drivers/md/raid1.c

   1 /*
   2  * raid1.c : Multiple Devices driver for Linux
   3  *
   4  * Copyright (C) 1999, 2000, 2001 Ingo Molnar, Red Hat
   5  *
   6  * Copyright (C) 1996, 1997, 1998 Ingo Molnar, Miguel de Icaza, Gadi Oxman
   7  *
   8  * RAID-1 management functions.
   9  *
  10  * Better read-balancing code written by Mika Kuoppala <miku@iki.fi>, 2000
  11  *
  12  * Fixes to reconstruction by Jakob Østergaard" <jakob@ostenfeld.dk>
  13  * Various fixes by Neil Brown <neilb@cse.unsw.edu.au>
  14  *
  15  * Changes by Peter T. Breuer <ptb@it.uc3m.es> 31/1/2003 to support
  16  * bitmapped intelligence in resync:
  17  *
  18  *      - bitmap marked during normal i/o
  19  *      - bitmap used to skip nondirty blocks during sync
  20  *
  21  * Additions to bitmap code, (C) 2003-2004 Paul Clements, SteelEye Technology:
  22  * - persistent bitmap code
  23  *
  24  * This program is free software; you can redistribute it and/or modify
  25  * it under the terms of the GNU General Public License as published by
  26  * the Free Software Foundation; either version 2, or (at your option)
  27  * any later version.
  28  *
  29  * You should have received a copy of the GNU General Public License
  30  * (for example /usr/src/linux/COPYING); if not, write to the Free
  31  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  32  */
  33
  34 #include <linux/slab.h>
  35 #include <linux/delay.h>
  36 #include <linux/blkdev.h>
  37 #include <linux/seq_file.h>
  38 #include <linux/ratelimit.h>
  39 #include "md.h"
  40 #include "raid1.h"
  41 #include "bitmap.h"
  42
  43 #define DEBUG 0
  44 #define PRINTK(x...) do { if (DEBUG) printk(x); } while (0)
  45
  46 /*
  47  * Number of guaranteed r1bios in case of extreme VM load:
  48  */
  49 #define NR_RAID1_BIOS 256
  50
  51
  52 static void allow_barrier(conf_t *conf);
  53 static void lower_barrier(conf_t *conf);
  54
  55 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
  56 {
  57         struct pool_info *pi = data;
  58         int size = offsetof(r1bio_t, bios[pi->raid_disks]);
  59
  60         /* allocate a r1bio with room for raid_disks entries in the bios array */
  61         return kzalloc(size, gfp_flags);
  62 }
  63
  64 static void r1bio_pool_free(void *r1_bio, void *data)
  65 {
  66         kfree(r1_bio);
  67 }
  68
  69 #define RESYNC_BLOCK_SIZE (64*1024)
  70 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
  71 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
  72 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  73 #define RESYNC_WINDOW (2048*1024)
  74
  75 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
  76 {
  77         struct pool_info *pi = data;
  78         struct page *page;
  79         r1bio_t *r1_bio;
  80         struct bio *bio;
  81         int i, j;
  82
  83         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
  84         if (!r1_bio)
  85                 return NULL;
  86
  87         /*
  88          * Allocate bios : 1 for reading, n-1 for writing
  89          */
  90         for (j = pi->raid_disks ; j-- ; ) {
  91                 bio = bio_kmalloc(gfp_flags, RESYNC_PAGES);
  92                 if (!bio)
  93                         goto out_free_bio;
  94                 r1_bio->bios[j] = bio;
  95         }
  96         /*
  97          * Allocate RESYNC_PAGES data pages and attach them to
  98          * the first bio.
  99          * If this is a user-requested check/repair, allocate
 100          * RESYNC_PAGES for each bio.
 101          */
 102         if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
 103                 j = pi->raid_disks;
 104         else
 105                 j = 1;
 106         while(j--) {
 107                 bio = r1_bio->bios[j];
 108                 for (i = 0; i < RESYNC_PAGES; i++) {
 109                         page = alloc_page(gfp_flags);
 110                         if (unlikely(!page))
 111                                 goto out_free_pages;
 112
 113                         bio->bi_io_vec[i].bv_page = page;
 114                         bio->bi_vcnt = i+1;
 115                 }
 116         }
 117         /* If not user-requests, copy the page pointers to all bios */
 118         if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
 119                 for (i=0; i<RESYNC_PAGES ; i++)
 120                         for (j=1; j<pi->raid_disks; j++)
 121                                 r1_bio->bios[j]->bi_io_vec[i].bv_page =
 122                                         r1_bio->bios[0]->bi_io_vec[i].bv_page;
 123         }
 124
 125         r1_bio->master_bio = NULL;
 126
 127         return r1_bio;
 128
 129 out_free_pages:
 130         for (j=0 ; j < pi->raid_disks; j++)
 131                 for (i=0; i < r1_bio->bios[j]->bi_vcnt ; i++)
 132                         put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
 133         j = -1;
 134 out_free_bio:
 135         while ( ++j < pi->raid_disks )
 136                 bio_put(r1_bio->bios[j]);
 137         r1bio_pool_free(r1_bio, data);
 138         return NULL;
 139 }
 140
 141 static void r1buf_pool_free(void *__r1_bio, void *data)
 142 {
 143         struct pool_info *pi = data;
 144         int i,j;
 145         r1bio_t *r1bio = __r1_bio;
 146
 147         for (i = 0; i < RESYNC_PAGES; i++)
 148                 for (j = pi->raid_disks; j-- ;) {
 149                         if (j == 0 ||
 150                             r1bio->bios[j]->bi_io_vec[i].bv_page !=
 151                             r1bio->bios[0]->bi_io_vec[i].bv_page)
 152                                 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
 153                 }
 154         for (i=0 ; i < pi->raid_disks; i++)
 155                 bio_put(r1bio->bios[i]);
 156
 157         r1bio_pool_free(r1bio, data);
 158 }
 159
 160 static void put_all_bios(conf_t *conf, r1bio_t *r1_bio)
 161 {
 162         int i;
 163
 164         for (i = 0; i < conf->raid_disks; i++) {
 165                 struct bio **bio = r1_bio->bios + i;
 166                 if (*bio && *bio != IO_BLOCKED)
 167                         bio_put(*bio);
 168                 *bio = NULL;
 169         }
 170 }
 171
 172 static void free_r1bio(r1bio_t *r1_bio)
 173 {
 174         conf_t *conf = r1_bio->mddev->private;
 175
 176         put_all_bios(conf, r1_bio);
 177         mempool_free(r1_bio, conf->r1bio_pool);
 178 }
 179
 180 static void put_buf(r1bio_t *r1_bio)
 181 {
 182         conf_t *conf = r1_bio->mddev->private;
 183         int i;
 184
 185         for (i=0; i<conf->raid_disks; i++) {
 186                 struct bio *bio = r1_bio->bios[i];
 187                 if (bio->bi_end_io)
 188                         rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
 189         }
 190
 191         mempool_free(r1_bio, conf->r1buf_pool);
 192
 193         lower_barrier(conf);
 194 }
 195
 196 static void reschedule_retry(r1bio_t *r1_bio)
 197 {
 198         unsigned long flags;
 199         mddev_t *mddev = r1_bio->mddev;
 200         conf_t *conf = mddev->private;
 201
 202         spin_lock_irqsave(&conf->device_lock, flags);
 203         list_add(&r1_bio->retry_list, &conf->retry_list);
 204         conf->nr_queued ++;
 205         spin_unlock_irqrestore(&conf->device_lock, flags);
 206
 207         wake_up(&conf->wait_barrier);
 208         md_wakeup_thread(mddev->thread);
 209 }
 210
 211 /*
 212  * raid_end_bio_io() is called when we have finished servicing a mirrored
 213  * operation and are ready to return a success/failure code to the buffer
 214  * cache layer.
 215  */
 216 static void call_bio_endio(r1bio_t *r1_bio)
 217 {
 218         struct bio *bio = r1_bio->master_bio;
 219         int done;
 220         conf_t *conf = r1_bio->mddev->private;
 221
 222         if (bio->bi_phys_segments) {
 223                 unsigned long flags;
 224                 spin_lock_irqsave(&conf->device_lock, flags);
 225                 bio->bi_phys_segments--;
 226                 done = (bio->bi_phys_segments == 0);
 227                 spin_unlock_irqrestore(&conf->device_lock, flags);
 228         } else
 229                 done = 1;
 230
 231         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
 232                 clear_bit(BIO_UPTODATE, &bio->bi_flags);
 233         if (done) {
 234                 bio_endio(bio, 0);
 235                 /*
 236                  * Wake up any possible resync thread that waits for the device
 237                  * to go idle.
 238                  */
 239                 allow_barrier(conf);
 240         }
 241 }
 242
 243 static void raid_end_bio_io(r1bio_t *r1_bio)
 244 {
 245         struct bio *bio = r1_bio->master_bio;
 246
 247         /* if nobody has done the final endio yet, do it now */
 248         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 249                 PRINTK(KERN_DEBUG "raid1: sync end %s on sectors %llu-%llu\n",
 250                         (bio_data_dir(bio) == WRITE) ? "write" : "read",
 251                         (unsigned long long) bio->bi_sector,
 252                         (unsigned long long) bio->bi_sector +
 253                                 (bio->bi_size >> 9) - 1);
 254
 255                 call_bio_endio(r1_bio);
 256         }
 257         free_r1bio(r1_bio);
 258 }
 259
 260 /*
 261  * Update disk head position estimator based on IRQ completion info.
 262  */
 263 static inline void update_head_pos(int disk, r1bio_t *r1_bio)
 264 {
 265         conf_t *conf = r1_bio->mddev->private;
 266
 267         conf->mirrors[disk].head_position =
 268                 r1_bio->sector + (r1_bio->sectors);
 269 }
 270
 271 static void raid1_end_read_request(struct bio *bio, int error)
 272 {
 273         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 274         r1bio_t *r1_bio = bio->bi_private;
 275         int mirror;
 276         conf_t *conf = r1_bio->mddev->private;
 277
 278         mirror = r1_bio->read_disk;
 279         /*
 280          * this branch is our 'one mirror IO has finished' event handler:
 281          */
 282         update_head_pos(mirror, r1_bio);
 283
 284         if (uptodate)
 285                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 286         else {
 287                 /* If all other devices have failed, we want to return
 288                  * the error upwards rather than fail the last device.
 289                  * Here we redefine "uptodate" to mean "Don't want to retry"
 290                  */
 291                 unsigned long flags;
 292                 spin_lock_irqsave(&conf->device_lock, flags);
 293                 if (r1_bio->mddev->degraded == conf->raid_disks ||
 294                     (r1_bio->mddev->degraded == conf->raid_disks-1 &&
 295                      !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
 296                         uptodate = 1;
 297                 spin_unlock_irqrestore(&conf->device_lock, flags);
 298         }
 299
 300         if (uptodate)
 301                 raid_end_bio_io(r1_bio);
 302         else {
 303                 /*
 304                  * oops, read error:
 305                  */
 306                 char b[BDEVNAME_SIZE];
 307                 printk_ratelimited(
 308                         KERN_ERR "md/raid1:%s: %s: "
 309                         "rescheduling sector %llu\n",
 310                         mdname(conf->mddev),
 311                         bdevname(conf->mirrors[mirror].rdev->bdev,
 312                                  b),
 313                         (unsigned long long)r1_bio->sector);
 314                 set_bit(R1BIO_ReadError, &r1_bio->state);
 315                 reschedule_retry(r1_bio);
 316         }
 317
 318         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 319 }
 320
 321 static void r1_bio_write_done(r1bio_t *r1_bio)
 322 {
 323         if (atomic_dec_and_test(&r1_bio->remaining))
 324         {
 325                 /* it really is the end of this request */
 326                 if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 327                         /* free extra copy of the data pages */
 328                         int i = r1_bio->behind_page_count;
 329                         while (i--)
 330                                 safe_put_page(r1_bio->behind_pages[i]);
 331                         kfree(r1_bio->behind_pages);
 332                         r1_bio->behind_pages = NULL;
 333                 }
 334                 /* clear the bitmap if all writes complete successfully */
 335                 bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 336                                 r1_bio->sectors,
 337                                 !test_bit(R1BIO_Degraded, &r1_bio->state),
 338                                 test_bit(R1BIO_BehindIO, &r1_bio->state));
 339                 md_write_end(r1_bio->mddev);
 340                 raid_end_bio_io(r1_bio);
 341         }
 342 }
 343
 344 static void raid1_end_write_request(struct bio *bio, int error)
 345 {
 346         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 347         r1bio_t *r1_bio = bio->bi_private;
 348         int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 349         conf_t *conf = r1_bio->mddev->private;
 350         struct bio *to_put = NULL;
 351
 352
 353         for (mirror = 0; mirror < conf->raid_disks; mirror++)
 354                 if (r1_bio->bios[mirror] == bio)
 355                         break;
 356
 357         /*
 358          * 'one mirror IO has finished' event handler:
 359          */
 360         r1_bio->bios[mirror] = NULL;
 361         to_put = bio;
 362         if (!uptodate) {
 363                 md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
 364                 /* an I/O failed, we can't clear the bitmap */
 365                 set_bit(R1BIO_Degraded, &r1_bio->state);
 366         } else
 367                 /*
 368                  * Set R1BIO_Uptodate in our master bio, so that we
 369                  * will return a good error code for to the higher
 370                  * levels even if IO on some other mirrored buffer
 371                  * fails.
 372                  *
 373                  * The 'master' represents the composite IO operation
 374                  * to user-side. So if something waits for IO, then it
 375                  * will wait for the 'master' bio.
 376                  */
 377                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 378
 379         update_head_pos(mirror, r1_bio);
 380
 381         if (behind) {
 382                 if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
 383                         atomic_dec(&r1_bio->behind_remaining);
 384
 385                 /*
 386                  * In behind mode, we ACK the master bio once the I/O
 387                  * has safely reached all non-writemostly
 388                  * disks. Setting the Returned bit ensures that this
 389                  * gets done only once -- we don't ever want to return
 390                  * -EIO here, instead we'll wait
 391                  */
 392                 if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 393                     test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 394                         /* Maybe we can return now */
 395                         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 396                                 struct bio *mbio = r1_bio->master_bio;
 397                                 PRINTK(KERN_DEBUG "raid1: behind end write sectors %llu-%llu\n",
 398                                        (unsigned long long) mbio->bi_sector,
 399                                        (unsigned long long) mbio->bi_sector +
 400                                        (mbio->bi_size >> 9) - 1);
 401                                 call_bio_endio(r1_bio);
 402                         }
 403                 }
 404         }
 405         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 406
 407         /*
 408          * Let's see if all mirrored write operations have finished
 409          * already.
 410          */
 411         r1_bio_write_done(r1_bio);
 412
 413         if (to_put)
 414                 bio_put(to_put);
 415 }
 416
 417
 418 /*
 419  * This routine returns the disk from which the requested read should
 420  * be done. There is a per-array 'next expected sequential IO' sector
 421  * number - if this matches on the next IO then we use the last disk.
 422  * There is also a per-disk 'last know head position' sector that is
 423  * maintained from IRQ contexts, both the normal and the resync IO
 424  * completion handlers update this position correctly. If there is no
 425  * perfect sequential match then we pick the disk whose head is closest.
 426  *
 427  * If there are 2 mirrors in the same 2 devices, performance degrades
 428  * because position is mirror, not device based.
 429  *
 430  * The rdev for the device selected will have nr_pending incremented.
 431  */
 432 static int read_balance(conf_t *conf, r1bio_t *r1_bio, int *max_sectors)
 433 {
 434         const sector_t this_sector = r1_bio->sector;
 435         int sectors;
 436         int best_good_sectors;
 437         int start_disk;
 438         int best_disk;
 439         int i;
 440         sector_t best_dist;
 441         mdk_rdev_t *rdev;
 442         int choose_first;
 443
 444         rcu_read_lock();
 445         /*
 446          * Check if we can balance. We can balance on the whole
 447          * device if no resync is going on, or below the resync window.
 448          * We take the first readable disk when above the resync window.
 449          */
 450  retry:
 451         sectors = r1_bio->sectors;
 452         best_disk = -1;
 453         best_dist = MaxSector;
 454         best_good_sectors = 0;
 455
 456         if (conf->mddev->recovery_cp < MaxSector &&
 457             (this_sector + sectors >= conf->next_resync)) {
 458                 choose_first = 1;
 459                 start_disk = 0;
 460         } else {
 461                 choose_first = 0;
 462                 start_disk = conf->last_used;
 463         }
 464
 465         for (i = 0 ; i < conf->raid_disks ; i++) {
 466                 sector_t dist;
 467                 sector_t first_bad;
 468                 int bad_sectors;
 469
 470                 int disk = start_disk + i;
 471                 if (disk >= conf->raid_disks)
 472                         disk -= conf->raid_disks;
 473
 474                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
 475                 if (r1_bio->bios[disk] == IO_BLOCKED
 476                     || rdev == NULL
 477                     || test_bit(Faulty, &rdev->flags))
 478                         continue;
 479                 if (!test_bit(In_sync, &rdev->flags) &&
 480                     rdev->recovery_offset < this_sector + sectors)
 481                         continue;
 482                 if (test_bit(WriteMostly, &rdev->flags)) {
 483                         /* Don't balance among write-mostly, just
 484                          * use the first as a last resort */
 485                         if (best_disk < 0)
 486                                 best_disk = disk;
 487                         continue;
 488                 }
 489                 /* This is a reasonable device to use.  It might
 490                  * even be best.
 491                  */
 492                 if (is_badblock(rdev, this_sector, sectors,
 493                                 &first_bad, &bad_sectors)) {
 494                         if (best_dist < MaxSector)
 495                                 /* already have a better device */
 496                                 continue;
 497                         if (first_bad <= this_sector) {
 498                                 /* cannot read here. If this is the 'primary'
 499                                  * device, then we must not read beyond
 500                                  * bad_sectors from another device..
 501                                  */
 502                                 bad_sectors -= (this_sector - first_bad);
 503                                 if (choose_first && sectors > bad_sectors)
 504                                         sectors = bad_sectors;
 505                                 if (best_good_sectors > sectors)
 506                                         best_good_sectors = sectors;
 507
 508                         } else {
 509                                 sector_t good_sectors = first_bad - this_sector;
 510                                 if (good_sectors > best_good_sectors) {
 511                                         best_good_sectors = good_sectors;
 512                                         best_disk = disk;
 513                                 }
 514                                 if (choose_first)
 515                                         break;
 516                         }
 517                         continue;
 518                 } else
 519                         best_good_sectors = sectors;
 520
 521                 dist = abs(this_sector - conf->mirrors[disk].head_position);
 522                 if (choose_first
 523                     /* Don't change to another disk for sequential reads */
 524                     || conf->next_seq_sect == this_sector
 525                     || dist == 0
 526                     /* If device is idle, use it */
 527                     || atomic_read(&rdev->nr_pending) == 0) {
 528                         best_disk = disk;
 529                         break;
 530                 }
 531                 if (dist < best_dist) {
 532                         best_dist = dist;
 533                         best_disk = disk;
 534                 }
 535         }
 536
 537         if (best_disk >= 0) {
 538                 rdev = rcu_dereference(conf->mirrors[best_disk].rdev);
 539                 if (!rdev)
 540                         goto retry;
 541                 atomic_inc(&rdev->nr_pending);
 542                 if (test_bit(Faulty, &rdev->flags)) {
 543                         /* cannot risk returning a device that failed
 544                          * before we inc'ed nr_pending
 545                          */
 546                         rdev_dec_pending(rdev, conf->mddev);
 547                         goto retry;
 548                 }
 549                 sectors = best_good_sectors;
 550                 conf->next_seq_sect = this_sector + sectors;
 551                 conf->last_used = best_disk;
 552         }
 553         rcu_read_unlock();
 554         *max_sectors = sectors;
 555
 556         return best_disk;
 557 }
 558
 559 int md_raid1_congested(mddev_t *mddev, int bits)
 560 {
 561         conf_t *conf = mddev->private;
 562         int i, ret = 0;
 563
 564         rcu_read_lock();
 565         for (i = 0; i < mddev->raid_disks; i++) {
 566                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 567                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 568                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 569
 570                         BUG_ON(!q);
 571
 572                         /* Note the '|| 1' - when read_balance prefers
 573                          * non-congested targets, it can be removed
 574                          */
 575                         if ((bits & (1<<BDI_async_congested)) || 1)
 576                                 ret |= bdi_congested(&q->backing_dev_info, bits);
 577                         else
 578                                 ret &= bdi_congested(&q->backing_dev_info, bits);
 579                 }
 580         }
 581         rcu_read_unlock();
 582         return ret;
 583 }
 584 EXPORT_SYMBOL_GPL(md_raid1_congested);
 585
 586 static int raid1_congested(void *data, int bits)
 587 {
 588         mddev_t *mddev = data;
 589
 590         return mddev_congested(mddev, bits) ||
 591                 md_raid1_congested(mddev, bits);
 592 }
 593
 594 static void flush_pending_writes(conf_t *conf)
 595 {
 596         /* Any writes that have been queued but are awaiting
 597          * bitmap updates get flushed here.
 598          */
 599         spin_lock_irq(&conf->device_lock);
 600
 601         if (conf->pending_bio_list.head) {
 602                 struct bio *bio;
 603                 bio = bio_list_get(&conf->pending_bio_list);
 604                 spin_unlock_irq(&conf->device_lock);
 605                 /* flush any pending bitmap writes to
 606                  * disk before proceeding w/ I/O */
 607                 bitmap_unplug(conf->mddev->bitmap);
 608
 609                 while (bio) { /* submit pending writes */
 610                         struct bio *next = bio->bi_next;
 611                         bio->bi_next = NULL;
 612                         generic_make_request(bio);
 613                         bio = next;
 614                 }
 615         } else
 616                 spin_unlock_irq(&conf->device_lock);
 617 }
 618
 619 /* Barriers....
 620  * Sometimes we need to suspend IO while we do something else,
 621  * either some resync/recovery, or reconfigure the array.
 622  * To do this we raise a 'barrier'.
 623  * The 'barrier' is a counter that can be raised multiple times
 624  * to count how many activities are happening which preclude
 625  * normal IO.
 626  * We can only raise the barrier if there is no pending IO.
 627  * i.e. if nr_pending == 0.
 628  * We choose only to raise the barrier if no-one is waiting for the
 629  * barrier to go down.  This means that as soon as an IO request
 630  * is ready, no other operations which require a barrier will start
 631  * until the IO request has had a chance.
 632  *
 633  * So: regular IO calls 'wait_barrier'.  When that returns there
 634  *    is no backgroup IO happening,  It must arrange to call
 635  *    allow_barrier when it has finished its IO.
 636  * backgroup IO calls must call raise_barrier.  Once that returns
 637  *    there is no normal IO happeing.  It must arrange to call
 638  *    lower_barrier when the particular background IO completes.
 639  */
 640 #define RESYNC_DEPTH 32
 641
 642 static void raise_barrier(conf_t *conf)
 643 {
 644         spin_lock_irq(&conf->resync_lock);
 645
 646         /* Wait until no block IO is waiting */
 647         wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
 648                             conf->resync_lock, );
 649
 650         /* block any new IO from starting */
 651         conf->barrier++;
 652
 653         /* Now wait for all pending IO to complete */
 654         wait_event_lock_irq(conf->wait_barrier,
 655                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 656                             conf->resync_lock, );
 657
 658         spin_unlock_irq(&conf->resync_lock);
 659 }
 660
 661 static void lower_barrier(conf_t *conf)
 662 {
 663         unsigned long flags;
 664         BUG_ON(conf->barrier <= 0);
 665         spin_lock_irqsave(&conf->resync_lock, flags);
 666         conf->barrier--;
 667         spin_unlock_irqrestore(&conf->resync_lock, flags);
 668         wake_up(&conf->wait_barrier);
 669 }
 670
 671 static void wait_barrier(conf_t *conf)
 672 {
 673         spin_lock_irq(&conf->resync_lock);
 674         if (conf->barrier) {
 675                 conf->nr_waiting++;
 676                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
 677                                     conf->resync_lock,
 678                                     );
 679                 conf->nr_waiting--;
 680         }
 681         conf->nr_pending++;
 682         spin_unlock_irq(&conf->resync_lock);
 683 }
 684
 685 static void allow_barrier(conf_t *conf)
 686 {
 687         unsigned long flags;
 688         spin_lock_irqsave(&conf->resync_lock, flags);
 689         conf->nr_pending--;
 690         spin_unlock_irqrestore(&conf->resync_lock, flags);
 691         wake_up(&conf->wait_barrier);
 692 }
 693
 694 static void freeze_array(conf_t *conf)
 695 {
 696         /* stop syncio and normal IO and wait for everything to
 697          * go quite.
 698          * We increment barrier and nr_waiting, and then
 699          * wait until nr_pending match nr_queued+1
 700          * This is called in the context of one normal IO request
 701          * that has failed. Thus any sync request that might be pending
 702          * will be blocked by nr_pending, and we need to wait for
 703          * pending IO requests to complete or be queued for re-try.
 704          * Thus the number queued (nr_queued) plus this request (1)
 705          * must match the number of pending IOs (nr_pending) before
 706          * we continue.
 707          */
 708         spin_lock_irq(&conf->resync_lock);
 709         conf->barrier++;
 710         conf->nr_waiting++;
 711         wait_event_lock_irq(conf->wait_barrier,
 712                             conf->nr_pending == conf->nr_queued+1,
 713                             conf->resync_lock,
 714                             flush_pending_writes(conf));
 715         spin_unlock_irq(&conf->resync_lock);
 716 }
 717 static void unfreeze_array(conf_t *conf)
 718 {
 719         /* reverse the effect of the freeze */
 720         spin_lock_irq(&conf->resync_lock);
 721         conf->barrier--;
 722         conf->nr_waiting--;
 723         wake_up(&conf->wait_barrier);
 724         spin_unlock_irq(&conf->resync_lock);
 725 }
 726
 727
 728 /* duplicate the data pages for behind I/O
 729  */
 730 static void alloc_behind_pages(struct bio *bio, r1bio_t *r1_bio)
 731 {
 732         int i;
 733         struct bio_vec *bvec;
 734         struct page **pages = kzalloc(bio->bi_vcnt * sizeof(struct page*),
 735                                         GFP_NOIO);
 736         if (unlikely(!pages))
 737                 return;
 738
 739         bio_for_each_segment(bvec, bio, i) {
 740                 pages[i] = alloc_page(GFP_NOIO);
 741                 if (unlikely(!pages[i]))
 742                         goto do_sync_io;
 743                 memcpy(kmap(pages[i]) + bvec->bv_offset,
 744                         kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
 745                 kunmap(pages[i]);
 746                 kunmap(bvec->bv_page);
 747         }
 748         r1_bio->behind_pages = pages;
 749         r1_bio->behind_page_count = bio->bi_vcnt;
 750         set_bit(R1BIO_BehindIO, &r1_bio->state);
 751         return;
 752
 753 do_sync_io:
 754         for (i = 0; i < bio->bi_vcnt; i++)
 755                 if (pages[i])
 756                         put_page(pages[i]);
 757         kfree(pages);
 758         PRINTK("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
 759 }
 760
 761 static int make_request(mddev_t *mddev, struct bio * bio)
 762 {
 763         conf_t *conf = mddev->private;
 764         mirror_info_t *mirror;
 765         r1bio_t *r1_bio;
 766         struct bio *read_bio;
 767         int i, targets = 0, disks;
 768         struct bitmap *bitmap;
 769         unsigned long flags;
 770         const int rw = bio_data_dir(bio);
 771         const unsigned long do_sync = (bio->bi_rw & REQ_SYNC);
 772         const unsigned long do_flush_fua = (bio->bi_rw & (REQ_FLUSH | REQ_FUA));
 773         mdk_rdev_t *blocked_rdev;
 774         int plugged;
 775
 776         /*
 777          * Register the new request and wait if the reconstruction
 778          * thread has put up a bar for new requests.
 779          * Continue immediately if no resync is active currently.
 780          */
 781
 782         md_write_start(mddev, bio); /* wait on superblock update early */
 783
 784         if (bio_data_dir(bio) == WRITE &&
 785             bio->bi_sector + bio->bi_size/512 > mddev->suspend_lo &&
 786             bio->bi_sector < mddev->suspend_hi) {
 787                 /* As the suspend_* range is controlled by
 788                  * userspace, we want an interruptible
 789                  * wait.
 790                  */
 791                 DEFINE_WAIT(w);
 792                 for (;;) {
 793                         flush_signals(current);
 794                         prepare_to_wait(&conf->wait_barrier,
 795                                         &w, TASK_INTERRUPTIBLE);
 796                         if (bio->bi_sector + bio->bi_size/512 <= mddev->suspend_lo ||
 797                             bio->bi_sector >= mddev->suspend_hi)
 798                                 break;
 799                         schedule();
 800                 }
 801                 finish_wait(&conf->wait_barrier, &w);
 802         }
 803
 804         wait_barrier(conf);
 805
 806         bitmap = mddev->bitmap;
 807
 808         /*
 809          * make_request() can abort the operation when READA is being
 810          * used and no empty request is available.
 811          *
 812          */
 813         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
 814
 815         r1_bio->master_bio = bio;
 816         r1_bio->sectors = bio->bi_size >> 9;
 817         r1_bio->state = 0;
 818         r1_bio->mddev = mddev;
 819         r1_bio->sector = bio->bi_sector;
 820
 821         /* We might need to issue multiple reads to different
 822          * devices if there are bad blocks around, so we keep
 823          * track of the number of reads in bio->bi_phys_segments.
 824          * If this is 0, there is only one r1_bio and no locking
 825          * will be needed when requests complete.  If it is
 826          * non-zero, then it is the number of not-completed requests.
 827          */
 828         bio->bi_phys_segments = 0;
 829         clear_bit(BIO_SEG_VALID, &bio->bi_flags);
 830
 831         if (rw == READ) {
 832                 /*
 833                  * read balancing logic:
 834                  */
 835                 int max_sectors;
 836                 int rdisk;
 837
 838 read_again:
 839                 rdisk = read_balance(conf, r1_bio, &max_sectors);
 840
 841                 if (rdisk < 0) {
 842                         /* couldn't find anywhere to read from */
 843                         raid_end_bio_io(r1_bio);
 844                         return 0;
 845                 }
 846                 mirror = conf->mirrors + rdisk;
 847
 848                 if (test_bit(WriteMostly, &mirror->rdev->flags) &&
 849                     bitmap) {
 850                         /* Reading from a write-mostly device must
 851                          * take care not to over-take any writes
 852                          * that are 'behind'
 853                          */
 854                         wait_event(bitmap->behind_wait,
 855                                    atomic_read(&bitmap->behind_writes) == 0);
 856                 }
 857                 r1_bio->read_disk = rdisk;
 858
 859                 read_bio = bio_clone_mddev(bio, GFP_NOIO, mddev);
 860                 md_trim_bio(read_bio, r1_bio->sector - bio->bi_sector,
 861                             max_sectors);
 862
 863                 r1_bio->bios[rdisk] = read_bio;
 864
 865                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
 866                 read_bio->bi_bdev = mirror->rdev->bdev;
 867                 read_bio->bi_end_io = raid1_end_read_request;
 868                 read_bio->bi_rw = READ | do_sync;
 869                 read_bio->bi_private = r1_bio;
 870
 871                 if (max_sectors < r1_bio->sectors) {
 872                         /* could not read all from this device, so we will
 873                          * need another r1_bio.
 874                          */
 875                         int sectors_handled;
 876
 877                         sectors_handled = (r1_bio->sector + max_sectors
 878                                            - bio->bi_sector);
 879                         r1_bio->sectors = max_sectors;
 880                         spin_lock_irq(&conf->device_lock);
 881                         if (bio->bi_phys_segments == 0)
 882                                 bio->bi_phys_segments = 2;
 883                         else
 884                                 bio->bi_phys_segments++;
 885                         spin_unlock_irq(&conf->device_lock);
 886                         /* Cannot call generic_make_request directly
 887                          * as that will be queued in __make_request
 888                          * and subsequent mempool_alloc might block waiting
 889                          * for it.  So hand bio over to raid1d.
 890                          */
 891                         reschedule_retry(r1_bio);
 892
 893                         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
 894
 895                         r1_bio->master_bio = bio;
 896                         r1_bio->sectors = (bio->bi_size >> 9) - sectors_handled;
 897                         r1_bio->state = 0;
 898                         r1_bio->mddev = mddev;
 899                         r1_bio->sector = bio->bi_sector + sectors_handled;
 900                         goto read_again;
 901                 } else
 902                         generic_make_request(read_bio);
 903                 return 0;
 904         }
 905
 906         /*
 907          * WRITE:
 908          */
 909         /* first select target devices under spinlock and
 910          * inc refcount on their rdev.  Record them by setting
 911          * bios[x] to bio
 912          */
 913         plugged = mddev_check_plugged(mddev);
 914
 915         disks = conf->raid_disks;
 916  retry_write:
 917         blocked_rdev = NULL;
 918         rcu_read_lock();
 919         for (i = 0;  i < disks; i++) {
 920                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 921                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
 922                         atomic_inc(&rdev->nr_pending);
 923                         blocked_rdev = rdev;
 924                         break;
 925                 }
 926                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 927                         atomic_inc(&rdev->nr_pending);
 928                         if (test_bit(Faulty, &rdev->flags)) {
 929                                 rdev_dec_pending(rdev, mddev);
 930                                 r1_bio->bios[i] = NULL;
 931                         } else {
 932                                 r1_bio->bios[i] = bio;
 933                                 targets++;
 934                         }
 935                 } else
 936                         r1_bio->bios[i] = NULL;
 937         }
 938         rcu_read_unlock();
 939
 940         if (unlikely(blocked_rdev)) {
 941                 /* Wait for this device to become unblocked */
 942                 int j;
 943
 944                 for (j = 0; j < i; j++)
 945                         if (r1_bio->bios[j])
 946                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
 947
 948                 allow_barrier(conf);
 949                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
 950                 wait_barrier(conf);
 951                 goto retry_write;
 952         }
 953
 954         if (targets < conf->raid_disks) {
 955                 /* array is degraded, we will not clear the bitmap
 956                  * on I/O completion (see raid1_end_write_request) */
 957                 set_bit(R1BIO_Degraded, &r1_bio->state);
 958         }
 959
 960         /* do behind I/O ?
 961          * Not if there are too many, or cannot allocate memory,
 962          * or a reader on WriteMostly is waiting for behind writes
 963          * to flush */
 964         if (bitmap &&
 965             (atomic_read(&bitmap->behind_writes)
 966              < mddev->bitmap_info.max_write_behind) &&
 967             !waitqueue_active(&bitmap->behind_wait))
 968                 alloc_behind_pages(bio, r1_bio);
 969
 970         atomic_set(&r1_bio->remaining, 1);
 971         atomic_set(&r1_bio->behind_remaining, 0);
 972
 973         bitmap_startwrite(bitmap, bio->bi_sector, r1_bio->sectors,
 974                                 test_bit(R1BIO_BehindIO, &r1_bio->state));
 975         for (i = 0; i < disks; i++) {
 976                 struct bio *mbio;
 977                 if (!r1_bio->bios[i])
 978                         continue;
 979
 980                 mbio = bio_clone_mddev(bio, GFP_NOIO, mddev);
 981                 r1_bio->bios[i] = mbio;
 982
 983                 mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
 984                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
 985                 mbio->bi_end_io = raid1_end_write_request;
 986                 mbio->bi_rw = WRITE | do_flush_fua | do_sync;
 987                 mbio->bi_private = r1_bio;
 988
 989                 if (r1_bio->behind_pages) {
 990                         struct bio_vec *bvec;
 991                         int j;
 992
 993                         /* Yes, I really want the '__' version so that
 994                          * we clear any unused pointer in the io_vec, rather
 995                          * than leave them unchanged.  This is important
 996                          * because when we come to free the pages, we won't
 997                          * know the original bi_idx, so we just free
 998                          * them all
 999                          */
1000                         __bio_for_each_segment(bvec, mbio, j, 0)
1001                                 bvec->bv_page = r1_bio->behind_pages[j];
1002                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
1003                                 atomic_inc(&r1_bio->behind_remaining);
1004                 }
1005
1006                 atomic_inc(&r1_bio->remaining);
1007                 spin_lock_irqsave(&conf->device_lock, flags);
1008                 bio_list_add(&conf->pending_bio_list, mbio);
1009                 spin_unlock_irqrestore(&conf->device_lock, flags);
1010         }
1011         r1_bio_write_done(r1_bio);
1012
1013         /* In case raid1d snuck in to freeze_array */
1014         wake_up(&conf->wait_barrier);
1015
1016         if (do_sync || !bitmap || !plugged)
1017                 md_wakeup_thread(mddev->thread);
1018
1019         return 0;
1020 }
1021
1022 static void status(struct seq_file *seq, mddev_t *mddev)
1023 {
1024         conf_t *conf = mddev->private;
1025         int i;
1026
1027         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
1028                    conf->raid_disks - mddev->degraded);
1029         rcu_read_lock();
1030         for (i = 0; i < conf->raid_disks; i++) {
1031                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1032                 seq_printf(seq, "%s",
1033                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1034         }
1035         rcu_read_unlock();
1036         seq_printf(seq, "]");
1037 }
1038
1039
1040 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1041 {
1042         char b[BDEVNAME_SIZE];
1043         conf_t *conf = mddev->private;
1044
1045         /*
1046          * If it is not operational, then we have already marked it as dead
1047          * else if it is the last working disks, ignore the error, let the
1048          * next level up know.
1049          * else mark the drive as failed
1050          */
1051         if (test_bit(In_sync, &rdev->flags)
1052             && (conf->raid_disks - mddev->degraded) == 1) {
1053                 /*
1054                  * Don't fail the drive, act as though we were just a
1055                  * normal single drive.
1056                  * However don't try a recovery from this drive as
1057                  * it is very likely to fail.
1058                  */
1059                 conf->recovery_disabled = mddev->recovery_disabled;
1060                 return;
1061         }
1062         set_bit(Blocked, &rdev->flags);
1063         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1064                 unsigned long flags;
1065                 spin_lock_irqsave(&conf->device_lock, flags);
1066                 mddev->degraded++;
1067                 set_bit(Faulty, &rdev->flags);
1068                 spin_unlock_irqrestore(&conf->device_lock, flags);
1069                 /*
1070                  * if recovery is running, make sure it aborts.
1071                  */
1072                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1073         } else
1074                 set_bit(Faulty, &rdev->flags);
1075         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1076         printk(KERN_ALERT
1077                "md/raid1:%s: Disk failure on %s, disabling device.\n"
1078                "md/raid1:%s: Operation continuing on %d devices.\n",
1079                mdname(mddev), bdevname(rdev->bdev, b),
1080                mdname(mddev), conf->raid_disks - mddev->degraded);
1081 }
1082
1083 static void print_conf(conf_t *conf)
1084 {
1085         int i;
1086
1087         printk(KERN_DEBUG "RAID1 conf printout:\n");
1088         if (!conf) {
1089                 printk(KERN_DEBUG "(!conf)\n");
1090                 return;
1091         }
1092         printk(KERN_DEBUG " --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1093                 conf->raid_disks);
1094
1095         rcu_read_lock();
1096         for (i = 0; i < conf->raid_disks; i++) {
1097                 char b[BDEVNAME_SIZE];
1098                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1099                 if (rdev)
1100                         printk(KERN_DEBUG " disk %d, wo:%d, o:%d, dev:%s\n",
1101                                i, !test_bit(In_sync, &rdev->flags),
1102                                !test_bit(Faulty, &rdev->flags),
1103                                bdevname(rdev->bdev,b));
1104         }
1105         rcu_read_unlock();
1106 }
1107
1108 static void close_sync(conf_t *conf)
1109 {
1110         wait_barrier(conf);
1111         allow_barrier(conf);
1112
1113         mempool_destroy(conf->r1buf_pool);
1114         conf->r1buf_pool = NULL;
1115 }
1116
1117 static int raid1_spare_active(mddev_t *mddev)
1118 {
1119         int i;
1120         conf_t *conf = mddev->private;
1121         int count = 0;
1122         unsigned long flags;
1123
1124         /*
1125          * Find all failed disks within the RAID1 configuration
1126          * and mark them readable.
1127          * Called under mddev lock, so rcu protection not needed.
1128          */
1129         for (i = 0; i < conf->raid_disks; i++) {
1130                 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
1131                 if (rdev
1132                     && !test_bit(Faulty, &rdev->flags)
1133                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1134                         count++;
1135                         sysfs_notify_dirent_safe(rdev->sysfs_state);
1136                 }
1137         }
1138         spin_lock_irqsave(&conf->device_lock, flags);
1139         mddev->degraded -= count;
1140         spin_unlock_irqrestore(&conf->device_lock, flags);
1141
1142         print_conf(conf);
1143         return count;
1144 }
1145
1146
1147 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1148 {
1149         conf_t *conf = mddev->private;
1150         int err = -EEXIST;
1151         int mirror = 0;
1152         mirror_info_t *p;
1153         int first = 0;
1154         int last = mddev->raid_disks - 1;
1155
1156         if (mddev->recovery_disabled == conf->recovery_disabled)
1157                 return -EBUSY;
1158
1159         if (rdev->raid_disk >= 0)
1160                 first = last = rdev->raid_disk;
1161
1162         for (mirror = first; mirror <= last; mirror++)
1163                 if ( !(p=conf->mirrors+mirror)->rdev) {
1164
1165                         disk_stack_limits(mddev->gendisk, rdev->bdev,
1166                                           rdev->data_offset << 9);
1167                         /* as we don't honour merge_bvec_fn, we must
1168                          * never risk violating it, so limit
1169                          * ->max_segments to one lying with a single
1170                          * page, as a one page request is never in
1171                          * violation.
1172                          */
1173                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
1174                                 blk_queue_max_segments(mddev->queue, 1);
1175                                 blk_queue_segment_boundary(mddev->queue,
1176                                                            PAGE_CACHE_SIZE - 1);
1177                         }
1178
1179                         p->head_position = 0;
1180                         rdev->raid_disk = mirror;
1181                         err = 0;
1182                         /* As all devices are equivalent, we don't need a full recovery
1183                          * if this was recently any drive of the array
1184                          */
1185                         if (rdev->saved_raid_disk < 0)
1186                                 conf->fullsync = 1;
1187                         rcu_assign_pointer(p->rdev, rdev);
1188                         break;
1189                 }
1190         md_integrity_add_rdev(rdev, mddev);
1191         print_conf(conf);
1192         return err;
1193 }
1194
1195 static int raid1_remove_disk(mddev_t *mddev, int number)
1196 {
1197         conf_t *conf = mddev->private;
1198         int err = 0;
1199         mdk_rdev_t *rdev;
1200         mirror_info_t *p = conf->mirrors+ number;
1201
1202         print_conf(conf);
1203         rdev = p->rdev;
1204         if (rdev) {
1205                 if (test_bit(In_sync, &rdev->flags) ||
1206                     atomic_read(&rdev->nr_pending)) {
1207                         err = -EBUSY;
1208                         goto abort;
1209                 }
1210                 /* Only remove non-faulty devices if recovery
1211                  * is not possible.
1212                  */
1213                 if (!test_bit(Faulty, &rdev->flags) &&
1214                     mddev->recovery_disabled != conf->recovery_disabled &&
1215                     mddev->degraded < conf->raid_disks) {
1216                         err = -EBUSY;
1217                         goto abort;
1218                 }
1219                 p->rdev = NULL;
1220                 synchronize_rcu();
1221                 if (atomic_read(&rdev->nr_pending)) {
1222                         /* lost the race, try later */
1223                         err = -EBUSY;
1224                         p->rdev = rdev;
1225                         goto abort;
1226                 }
1227                 err = md_integrity_register(mddev);
1228         }
1229 abort:
1230
1231         print_conf(conf);
1232         return err;
1233 }
1234
1235
1236 static void end_sync_read(struct bio *bio, int error)
1237 {
1238         r1bio_t *r1_bio = bio->bi_private;
1239         int i;
1240
1241         for (i=r1_bio->mddev->raid_disks; i--; )
1242                 if (r1_bio->bios[i] == bio)
1243                         break;
1244         BUG_ON(i < 0);
1245         update_head_pos(i, r1_bio);
1246         /*
1247          * we have read a block, now it needs to be re-written,
1248          * or re-read if the read failed.
1249          * We don't do much here, just schedule handling by raid1d
1250          */
1251         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1252                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1253
1254         if (atomic_dec_and_test(&r1_bio->remaining))
1255                 reschedule_retry(r1_bio);
1256 }
1257
1258 static void end_sync_write(struct bio *bio, int error)
1259 {
1260         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1261         r1bio_t *r1_bio = bio->bi_private;
1262         mddev_t *mddev = r1_bio->mddev;
1263         conf_t *conf = mddev->private;
1264         int i;
1265         int mirror=0;
1266
1267         for (i = 0; i < conf->raid_disks; i++)
1268                 if (r1_bio->bios[i] == bio) {
1269                         mirror = i;
1270                         break;
1271                 }
1272         if (!uptodate) {
1273                 sector_t sync_blocks = 0;
1274                 sector_t s = r1_bio->sector;
1275                 long sectors_to_go = r1_bio->sectors;
1276                 /* make sure these bits doesn't get cleared. */
1277                 do {
1278                         bitmap_end_sync(mddev->bitmap, s,
1279                                         &sync_blocks, 1);
1280                         s += sync_blocks;
1281                         sectors_to_go -= sync_blocks;
1282                 } while (sectors_to_go > 0);
1283                 md_error(mddev, conf->mirrors[mirror].rdev);
1284         }
1285
1286         update_head_pos(mirror, r1_bio);
1287
1288         if (atomic_dec_and_test(&r1_bio->remaining)) {
1289                 sector_t s = r1_bio->sectors;
1290                 put_buf(r1_bio);
1291                 md_done_sync(mddev, s, uptodate);
1292         }
1293 }
1294
1295 static int fix_sync_read_error(r1bio_t *r1_bio)
1296 {
1297         /* Try some synchronous reads of other devices to get
1298          * good data, much like with normal read errors.  Only
1299          * read into the pages we already have so we don't
1300          * need to re-issue the read request.
1301          * We don't need to freeze the array, because being in an
1302          * active sync request, there is no normal IO, and
1303          * no overlapping syncs.
1304          * We don't need to check is_badblock() again as we
1305          * made sure that anything with a bad block in range
1306          * will have bi_end_io clear.
1307          */
1308         mddev_t *mddev = r1_bio->mddev;
1309         conf_t *conf = mddev->private;
1310         struct bio *bio = r1_bio->bios[r1_bio->read_disk];
1311         sector_t sect = r1_bio->sector;
1312         int sectors = r1_bio->sectors;
1313         int idx = 0;
1314
1315         while(sectors) {
1316                 int s = sectors;
1317                 int d = r1_bio->read_disk;
1318                 int success = 0;
1319                 mdk_rdev_t *rdev;
1320                 int start;
1321
1322                 if (s > (PAGE_SIZE>>9))
1323                         s = PAGE_SIZE >> 9;
1324                 do {
1325                         if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1326                                 /* No rcu protection needed here devices
1327                                  * can only be removed when no resync is
1328                                  * active, and resync is currently active
1329                                  */
1330                                 rdev = conf->mirrors[d].rdev;
1331                                 if (sync_page_io(rdev, sect, s<<9,
1332                                                  bio->bi_io_vec[idx].bv_page,
1333                                                  READ, false)) {
1334                                         success = 1;
1335                                         break;
1336                                 }
1337                         }
1338                         d++;
1339                         if (d == conf->raid_disks)
1340                                 d = 0;
1341                 } while (!success && d != r1_bio->read_disk);
1342
1343                 if (!success) {
1344                         char b[BDEVNAME_SIZE];
1345                         /* Cannot read from anywhere, array is toast */
1346                         md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1347                         printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O read error"
1348                                " for block %llu\n",
1349                                mdname(mddev),
1350                                bdevname(bio->bi_bdev, b),
1351                                (unsigned long long)r1_bio->sector);
1352                         md_done_sync(mddev, r1_bio->sectors, 0);
1353                         put_buf(r1_bio);
1354                         return 0;
1355                 }
1356
1357                 start = d;
1358                 /* write it back and re-read */
1359                 while (d != r1_bio->read_disk) {
1360                         if (d == 0)
1361                                 d = conf->raid_disks;
1362                         d--;
1363                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1364                                 continue;
1365                         rdev = conf->mirrors[d].rdev;
1366                         if (sync_page_io(rdev, sect, s<<9,
1367                                          bio->bi_io_vec[idx].bv_page,
1368                                          WRITE, false) == 0) {
1369                                 r1_bio->bios[d]->bi_end_io = NULL;
1370                                 rdev_dec_pending(rdev, mddev);
1371                                 md_error(mddev, rdev);
1372                         }
1373                 }
1374                 d = start;
1375                 while (d != r1_bio->read_disk) {
1376                         if (d == 0)
1377                                 d = conf->raid_disks;
1378                         d--;
1379                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1380                                 continue;
1381                         rdev = conf->mirrors[d].rdev;
1382                         if (sync_page_io(rdev, sect, s<<9,
1383                                          bio->bi_io_vec[idx].bv_page,
1384                                          READ, false) == 0)
1385                                 md_error(mddev, rdev);
1386                         else
1387                                 atomic_add(s, &rdev->corrected_errors);
1388                 }
1389                 sectors -= s;
1390                 sect += s;
1391                 idx ++;
1392         }
1393         set_bit(R1BIO_Uptodate, &r1_bio->state);
1394         set_bit(BIO_UPTODATE, &bio->bi_flags);
1395         return 1;
1396 }
1397
1398 static int process_checks(r1bio_t *r1_bio)
1399 {
1400         /* We have read all readable devices.  If we haven't
1401          * got the block, then there is no hope left.
1402          * If we have, then we want to do a comparison
1403          * and skip the write if everything is the same.
1404          * If any blocks failed to read, then we need to
1405          * attempt an over-write
1406          */
1407         mddev_t *mddev = r1_bio->mddev;
1408         conf_t *conf = mddev->private;
1409         int primary;
1410         int i;
1411
1412         for (primary = 0; primary < conf->raid_disks; primary++)
1413                 if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1414                     test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1415                         r1_bio->bios[primary]->bi_end_io = NULL;
1416                         rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1417                         break;
1418                 }
1419         r1_bio->read_disk = primary;
1420         for (i = 0; i < conf->raid_disks; i++) {
1421                 int j;
1422                 int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1423                 struct bio *pbio = r1_bio->bios[primary];
1424                 struct bio *sbio = r1_bio->bios[i];
1425                 int size;
1426
1427                 if (r1_bio->bios[i]->bi_end_io != end_sync_read)
1428                         continue;
1429
1430                 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1431                         for (j = vcnt; j-- ; ) {
1432                                 struct page *p, *s;
1433                                 p = pbio->bi_io_vec[j].bv_page;
1434                                 s = sbio->bi_io_vec[j].bv_page;
1435                                 if (memcmp(page_address(p),
1436                                            page_address(s),
1437                                            PAGE_SIZE))
1438                                         break;
1439                         }
1440                 } else
1441                         j = 0;
1442                 if (j >= 0)
1443                         mddev->resync_mismatches += r1_bio->sectors;
1444                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1445                               && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1446                         /* No need to write to this device. */
1447                         sbio->bi_end_io = NULL;
1448                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1449                         continue;
1450                 }
1451                 /* fixup the bio for reuse */
1452                 sbio->bi_vcnt = vcnt;
1453                 sbio->bi_size = r1_bio->sectors << 9;
1454                 sbio->bi_idx = 0;
1455                 sbio->bi_phys_segments = 0;
1456                 sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1457                 sbio->bi_flags |= 1 << BIO_UPTODATE;
1458                 sbio->bi_next = NULL;
1459                 sbio->bi_sector = r1_bio->sector +
1460                         conf->mirrors[i].rdev->data_offset;
1461                 sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1462                 size = sbio->bi_size;
1463                 for (j = 0; j < vcnt ; j++) {
1464                         struct bio_vec *bi;
1465                         bi = &sbio->bi_io_vec[j];
1466                         bi->bv_offset = 0;
1467                         if (size > PAGE_SIZE)
1468                                 bi->bv_len = PAGE_SIZE;
1469                         else
1470                                 bi->bv_len = size;
1471                         size -= PAGE_SIZE;
1472                         memcpy(page_address(bi->bv_page),
1473                                page_address(pbio->bi_io_vec[j].bv_page),
1474                                PAGE_SIZE);
1475                 }
1476         }
1477         return 0;
1478 }
1479
1480 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
1481 {
1482         conf_t *conf = mddev->private;
1483         int i;
1484         int disks = conf->raid_disks;
1485         struct bio *bio, *wbio;
1486
1487         bio = r1_bio->bios[r1_bio->read_disk];
1488
1489         if (!test_bit(R1BIO_Uptodate, &r1_bio->state))
1490                 /* ouch - failed to read all of that. */
1491                 if (!fix_sync_read_error(r1_bio))
1492                         return;
1493
1494         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1495                 if (process_checks(r1_bio) < 0)
1496                         return;
1497         /*
1498          * schedule writes
1499          */
1500         atomic_set(&r1_bio->remaining, 1);
1501         for (i = 0; i < disks ; i++) {
1502                 wbio = r1_bio->bios[i];
1503                 if (wbio->bi_end_io == NULL ||
1504                     (wbio->bi_end_io == end_sync_read &&
1505                      (i == r1_bio->read_disk ||
1506                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1507                         continue;
1508
1509                 wbio->bi_rw = WRITE;
1510                 wbio->bi_end_io = end_sync_write;
1511                 atomic_inc(&r1_bio->remaining);
1512                 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1513
1514                 generic_make_request(wbio);
1515         }
1516
1517         if (atomic_dec_and_test(&r1_bio->remaining)) {
1518                 /* if we're here, all write(s) have completed, so clean up */
1519                 md_done_sync(mddev, r1_bio->sectors, 1);
1520                 put_buf(r1_bio);
1521         }
1522 }
1523
1524 /*
1525  * This is a kernel thread which:
1526  *
1527  *      1.      Retries failed read operations on working mirrors.
1528  *      2.      Updates the raid superblock when problems encounter.
1529  *      3.      Performs writes following reads for array synchronising.
1530  */
1531
1532 static void fix_read_error(conf_t *conf, int read_disk,
1533                            sector_t sect, int sectors)
1534 {
1535         mddev_t *mddev = conf->mddev;
1536         while(sectors) {
1537                 int s = sectors;
1538                 int d = read_disk;
1539                 int success = 0;
1540                 int start;
1541                 mdk_rdev_t *rdev;
1542
1543                 if (s > (PAGE_SIZE>>9))
1544                         s = PAGE_SIZE >> 9;
1545
1546                 do {
1547                         /* Note: no rcu protection needed here
1548                          * as this is synchronous in the raid1d thread
1549                          * which is the thread that might remove
1550                          * a device.  If raid1d ever becomes multi-threaded....
1551                          */
1552                         sector_t first_bad;
1553                         int bad_sectors;
1554
1555                         rdev = conf->mirrors[d].rdev;
1556                         if (rdev &&
1557                             test_bit(In_sync, &rdev->flags) &&
1558                             is_badblock(rdev, sect, s,
1559                                         &first_bad, &bad_sectors) == 0 &&
1560                             sync_page_io(rdev, sect, s<<9,
1561                                          conf->tmppage, READ, false))
1562                                 success = 1;
1563                         else {
1564                                 d++;
1565                                 if (d == conf->raid_disks)
1566                                         d = 0;
1567                         }
1568                 } while (!success && d != read_disk);
1569
1570                 if (!success) {
1571                         /* Cannot read from anywhere -- bye bye array */
1572                         md_error(mddev, conf->mirrors[read_disk].rdev);
1573                         break;
1574                 }
1575                 /* write it back and re-read */
1576                 start = d;
1577                 while (d != read_disk) {
1578                         if (d==0)
1579                                 d = conf->raid_disks;
1580                         d--;
1581                         rdev = conf->mirrors[d].rdev;
1582                         if (rdev &&
1583                             test_bit(In_sync, &rdev->flags)) {
1584                                 if (sync_page_io(rdev, sect, s<<9,
1585                                                  conf->tmppage, WRITE, false)
1586                                     == 0)
1587                                         /* Well, this device is dead */
1588                                         md_error(mddev, rdev);
1589                         }
1590                 }
1591                 d = start;
1592                 while (d != read_disk) {
1593                         char b[BDEVNAME_SIZE];
1594                         if (d==0)
1595                                 d = conf->raid_disks;
1596                         d--;
1597                         rdev = conf->mirrors[d].rdev;
1598                         if (rdev &&
1599                             test_bit(In_sync, &rdev->flags)) {
1600                                 if (sync_page_io(rdev, sect, s<<9,
1601                                                  conf->tmppage, READ, false)
1602                                     == 0)
1603                                         /* Well, this device is dead */
1604                                         md_error(mddev, rdev);
1605                                 else {
1606                                         atomic_add(s, &rdev->corrected_errors);
1607                                         printk(KERN_INFO
1608                                                "md/raid1:%s: read error corrected "
1609                                                "(%d sectors at %llu on %s)\n",
1610                                                mdname(mddev), s,
1611                                                (unsigned long long)(sect +
1612                                                    rdev->data_offset),
1613                                                bdevname(rdev->bdev, b));
1614                                 }
1615                         }
1616                 }
1617                 sectors -= s;
1618                 sect += s;
1619         }
1620 }
1621
1622 static void raid1d(mddev_t *mddev)
1623 {
1624         r1bio_t *r1_bio;
1625         struct bio *bio;
1626         unsigned long flags;
1627         conf_t *conf = mddev->private;
1628         struct list_head *head = &conf->retry_list;
1629         mdk_rdev_t *rdev;
1630         struct blk_plug plug;
1631
1632         md_check_recovery(mddev);
1633
1634         blk_start_plug(&plug);
1635         for (;;) {
1636                 char b[BDEVNAME_SIZE];
1637
1638                 if (atomic_read(&mddev->plug_cnt) == 0)
1639                         flush_pending_writes(conf);
1640
1641                 spin_lock_irqsave(&conf->device_lock, flags);
1642                 if (list_empty(head)) {
1643                         spin_unlock_irqrestore(&conf->device_lock, flags);
1644                         break;
1645                 }
1646                 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
1647                 list_del(head->prev);
1648                 conf->nr_queued--;
1649                 spin_unlock_irqrestore(&conf->device_lock, flags);
1650
1651                 mddev = r1_bio->mddev;
1652                 conf = mddev->private;
1653                 if (test_bit(R1BIO_IsSync, &r1_bio->state))
1654                         sync_request_write(mddev, r1_bio);
1655                 else if (test_bit(R1BIO_ReadError, &r1_bio->state)) {
1656                         int disk;
1657                         int max_sectors;
1658
1659                         clear_bit(R1BIO_ReadError, &r1_bio->state);
1660                         /* we got a read error. Maybe the drive is bad.  Maybe just
1661                          * the block and we can fix it.
1662                          * We freeze all other IO, and try reading the block from
1663                          * other devices.  When we find one, we re-write
1664                          * and check it that fixes the read error.
1665                          * This is all done synchronously while the array is
1666                          * frozen
1667                          */
1668                         if (mddev->ro == 0) {
1669                                 freeze_array(conf);
1670                                 fix_read_error(conf, r1_bio->read_disk,
1671                                                r1_bio->sector,
1672                                                r1_bio->sectors);
1673                                 unfreeze_array(conf);
1674                         } else
1675                                 md_error(mddev,
1676                                          conf->mirrors[r1_bio->read_disk].rdev);
1677
1678                         bio = r1_bio->bios[r1_bio->read_disk];
1679                         bdevname(bio->bi_bdev, b);
1680 read_more:
1681                         disk = read_balance(conf, r1_bio, &max_sectors);
1682                         if (disk == -1) {
1683                                 printk(KERN_ALERT "md/raid1:%s: %s: unrecoverable I/O"
1684                                        " read error for block %llu\n",
1685                                        mdname(mddev), b,
1686                                        (unsigned long long)r1_bio->sector);
1687                                 raid_end_bio_io(r1_bio);
1688                         } else {
1689                                 const unsigned long do_sync = r1_bio->master_bio->bi_rw & REQ_SYNC;
1690                                 if (bio) {
1691                                         r1_bio->bios[r1_bio->read_disk] =
1692                                                 mddev->ro ? IO_BLOCKED : NULL;
1693                                         bio_put(bio);
1694                                 }
1695                                 r1_bio->read_disk = disk;
1696                                 bio = bio_clone_mddev(r1_bio->master_bio,
1697                                                       GFP_NOIO, mddev);
1698                                 md_trim_bio(bio,
1699                                             r1_bio->sector - bio->bi_sector,
1700                                             max_sectors);
1701                                 r1_bio->bios[r1_bio->read_disk] = bio;
1702                                 rdev = conf->mirrors[disk].rdev;
1703                                 printk_ratelimited(
1704                                         KERN_ERR
1705                                         "md/raid1:%s: redirecting sector %llu"
1706                                         " to other mirror: %s\n",
1707                                         mdname(mddev),
1708                                         (unsigned long long)r1_bio->sector,
1709                                         bdevname(rdev->bdev, b));
1710                                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
1711                                 bio->bi_bdev = rdev->bdev;
1712                                 bio->bi_end_io = raid1_end_read_request;
1713                                 bio->bi_rw = READ | do_sync;
1714                                 bio->bi_private = r1_bio;
1715                                 if (max_sectors < r1_bio->sectors) {
1716                                         /* Drat - have to split this up more */
1717                                         struct bio *mbio = r1_bio->master_bio;
1718                                         int sectors_handled =
1719                                                 r1_bio->sector + max_sectors
1720                                                 - mbio->bi_sector;
1721                                         r1_bio->sectors = max_sectors;
1722                                         spin_lock_irq(&conf->device_lock);
1723                                         if (mbio->bi_phys_segments == 0)
1724                                                 mbio->bi_phys_segments = 2;
1725                                         else
1726                                                 mbio->bi_phys_segments++;
1727                                         spin_unlock_irq(&conf->device_lock);
1728                                         generic_make_request(bio);
1729                                         bio = NULL;
1730
1731                                         r1_bio = mempool_alloc(conf->r1bio_pool,
1732                                                                GFP_NOIO);
1733
1734                                         r1_bio->master_bio = mbio;
1735                                         r1_bio->sectors = (mbio->bi_size >> 9)
1736                                                 - sectors_handled;
1737                                         r1_bio->state = 0;
1738                                         set_bit(R1BIO_ReadError,
1739                                                 &r1_bio->state);
1740                                         r1_bio->mddev = mddev;
1741                                         r1_bio->sector = mbio->bi_sector
1742                                                 + sectors_handled;
1743
1744                                         goto read_more;
1745                                 } else
1746                                         generic_make_request(bio);
1747                         }
1748                 } else {
1749                         /* just a partial read to be scheduled from separate
1750                          * context
1751                          */
1752                         generic_make_request(r1_bio->bios[r1_bio->read_disk]);
1753                 }
1754                 cond_resched();
1755                 if (mddev->flags & ~(1<<MD_CHANGE_PENDING))
1756                         md_check_recovery(mddev);
1757         }
1758         blk_finish_plug(&plug);
1759 }
1760
1761
1762 static int init_resync(conf_t *conf)
1763 {
1764         int buffs;
1765
1766         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1767         BUG_ON(conf->r1buf_pool);
1768         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
1769                                           conf->poolinfo);
1770         if (!conf->r1buf_pool)
1771                 return -ENOMEM;
1772         conf->next_resync = 0;
1773         return 0;
1774 }
1775
1776 /*
1777  * perform a "sync" on one "block"
1778  *
1779  * We need to make sure that no normal I/O request - particularly write
1780  * requests - conflict with active sync requests.
1781  *
1782  * This is achieved by tracking pending requests and a 'barrier' concept
1783  * that can be installed to exclude normal IO requests.
1784  */
1785
1786 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1787 {
1788         conf_t *conf = mddev->private;
1789         r1bio_t *r1_bio;
1790         struct bio *bio;
1791         sector_t max_sector, nr_sectors;
1792         int disk = -1;
1793         int i;
1794         int wonly = -1;
1795         int write_targets = 0, read_targets = 0;
1796         sector_t sync_blocks;
1797         int still_degraded = 0;
1798         int good_sectors = RESYNC_SECTORS;
1799         int min_bad = 0; /* number of sectors that are bad in all devices */
1800
1801         if (!conf->r1buf_pool)
1802                 if (init_resync(conf))
1803                         return 0;
1804
1805         max_sector = mddev->dev_sectors;
1806         if (sector_nr >= max_sector) {
1807                 /* If we aborted, we need to abort the
1808                  * sync on the 'current' bitmap chunk (there will
1809                  * only be one in raid1 resync.
1810                  * We can find the current addess in mddev->curr_resync
1811                  */
1812                 if (mddev->curr_resync < max_sector) /* aborted */
1813                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1814                                                 &sync_blocks, 1);
1815                 else /* completed sync */
1816                         conf->fullsync = 0;
1817
1818                 bitmap_close_sync(mddev->bitmap);
1819                 close_sync(conf);
1820                 return 0;
1821         }
1822
1823         if (mddev->bitmap == NULL &&
1824             mddev->recovery_cp == MaxSector &&
1825             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
1826             conf->fullsync == 0) {
1827                 *skipped = 1;
1828                 return max_sector - sector_nr;
1829         }
1830         /* before building a request, check if we can skip these blocks..
1831          * This call the bitmap_start_sync doesn't actually record anything
1832          */
1833         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1834             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1835                 /* We can skip this block, and probably several more */
1836                 *skipped = 1;
1837                 return sync_blocks;
1838         }
1839         /*
1840          * If there is non-resync activity waiting for a turn,
1841          * and resync is going fast enough,
1842          * then let it though before starting on this new sync request.
1843          */
1844         if (!go_faster && conf->nr_waiting)
1845                 msleep_interruptible(1000);
1846
1847         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1848         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1849         raise_barrier(conf);
1850
1851         conf->next_resync = sector_nr;
1852
1853         rcu_read_lock();
1854         /*
1855          * If we get a correctably read error during resync or recovery,
1856          * we might want to read from a different device.  So we
1857          * flag all drives that could conceivably be read from for READ,
1858          * and any others (which will be non-In_sync devices) for WRITE.
1859          * If a read fails, we try reading from something else for which READ
1860          * is OK.
1861          */
1862
1863         r1_bio->mddev = mddev;
1864         r1_bio->sector = sector_nr;
1865         r1_bio->state = 0;
1866         set_bit(R1BIO_IsSync, &r1_bio->state);
1867
1868         for (i=0; i < conf->raid_disks; i++) {
1869                 mdk_rdev_t *rdev;
1870                 bio = r1_bio->bios[i];
1871
1872                 /* take from bio_init */
1873                 bio->bi_next = NULL;
1874                 bio->bi_flags &= ~(BIO_POOL_MASK-1);
1875                 bio->bi_flags |= 1 << BIO_UPTODATE;
1876                 bio->bi_comp_cpu = -1;
1877                 bio->bi_rw = READ;
1878                 bio->bi_vcnt = 0;
1879                 bio->bi_idx = 0;
1880                 bio->bi_phys_segments = 0;
1881                 bio->bi_size = 0;
1882                 bio->bi_end_io = NULL;
1883                 bio->bi_private = NULL;
1884
1885                 rdev = rcu_dereference(conf->mirrors[i].rdev);
1886                 if (rdev == NULL ||
1887                     test_bit(Faulty, &rdev->flags)) {
1888                         still_degraded = 1;
1889                 } else if (!test_bit(In_sync, &rdev->flags)) {
1890                         bio->bi_rw = WRITE;
1891                         bio->bi_end_io = end_sync_write;
1892                         write_targets ++;
1893                 } else {
1894                         /* may need to read from here */
1895                         sector_t first_bad = MaxSector;
1896                         int bad_sectors;
1897
1898                         if (is_badblock(rdev, sector_nr, good_sectors,
1899                                         &first_bad, &bad_sectors)) {
1900                                 if (first_bad > sector_nr)
1901                                         good_sectors = first_bad - sector_nr;
1902                                 else {
1903                                         bad_sectors -= (sector_nr - first_bad);
1904                                         if (min_bad == 0 ||
1905                                             min_bad > bad_sectors)
1906                                                 min_bad = bad_sectors;
1907                                 }
1908                         }
1909                         if (sector_nr < first_bad) {
1910                                 if (test_bit(WriteMostly, &rdev->flags)) {
1911                                         if (wonly < 0)
1912                                                 wonly = i;
1913                                 } else {
1914                                         if (disk < 0)
1915                                                 disk = i;
1916                                 }
1917                                 bio->bi_rw = READ;
1918                                 bio->bi_end_io = end_sync_read;
1919                                 read_targets++;
1920                         }
1921                 }
1922                 if (bio->bi_end_io) {
1923                         atomic_inc(&rdev->nr_pending);
1924                         bio->bi_sector = sector_nr + rdev->data_offset;
1925                         bio->bi_bdev = rdev->bdev;
1926                         bio->bi_private = r1_bio;
1927                 }
1928         }
1929         rcu_read_unlock();
1930         if (disk < 0)
1931                 disk = wonly;
1932         r1_bio->read_disk = disk;
1933
1934         if (read_targets == 0 && min_bad > 0) {
1935                 /* These sectors are bad on all InSync devices, so we
1936                  * need to mark them bad on all write targets
1937                  */
1938                 int ok = 1;
1939                 for (i = 0 ; i < conf->raid_disks ; i++)
1940                         if (r1_bio->bios[i]->bi_end_io == end_sync_write) {
1941                                 mdk_rdev_t *rdev =
1942                                         rcu_dereference(conf->mirrors[i].rdev);
1943                                 ok = rdev_set_badblocks(rdev, sector_nr,
1944                                                         min_bad, 0
1945                                         ) && ok;
1946                         }
1947                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
1948                 *skipped = 1;
1949                 put_buf(r1_bio);
1950
1951                 if (!ok) {
1952                         /* Cannot record the badblocks, so need to
1953                          * abort the resync.
1954                          * If there are multiple read targets, could just
1955                          * fail the really bad ones ???
1956                          */
1957                         conf->recovery_disabled = mddev->recovery_disabled;
1958                         set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1959                         return 0;
1960                 } else
1961                         return min_bad;
1962
1963         }
1964         if (min_bad > 0 && min_bad < good_sectors) {
1965                 /* only resync enough to reach the next bad->good
1966                  * transition */
1967                 good_sectors = min_bad;
1968         }
1969
1970         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
1971                 /* extra read targets are also write targets */
1972                 write_targets += read_targets-1;
1973
1974         if (write_targets == 0 || read_targets == 0) {
1975                 /* There is nowhere to write, so all non-sync
1976                  * drives must be failed - so we are finished
1977                  */
1978                 sector_t rv = max_sector - sector_nr;
1979                 *skipped = 1;
1980                 put_buf(r1_bio);
1981                 return rv;
1982         }
1983
1984         if (max_sector > mddev->resync_max)
1985                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1986         if (max_sector > sector_nr + good_sectors)
1987                 max_sector = sector_nr + good_sectors;
1988         nr_sectors = 0;
1989         sync_blocks = 0;
1990         do {
1991                 struct page *page;
1992                 int len = PAGE_SIZE;
1993                 if (sector_nr + (len>>9) > max_sector)
1994                         len = (max_sector - sector_nr) << 9;
1995                 if (len == 0)
1996                         break;
1997                 if (sync_blocks == 0) {
1998                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1999                                                &sync_blocks, still_degraded) &&
2000                             !conf->fullsync &&
2001                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
2002                                 break;
2003                         BUG_ON(sync_blocks < (PAGE_SIZE>>9));
2004                         if ((len >> 9) > sync_blocks)
2005                                 len = sync_blocks<<9;
2006                 }
2007
2008                 for (i=0 ; i < conf->raid_disks; i++) {
2009                         bio = r1_bio->bios[i];
2010                         if (bio->bi_end_io) {
2011                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
2012                                 if (bio_add_page(bio, page, len, 0) == 0) {
2013                                         /* stop here */
2014                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
2015                                         while (i > 0) {
2016                                                 i--;
2017                                                 bio = r1_bio->bios[i];
2018                                                 if (bio->bi_end_io==NULL)
2019                                                         continue;
2020                                                 /* remove last page from this bio */
2021                                                 bio->bi_vcnt--;
2022                                                 bio->bi_size -= len;
2023                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
2024                                         }
2025                                         goto bio_full;
2026                                 }
2027                         }
2028                 }
2029                 nr_sectors += len>>9;
2030                 sector_nr += len>>9;
2031                 sync_blocks -= (len>>9);
2032         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
2033  bio_full:
2034         r1_bio->sectors = nr_sectors;
2035
2036         /* For a user-requested sync, we read all readable devices and do a
2037          * compare
2038          */
2039         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
2040                 atomic_set(&r1_bio->remaining, read_targets);
2041                 for (i=0; i<conf->raid_disks; i++) {
2042                         bio = r1_bio->bios[i];
2043                         if (bio->bi_end_io == end_sync_read) {
2044                                 md_sync_acct(bio->bi_bdev, nr_sectors);
2045                                 generic_make_request(bio);
2046                         }
2047                 }
2048         } else {
2049                 atomic_set(&r1_bio->remaining, 1);
2050                 bio = r1_bio->bios[r1_bio->read_disk];
2051                 md_sync_acct(bio->bi_bdev, nr_sectors);
2052                 generic_make_request(bio);
2053
2054         }
2055         return nr_sectors;
2056 }
2057
2058 static sector_t raid1_size(mddev_t *mddev, sector_t sectors, int raid_disks)
2059 {
2060         if (sectors)
2061                 return sectors;
2062
2063         return mddev->dev_sectors;
2064 }
2065
2066 static conf_t *setup_conf(mddev_t *mddev)
2067 {
2068         conf_t *conf;
2069         int i;
2070         mirror_info_t *disk;
2071         mdk_rdev_t *rdev;
2072         int err = -ENOMEM;
2073
2074         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
2075         if (!conf)
2076                 goto abort;
2077
2078         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
2079                                  GFP_KERNEL);
2080         if (!conf->mirrors)
2081                 goto abort;
2082
2083         conf->tmppage = alloc_page(GFP_KERNEL);
2084         if (!conf->tmppage)
2085                 goto abort;
2086
2087         conf->poolinfo = kzalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
2088         if (!conf->poolinfo)
2089                 goto abort;
2090         conf->poolinfo->raid_disks = mddev->raid_disks;
2091         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2092                                           r1bio_pool_free,
2093                                           conf->poolinfo);
2094         if (!conf->r1bio_pool)
2095                 goto abort;
2096
2097         conf->poolinfo->mddev = mddev;
2098
2099         spin_lock_init(&conf->device_lock);
2100         list_for_each_entry(rdev, &mddev->disks, same_set) {
2101                 int disk_idx = rdev->raid_disk;
2102                 if (disk_idx >= mddev->raid_disks
2103                     || disk_idx < 0)
2104                         continue;
2105                 disk = conf->mirrors + disk_idx;
2106
2107                 disk->rdev = rdev;
2108
2109                 disk->head_position = 0;
2110         }
2111         conf->raid_disks = mddev->raid_disks;
2112         conf->mddev = mddev;
2113         INIT_LIST_HEAD(&conf->retry_list);
2114
2115         spin_lock_init(&conf->resync_lock);
2116         init_waitqueue_head(&conf->wait_barrier);
2117
2118         bio_list_init(&conf->pending_bio_list);
2119
2120         conf->last_used = -1;
2121         for (i = 0; i < conf->raid_disks; i++) {
2122
2123                 disk = conf->mirrors + i;
2124
2125                 if (!disk->rdev ||
2126                     !test_bit(In_sync, &disk->rdev->flags)) {
2127                         disk->head_position = 0;
2128                         if (disk->rdev)
2129                                 conf->fullsync = 1;
2130                 } else if (conf->last_used < 0)
2131                         /*
2132                          * The first working device is used as a
2133                          * starting point to read balancing.
2134                          */
2135                         conf->last_used = i;
2136         }
2137
2138         err = -EIO;
2139         if (conf->last_used < 0) {
2140                 printk(KERN_ERR "md/raid1:%s: no operational mirrors\n",
2141                        mdname(mddev));
2142                 goto abort;
2143         }
2144         err = -ENOMEM;
2145         conf->thread = md_register_thread(raid1d, mddev, NULL);
2146         if (!conf->thread) {
2147                 printk(KERN_ERR
2148                        "md/raid1:%s: couldn't allocate thread\n",
2149                        mdname(mddev));
2150                 goto abort;
2151         }
2152
2153         return conf;
2154
2155  abort:
2156         if (conf) {
2157                 if (conf->r1bio_pool)
2158                         mempool_destroy(conf->r1bio_pool);
2159                 kfree(conf->mirrors);
2160                 safe_put_page(conf->tmppage);
2161                 kfree(conf->poolinfo);
2162                 kfree(conf);
2163         }
2164         return ERR_PTR(err);
2165 }
2166
2167 static int run(mddev_t *mddev)
2168 {
2169         conf_t *conf;
2170         int i;
2171         mdk_rdev_t *rdev;
2172
2173         if (mddev->level != 1) {
2174                 printk(KERN_ERR "md/raid1:%s: raid level not set to mirroring (%d)\n",
2175                        mdname(mddev), mddev->level);
2176                 return -EIO;
2177         }
2178         if (mddev->reshape_position != MaxSector) {
2179                 printk(KERN_ERR "md/raid1:%s: reshape_position set but not supported\n",
2180                        mdname(mddev));
2181                 return -EIO;
2182         }
2183         /*
2184          * copy the already verified devices into our private RAID1
2185          * bookkeeping area. [whatever we allocate in run(),
2186          * should be freed in stop()]
2187          */
2188         if (mddev->private == NULL)
2189                 conf = setup_conf(mddev);
2190         else
2191                 conf = mddev->private;
2192
2193         if (IS_ERR(conf))
2194                 return PTR_ERR(conf);
2195
2196         list_for_each_entry(rdev, &mddev->disks, same_set) {
2197                 if (!mddev->gendisk)
2198                         continue;
2199                 disk_stack_limits(mddev->gendisk, rdev->bdev,
2200                                   rdev->data_offset << 9);
2201                 /* as we don't honour merge_bvec_fn, we must never risk
2202                  * violating it, so limit ->max_segments to 1 lying within
2203                  * a single page, as a one page request is never in violation.
2204                  */
2205                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn) {
2206                         blk_queue_max_segments(mddev->queue, 1);
2207                         blk_queue_segment_boundary(mddev->queue,
2208                                                    PAGE_CACHE_SIZE - 1);
2209                 }
2210         }
2211
2212         mddev->degraded = 0;
2213         for (i=0; i < conf->raid_disks; i++)
2214                 if (conf->mirrors[i].rdev == NULL ||
2215                     !test_bit(In_sync, &conf->mirrors[i].rdev->flags) ||
2216                     test_bit(Faulty, &conf->mirrors[i].rdev->flags))
2217                         mddev->degraded++;
2218
2219         if (conf->raid_disks - mddev->degraded == 1)
2220                 mddev->recovery_cp = MaxSector;
2221
2222         if (mddev->recovery_cp != MaxSector)
2223                 printk(KERN_NOTICE "md/raid1:%s: not clean"
2224                        " -- starting background reconstruction\n",
2225                        mdname(mddev));
2226         printk(KERN_INFO
2227                 "md/raid1:%s: active with %d out of %d mirrors\n",
2228                 mdname(mddev), mddev->raid_disks - mddev->degraded,
2229                 mddev->raid_disks);
2230
2231         /*
2232          * Ok, everything is just fine now
2233          */
2234         mddev->thread = conf->thread;
2235         conf->thread = NULL;
2236         mddev->private = conf;
2237
2238         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2239
2240         if (mddev->queue) {
2241                 mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2242                 mddev->queue->backing_dev_info.congested_data = mddev;
2243         }
2244         return md_integrity_register(mddev);
2245 }
2246
2247 static int stop(mddev_t *mddev)
2248 {
2249         conf_t *conf = mddev->private;
2250         struct bitmap *bitmap = mddev->bitmap;
2251
2252         /* wait for behind writes to complete */
2253         if (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2254                 printk(KERN_INFO "md/raid1:%s: behind writes in progress - waiting to stop.\n",
2255                        mdname(mddev));
2256                 /* need to kick something here to make sure I/O goes? */
2257                 wait_event(bitmap->behind_wait,
2258                            atomic_read(&bitmap->behind_writes) == 0);
2259         }
2260
2261         raise_barrier(conf);
2262         lower_barrier(conf);
2263
2264         md_unregister_thread(mddev->thread);
2265         mddev->thread = NULL;
2266         if (conf->r1bio_pool)
2267                 mempool_destroy(conf->r1bio_pool);
2268         kfree(conf->mirrors);
2269         kfree(conf->poolinfo);
2270         kfree(conf);
2271         mddev->private = NULL;
2272         return 0;
2273 }
2274
2275 static int raid1_resize(mddev_t *mddev, sector_t sectors)
2276 {
2277         /* no resync is happening, and there is enough space
2278          * on all devices, so we can resize.
2279          * We need to make sure resync covers any new space.
2280          * If the array is shrinking we should possibly wait until
2281          * any io in the removed space completes, but it hardly seems
2282          * worth it.
2283          */
2284         md_set_array_sectors(mddev, raid1_size(mddev, sectors, 0));
2285         if (mddev->array_sectors > raid1_size(mddev, sectors, 0))
2286                 return -EINVAL;
2287         set_capacity(mddev->gendisk, mddev->array_sectors);
2288         revalidate_disk(mddev->gendisk);
2289         if (sectors > mddev->dev_sectors &&
2290             mddev->recovery_cp > mddev->dev_sectors) {
2291                 mddev->recovery_cp = mddev->dev_sectors;
2292                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2293         }
2294         mddev->dev_sectors = sectors;
2295         mddev->resync_max_sectors = sectors;
2296         return 0;
2297 }
2298
2299 static int raid1_reshape(mddev_t *mddev)
2300 {
2301         /* We need to:
2302          * 1/ resize the r1bio_pool
2303          * 2/ resize conf->mirrors
2304          *
2305          * We allocate a new r1bio_pool if we can.
2306          * Then raise a device barrier and wait until all IO stops.
2307          * Then resize conf->mirrors and swap in the new r1bio pool.
2308          *
2309          * At the same time, we "pack" the devices so that all the missing
2310          * devices have the higher raid_disk numbers.
2311          */
2312         mempool_t *newpool, *oldpool;
2313         struct pool_info *newpoolinfo;
2314         mirror_info_t *newmirrors;
2315         conf_t *conf = mddev->private;
2316         int cnt, raid_disks;
2317         unsigned long flags;
2318         int d, d2, err;
2319
2320         /* Cannot change chunk_size, layout, or level */
2321         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2322             mddev->layout != mddev->new_layout ||
2323             mddev->level != mddev->new_level) {
2324                 mddev->new_chunk_sectors = mddev->chunk_sectors;
2325                 mddev->new_layout = mddev->layout;
2326                 mddev->new_level = mddev->level;
2327                 return -EINVAL;
2328         }
2329
2330         err = md_allow_write(mddev);
2331         if (err)
2332                 return err;
2333
2334         raid_disks = mddev->raid_disks + mddev->delta_disks;
2335
2336         if (raid_disks < conf->raid_disks) {
2337                 cnt=0;
2338                 for (d= 0; d < conf->raid_disks; d++)
2339                         if (conf->mirrors[d].rdev)
2340                                 cnt++;
2341                 if (cnt > raid_disks)
2342                         return -EBUSY;
2343         }
2344
2345         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2346         if (!newpoolinfo)
2347                 return -ENOMEM;
2348         newpoolinfo->mddev = mddev;
2349         newpoolinfo->raid_disks = raid_disks;
2350
2351         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2352                                  r1bio_pool_free, newpoolinfo);
2353         if (!newpool) {
2354                 kfree(newpoolinfo);
2355                 return -ENOMEM;
2356         }
2357         newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
2358         if (!newmirrors) {
2359                 kfree(newpoolinfo);
2360                 mempool_destroy(newpool);
2361                 return -ENOMEM;
2362         }
2363
2364         raise_barrier(conf);
2365
2366         /* ok, everything is stopped */
2367         oldpool = conf->r1bio_pool;
2368         conf->r1bio_pool = newpool;
2369
2370         for (d = d2 = 0; d < conf->raid_disks; d++) {
2371                 mdk_rdev_t *rdev = conf->mirrors[d].rdev;
2372                 if (rdev && rdev->raid_disk != d2) {
2373                         sysfs_unlink_rdev(mddev, rdev);
2374                         rdev->raid_disk = d2;
2375                         sysfs_unlink_rdev(mddev, rdev);
2376                         if (sysfs_link_rdev(mddev, rdev))
2377                                 printk(KERN_WARNING
2378                                        "md/raid1:%s: cannot register rd%d\n",
2379                                        mdname(mddev), rdev->raid_disk);
2380                 }
2381                 if (rdev)
2382                         newmirrors[d2++].rdev = rdev;
2383         }
2384         kfree(conf->mirrors);
2385         conf->mirrors = newmirrors;
2386         kfree(conf->poolinfo);
2387         conf->poolinfo = newpoolinfo;
2388
2389         spin_lock_irqsave(&conf->device_lock, flags);
2390         mddev->degraded += (raid_disks - conf->raid_disks);
2391         spin_unlock_irqrestore(&conf->device_lock, flags);
2392         conf->raid_disks = mddev->raid_disks = raid_disks;
2393         mddev->delta_disks = 0;
2394
2395         conf->last_used = 0; /* just make sure it is in-range */
2396         lower_barrier(conf);
2397
2398         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2399         md_wakeup_thread(mddev->thread);
2400
2401         mempool_destroy(oldpool);
2402         return 0;
2403 }
2404
2405 static void raid1_quiesce(mddev_t *mddev, int state)
2406 {
2407         conf_t *conf = mddev->private;
2408
2409         switch(state) {
2410         case 2: /* wake for suspend */
2411                 wake_up(&conf->wait_barrier);
2412                 break;
2413         case 1:
2414                 raise_barrier(conf);
2415                 break;
2416         case 0:
2417                 lower_barrier(conf);
2418                 break;
2419         }
2420 }
2421
2422 static void *raid1_takeover(mddev_t *mddev)
2423 {
2424         /* raid1 can take over:
2425          *  raid5 with 2 devices, any layout or chunk size
2426          */
2427         if (mddev->level == 5 && mddev->raid_disks == 2) {
2428                 conf_t *conf;
2429                 mddev->new_level = 1;
2430                 mddev->new_layout = 0;
2431                 mddev->new_chunk_sectors = 0;
2432                 conf = setup_conf(mddev);
2433                 if (!IS_ERR(conf))
2434                         conf->barrier = 1;
2435                 return conf;
2436         }
2437         return ERR_PTR(-EINVAL);
2438 }
2439
2440 static struct mdk_personality raid1_personality =
2441 {
2442         .name           = "raid1",
2443         .level          = 1,
2444         .owner          = THIS_MODULE,
2445         .make_request   = make_request,
2446         .run            = run,
2447         .stop           = stop,
2448         .status         = status,
2449         .error_handler  = error,
2450         .hot_add_disk   = raid1_add_disk,
2451         .hot_remove_disk= raid1_remove_disk,
2452         .spare_active   = raid1_spare_active,
2453         .sync_request   = sync_request,
2454         .resize         = raid1_resize,
2455         .size           = raid1_size,
2456         .check_reshape  = raid1_reshape,
2457         .quiesce        = raid1_quiesce,
2458         .takeover       = raid1_takeover,
2459 };
2460
2461 static int __init raid_init(void)
2462 {
2463         return register_md_personality(&raid1_personality);
2464 }
2465
2466 static void raid_exit(void)
2467 {
2468         unregister_md_personality(&raid1_personality);
2469 }
2470
2471 module_init(raid_init);
2472 module_exit(raid_exit);
2473 MODULE_LICENSE("GPL");
2474 MODULE_DESCRIPTION("RAID1 (mirroring) personality for MD");
2475 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2476 MODULE_ALIAS("md-raid1");
2477 MODULE_ALIAS("md-level-1");