1004 }
1005
1006 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1007 {
1008         char b[BDEVNAME_SIZE];
1009         conf_t *conf = mddev->private;
1010
1011         /*
1012          * If it is not operational, then we have already marked it as dead
1013          * else if it is the last working disks, ignore the error, let the
1014          * next level up know.
1015          * else mark the drive as failed
1016          */
1017         if (test_bit(In_sync, &rdev->flags)
1018             && conf->raid_disks-mddev->degraded == 1)
1019                 /*
1020                  * Don't fail the drive, just return an IO error.
1021                  * The test should really be more sophisticated than
1022                  * "working_disks == 1", but it isn't critical, and
1023                  * can wait until we do more sophisticated "is the drive
1024                  * really dead" tests...
1025                  */
1026                 return;
1027         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1028                 unsigned long flags;
1029                 spin_lock_irqsave(&conf->device_lock, flags);
1030                 mddev->degraded++;
1031                 spin_unlock_irqrestore(&conf->device_lock, flags);
1032                 /*
1033                  * if recovery is running, make sure it aborts.
1034                  */
1035                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1036         }
1037         set_bit(Faulty, &rdev->flags);
1038         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1039         printk(KERN_ALERT "raid10: Disk failure on %s, disabling device.\n"
1040                 "raid10: Operation continuing on %d devices.\n",
1041                 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1042 }
1043
1044 static void print_conf(conf_t *conf)
1045 {
1046         int i;
1047         mirror_info_t *tmp;
1048
1049         printk("RAID10 conf printout:\n");
1050         if (!conf) {
1051                 printk("(!conf)\n");
1052                 return;
1053         }
1054         printk(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1055                 conf->raid_disks);
1056
1057         for (i = 0; i < conf->raid_disks; i++) {
1058                 char b[BDEVNAME_SIZE];
1059                 tmp = conf->mirrors + i;
1060                 if (tmp->rdev)
1061                         printk(" disk %d, wo:%d, o:%d, dev:%s\n",
1062                                 i, !test_bit(In_sync, &tmp->rdev->flags),
1063                                 !test_bit(Faulty, &tmp->rdev->flags),
1064                                 bdevname(tmp->rdev->bdev,b));
1065         }
1066 }
1067
1068 static void close_sync(conf_t *conf)
1069 {
1070         wait_barrier(conf);
1071         allow_barrier(conf);
1072
1073         mempool_destroy(conf->r10buf_pool);
1074         conf->r10buf_pool = NULL;
1075 }
1076
1077 /* check if there are enough drives for
1078  * every block to appear on atleast one
1079  */
1080 static int enough(conf_t *conf)
1081 {
1082         int first = 0;
1083
1084         do {
1085                 int n = conf->copies;
1086                 int cnt = 0;
1087                 while (n--) {
1088                         if (conf->mirrors[first].rdev)
1089                                 cnt++;
1090                         first = (first+1) % conf->raid_disks;
1091                 }
1092                 if (cnt == 0)
1093                         return 0;
1094         } while (first != 0);
1095         return 1;
1096 }
1097
1098 static int raid10_spare_active(mddev_t *mddev)
1099 {
1100         int i;
1101         conf_t *conf = mddev->private;
1102         mirror_info_t *tmp;
1103
1104         /*
1105          * Find all non-in_sync disks within the RAID10 configuration
1106          * and mark them in_sync
1107          */
1108         for (i = 0; i < conf->raid_disks; i++) {
1109                 tmp = conf->mirrors + i;
1110                 if (tmp->rdev
1111                     && !test_bit(Faulty, &tmp->rdev->flags)
1112                     && !test_and_set_bit(In_sync, &tmp->rdev->flags)) {
1113                         unsigned long flags;
1114                         spin_lock_irqsave(&conf->device_lock, flags);
1115                         mddev->degraded--;
1116                         spin_unlock_irqrestore(&conf->device_lock, flags);
1117                 }
1118         }
1119
1120         print_conf(conf);
1121         return 0;
1122 }
1123
1124
1125 static int raid10_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1126 {
1127         conf_t *conf = mddev->private;
1128         int err = -EEXIST;
1129         int mirror;
1130         mirror_info_t *p;
1131         int first = 0;
1132         int last = mddev->raid_disks - 1;
1133
1134         if (mddev->recovery_cp < MaxSector)
1135                 /* only hot-add to in-sync arrays, as recovery is
1136                  * very different from resync
1137                  */
1138                 return -EBUSY;
1139         if (!enough(conf))
1140                 return -EINVAL;
1141
1142         if (rdev->raid_disk >= 0)
1143                 first = last = rdev->raid_disk;
1144
1145         if (rdev->saved_raid_disk >= 0 &&
1146             rdev->saved_raid_disk >= first &&
1147             conf->mirrors[rdev->saved_raid_disk].rdev == NULL)
1148                 mirror = rdev->saved_raid_disk;
1149         else
1150                 mirror = first;
1151         for ( ; mirror <= last ; mirror++)
1152                 if ( !(p=conf->mirrors+mirror)->rdev) {
1153
1154                         disk_stack_limits(mddev->gendisk, rdev->bdev,
1155                                           rdev->data_offset << 9);
1156                         /* as we don't honour merge_bvec_fn, we must never risk
1157                          * violating it, so limit ->max_sector to one PAGE, as
1158                          * a one page request is never in violation.
1159                          */
1160                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1161                             queue_max_sectors(mddev->queue) > (PAGE_SIZE>>9))
1162                                 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
1163
1164                         p->head_position = 0;
1165                         rdev->raid_disk = mirror;
1166                         err = 0;
1167                         if (rdev->saved_raid_disk != mirror)
1168                                 conf->fullsync = 1;
1169                         rcu_assign_pointer(p->rdev, rdev);
1170                         break;
1171                 }
1172
1173         print_conf(conf);
1174         return err;
1175 }
1176
1177 static int raid10_remove_disk(mddev_t *mddev, int number)
1178 {
1179         conf_t *conf = mddev->private;
1180         int err = 0;
1181         mdk_rdev_t *rdev;
1182         mirror_info_t *p = conf->mirrors+ number;
1183
1184         print_conf(conf);
1185         rdev = p->rdev;
1186         if (rdev) {
1187                 if (test_bit(In_sync, &rdev->flags) ||
1188                     atomic_read(&rdev->nr_pending)) {
1189                         err = -EBUSY;
1190                         goto abort;
1191                 }
1192                 /* Only remove faulty devices in recovery
1193                  * is not possible.
1194                  */
1195                 if (!test_bit(Faulty, &rdev->flags) &&
1196                     enough(conf)) {
1197                         err = -EBUSY;
1198                         goto abort;
1199                 }
1200                 p->rdev = NULL;
1201                 synchronize_rcu();
1202                 if (atomic_read(&rdev->nr_pending)) {
1203                         /* lost the race, try later */
1204                         err = -EBUSY;
1205                         p->rdev = rdev;
1206                 }
1207         }
1208 abort:
1209
1210         print_conf(conf);
1211         return err;
1212 }
1213
1214
1215 static void end_sync_read(struct bio *bio, int error)
1216 {
1217         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1218         conf_t *conf = r10_bio->mddev->private;
1219         int i,d;
1220
1221         for (i=0; i<conf->copies; i++)
1222                 if (r10_bio->devs[i].bio == bio)
1223                         break;
1224         BUG_ON(i == conf->copies);
1225         update_head_pos(i, r10_bio);
1226         d = r10_bio->devs[i].devnum;
1227
1228         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1229                 set_bit(R10BIO_Uptodate, &r10_bio->state);
1230         else {
1231                 atomic_add(r10_bio->sectors,
1232                            &conf->mirrors[d].rdev->corrected_errors);
1233                 if (!test_bit(MD_RECOVERY_SYNC, &conf->mddev->recovery))
1234                         md_error(r10_bio->mddev,
1235                                  conf->mirrors[d].rdev);
1236         }
1237
1238         /* for reconstruct, we always reschedule after a read.
1239          * for resync, only after all reads
1240          */
1241         rdev_dec_pending(conf->mirrors[d].rdev, conf->mddev);
1242         if (test_bit(R10BIO_IsRecover, &r10_bio->state) ||
1243             atomic_dec_and_test(&r10_bio->remaining)) {
1244                 /* we have read all the blocks,
1245                  * do the comparison in process context in raid10d
1246                  */
1247                 reschedule_retry(r10_bio);
1248         }
1249 }
1250
1251 static void end_sync_write(struct bio *bio, int error)
1252 {
1253         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1254         r10bio_t * r10_bio = (r10bio_t *)(bio->bi_private);
1255         mddev_t *mddev = r10_bio->mddev;
1256         conf_t *conf = mddev->private;
1257         int i,d;
1258
1259         for (i = 0; i < conf->copies; i++)
1260                 if (r10_bio->devs[i].bio == bio)
1261                         break;
1262         d = r10_bio->devs[i].devnum;
1263
1264         if (!uptodate)
1265                 md_error(mddev, conf->mirrors[d].rdev);
1266
1267         update_head_pos(i, r10_bio);
1268
1269         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1270         while (atomic_dec_and_test(&r10_bio->remaining)) {
1271                 if (r10_bio->master_bio == NULL) {
1272                         /* the primary of several recovery bios */
1273                         sector_t s = r10_bio->sectors;
1274                         put_buf(r10_bio);
1275                         md_done_sync(mddev, s, 1);
1276                         break;
1277                 } else {
1278                         r10bio_t *r10_bio2 = (r10bio_t *)r10_bio->master_bio;
1279                         put_buf(r10_bio);
1280                         r10_bio = r10_bio2;
1281                 }
1282         }
1283 }
1284
1285 /*
1286  * Note: sync and recover and handled very differently for raid10
1287  * This code is for resync.
1288  * For resync, we read through virtual addresses and read all blocks.
1289  * If there is any error, we schedule a write.  The lowest numbered
1290  * drive is authoritative.
1291  * However requests come for physical address, so we need to map.
1292  * For every physical address there are raid_disks/copies virtual addresses,
1293  * which is always are least one, but is not necessarly an integer.
1294  * This means that a physical address can span multiple chunks, so we may
1295  * have to submit multiple io requests for a single sync request.
1296  */
1297 /*
1298  * We check if all blocks are in-sync and only write to blocks that
1299  * aren't in sync
1300  */
1301 static void sync_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1302 {
1303         conf_t *conf = mddev->private;
1304         int i, first;
1305         struct bio *tbio, *fbio;
1306
1307         atomic_set(&r10_bio->remaining, 1);
1308
1309         /* find the first device with a block */
1310         for (i=0; i<conf->copies; i++)
1311                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags))
1312                         break;
1313
1314         if (i == conf->copies)
1315                 goto done;
1316
1317         first = i;
1318         fbio = r10_bio->devs[i].bio;
1319
1320         /* now find blocks with errors */
1321         for (i=0 ; i < conf->copies ; i++) {
1322                 int  j, d;
1323                 int vcnt = r10_bio->sectors >> (PAGE_SHIFT-9);
1324
1325                 tbio = r10_bio->devs[i].bio;
1326
1327                 if (tbio->bi_end_io != end_sync_read)
1328                         continue;
1329                 if (i == first)
1330                         continue;
1331                 if (test_bit(BIO_UPTODATE, &r10_bio->devs[i].bio->bi_flags)) {
1332                         /* We know that the bi_io_vec layout is the same for
1333                          * both 'first' and 'i', so we just compare them.
1334                          * All vec entries are PAGE_SIZE;
1335                          */
1336                         for (j = 0; j < vcnt; j++)
1337                                 if (memcmp(page_address(fbio->bi_io_vec[j].bv_page),
1338                                            page_address(tbio->bi_io_vec[j].bv_page),
1339                                            PAGE_SIZE))
1340                                         break;
1341                         if (j == vcnt)
1342                                 continue;
1343                         mddev->resync_mismatches += r10_bio->sectors;
1344                 }
1345                 if (test_bit(MD_RECOVERY_CHECK, &mddev->recovery))
1346                         /* Don't fix anything. */
1347                         continue;
1348                 /* Ok, we need to write this bio
1349                  * First we need to fixup bv_offset, bv_len and
1350                  * bi_vecs, as the read request might have corrupted these
1351                  */
1352                 tbio->bi_vcnt = vcnt;
1353                 tbio->bi_size = r10_bio->sectors << 9;
1354                 tbio->bi_idx = 0;
1355                 tbio->bi_phys_segments = 0;
1356                 tbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1357                 tbio->bi_flags |= 1 << BIO_UPTODATE;
1358                 tbio->bi_next = NULL;
1359                 tbio->bi_rw = WRITE;
1360                 tbio->bi_private = r10_bio;
1361                 tbio->bi_sector = r10_bio->devs[i].addr;
1362
1363                 for (j=0; j < vcnt ; j++) {
1364                         tbio->bi_io_vec[j].bv_offset = 0;
1365                         tbio->bi_io_vec[j].bv_len = PAGE_SIZE;
1366
1367                         memcpy(page_address(tbio->bi_io_vec[j].bv_page),
1368                                page_address(fbio->bi_io_vec[j].bv_page),
1369                                PAGE_SIZE);
1370                 }
1371                 tbio->bi_end_io = end_sync_write;
1372
1373                 d = r10_bio->devs[i].devnum;
1374                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1375                 atomic_inc(&r10_bio->remaining);
1376                 md_sync_acct(conf->mirrors[d].rdev->bdev, tbio->bi_size >> 9);
1377
1378                 tbio->bi_sector += conf->mirrors[d].rdev->data_offset;
1379                 tbio->bi_bdev = conf->mirrors[d].rdev->bdev;
1380                 generic_make_request(tbio);
1381         }
1382
1383 done:
1384         if (atomic_dec_and_test(&r10_bio->remaining)) {
1385                 md_done_sync(mddev, r10_bio->sectors, 1);
1386                 put_buf(r10_bio);
1387         }
1388 }
1389
1390 /*
1391  * Now for the recovery code.
1392  * Recovery happens across physical sectors.
1393  * We recover all non-is_sync drives by finding the virtual address of
1394  * each, and then choose a working drive that also has that virt address.
1395  * There is a separate r10_bio for each non-in_sync drive.
1396  * Only the first two slots are in use. The first for reading,
1397  * The second for writing.
1398  *
1399  */
1400
1401 static void recovery_request_write(mddev_t *mddev, r10bio_t *r10_bio)
1402 {
1403         conf_t *conf = mddev->private;
1404         int i, d;
1405         struct bio *bio, *wbio;
1406
1407
1408         /* move the pages across to the second bio
1409          * and submit the write request
1410          */
1411         bio = r10_bio->devs[0].bio;
1412         wbio = r10_bio->devs[1].bio;
1413         for (i=0; i < wbio->bi_vcnt; i++) {
1414                 struct page *p = bio->bi_io_vec[i].bv_page;
1415                 bio->bi_io_vec[i].bv_page = wbio->bi_io_vec[i].bv_page;
1416                 wbio->bi_io_vec[i].bv_page = p;
1417         }
1418         d = r10_bio->devs[1].devnum;
1419
1420         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1421         md_sync_acct(conf->mirrors[d].rdev->bdev, wbio->bi_size >> 9);
1422         if (test_bit(R10BIO_Uptodate, &r10_bio->state))
1423                 generic_make_request(wbio);
1424         else
1425                 bio_endio(wbio, -EIO);
1426 }
1427
1428
1429 /*
1430  * This is a kernel thread which:
1431  *
1432  *      1.      Retries failed read operations on working mirrors.
1433  *      2.      Updates the raid superblock when problems encounter.
1434  *      3.      Performs writes following reads for array synchronising.
1435  */
1436
1437 static void fix_read_error(conf_t *conf, mddev_t *mddev, r10bio_t *r10_bio)
1438 {
1439         int sect = 0; /* Offset from r10_bio->sector */
1440         int sectors = r10_bio->sectors;
1441         mdk_rdev_t*rdev;
1442         while(sectors) {
1443                 int s = sectors;
1444                 int sl = r10_bio->read_slot;
1445                 int success = 0;
1446                 int start;
1447
1448                 if (s > (PAGE_SIZE>>9))
1449                         s = PAGE_SIZE >> 9;
1450
1451                 rcu_read_lock();
1452                 do {
1453                         int d = r10_bio->devs[sl].devnum;
1454                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1455                         if (rdev &&
1456                             test_bit(In_sync, &rdev->flags)) {
1457                                 atomic_inc(&rdev->nr_pending);
1458                                 rcu_read_unlock();
1459                                 success = sync_page_io(rdev->bdev,
1460                                                        r10_bio->devs[sl].addr +
1461                                                        sect + rdev->data_offset,
1462                                                        s<<9,
1463                                                        conf->tmppage, READ);
1464                                 rdev_dec_pending(rdev, mddev);
1465                                 rcu_read_lock();
1466                                 if (success)
1467                                         break;
1468                         }
1469                         sl++;
1470                         if (sl == conf->copies)
1471                                 sl = 0;
1472                 } while (!success && sl != r10_bio->read_slot);
1473                 rcu_read_unlock();
1474
1475                 if (!success) {
1476                         /* Cannot read from anywhere -- bye bye array */
1477                         int dn = r10_bio->devs[r10_bio->read_slot].devnum;
1478                         md_error(mddev, conf->mirrors[dn].rdev);
1479                         break;
1480                 }
1481
1482                 start = sl;
1483                 /* write it back and re-read */
1484                 rcu_read_lock();
1485                 while (sl != r10_bio->read_slot) {
1486                         int d;
1487                         if (sl==0)
1488                                 sl = conf->copies;
1489                         sl--;
1490                         d = r10_bio->devs[sl].devnum;
1491                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1492                         if (rdev &&
1493                             test_bit(In_sync, &rdev->flags)) {
1494                                 atomic_inc(&rdev->nr_pending);
1495                                 rcu_read_unlock();
1496                                 atomic_add(s, &rdev->corrected_errors);
1497                                 if (sync_page_io(rdev->bdev,
1498                                                  r10_bio->devs[sl].addr +
1499                                                  sect + rdev->data_offset,
1500                                                  s<<9, conf->tmppage, WRITE)
1501                                     == 0)
1502                                         /* Well, this device is dead */
1503                                         md_error(mddev, rdev);
1504                                 rdev_dec_pending(rdev, mddev);
1505                                 rcu_read_lock();
1506                         }
1507                 }
1508                 sl = start;
1509                 while (sl != r10_bio->read_slot) {
1510                         int d;
1511                         if (sl==0)
1512                                 sl = conf->copies;
1513                         sl--;
1514                         d = r10_bio->devs[sl].devnum;
1515                         rdev = rcu_dereference(conf->mirrors[d].rdev);
1516                         if (rdev &&
1517                             test_bit(In_sync, &rdev->flags)) {
1518                                 char b[BDEVNAME_SIZE];
1519                                 atomic_inc(&rdev->nr_pending);
1520                                 rcu_read_unlock();
1521                                 if (sync_page_io(rdev->bdev,
1522                                                  r10_bio->devs[sl].addr +
1523                                                  sect + rdev->data_offset,
1524                                                  s<<9, conf->tmppage, READ) == 0)
1525                                         /* Well, this device is dead */
1526                                         md_error(mddev, rdev);
1527                                 else
1528                                         printk(KERN_INFO
1529                                                "raid10:%s: read error corrected"
1530                                                " (%d sectors at %llu on %s)\n",
1531                                                mdname(mddev), s,
1532                                                (unsigned long long)(sect+
1533                                                     rdev->data_offset),
1534                                                bdevname(rdev->bdev, b));
1535
1536                                 rdev_dec_pending(rdev, mddev);
1537                                 rcu_read_lock();
1538                         }
1539                 }
1540                 rcu_read_unlock();
1541
1542                 sectors -= s;
1543                 sect += s;
1544         }
1545 }
1546
1547 static void raid10d(mddev_t *mddev)
1548 {
1549         r10bio_t *r10_bio;
1550         struct bio *bio;
1551         unsigned long flags;
1552         conf_t *conf = mddev->private;
1553         struct list_head *head = &conf->retry_list;
1554         int unplug=0;
1555         mdk_rdev_t *rdev;
1556
1557         md_check_recovery(mddev);
1558
1559         for (;;) {
1560                 char b[BDEVNAME_SIZE];
1561
1562                 unplug += flush_pending_writes(conf);
1563
1564                 spin_lock_irqsave(&conf->device_lock, flags);
1565                 if (list_empty(head)) {
1566                         spin_unlock_irqrestore(&conf->device_lock, flags);
1567                         break;
1568                 }
1569                 r10_bio = list_entry(head->prev, r10bio_t, retry_list);
1570                 list_del(head->prev);
1571                 conf->nr_queued--;
1572                 spin_unlock_irqrestore(&conf->device_lock, flags);
1573
1574                 mddev = r10_bio->mddev;
1575                 conf = mddev->private;
1576                 if (test_bit(R10BIO_IsSync, &r10_bio->state)) {
1577                         sync_request_write(mddev, r10_bio);
1578                         unplug = 1;
1579                 } else  if (test_bit(R10BIO_IsRecover, &r10_bio->state)) {
1580                         recovery_request_write(mddev, r10_bio);
1581                         unplug = 1;
1582                 } else {
1583                         int mirror;
1584                         /* we got a read error. Maybe the drive is bad.  Maybe just
1585                          * the block and we can fix it.
1586                          * We freeze all other IO, and try reading the block from
1587                          * other devices.  When we find one, we re-write
1588                          * and check it that fixes the read error.
1589                          * This is all done synchronously while the array is
1590                          * frozen.
1591                          */
1592                         if (mddev->ro == 0) {
1593                                 freeze_array(conf);
1594                                 fix_read_error(conf, mddev, r10_bio);
1595                                 unfreeze_array(conf);
1596                         }
1597
1598                         bio = r10_bio->devs[r10_bio->read_slot].bio;
1599                         r10_bio->devs[r10_bio->read_slot].bio =
1600                                 mddev->ro ? IO_BLOCKED : NULL;
1601                         mirror = read_balance(conf, r10_bio);
1602                         if (mirror == -1) {
1603                                 printk(KERN_ALERT "raid10: %s: unrecoverable I/O"
1604                                        " read error for block %llu\n",
1605                                        bdevname(bio->bi_bdev,b),
1606                                        (unsigned long long)r10_bio->sector);
1607                                 raid_end_bio_io(r10_bio);
1608                                 bio_put(bio);
1609                         } else {
1610                                 const int do_sync = bio_sync(r10_bio->master_bio);
1611                                 bio_put(bio);
1612                                 rdev = conf->mirrors[mirror].rdev;
1613                                 if (printk_ratelimit())
1614                                         printk(KERN_ERR "raid10: %s: redirecting sector %llu to"
1615                                                " another mirror\n",
1616                                                bdevname(rdev->bdev,b),
1617                                                (unsigned long long)r10_bio->sector);
1618                                 bio = bio_clone(r10_bio->master_bio, GFP_NOIO);
1619                                 r10_bio->devs[r10_bio->read_slot].bio = bio;
1620                                 bio->bi_sector = r10_bio->devs[r10_bio->read_slot].addr
1621                                         + rdev->data_offset;
1622                                 bio->bi_bdev = rdev->bdev;
1623                                 bio->bi_rw = READ | do_sync;
1624                                 bio->bi_private = r10_bio;
1625                                 bio->bi_end_io = raid10_end_read_request;
1626                                 unplug = 1;
1627                                 generic_make_request(bio);
1628                         }
1629                 }
1630         }
1631         if (unplug)
1632                 unplug_slaves(mddev);
1633 }
1634
1635
1636 static int init_resync(conf_t *conf)
1637 {
1638         int buffs;
1639
1640         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1641         BUG_ON(conf->r10buf_pool);
1642         conf->r10buf_pool = mempool_create(buffs, r10buf_pool_alloc, r10buf_pool_free, conf);
1643         if (!conf->r10buf_pool)
1644                 return -ENOMEM;
1645         conf->next_resync = 0;
1646         return 0;
1647 }
1648
1649 /*
1650  * perform a "sync" on one "block"
1651  *
1652  * We need to make sure that no normal I/O request - particularly write
1653  * requests - conflict with active sync requests.
1654  *
1655  * This is achieved by tracking pending requests and a 'barrier' concept
1656  * that can be installed to exclude normal IO requests.
1657  *
1658  * Resync and recovery are handled very differently.
1659  * We differentiate by looking at MD_RECOVERY_SYNC in mddev->recovery.
1660  *
1661  * For resync, we iterate over virtual addresses, read all copies,
1662  * and update if there are differences.  If only one copy is live,
1663  * skip it.
1664  * For recovery, we iterate over physical addresses, read a good
1665  * value for each non-in_sync drive, and over-write.
1666  *
1667  * So, for recovery we may have several outstanding complex requests for a
1668  * given address, one for each out-of-sync device.  We model this by allocating
1669  * a number of r10_bio structures, one for each out-of-sync device.
1670  * As we setup these structures, we collect all bio's together into a list
1671  * which we then process collectively to add pages, and then process again
1672  * to pass to generic_make_request.
1673  *
1674  * The r10_bio structures are linked using a borrowed master_bio pointer.
1675  * This link is counted in ->remaining.  When the r10_bio that points to NULL
1676  * has its remaining count decremented to 0, the whole complex operation
1677  * is complete.
1678  *
1679  */
1680
1681 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1682 {
1683         conf_t *conf = mddev->private;
1684         r10bio_t *r10_bio;
1685         struct bio *biolist = NULL, *bio;
1686         sector_t max_sector, nr_sectors;
1687         int disk;
1688         int i;
1689         int max_sync;
1690         int sync_blocks;
1691
1692         sector_t sectors_skipped = 0;
1693         int chunks_skipped = 0;
1694
1695         if (!conf->r10buf_pool)
1696                 if (init_resync(conf))
1697                         return 0;
1698
1699  skipped:
1700         max_sector = mddev->dev_sectors;
1701         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1702                 max_sector = mddev->resync_max_sectors;
1703         if (sector_nr >= max_sector) {
1704                 /* If we aborted, we need to abort the
1705                  * sync on the 'current' bitmap chucks (there can
1706                  * be several when recovering multiple devices).
1707                  * as we may have started syncing it but not finished.
1708                  * We can find the current address in
1709                  * mddev->curr_resync, but for recovery,
1710                  * we need to convert that to several
1711                  * virtual addresses.
1712                  */
1713                 if (mddev->curr_resync < max_sector) { /* aborted */
1714                         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery))
1715                                 bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1716                                                 &sync_blocks, 1);
1717                         else for (i=0; i<conf->raid_disks; i++) {
1718                                 sector_t sect =
1719                                         raid10_find_virt(conf, mddev->curr_resync, i);
1720                                 bitmap_end_sync(mddev->bitmap, sect,
1721                                                 &sync_blocks, 1);
1722                         }
1723                 } else /* completed sync */
1724                         conf->fullsync = 0;
1725
1726                 bitmap_close_sync(mddev->bitmap);
1727                 close_sync(conf);
1728                 *skipped = 1;
1729                 return sectors_skipped;
1730         }
1731         if (chunks_skipped >= conf->raid_disks) {
1732                 /* if there has been nothing to do on any drive,
1733                  * then there is nothing to do at all..
1734                  */
1735                 *skipped = 1;
1736                 return (max_sector - sector_nr) + sectors_skipped;
1737         }
1738
1739         if (max_sector > mddev->resync_max)
1740                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1741
1742         /* make sure whole request will fit in a chunk - if chunks
1743          * are meaningful
1744          */
1745         if (conf->near_copies < conf->raid_disks &&
1746             max_sector > (sector_nr | conf->chunk_mask))
1747                 max_sector = (sector_nr | conf->chunk_mask) + 1;
1748         /*
1749          * If there is non-resync activity waiting for us then
1750          * put in a delay to throttle resync.
1751          */
1752         if (!go_faster && conf->nr_waiting)
1753                 msleep_interruptible(1000);
1754
1755         /* Again, very different code for resync and recovery.
1756          * Both must result in an r10bio with a list of bios that
1757          * have bi_end_io, bi_sector, bi_bdev set,
1758          * and bi_private set to the r10bio.
1759          * For recovery, we may actually create several r10bios
1760          * with 2 bios in each, that correspond to the bios in the main one.
1761          * In this case, the subordinate r10bios link back through a
1762          * borrowed master_bio pointer, and the counter in the master
1763          * includes a ref from each subordinate.
1764          */
1765         /* First, we decide what to do and set ->bi_end_io
1766          * To end_sync_read if we want to read, and
1767          * end_sync_write if we will want to write.
1768          */
1769
1770         max_sync = RESYNC_PAGES << (PAGE_SHIFT-9);
1771         if (!test_bit(MD_RECOVERY_SYNC, &mddev->recovery)) {
1772                 /* recovery... the complicated one */
1773                 int i, j, k;
1774                 r10_bio = NULL;
1775
1776                 for (i=0 ; i<conf->raid_disks; i++)
1777                         if (conf->mirrors[i].rdev &&
1778                             !test_bit(In_sync, &conf->mirrors[i].rdev->flags)) {
1779                                 int still_degraded = 0;
1780                                 /* want to reconstruct this device */
1781                                 r10bio_t *rb2 = r10_bio;
1782                                 sector_t sect = raid10_find_virt(conf, sector_nr, i);
1783                                 int must_sync;
1784                                 /* Unless we are doing a full sync, we only need
1785                                  * to recover the block if it is set in the bitmap
1786                                  */
1787                                 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1788                                                               &sync_blocks, 1);
1789                                 if (sync_blocks < max_sync)
1790                                         max_sync = sync_blocks;
1791                                 if (!must_sync &&
1792                                     !conf->fullsync) {
1793                                         /* yep, skip the sync_blocks here, but don't assume
1794                                          * that there will never be anything to do here
1795                                          */
1796                                         chunks_skipped = -1;
1797                                         continue;
1798                                 }
1799
1800                                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1801                                 raise_barrier(conf, rb2 != NULL);
1802                                 atomic_set(&r10_bio->remaining, 0);
1803
1804                                 r10_bio->master_bio = (struct bio*)rb2;
1805                                 if (rb2)
1806                                         atomic_inc(&rb2->remaining);
1807                                 r10_bio->mddev = mddev;
1808                                 set_bit(R10BIO_IsRecover, &r10_bio->state);
1809                                 r10_bio->sector = sect;
1810
1811                                 raid10_find_phys(conf, r10_bio);
1812
1813                                 /* Need to check if the array will still be
1814                                  * degraded
1815                                  */
1816                                 for (j=0; j<conf->raid_disks; j++)
1817                                         if (conf->mirrors[j].rdev == NULL ||
1818                                             test_bit(Faulty, &conf->mirrors[j].rdev->flags)) {
1819                                                 still_degraded = 1;
1820                                                 break;
1821                                         }
1822
1823                                 must_sync = bitmap_start_sync(mddev->bitmap, sect,
1824                                                               &sync_blocks, still_degraded);
1825
1826                                 for (j=0; j<conf->copies;j++) {
1827                                         int d = r10_bio->devs[j].devnum;
1828                                         if (conf->mirrors[d].rdev &&
1829                                             test_bit(In_sync, &conf->mirrors[d].rdev->flags)) {
1830                                                 /* This is where we read from */
1831                                                 bio = r10_bio->devs[0].bio;
1832                                                 bio->bi_next = biolist;
1833                                                 biolist = bio;
1834                                                 bio->bi_private = r10_bio;
1835                                                 bio->bi_end_io = end_sync_read;
1836                                                 bio->bi_rw = READ;
1837                                                 bio->bi_sector = r10_bio->devs[j].addr +
1838                                                         conf->mirrors[d].rdev->data_offset;
1839                                                 bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1840                                                 atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1841                                                 atomic_inc(&r10_bio->remaining);
1842                                                 /* and we write to 'i' */
1843
1844                                                 for (k=0; k<conf->copies; k++)
1845                                                         if (r10_bio->devs[k].devnum == i)
1846                                                                 break;
1847                                                 BUG_ON(k == conf->copies);
1848                                                 bio = r10_bio->devs[1].bio;
1849                                                 bio->bi_next = biolist;
1850                                                 biolist = bio;
1851                                                 bio->bi_private = r10_bio;
1852                                                 bio->bi_end_io = end_sync_write;
1853                                                 bio->bi_rw = WRITE;
1854                                                 bio->bi_sector = r10_bio->devs[k].addr +
1855                                                         conf->mirrors[i].rdev->data_offset;
1856                                                 bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1857
1858                                                 r10_bio->devs[0].devnum = d;
1859                                                 r10_bio->devs[1].devnum = i;
1860
1861                                                 break;
1862                                         }
1863                                 }
1864                                 if (j == conf->copies) {
1865                                         /* Cannot recover, so abort the recovery */
1866                                         put_buf(r10_bio);
1867                                         if (rb2)
1868                                                 atomic_dec(&rb2->remaining);
1869                                         r10_bio = rb2;
1870                                         if (!test_and_set_bit(MD_RECOVERY_INTR,
1871                                                               &mddev->recovery))
1872                                                 printk(KERN_INFO "raid10: %s: insufficient working devices for recovery.\n",
1873                                                        mdname(mddev));
1874                                         break;
1875                                 }
1876                         }
1877                 if (biolist == NULL) {
1878                         while (r10_bio) {
1879                                 r10bio_t *rb2 = r10_bio;
1880                                 r10_bio = (r10bio_t*) rb2->master_bio;
1881                                 rb2->master_bio = NULL;
1882                                 put_buf(rb2);
1883                         }
1884                         goto giveup;
1885                 }
1886         } else {
1887                 /* resync. Schedule a read for every block at this virt offset */
1888                 int count = 0;
1889
1890                 bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1891
1892                 if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1893                                        &sync_blocks, mddev->degraded) &&
1894                     !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1895                         /* We can skip this block */
1896                         *skipped = 1;
1897                         return sync_blocks + sectors_skipped;
1898                 }
1899                 if (sync_blocks < max_sync)
1900                         max_sync = sync_blocks;
1901                 r10_bio = mempool_alloc(conf->r10buf_pool, GFP_NOIO);
1902
1903                 r10_bio->mddev = mddev;
1904                 atomic_set(&r10_bio->remaining, 0);
1905                 raise_barrier(conf, 0);
1906                 conf->next_resync = sector_nr;
1907
1908                 r10_bio->master_bio = NULL;
1909                 r10_bio->sector = sector_nr;
1910                 set_bit(R10BIO_IsSync, &r10_bio->state);
1911                 raid10_find_phys(conf, r10_bio);
1912                 r10_bio->sectors = (sector_nr | conf->chunk_mask) - sector_nr +1;
1913
1914                 for (i=0; i<conf->copies; i++) {
1915                         int d = r10_bio->devs[i].devnum;
1916                         bio = r10_bio->devs[i].bio;
1917                         bio->bi_end_io = NULL;
1918                         clear_bit(BIO_UPTODATE, &bio->bi_flags);
1919                         if (conf->mirrors[d].rdev == NULL ||
1920                             test_bit(Faulty, &conf->mirrors[d].rdev->flags))
1921                                 continue;
1922                         atomic_inc(&conf->mirrors[d].rdev->nr_pending);
1923                         atomic_inc(&r10_bio->remaining);
1924                         bio->bi_next = biolist;
1925                         biolist = bio;
1926                         bio->bi_private = r10_bio;
1927                         bio->bi_end_io = end_sync_read;
1928                         bio->bi_rw = READ;
1929                         bio->bi_sector = r10_bio->devs[i].addr +
1930                                 conf->mirrors[d].rdev->data_offset;
1931                         bio->bi_bdev = conf->mirrors[d].rdev->bdev;
1932                         count++;
1933                 }
1934
1935                 if (count < 2) {
1936                         for (i=0; i<conf->copies; i++) {
1937                                 int d = r10_bio->devs[i].devnum;
1938                                 if (r10_bio->devs[i].bio->bi_end_io)
1939                                         rdev_dec_pending(conf->mirrors[d].rdev, mddev);
1940                         }
1941                         put_buf(r10_bio);
1942                         biolist = NULL;
1943                         goto giveup;
1944                 }
1945         }
1946
1947         for (bio = biolist; bio ; bio=bio->bi_next) {
1948
1949                 bio->bi_flags &= ~(BIO_POOL_MASK - 1);
1950                 if (bio->bi_end_io)
1951                         bio->bi_flags |= 1 << BIO_UPTODATE;
1952                 bio->bi_vcnt = 0;
1953                 bio->bi_idx = 0;
1954                 bio->bi_phys_segments = 0;
1955                 bio->bi_size = 0;
1956         }
1957
1958         nr_sectors = 0;
1959         if (sector_nr + max_sync < max_sector)
1960                 max_sector = sector_nr + max_sync;
1961         do {
1962                 struct page *page;
1963                 int len = PAGE_SIZE;
1964                 disk = 0;
1965                 if (sector_nr + (len>>9) > max_sector)
1966                         len = (max_sector - sector_nr) << 9;
1967                 if (len == 0)
1968                         break;
1969                 for (bio= biolist ; bio ; bio=bio->bi_next) {
1970                         page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1971                         if (bio_add_page(bio, page, len, 0) == 0) {
1972                                 /* stop here */
1973                                 struct bio *bio2;
1974                                 bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1975                                 for (bio2 = biolist; bio2 && bio2 != bio; bio2 = bio2->bi_next) {
1976                                         /* remove last page from this bio */
1977                                         bio2->bi_vcnt--;
1978                                         bio2->bi_size -= len;
1979                                         bio2->bi_flags &= ~(1<< BIO_SEG_VALID);
1980                                 }
1981                                 goto bio_full;
1982                         }
1983                         disk = i;
1984                 }
1985                 nr_sectors += len>>9;
1986                 sector_nr += len>>9;
1987         } while (biolist->bi_vcnt < RESYNC_PAGES);
1988  bio_full:
1989         r10_bio->sectors = nr_sectors;
1990
1991         while (biolist) {
1992                 bio = biolist;
1993                 biolist = biolist->bi_next;
1994
1995                 bio->bi_next = NULL;
1996                 r10_bio = bio->bi_private;
1997                 r10_bio->sectors = nr_sectors;
1998
1999                 if (bio->bi_end_io == end_sync_read) {
2000                         md_sync_acct(bio->bi_bdev, nr_sectors);
2001                         generic_make_request(bio);
2002                 }
2003         }
2004
2005         if (sectors_skipped)
2006                 /* pretend they weren't skipped, it makes
2007                  * no important difference in this case
2008                  */
2009                 md_done_sync(mddev, sectors_skipped, 1);
2010
2011         return sectors_skipped + nr_sectors;
2012  giveup:
2013         /* There is nowhere to write, so all non-sync
2014          * drives must be failed, so try the next chunk...
2015          */
2016         if (sector_nr + max_sync < max_sector)
2017                 max_sector = sector_nr + max_sync;
2018
2019         sectors_skipped += (max_sector - sector_nr);
2020         chunks_skipped ++;
2021         sector_nr = max_sector;
2022         goto skipped;
2023 }
2024
2025 static sector_t
2026 raid10_size(mddev_t *mddev, sector_t sectors, int raid_disks)
2027 {
2028         sector_t size;
2029         conf_t *conf = mddev->private;
2030
2031         if (!raid_disks)
2032                 raid_disks = mddev->raid_disks;
2033         if (!sectors)
2034                 sectors = mddev->dev_sectors;
2035
2036         size = sectors >> conf->chunk_shift;
2037         sector_div(size, conf->far_copies);
2038         size = size * raid_disks;
2039         sector_div(size, conf->near_copies);
2040
2041         return size << conf->chunk_shift;
2042 }
2043
2044 static int run(mddev_t *mddev)
2045 {
2046         conf_t *conf;
2047         int i, disk_idx, chunk_size;
2048         mirror_info_t *disk;
2049         mdk_rdev_t *rdev;
2050         int nc, fc, fo;
2051         sector_t stride, size;
2052
2053         if (mddev->chunk_sectors < (PAGE_SIZE >> 9) ||
2054             !is_power_of_2(mddev->chunk_sectors)) {
2055                 printk(KERN_ERR "md/raid10: chunk size must be "
2056                        "at least PAGE_SIZE(%ld) and be a power of 2.\n", PAGE_SIZE);
2057                 return -EINVAL;
2058         }
2059
2060         nc = mddev->layout & 255;
2061         fc = (mddev->layout >> 8) & 255;
2062         fo = mddev->layout & (1<<16);
2063         if ((nc*fc) <2 || (nc*fc) > mddev->raid_disks ||
2064             (mddev->layout >> 17)) {
2065                 printk(KERN_ERR "raid10: %s: unsupported raid10 layout: 0x%8x\n",
2066                        mdname(mddev), mddev->layout);
2067                 goto out;
2068         }
2069         /*
2070          * copy the already verified devices into our private RAID10
2071          * bookkeeping area. [whatever we allocate in run(),
2072          * should be freed in stop()]
2073          */
2074         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
2075         mddev->private = conf;
2076         if (!conf) {
2077                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2078                         mdname(mddev));
2079                 goto out;
2080         }
2081         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
2082                                  GFP_KERNEL);
2083         if (!conf->mirrors) {
2084                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2085                        mdname(mddev));
2086                 goto out_free_conf;
2087         }
2088
2089         conf->tmppage = alloc_page(GFP_KERNEL);
2090         if (!conf->tmppage)
2091                 goto out_free_conf;
2092
2093         conf->mddev = mddev;
2094         conf->raid_disks = mddev->raid_disks;
2095         conf->near_copies = nc;
2096         conf->far_copies = fc;
2097         conf->copies = nc*fc;
2098         conf->far_offset = fo;
2099         conf->chunk_mask = mddev->chunk_sectors - 1;
2100         conf->chunk_shift = ffz(~mddev->chunk_sectors);
2101         size = mddev->dev_sectors >> conf->chunk_shift;
2102         sector_div(size, fc);
2103         size = size * conf->raid_disks;
2104         sector_div(size, nc);
2105         /* 'size' is now the number of chunks in the array */
2106         /* calculate "used chunks per device" in 'stride' */
2107         stride = size * conf->copies;
2108
2109         /* We need to round up when dividing by raid_disks to
2110          * get the stride size.
2111          */
2112         stride += conf->raid_disks - 1;
2113         sector_div(stride, conf->raid_disks);
2114         mddev->dev_sectors = stride << conf->chunk_shift;
2115
2116         if (fo)
2117                 stride = 1;
2118         else
2119                 sector_div(stride, fc);
2120         conf->stride = stride << conf->chunk_shift;
2121
2122         conf->r10bio_pool = mempool_create(NR_RAID10_BIOS, r10bio_pool_alloc,
2123                                                 r10bio_pool_free, conf);
2124         if (!conf->r10bio_pool) {
2125                 printk(KERN_ERR "raid10: couldn't allocate memory for %s\n",
2126                         mdname(mddev));
2127                 goto out_free_conf;
2128         }
2129
2130         spin_lock_init(&conf->device_lock);
2131         mddev->queue->queue_lock = &conf->device_lock;
2132
2133         chunk_size = mddev->chunk_sectors << 9;
2134         blk_queue_io_min(mddev->queue, chunk_size);
2135         if (conf->raid_disks % conf->near_copies)
2136                 blk_queue_io_opt(mddev->queue, chunk_size * conf->raid_disks);
2137         else
2138                 blk_queue_io_opt(mddev->queue, chunk_size *
2139                                  (conf->raid_disks / conf->near_copies));
2140
2141         list_for_each_entry(rdev, &mddev->disks, same_set) {
2142                 disk_idx = rdev->raid_disk;
2143                 if (disk_idx >= mddev->raid_disks
2144                     || disk_idx < 0)
2145                         continue;
2146                 disk = conf->mirrors + disk_idx;
2147
2148                 disk->rdev = rdev;
2149                 disk_stack_limits(mddev->gendisk, rdev->bdev,
2150                                   rdev->data_offset << 9);
2151                 /* as we don't honour merge_bvec_fn, we must never risk
2152                  * violating it, so limit ->max_sector to one PAGE, as
2153                  * a one page request is never in violation.
2154                  */
2155                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
2156                     queue_max_sectors(mddev->queue) > (PAGE_SIZE>>9))
2157                         blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
2158
2159                 disk->head_position = 0;
2160         }
2161         INIT_LIST_HEAD(&conf->retry_list);
2162
2163         spin_lock_init(&conf->resync_lock);
2164         init_waitqueue_head(&conf->wait_barrier);
2165
2166         /* need to check that every block has at least one working mirror */
2167         if (!enough(conf)) {
2168                 printk(KERN_ERR "raid10: not enough operational mirrors for %s\n",
2169                        mdname(mddev));
2170                 goto out_free_conf;
2171         }
2172
2173         mddev->degraded = 0;
2174         for (i = 0; i < conf->raid_disks; i++) {
2175
2176                 disk = conf->mirrors + i;
2177
2178                 if (!disk->rdev ||
2179                     !test_bit(In_sync, &disk->rdev->flags)) {
2180                         disk->head_position = 0;
2181                         mddev->degraded++;
2182                         if (disk->rdev)
2183                                 conf->fullsync = 1;
2184                 }
2185         }
2186
2187
2188         mddev->thread = md_register_thread(raid10d, mddev, "%s_raid10");
2189         if (!mddev->thread) {
2190                 printk(KERN_ERR
2191                        "raid10: couldn't allocate thread for %s\n",
2192                        mdname(mddev));
2193                 goto out_free_conf;
2194         }
2195
2196         if (mddev->recovery_cp != MaxSector)
2197                 printk(KERN_NOTICE "raid10: %s is not clean"
2198                        " -- starting background reconstruction\n",
2199                        mdname(mddev));
2200         printk(KERN_INFO
2201                 "raid10: raid set %s active with %d out of %d devices\n",
2202                 mdname(mddev), mddev->raid_disks - mddev->degraded,
2203                 mddev->raid_disks);
2204         /*
2205          * Ok, everything is just fine now
2206          */
2207         md_set_array_sectors(mddev, raid10_size(mddev, 0, 0));
2208         mddev->resync_max_sectors = raid10_size(mddev, 0, 0);
2209
2210         mddev->queue->unplug_fn = raid10_unplug;
2211         mddev->queue->backing_dev_info.congested_fn = raid10_congested;
2212         mddev->queue->backing_dev_info.congested_data = mddev;
2213
2214         /* Calculate max read-ahead size.
2215          * We need to readahead at least twice a whole stripe....
2216          * maybe...
2217          */
2218         {
2219                 int stripe = conf->raid_disks *
2220                         ((mddev->chunk_sectors << 9) / PAGE_SIZE);
2221                 stripe /= conf->near_copies;
2222                 if (mddev->queue->backing_dev_info.ra_pages < 2* stripe)
2223                         mddev->queue->backing_dev_info.ra_pages = 2* stripe;
2224         }
2225
2226         if (conf->near_copies < mddev->raid_disks)
2227                 blk_queue_merge_bvec(mddev->queue, raid10_mergeable_bvec);
2228         return 0;
2229
2230 out_free_conf:
2231         if (conf->r10bio_pool)
2232                 mempool_destroy(conf->r10bio_pool);
2233         safe_put_page(conf->tmppage);
2234         kfree(conf->mirrors);
2235         kfree(conf);
2236         mddev->private = NULL;
2237 out:
2238         return -EIO;
2239 }
2240
2241 static int stop(mddev_t *mddev)
2242 {
2243         conf_t *conf = mddev->private;
2244
2245         raise_barrier(conf, 0);
2246         lower_barrier(conf);
2247
2248         md_unregister_thread(mddev->thread);
2249         mddev->thread = NULL;
2250         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2251         if (conf->r10bio_pool)
2252                 mempool_destroy(conf->r10bio_pool);
2253         kfree(conf->mirrors);
2254         kfree(conf);
2255         mddev->private = NULL;
2256         return 0;
2257 }
2258
2259 static void raid10_quiesce(mddev_t *mddev, int state)
2260 {
2261         conf_t *conf = mddev->private;
2262
2263         switch(state) {
2264         case 1:
2265                 raise_barrier(conf, 0);
2266                 break;
2267         case 0:
2268                 lower_barrier(conf);
2269                 break;
2270         }
2271         if (mddev->thread) {
2272                 if (mddev->bitmap)
2273                         mddev->thread->timeout = mddev->bitmap->daemon_sleep * HZ;
2274                 else
2275                         mddev->thread->timeout = MAX_SCHEDULE_TIMEOUT;
2276                 md_wakeup_thread(mddev->thread);
2277         }
2278 }
2279
2280 static struct mdk_personality raid10_personality =
2281 {
2282         .name           = "raid10",
2283         .level          = 10,
2284         .owner          = THIS_MODULE,
2285         .make_request   = make_request,
2286         .run            = run,
2287         .stop           = stop,
2288         .status         = status,
2289         .error_handler  = error,
2290         .hot_add_disk   = raid10_add_disk,
2291         .hot_remove_disk= raid10_remove_disk,
2292         .spare_active   = raid10_spare_active,
2293         .sync_request   = sync_request,
2294         .quiesce        = raid10_quiesce,
2295         .size           = raid10_size,
2296 };
2297
2298 static int __init raid_init(void)
2299 {
2300         return register_md_personality(&raid10_personality);
2301 }
2302
2303 static void raid_exit(void)
2304 {
2305         unregister_md_personality(&raid10_personality);
2306 }
2307
2308 module_init(raid_init);
2309 module_exit(raid_exit);
2310 MODULE_LICENSE("GPL");
2311 MODULE_ALIAS("md-personality-9"); /* RAID10 */
2312 MODULE_ALIAS("md-raid10");
2313 MODULE_ALIAS("md-level-10");