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/delay.h>
  35 #include <linux/blkdev.h>
  36 #include <linux/seq_file.h>
  37 #include "md.h"
  38 #include "raid1.h"
  39 #include "bitmap.h"
  40
  41 #define DEBUG 0
  42 #if DEBUG
  43 #define PRINTK(x...) printk(x)
  44 #else
  45 #define PRINTK(x...)
  46 #endif
  47
  48 /*
  49  * Number of guaranteed r1bios in case of extreme VM load:
  50  */
  51 #define NR_RAID1_BIOS 256
  52
  53
  54 static void unplug_slaves(mddev_t *mddev);
  55
  56 static void allow_barrier(conf_t *conf);
  57 static void lower_barrier(conf_t *conf);
  58
  59 static void * r1bio_pool_alloc(gfp_t gfp_flags, void *data)
  60 {
  61         struct pool_info *pi = data;
  62         r1bio_t *r1_bio;
  63         int size = offsetof(r1bio_t, bios[pi->raid_disks]);
  64
  65         /* allocate a r1bio with room for raid_disks entries in the bios array */
  66         r1_bio = kzalloc(size, gfp_flags);
  67         if (!r1_bio)
  68                 unplug_slaves(pi->mddev);
  69
  70         return r1_bio;
  71 }
  72
  73 static void r1bio_pool_free(void *r1_bio, void *data)
  74 {
  75         kfree(r1_bio);
  76 }
  77
  78 #define RESYNC_BLOCK_SIZE (64*1024)
  79 //#define RESYNC_BLOCK_SIZE PAGE_SIZE
  80 #define RESYNC_SECTORS (RESYNC_BLOCK_SIZE >> 9)
  81 #define RESYNC_PAGES ((RESYNC_BLOCK_SIZE + PAGE_SIZE-1) / PAGE_SIZE)
  82 #define RESYNC_WINDOW (2048*1024)
  83
  84 static void * r1buf_pool_alloc(gfp_t gfp_flags, void *data)
  85 {
  86         struct pool_info *pi = data;
  87         struct page *page;
  88         r1bio_t *r1_bio;
  89         struct bio *bio;
  90         int i, j;
  91
  92         r1_bio = r1bio_pool_alloc(gfp_flags, pi);
  93         if (!r1_bio) {
  94                 unplug_slaves(pi->mddev);
  95                 return NULL;
  96         }
  97
  98         /*
  99          * Allocate bios : 1 for reading, n-1 for writing
 100          */
 101         for (j = pi->raid_disks ; j-- ; ) {
 102                 bio = bio_alloc(gfp_flags, RESYNC_PAGES);
 103                 if (!bio)
 104                         goto out_free_bio;
 105                 r1_bio->bios[j] = bio;
 106         }
 107         /*
 108          * Allocate RESYNC_PAGES data pages and attach them to
 109          * the first bio.
 110          * If this is a user-requested check/repair, allocate
 111          * RESYNC_PAGES for each bio.
 112          */
 113         if (test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery))
 114                 j = pi->raid_disks;
 115         else
 116                 j = 1;
 117         while(j--) {
 118                 bio = r1_bio->bios[j];
 119                 for (i = 0; i < RESYNC_PAGES; i++) {
 120                         page = alloc_page(gfp_flags);
 121                         if (unlikely(!page))
 122                                 goto out_free_pages;
 123
 124                         bio->bi_io_vec[i].bv_page = page;
 125                         bio->bi_vcnt = i+1;
 126                 }
 127         }
 128         /* If not user-requests, copy the page pointers to all bios */
 129         if (!test_bit(MD_RECOVERY_REQUESTED, &pi->mddev->recovery)) {
 130                 for (i=0; i<RESYNC_PAGES ; i++)
 131                         for (j=1; j<pi->raid_disks; j++)
 132                                 r1_bio->bios[j]->bi_io_vec[i].bv_page =
 133                                         r1_bio->bios[0]->bi_io_vec[i].bv_page;
 134         }
 135
 136         r1_bio->master_bio = NULL;
 137
 138         return r1_bio;
 139
 140 out_free_pages:
 141         for (j=0 ; j < pi->raid_disks; j++)
 142                 for (i=0; i < r1_bio->bios[j]->bi_vcnt ; i++)
 143                         put_page(r1_bio->bios[j]->bi_io_vec[i].bv_page);
 144         j = -1;
 145 out_free_bio:
 146         while ( ++j < pi->raid_disks )
 147                 bio_put(r1_bio->bios[j]);
 148         r1bio_pool_free(r1_bio, data);
 149         return NULL;
 150 }
 151
 152 static void r1buf_pool_free(void *__r1_bio, void *data)
 153 {
 154         struct pool_info *pi = data;
 155         int i,j;
 156         r1bio_t *r1bio = __r1_bio;
 157
 158         for (i = 0; i < RESYNC_PAGES; i++)
 159                 for (j = pi->raid_disks; j-- ;) {
 160                         if (j == 0 ||
 161                             r1bio->bios[j]->bi_io_vec[i].bv_page !=
 162                             r1bio->bios[0]->bi_io_vec[i].bv_page)
 163                                 safe_put_page(r1bio->bios[j]->bi_io_vec[i].bv_page);
 164                 }
 165         for (i=0 ; i < pi->raid_disks; i++)
 166                 bio_put(r1bio->bios[i]);
 167
 168         r1bio_pool_free(r1bio, data);
 169 }
 170
 171 static void put_all_bios(conf_t *conf, r1bio_t *r1_bio)
 172 {
 173         int i;
 174
 175         for (i = 0; i < conf->raid_disks; i++) {
 176                 struct bio **bio = r1_bio->bios + i;
 177                 if (*bio && *bio != IO_BLOCKED)
 178                         bio_put(*bio);
 179                 *bio = NULL;
 180         }
 181 }
 182
 183 static void free_r1bio(r1bio_t *r1_bio)
 184 {
 185         conf_t *conf = r1_bio->mddev->private;
 186
 187         /*
 188          * Wake up any possible resync thread that waits for the device
 189          * to go idle.
 190          */
 191         allow_barrier(conf);
 192
 193         put_all_bios(conf, r1_bio);
 194         mempool_free(r1_bio, conf->r1bio_pool);
 195 }
 196
 197 static void put_buf(r1bio_t *r1_bio)
 198 {
 199         conf_t *conf = r1_bio->mddev->private;
 200         int i;
 201
 202         for (i=0; i<conf->raid_disks; i++) {
 203                 struct bio *bio = r1_bio->bios[i];
 204                 if (bio->bi_end_io)
 205                         rdev_dec_pending(conf->mirrors[i].rdev, r1_bio->mddev);
 206         }
 207
 208         mempool_free(r1_bio, conf->r1buf_pool);
 209
 210         lower_barrier(conf);
 211 }
 212
 213 static void reschedule_retry(r1bio_t *r1_bio)
 214 {
 215         unsigned long flags;
 216         mddev_t *mddev = r1_bio->mddev;
 217         conf_t *conf = mddev->private;
 218
 219         spin_lock_irqsave(&conf->device_lock, flags);
 220         list_add(&r1_bio->retry_list, &conf->retry_list);
 221         conf->nr_queued ++;
 222         spin_unlock_irqrestore(&conf->device_lock, flags);
 223
 224         wake_up(&conf->wait_barrier);
 225         md_wakeup_thread(mddev->thread);
 226 }
 227
 228 /*
 229  * raid_end_bio_io() is called when we have finished servicing a mirrored
 230  * operation and are ready to return a success/failure code to the buffer
 231  * cache layer.
 232  */
 233 static void raid_end_bio_io(r1bio_t *r1_bio)
 234 {
 235         struct bio *bio = r1_bio->master_bio;
 236
 237         /* if nobody has done the final endio yet, do it now */
 238         if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 239                 PRINTK(KERN_DEBUG "raid1: sync end %s on sectors %llu-%llu\n",
 240                         (bio_data_dir(bio) == WRITE) ? "write" : "read",
 241                         (unsigned long long) bio->bi_sector,
 242                         (unsigned long long) bio->bi_sector +
 243                                 (bio->bi_size >> 9) - 1);
 244
 245                 bio_endio(bio,
 246                         test_bit(R1BIO_Uptodate, &r1_bio->state) ? 0 : -EIO);
 247         }
 248         free_r1bio(r1_bio);
 249 }
 250
 251 /*
 252  * Update disk head position estimator based on IRQ completion info.
 253  */
 254 static inline void update_head_pos(int disk, r1bio_t *r1_bio)
 255 {
 256         conf_t *conf = r1_bio->mddev->private;
 257
 258         conf->mirrors[disk].head_position =
 259                 r1_bio->sector + (r1_bio->sectors);
 260 }
 261
 262 static void raid1_end_read_request(struct bio *bio, int error)
 263 {
 264         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 265         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
 266         int mirror;
 267         conf_t *conf = r1_bio->mddev->private;
 268
 269         mirror = r1_bio->read_disk;
 270         /*
 271          * this branch is our 'one mirror IO has finished' event handler:
 272          */
 273         update_head_pos(mirror, r1_bio);
 274
 275         if (uptodate)
 276                 set_bit(R1BIO_Uptodate, &r1_bio->state);
 277         else {
 278                 /* If all other devices have failed, we want to return
 279                  * the error upwards rather than fail the last device.
 280                  * Here we redefine "uptodate" to mean "Don't want to retry"
 281                  */
 282                 unsigned long flags;
 283                 spin_lock_irqsave(&conf->device_lock, flags);
 284                 if (r1_bio->mddev->degraded == conf->raid_disks ||
 285                     (r1_bio->mddev->degraded == conf->raid_disks-1 &&
 286                      !test_bit(Faulty, &conf->mirrors[mirror].rdev->flags)))
 287                         uptodate = 1;
 288                 spin_unlock_irqrestore(&conf->device_lock, flags);
 289         }
 290
 291         if (uptodate)
 292                 raid_end_bio_io(r1_bio);
 293         else {
 294                 /*
 295                  * oops, read error:
 296                  */
 297                 char b[BDEVNAME_SIZE];
 298                 if (printk_ratelimit())
 299                         printk(KERN_ERR "raid1: %s: rescheduling sector %llu\n",
 300                                bdevname(conf->mirrors[mirror].rdev->bdev,b), (unsigned long long)r1_bio->sector);
 301                 reschedule_retry(r1_bio);
 302         }
 303
 304         rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 305 }
 306
 307 static void raid1_end_write_request(struct bio *bio, int error)
 308 {
 309         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
 310         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
 311         int mirror, behind = test_bit(R1BIO_BehindIO, &r1_bio->state);
 312         conf_t *conf = r1_bio->mddev->private;
 313         struct bio *to_put = NULL;
 314
 315
 316         for (mirror = 0; mirror < conf->raid_disks; mirror++)
 317                 if (r1_bio->bios[mirror] == bio)
 318                         break;
 319
 320         if (error == -EOPNOTSUPP && test_bit(R1BIO_Barrier, &r1_bio->state)) {
 321                 set_bit(BarriersNotsupp, &conf->mirrors[mirror].rdev->flags);
 322                 set_bit(R1BIO_BarrierRetry, &r1_bio->state);
 323                 r1_bio->mddev->barriers_work = 0;
 324                 /* Don't rdev_dec_pending in this branch - keep it for the retry */
 325         } else {
 326                 /*
 327                  * this branch is our 'one mirror IO has finished' event handler:
 328                  */
 329                 r1_bio->bios[mirror] = NULL;
 330                 to_put = bio;
 331                 if (!uptodate) {
 332                         md_error(r1_bio->mddev, conf->mirrors[mirror].rdev);
 333                         /* an I/O failed, we can't clear the bitmap */
 334                         set_bit(R1BIO_Degraded, &r1_bio->state);
 335                 } else
 336                         /*
 337                          * Set R1BIO_Uptodate in our master bio, so that
 338                          * we will return a good error code for to the higher
 339                          * levels even if IO on some other mirrored buffer fails.
 340                          *
 341                          * The 'master' represents the composite IO operation to
 342                          * user-side. So if something waits for IO, then it will
 343                          * wait for the 'master' bio.
 344                          */
 345                         set_bit(R1BIO_Uptodate, &r1_bio->state);
 346
 347                 update_head_pos(mirror, r1_bio);
 348
 349                 if (behind) {
 350                         if (test_bit(WriteMostly, &conf->mirrors[mirror].rdev->flags))
 351                                 atomic_dec(&r1_bio->behind_remaining);
 352
 353                         /* In behind mode, we ACK the master bio once the I/O has safely
 354                          * reached all non-writemostly disks. Setting the Returned bit
 355                          * ensures that this gets done only once -- we don't ever want to
 356                          * return -EIO here, instead we'll wait */
 357
 358                         if (atomic_read(&r1_bio->behind_remaining) >= (atomic_read(&r1_bio->remaining)-1) &&
 359                             test_bit(R1BIO_Uptodate, &r1_bio->state)) {
 360                                 /* Maybe we can return now */
 361                                 if (!test_and_set_bit(R1BIO_Returned, &r1_bio->state)) {
 362                                         struct bio *mbio = r1_bio->master_bio;
 363                                         PRINTK(KERN_DEBUG "raid1: behind end write sectors %llu-%llu\n",
 364                                                (unsigned long long) mbio->bi_sector,
 365                                                (unsigned long long) mbio->bi_sector +
 366                                                (mbio->bi_size >> 9) - 1);
 367                                         bio_endio(mbio, 0);
 368                                 }
 369                         }
 370                 }
 371                 rdev_dec_pending(conf->mirrors[mirror].rdev, conf->mddev);
 372         }
 373         /*
 374          *
 375          * Let's see if all mirrored write operations have finished
 376          * already.
 377          */
 378         if (atomic_dec_and_test(&r1_bio->remaining)) {
 379                 if (test_bit(R1BIO_BarrierRetry, &r1_bio->state))
 380                         reschedule_retry(r1_bio);
 381                 else {
 382                         /* it really is the end of this request */
 383                         if (test_bit(R1BIO_BehindIO, &r1_bio->state)) {
 384                                 /* free extra copy of the data pages */
 385                                 int i = bio->bi_vcnt;
 386                                 while (i--)
 387                                         safe_put_page(bio->bi_io_vec[i].bv_page);
 388                         }
 389                         /* clear the bitmap if all writes complete successfully */
 390                         bitmap_endwrite(r1_bio->mddev->bitmap, r1_bio->sector,
 391                                         r1_bio->sectors,
 392                                         !test_bit(R1BIO_Degraded, &r1_bio->state),
 393                                         behind);
 394                         md_write_end(r1_bio->mddev);
 395                         raid_end_bio_io(r1_bio);
 396                 }
 397         }
 398
 399         if (to_put)
 400                 bio_put(to_put);
 401 }
 402
 403
 404 /*
 405  * This routine returns the disk from which the requested read should
 406  * be done. There is a per-array 'next expected sequential IO' sector
 407  * number - if this matches on the next IO then we use the last disk.
 408  * There is also a per-disk 'last know head position' sector that is
 409  * maintained from IRQ contexts, both the normal and the resync IO
 410  * completion handlers update this position correctly. If there is no
 411  * perfect sequential match then we pick the disk whose head is closest.
 412  *
 413  * If there are 2 mirrors in the same 2 devices, performance degrades
 414  * because position is mirror, not device based.
 415  *
 416  * The rdev for the device selected will have nr_pending incremented.
 417  */
 418 static int read_balance(conf_t *conf, r1bio_t *r1_bio)
 419 {
 420         const unsigned long this_sector = r1_bio->sector;
 421         int new_disk = conf->last_used, disk = new_disk;
 422         int wonly_disk = -1;
 423         const int sectors = r1_bio->sectors;
 424         sector_t new_distance, current_distance;
 425         mdk_rdev_t *rdev;
 426
 427         rcu_read_lock();
 428         /*
 429          * Check if we can balance. We can balance on the whole
 430          * device if no resync is going on, or below the resync window.
 431          * We take the first readable disk when above the resync window.
 432          */
 433  retry:
 434         if (conf->mddev->recovery_cp < MaxSector &&
 435             (this_sector + sectors >= conf->next_resync)) {
 436                 /* Choose the first operation device, for consistancy */
 437                 new_disk = 0;
 438
 439                 for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
 440                      r1_bio->bios[new_disk] == IO_BLOCKED ||
 441                      !rdev || !test_bit(In_sync, &rdev->flags)
 442                              || test_bit(WriteMostly, &rdev->flags);
 443                      rdev = rcu_dereference(conf->mirrors[++new_disk].rdev)) {
 444
 445                         if (rdev && test_bit(In_sync, &rdev->flags) &&
 446                                 r1_bio->bios[new_disk] != IO_BLOCKED)
 447                                 wonly_disk = new_disk;
 448
 449                         if (new_disk == conf->raid_disks - 1) {
 450                                 new_disk = wonly_disk;
 451                                 break;
 452                         }
 453                 }
 454                 goto rb_out;
 455         }
 456
 457
 458         /* make sure the disk is operational */
 459         for (rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
 460              r1_bio->bios[new_disk] == IO_BLOCKED ||
 461              !rdev || !test_bit(In_sync, &rdev->flags) ||
 462                      test_bit(WriteMostly, &rdev->flags);
 463              rdev = rcu_dereference(conf->mirrors[new_disk].rdev)) {
 464
 465                 if (rdev && test_bit(In_sync, &rdev->flags) &&
 466                     r1_bio->bios[new_disk] != IO_BLOCKED)
 467                         wonly_disk = new_disk;
 468
 469                 if (new_disk <= 0)
 470                         new_disk = conf->raid_disks;
 471                 new_disk--;
 472                 if (new_disk == disk) {
 473                         new_disk = wonly_disk;
 474                         break;
 475                 }
 476         }
 477
 478         if (new_disk < 0)
 479                 goto rb_out;
 480
 481         disk = new_disk;
 482         /* now disk == new_disk == starting point for search */
 483
 484         /*
 485          * Don't change to another disk for sequential reads:
 486          */
 487         if (conf->next_seq_sect == this_sector)
 488                 goto rb_out;
 489         if (this_sector == conf->mirrors[new_disk].head_position)
 490                 goto rb_out;
 491
 492         current_distance = abs(this_sector - conf->mirrors[disk].head_position);
 493
 494         /* Find the disk whose head is closest */
 495
 496         do {
 497                 if (disk <= 0)
 498                         disk = conf->raid_disks;
 499                 disk--;
 500
 501                 rdev = rcu_dereference(conf->mirrors[disk].rdev);
 502
 503                 if (!rdev || r1_bio->bios[disk] == IO_BLOCKED ||
 504                     !test_bit(In_sync, &rdev->flags) ||
 505                     test_bit(WriteMostly, &rdev->flags))
 506                         continue;
 507
 508                 if (!atomic_read(&rdev->nr_pending)) {
 509                         new_disk = disk;
 510                         break;
 511                 }
 512                 new_distance = abs(this_sector - conf->mirrors[disk].head_position);
 513                 if (new_distance < current_distance) {
 514                         current_distance = new_distance;
 515                         new_disk = disk;
 516                 }
 517         } while (disk != conf->last_used);
 518
 519  rb_out:
 520
 521
 522         if (new_disk >= 0) {
 523                 rdev = rcu_dereference(conf->mirrors[new_disk].rdev);
 524                 if (!rdev)
 525                         goto retry;
 526                 atomic_inc(&rdev->nr_pending);
 527                 if (!test_bit(In_sync, &rdev->flags)) {
 528                         /* cannot risk returning a device that failed
 529                          * before we inc'ed nr_pending
 530                          */
 531                         rdev_dec_pending(rdev, conf->mddev);
 532                         goto retry;
 533                 }
 534                 conf->next_seq_sect = this_sector + sectors;
 535                 conf->last_used = new_disk;
 536         }
 537         rcu_read_unlock();
 538
 539         return new_disk;
 540 }
 541
 542 static void unplug_slaves(mddev_t *mddev)
 543 {
 544         conf_t *conf = mddev->private;
 545         int i;
 546
 547         rcu_read_lock();
 548         for (i=0; i<mddev->raid_disks; i++) {
 549                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 550                 if (rdev && !test_bit(Faulty, &rdev->flags) && atomic_read(&rdev->nr_pending)) {
 551                         struct request_queue *r_queue = bdev_get_queue(rdev->bdev);
 552
 553                         atomic_inc(&rdev->nr_pending);
 554                         rcu_read_unlock();
 555
 556                         blk_unplug(r_queue);
 557
 558                         rdev_dec_pending(rdev, mddev);
 559                         rcu_read_lock();
 560                 }
 561         }
 562         rcu_read_unlock();
 563 }
 564
 565 static void raid1_unplug(struct request_queue *q)
 566 {
 567         mddev_t *mddev = q->queuedata;
 568
 569         unplug_slaves(mddev);
 570         md_wakeup_thread(mddev->thread);
 571 }
 572
 573 static int raid1_congested(void *data, int bits)
 574 {
 575         mddev_t *mddev = data;
 576         conf_t *conf = mddev->private;
 577         int i, ret = 0;
 578
 579         rcu_read_lock();
 580         for (i = 0; i < mddev->raid_disks; i++) {
 581                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 582                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 583                         struct request_queue *q = bdev_get_queue(rdev->bdev);
 584
 585                         /* Note the '|| 1' - when read_balance prefers
 586                          * non-congested targets, it can be removed
 587                          */
 588                         if ((bits & (1<<BDI_async_congested)) || 1)
 589                                 ret |= bdi_congested(&q->backing_dev_info, bits);
 590                         else
 591                                 ret &= bdi_congested(&q->backing_dev_info, bits);
 592                 }
 593         }
 594         rcu_read_unlock();
 595         return ret;
 596 }
 597
 598
 599 static int flush_pending_writes(conf_t *conf)
 600 {
 601         /* Any writes that have been queued but are awaiting
 602          * bitmap updates get flushed here.
 603          * We return 1 if any requests were actually submitted.
 604          */
 605         int rv = 0;
 606
 607         spin_lock_irq(&conf->device_lock);
 608
 609         if (conf->pending_bio_list.head) {
 610                 struct bio *bio;
 611                 bio = bio_list_get(&conf->pending_bio_list);
 612                 blk_remove_plug(conf->mddev->queue);
 613                 spin_unlock_irq(&conf->device_lock);
 614                 /* flush any pending bitmap writes to
 615                  * disk before proceeding w/ I/O */
 616                 bitmap_unplug(conf->mddev->bitmap);
 617
 618                 while (bio) { /* submit pending writes */
 619                         struct bio *next = bio->bi_next;
 620                         bio->bi_next = NULL;
 621                         generic_make_request(bio);
 622                         bio = next;
 623                 }
 624                 rv = 1;
 625         } else
 626                 spin_unlock_irq(&conf->device_lock);
 627         return rv;
 628 }
 629
 630 /* Barriers....
 631  * Sometimes we need to suspend IO while we do something else,
 632  * either some resync/recovery, or reconfigure the array.
 633  * To do this we raise a 'barrier'.
 634  * The 'barrier' is a counter that can be raised multiple times
 635  * to count how many activities are happening which preclude
 636  * normal IO.
 637  * We can only raise the barrier if there is no pending IO.
 638  * i.e. if nr_pending == 0.
 639  * We choose only to raise the barrier if no-one is waiting for the
 640  * barrier to go down.  This means that as soon as an IO request
 641  * is ready, no other operations which require a barrier will start
 642  * until the IO request has had a chance.
 643  *
 644  * So: regular IO calls 'wait_barrier'.  When that returns there
 645  *    is no backgroup IO happening,  It must arrange to call
 646  *    allow_barrier when it has finished its IO.
 647  * backgroup IO calls must call raise_barrier.  Once that returns
 648  *    there is no normal IO happeing.  It must arrange to call
 649  *    lower_barrier when the particular background IO completes.
 650  */
 651 #define RESYNC_DEPTH 32
 652
 653 static void raise_barrier(conf_t *conf)
 654 {
 655         spin_lock_irq(&conf->resync_lock);
 656
 657         /* Wait until no block IO is waiting */
 658         wait_event_lock_irq(conf->wait_barrier, !conf->nr_waiting,
 659                             conf->resync_lock,
 660                             raid1_unplug(conf->mddev->queue));
 661
 662         /* block any new IO from starting */
 663         conf->barrier++;
 664
 665         /* No wait for all pending IO to complete */
 666         wait_event_lock_irq(conf->wait_barrier,
 667                             !conf->nr_pending && conf->barrier < RESYNC_DEPTH,
 668                             conf->resync_lock,
 669                             raid1_unplug(conf->mddev->queue));
 670
 671         spin_unlock_irq(&conf->resync_lock);
 672 }
 673
 674 static void lower_barrier(conf_t *conf)
 675 {
 676         unsigned long flags;
 677         spin_lock_irqsave(&conf->resync_lock, flags);
 678         conf->barrier--;
 679         spin_unlock_irqrestore(&conf->resync_lock, flags);
 680         wake_up(&conf->wait_barrier);
 681 }
 682
 683 static void wait_barrier(conf_t *conf)
 684 {
 685         spin_lock_irq(&conf->resync_lock);
 686         if (conf->barrier) {
 687                 conf->nr_waiting++;
 688                 wait_event_lock_irq(conf->wait_barrier, !conf->barrier,
 689                                     conf->resync_lock,
 690                                     raid1_unplug(conf->mddev->queue));
 691                 conf->nr_waiting--;
 692         }
 693         conf->nr_pending++;
 694         spin_unlock_irq(&conf->resync_lock);
 695 }
 696
 697 static void allow_barrier(conf_t *conf)
 698 {
 699         unsigned long flags;
 700         spin_lock_irqsave(&conf->resync_lock, flags);
 701         conf->nr_pending--;
 702         spin_unlock_irqrestore(&conf->resync_lock, flags);
 703         wake_up(&conf->wait_barrier);
 704 }
 705
 706 static void freeze_array(conf_t *conf)
 707 {
 708         /* stop syncio and normal IO and wait for everything to
 709          * go quite.
 710          * We increment barrier and nr_waiting, and then
 711          * wait until nr_pending match nr_queued+1
 712          * This is called in the context of one normal IO request
 713          * that has failed. Thus any sync request that might be pending
 714          * will be blocked by nr_pending, and we need to wait for
 715          * pending IO requests to complete or be queued for re-try.
 716          * Thus the number queued (nr_queued) plus this request (1)
 717          * must match the number of pending IOs (nr_pending) before
 718          * we continue.
 719          */
 720         spin_lock_irq(&conf->resync_lock);
 721         conf->barrier++;
 722         conf->nr_waiting++;
 723         wait_event_lock_irq(conf->wait_barrier,
 724                             conf->nr_pending == conf->nr_queued+1,
 725                             conf->resync_lock,
 726                             ({ flush_pending_writes(conf);
 727                                raid1_unplug(conf->mddev->queue); }));
 728         spin_unlock_irq(&conf->resync_lock);
 729 }
 730 static void unfreeze_array(conf_t *conf)
 731 {
 732         /* reverse the effect of the freeze */
 733         spin_lock_irq(&conf->resync_lock);
 734         conf->barrier--;
 735         conf->nr_waiting--;
 736         wake_up(&conf->wait_barrier);
 737         spin_unlock_irq(&conf->resync_lock);
 738 }
 739
 740
 741 /* duplicate the data pages for behind I/O */
 742 static struct page **alloc_behind_pages(struct bio *bio)
 743 {
 744         int i;
 745         struct bio_vec *bvec;
 746         struct page **pages = kzalloc(bio->bi_vcnt * sizeof(struct page *),
 747                                         GFP_NOIO);
 748         if (unlikely(!pages))
 749                 goto do_sync_io;
 750
 751         bio_for_each_segment(bvec, bio, i) {
 752                 pages[i] = alloc_page(GFP_NOIO);
 753                 if (unlikely(!pages[i]))
 754                         goto do_sync_io;
 755                 memcpy(kmap(pages[i]) + bvec->bv_offset,
 756                         kmap(bvec->bv_page) + bvec->bv_offset, bvec->bv_len);
 757                 kunmap(pages[i]);
 758                 kunmap(bvec->bv_page);
 759         }
 760
 761         return pages;
 762
 763 do_sync_io:
 764         if (pages)
 765                 for (i = 0; i < bio->bi_vcnt && pages[i]; i++)
 766                         put_page(pages[i]);
 767         kfree(pages);
 768         PRINTK("%dB behind alloc failed, doing sync I/O\n", bio->bi_size);
 769         return NULL;
 770 }
 771
 772 static int make_request(struct request_queue *q, struct bio * bio)
 773 {
 774         mddev_t *mddev = q->queuedata;
 775         conf_t *conf = mddev->private;
 776         mirror_info_t *mirror;
 777         r1bio_t *r1_bio;
 778         struct bio *read_bio;
 779         int i, targets = 0, disks;
 780         struct bitmap *bitmap;
 781         unsigned long flags;
 782         struct bio_list bl;
 783         struct page **behind_pages = NULL;
 784         const int rw = bio_data_dir(bio);
 785         const int do_sync = bio_sync(bio);
 786         int cpu, do_barriers;
 787         mdk_rdev_t *blocked_rdev;
 788
 789         /*
 790          * Register the new request and wait if the reconstruction
 791          * thread has put up a bar for new requests.
 792          * Continue immediately if no resync is active currently.
 793          * We test barriers_work *after* md_write_start as md_write_start
 794          * may cause the first superblock write, and that will check out
 795          * if barriers work.
 796          */
 797
 798         md_write_start(mddev, bio); /* wait on superblock update early */
 799
 800         if (unlikely(!mddev->barriers_work && bio_barrier(bio))) {
 801                 if (rw == WRITE)
 802                         md_write_end(mddev);
 803                 bio_endio(bio, -EOPNOTSUPP);
 804                 return 0;
 805         }
 806
 807         wait_barrier(conf);
 808
 809         bitmap = mddev->bitmap;
 810
 811         cpu = part_stat_lock();
 812         part_stat_inc(cpu, &mddev->gendisk->part0, ios[rw]);
 813         part_stat_add(cpu, &mddev->gendisk->part0, sectors[rw],
 814                       bio_sectors(bio));
 815         part_stat_unlock();
 816
 817         /*
 818          * make_request() can abort the operation when READA is being
 819          * used and no empty request is available.
 820          *
 821          */
 822         r1_bio = mempool_alloc(conf->r1bio_pool, GFP_NOIO);
 823
 824         r1_bio->master_bio = bio;
 825         r1_bio->sectors = bio->bi_size >> 9;
 826         r1_bio->state = 0;
 827         r1_bio->mddev = mddev;
 828         r1_bio->sector = bio->bi_sector;
 829
 830         if (rw == READ) {
 831                 /*
 832                  * read balancing logic:
 833                  */
 834                 int rdisk = read_balance(conf, r1_bio);
 835
 836                 if (rdisk < 0) {
 837                         /* couldn't find anywhere to read from */
 838                         raid_end_bio_io(r1_bio);
 839                         return 0;
 840                 }
 841                 mirror = conf->mirrors + rdisk;
 842
 843                 r1_bio->read_disk = rdisk;
 844
 845                 read_bio = bio_clone(bio, GFP_NOIO);
 846
 847                 r1_bio->bios[rdisk] = read_bio;
 848
 849                 read_bio->bi_sector = r1_bio->sector + mirror->rdev->data_offset;
 850                 read_bio->bi_bdev = mirror->rdev->bdev;
 851                 read_bio->bi_end_io = raid1_end_read_request;
 852                 read_bio->bi_rw = READ | do_sync;
 853                 read_bio->bi_private = r1_bio;
 854
 855                 generic_make_request(read_bio);
 856                 return 0;
 857         }
 858
 859         /*
 860          * WRITE:
 861          */
 862         /* first select target devices under spinlock and
 863          * inc refcount on their rdev.  Record them by setting
 864          * bios[x] to bio
 865          */
 866         disks = conf->raid_disks;
 867 #if 0
 868         { static int first=1;
 869         if (first) printk("First Write sector %llu disks %d\n",
 870                           (unsigned long long)r1_bio->sector, disks);
 871         first = 0;
 872         }
 873 #endif
 874  retry_write:
 875         blocked_rdev = NULL;
 876         rcu_read_lock();
 877         for (i = 0;  i < disks; i++) {
 878                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
 879                 if (rdev && unlikely(test_bit(Blocked, &rdev->flags))) {
 880                         atomic_inc(&rdev->nr_pending);
 881                         blocked_rdev = rdev;
 882                         break;
 883                 }
 884                 if (rdev && !test_bit(Faulty, &rdev->flags)) {
 885                         atomic_inc(&rdev->nr_pending);
 886                         if (test_bit(Faulty, &rdev->flags)) {
 887                                 rdev_dec_pending(rdev, mddev);
 888                                 r1_bio->bios[i] = NULL;
 889                         } else
 890                                 r1_bio->bios[i] = bio;
 891                         targets++;
 892                 } else
 893                         r1_bio->bios[i] = NULL;
 894         }
 895         rcu_read_unlock();
 896
 897         if (unlikely(blocked_rdev)) {
 898                 /* Wait for this device to become unblocked */
 899                 int j;
 900
 901                 for (j = 0; j < i; j++)
 902                         if (r1_bio->bios[j])
 903                                 rdev_dec_pending(conf->mirrors[j].rdev, mddev);
 904
 905                 allow_barrier(conf);
 906                 md_wait_for_blocked_rdev(blocked_rdev, mddev);
 907                 wait_barrier(conf);
 908                 goto retry_write;
 909         }
 910
 911         BUG_ON(targets == 0); /* we never fail the last device */
 912
 913         if (targets < conf->raid_disks) {
 914                 /* array is degraded, we will not clear the bitmap
 915                  * on I/O completion (see raid1_end_write_request) */
 916                 set_bit(R1BIO_Degraded, &r1_bio->state);
 917         }
 918
 919         /* do behind I/O ? */
 920         if (bitmap &&
 921             atomic_read(&bitmap->behind_writes) < bitmap->max_write_behind &&
 922             (behind_pages = alloc_behind_pages(bio)) != NULL)
 923                 set_bit(R1BIO_BehindIO, &r1_bio->state);
 924
 925         atomic_set(&r1_bio->remaining, 0);
 926         atomic_set(&r1_bio->behind_remaining, 0);
 927
 928         do_barriers = bio_barrier(bio);
 929         if (do_barriers)
 930                 set_bit(R1BIO_Barrier, &r1_bio->state);
 931
 932         bio_list_init(&bl);
 933         for (i = 0; i < disks; i++) {
 934                 struct bio *mbio;
 935                 if (!r1_bio->bios[i])
 936                         continue;
 937
 938                 mbio = bio_clone(bio, GFP_NOIO);
 939                 r1_bio->bios[i] = mbio;
 940
 941                 mbio->bi_sector = r1_bio->sector + conf->mirrors[i].rdev->data_offset;
 942                 mbio->bi_bdev = conf->mirrors[i].rdev->bdev;
 943                 mbio->bi_end_io = raid1_end_write_request;
 944                 mbio->bi_rw = WRITE | do_barriers | do_sync;
 945                 mbio->bi_private = r1_bio;
 946
 947                 if (behind_pages) {
 948                         struct bio_vec *bvec;
 949                         int j;
 950
 951                         /* Yes, I really want the '__' version so that
 952                          * we clear any unused pointer in the io_vec, rather
 953                          * than leave them unchanged.  This is important
 954                          * because when we come to free the pages, we won't
 955                          * know the originial bi_idx, so we just free
 956                          * them all
 957                          */
 958                         __bio_for_each_segment(bvec, mbio, j, 0)
 959                                 bvec->bv_page = behind_pages[j];
 960                         if (test_bit(WriteMostly, &conf->mirrors[i].rdev->flags))
 961                                 atomic_inc(&r1_bio->behind_remaining);
 962                 }
 963
 964                 atomic_inc(&r1_bio->remaining);
 965
 966                 bio_list_add(&bl, mbio);
 967         }
 968         kfree(behind_pages); /* the behind pages are attached to the bios now */
 969
 970         bitmap_startwrite(bitmap, bio->bi_sector, r1_bio->sectors,
 971                                 test_bit(R1BIO_BehindIO, &r1_bio->state));
 972         spin_lock_irqsave(&conf->device_lock, flags);
 973         bio_list_merge(&conf->pending_bio_list, &bl);
 974         bio_list_init(&bl);
 975
 976         blk_plug_device(mddev->queue);
 977         spin_unlock_irqrestore(&conf->device_lock, flags);
 978
 979         /* In case raid1d snuck into freeze_array */
 980         wake_up(&conf->wait_barrier);
 981
 982         if (do_sync)
 983                 md_wakeup_thread(mddev->thread);
 984 #if 0
 985         while ((bio = bio_list_pop(&bl)) != NULL)
 986                 generic_make_request(bio);
 987 #endif
 988
 989         return 0;
 990 }
 991
 992 static void status(struct seq_file *seq, mddev_t *mddev)
 993 {
 994         conf_t *conf = mddev->private;
 995         int i;
 996
 997         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
 998                    conf->raid_disks - mddev->degraded);
 999         rcu_read_lock();
1000         for (i = 0; i < conf->raid_disks; i++) {
1001                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1002                 seq_printf(seq, "%s",
1003                            rdev && test_bit(In_sync, &rdev->flags) ? "U" : "_");
1004         }
1005         rcu_read_unlock();
1006         seq_printf(seq, "]");
1007 }
1008
1009
1010 static void error(mddev_t *mddev, mdk_rdev_t *rdev)
1011 {
1012         char b[BDEVNAME_SIZE];
1013         conf_t *conf = mddev->private;
1014
1015         /*
1016          * If it is not operational, then we have already marked it as dead
1017          * else if it is the last working disks, ignore the error, let the
1018          * next level up know.
1019          * else mark the drive as failed
1020          */
1021         if (test_bit(In_sync, &rdev->flags)
1022             && (conf->raid_disks - mddev->degraded) == 1) {
1023                 /*
1024                  * Don't fail the drive, act as though we were just a
1025                  * normal single drive.
1026                  * However don't try a recovery from this drive as
1027                  * it is very likely to fail.
1028                  */
1029                 mddev->recovery_disabled = 1;
1030                 return;
1031         }
1032         if (test_and_clear_bit(In_sync, &rdev->flags)) {
1033                 unsigned long flags;
1034                 spin_lock_irqsave(&conf->device_lock, flags);
1035                 mddev->degraded++;
1036                 set_bit(Faulty, &rdev->flags);
1037                 spin_unlock_irqrestore(&conf->device_lock, flags);
1038                 /*
1039                  * if recovery is running, make sure it aborts.
1040                  */
1041                 set_bit(MD_RECOVERY_INTR, &mddev->recovery);
1042         } else
1043                 set_bit(Faulty, &rdev->flags);
1044         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1045         printk(KERN_ALERT "raid1: Disk failure on %s, disabling device.\n"
1046                 "raid1: Operation continuing on %d devices.\n",
1047                 bdevname(rdev->bdev,b), conf->raid_disks - mddev->degraded);
1048 }
1049
1050 static void print_conf(conf_t *conf)
1051 {
1052         int i;
1053
1054         printk("RAID1 conf printout:\n");
1055         if (!conf) {
1056                 printk("(!conf)\n");
1057                 return;
1058         }
1059         printk(" --- wd:%d rd:%d\n", conf->raid_disks - conf->mddev->degraded,
1060                 conf->raid_disks);
1061
1062         rcu_read_lock();
1063         for (i = 0; i < conf->raid_disks; i++) {
1064                 char b[BDEVNAME_SIZE];
1065                 mdk_rdev_t *rdev = rcu_dereference(conf->mirrors[i].rdev);
1066                 if (rdev)
1067                         printk(" disk %d, wo:%d, o:%d, dev:%s\n",
1068                                i, !test_bit(In_sync, &rdev->flags),
1069                                !test_bit(Faulty, &rdev->flags),
1070                                bdevname(rdev->bdev,b));
1071         }
1072         rcu_read_unlock();
1073 }
1074
1075 static void close_sync(conf_t *conf)
1076 {
1077         wait_barrier(conf);
1078         allow_barrier(conf);
1079
1080         mempool_destroy(conf->r1buf_pool);
1081         conf->r1buf_pool = NULL;
1082 }
1083
1084 static int raid1_spare_active(mddev_t *mddev)
1085 {
1086         int i;
1087         conf_t *conf = mddev->private;
1088
1089         /*
1090          * Find all failed disks within the RAID1 configuration
1091          * and mark them readable.
1092          * Called under mddev lock, so rcu protection not needed.
1093          */
1094         for (i = 0; i < conf->raid_disks; i++) {
1095                 mdk_rdev_t *rdev = conf->mirrors[i].rdev;
1096                 if (rdev
1097                     && !test_bit(Faulty, &rdev->flags)
1098                     && !test_and_set_bit(In_sync, &rdev->flags)) {
1099                         unsigned long flags;
1100                         spin_lock_irqsave(&conf->device_lock, flags);
1101                         mddev->degraded--;
1102                         spin_unlock_irqrestore(&conf->device_lock, flags);
1103                 }
1104         }
1105
1106         print_conf(conf);
1107         return 0;
1108 }
1109
1110
1111 static int raid1_add_disk(mddev_t *mddev, mdk_rdev_t *rdev)
1112 {
1113         conf_t *conf = mddev->private;
1114         int err = -EEXIST;
1115         int mirror = 0;
1116         mirror_info_t *p;
1117         int first = 0;
1118         int last = mddev->raid_disks - 1;
1119
1120         if (rdev->raid_disk >= 0)
1121                 first = last = rdev->raid_disk;
1122
1123         for (mirror = first; mirror <= last; mirror++)
1124                 if ( !(p=conf->mirrors+mirror)->rdev) {
1125
1126                         blk_queue_stack_limits(mddev->queue,
1127                                                rdev->bdev->bd_disk->queue);
1128                         /* as we don't honour merge_bvec_fn, we must never risk
1129                          * violating it, so limit ->max_sector to one PAGE, as
1130                          * a one page request is never in violation.
1131                          */
1132                         if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1133                             queue_max_sectors(mddev->queue) > (PAGE_SIZE>>9))
1134                                 blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
1135
1136                         p->head_position = 0;
1137                         rdev->raid_disk = mirror;
1138                         err = 0;
1139                         /* As all devices are equivalent, we don't need a full recovery
1140                          * if this was recently any drive of the array
1141                          */
1142                         if (rdev->saved_raid_disk < 0)
1143                                 conf->fullsync = 1;
1144                         rcu_assign_pointer(p->rdev, rdev);
1145                         break;
1146                 }
1147
1148         print_conf(conf);
1149         return err;
1150 }
1151
1152 static int raid1_remove_disk(mddev_t *mddev, int number)
1153 {
1154         conf_t *conf = mddev->private;
1155         int err = 0;
1156         mdk_rdev_t *rdev;
1157         mirror_info_t *p = conf->mirrors+ number;
1158
1159         print_conf(conf);
1160         rdev = p->rdev;
1161         if (rdev) {
1162                 if (test_bit(In_sync, &rdev->flags) ||
1163                     atomic_read(&rdev->nr_pending)) {
1164                         err = -EBUSY;
1165                         goto abort;
1166                 }
1167                 /* Only remove non-faulty devices is recovery
1168                  * is not possible.
1169                  */
1170                 if (!test_bit(Faulty, &rdev->flags) &&
1171                     mddev->degraded < conf->raid_disks) {
1172                         err = -EBUSY;
1173                         goto abort;
1174                 }
1175                 p->rdev = NULL;
1176                 synchronize_rcu();
1177                 if (atomic_read(&rdev->nr_pending)) {
1178                         /* lost the race, try later */
1179                         err = -EBUSY;
1180                         p->rdev = rdev;
1181                 }
1182         }
1183 abort:
1184
1185         print_conf(conf);
1186         return err;
1187 }
1188
1189
1190 static void end_sync_read(struct bio *bio, int error)
1191 {
1192         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1193         int i;
1194
1195         for (i=r1_bio->mddev->raid_disks; i--; )
1196                 if (r1_bio->bios[i] == bio)
1197                         break;
1198         BUG_ON(i < 0);
1199         update_head_pos(i, r1_bio);
1200         /*
1201          * we have read a block, now it needs to be re-written,
1202          * or re-read if the read failed.
1203          * We don't do much here, just schedule handling by raid1d
1204          */
1205         if (test_bit(BIO_UPTODATE, &bio->bi_flags))
1206                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1207
1208         if (atomic_dec_and_test(&r1_bio->remaining))
1209                 reschedule_retry(r1_bio);
1210 }
1211
1212 static void end_sync_write(struct bio *bio, int error)
1213 {
1214         int uptodate = test_bit(BIO_UPTODATE, &bio->bi_flags);
1215         r1bio_t * r1_bio = (r1bio_t *)(bio->bi_private);
1216         mddev_t *mddev = r1_bio->mddev;
1217         conf_t *conf = mddev->private;
1218         int i;
1219         int mirror=0;
1220
1221         for (i = 0; i < conf->raid_disks; i++)
1222                 if (r1_bio->bios[i] == bio) {
1223                         mirror = i;
1224                         break;
1225                 }
1226         if (!uptodate) {
1227                 int sync_blocks = 0;
1228                 sector_t s = r1_bio->sector;
1229                 long sectors_to_go = r1_bio->sectors;
1230                 /* make sure these bits doesn't get cleared. */
1231                 do {
1232                         bitmap_end_sync(mddev->bitmap, s,
1233                                         &sync_blocks, 1);
1234                         s += sync_blocks;
1235                         sectors_to_go -= sync_blocks;
1236                 } while (sectors_to_go > 0);
1237                 md_error(mddev, conf->mirrors[mirror].rdev);
1238         }
1239
1240         update_head_pos(mirror, r1_bio);
1241
1242         if (atomic_dec_and_test(&r1_bio->remaining)) {
1243                 sector_t s = r1_bio->sectors;
1244                 put_buf(r1_bio);
1245                 md_done_sync(mddev, s, uptodate);
1246         }
1247 }
1248
1249 static void sync_request_write(mddev_t *mddev, r1bio_t *r1_bio)
1250 {
1251         conf_t *conf = mddev->private;
1252         int i;
1253         int disks = conf->raid_disks;
1254         struct bio *bio, *wbio;
1255
1256         bio = r1_bio->bios[r1_bio->read_disk];
1257
1258
1259         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1260                 /* We have read all readable devices.  If we haven't
1261                  * got the block, then there is no hope left.
1262                  * If we have, then we want to do a comparison
1263                  * and skip the write if everything is the same.
1264                  * If any blocks failed to read, then we need to
1265                  * attempt an over-write
1266                  */
1267                 int primary;
1268                 if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1269                         for (i=0; i<mddev->raid_disks; i++)
1270                                 if (r1_bio->bios[i]->bi_end_io == end_sync_read)
1271                                         md_error(mddev, conf->mirrors[i].rdev);
1272
1273                         md_done_sync(mddev, r1_bio->sectors, 1);
1274                         put_buf(r1_bio);
1275                         return;
1276                 }
1277                 for (primary=0; primary<mddev->raid_disks; primary++)
1278                         if (r1_bio->bios[primary]->bi_end_io == end_sync_read &&
1279                             test_bit(BIO_UPTODATE, &r1_bio->bios[primary]->bi_flags)) {
1280                                 r1_bio->bios[primary]->bi_end_io = NULL;
1281                                 rdev_dec_pending(conf->mirrors[primary].rdev, mddev);
1282                                 break;
1283                         }
1284                 r1_bio->read_disk = primary;
1285                 for (i=0; i<mddev->raid_disks; i++)
1286                         if (r1_bio->bios[i]->bi_end_io == end_sync_read) {
1287                                 int j;
1288                                 int vcnt = r1_bio->sectors >> (PAGE_SHIFT- 9);
1289                                 struct bio *pbio = r1_bio->bios[primary];
1290                                 struct bio *sbio = r1_bio->bios[i];
1291
1292                                 if (test_bit(BIO_UPTODATE, &sbio->bi_flags)) {
1293                                         for (j = vcnt; j-- ; ) {
1294                                                 struct page *p, *s;
1295                                                 p = pbio->bi_io_vec[j].bv_page;
1296                                                 s = sbio->bi_io_vec[j].bv_page;
1297                                                 if (memcmp(page_address(p),
1298                                                            page_address(s),
1299                                                            PAGE_SIZE))
1300                                                         break;
1301                                         }
1302                                 } else
1303                                         j = 0;
1304                                 if (j >= 0)
1305                                         mddev->resync_mismatches += r1_bio->sectors;
1306                                 if (j < 0 || (test_bit(MD_RECOVERY_CHECK, &mddev->recovery)
1307                                               && test_bit(BIO_UPTODATE, &sbio->bi_flags))) {
1308                                         sbio->bi_end_io = NULL;
1309                                         rdev_dec_pending(conf->mirrors[i].rdev, mddev);
1310                                 } else {
1311                                         /* fixup the bio for reuse */
1312                                         int size;
1313                                         sbio->bi_vcnt = vcnt;
1314                                         sbio->bi_size = r1_bio->sectors << 9;
1315                                         sbio->bi_idx = 0;
1316                                         sbio->bi_phys_segments = 0;
1317                                         sbio->bi_flags &= ~(BIO_POOL_MASK - 1);
1318                                         sbio->bi_flags |= 1 << BIO_UPTODATE;
1319                                         sbio->bi_next = NULL;
1320                                         sbio->bi_sector = r1_bio->sector +
1321                                                 conf->mirrors[i].rdev->data_offset;
1322                                         sbio->bi_bdev = conf->mirrors[i].rdev->bdev;
1323                                         size = sbio->bi_size;
1324                                         for (j = 0; j < vcnt ; j++) {
1325                                                 struct bio_vec *bi;
1326                                                 bi = &sbio->bi_io_vec[j];
1327                                                 bi->bv_offset = 0;
1328                                                 if (size > PAGE_SIZE)
1329                                                         bi->bv_len = PAGE_SIZE;
1330                                                 else
1331                                                         bi->bv_len = size;
1332                                                 size -= PAGE_SIZE;
1333                                                 memcpy(page_address(bi->bv_page),
1334                                                        page_address(pbio->bi_io_vec[j].bv_page),
1335                                                        PAGE_SIZE);
1336                                         }
1337
1338                                 }
1339                         }
1340         }
1341         if (!test_bit(R1BIO_Uptodate, &r1_bio->state)) {
1342                 /* ouch - failed to read all of that.
1343                  * Try some synchronous reads of other devices to get
1344                  * good data, much like with normal read errors.  Only
1345                  * read into the pages we already have so we don't
1346                  * need to re-issue the read request.
1347                  * We don't need to freeze the array, because being in an
1348                  * active sync request, there is no normal IO, and
1349                  * no overlapping syncs.
1350                  */
1351                 sector_t sect = r1_bio->sector;
1352                 int sectors = r1_bio->sectors;
1353                 int idx = 0;
1354
1355                 while(sectors) {
1356                         int s = sectors;
1357                         int d = r1_bio->read_disk;
1358                         int success = 0;
1359                         mdk_rdev_t *rdev;
1360
1361                         if (s > (PAGE_SIZE>>9))
1362                                 s = PAGE_SIZE >> 9;
1363                         do {
1364                                 if (r1_bio->bios[d]->bi_end_io == end_sync_read) {
1365                                         /* No rcu protection needed here devices
1366                                          * can only be removed when no resync is
1367                                          * active, and resync is currently active
1368                                          */
1369                                         rdev = conf->mirrors[d].rdev;
1370                                         if (sync_page_io(rdev->bdev,
1371                                                          sect + rdev->data_offset,
1372                                                          s<<9,
1373                                                          bio->bi_io_vec[idx].bv_page,
1374                                                          READ)) {
1375                                                 success = 1;
1376                                                 break;
1377                                         }
1378                                 }
1379                                 d++;
1380                                 if (d == conf->raid_disks)
1381                                         d = 0;
1382                         } while (!success && d != r1_bio->read_disk);
1383
1384                         if (success) {
1385                                 int start = d;
1386                                 /* write it back and re-read */
1387                                 set_bit(R1BIO_Uptodate, &r1_bio->state);
1388                                 while (d != r1_bio->read_disk) {
1389                                         if (d == 0)
1390                                                 d = conf->raid_disks;
1391                                         d--;
1392                                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1393                                                 continue;
1394                                         rdev = conf->mirrors[d].rdev;
1395                                         atomic_add(s, &rdev->corrected_errors);
1396                                         if (sync_page_io(rdev->bdev,
1397                                                          sect + rdev->data_offset,
1398                                                          s<<9,
1399                                                          bio->bi_io_vec[idx].bv_page,
1400                                                          WRITE) == 0)
1401                                                 md_error(mddev, rdev);
1402                                 }
1403                                 d = start;
1404                                 while (d != r1_bio->read_disk) {
1405                                         if (d == 0)
1406                                                 d = conf->raid_disks;
1407                                         d--;
1408                                         if (r1_bio->bios[d]->bi_end_io != end_sync_read)
1409                                                 continue;
1410                                         rdev = conf->mirrors[d].rdev;
1411                                         if (sync_page_io(rdev->bdev,
1412                                                          sect + rdev->data_offset,
1413                                                          s<<9,
1414                                                          bio->bi_io_vec[idx].bv_page,
1415                                                          READ) == 0)
1416                                                 md_error(mddev, rdev);
1417                                 }
1418                         } else {
1419                                 char b[BDEVNAME_SIZE];
1420                                 /* Cannot read from anywhere, array is toast */
1421                                 md_error(mddev, conf->mirrors[r1_bio->read_disk].rdev);
1422                                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O read error"
1423                                        " for block %llu\n",
1424                                        bdevname(bio->bi_bdev,b),
1425                                        (unsigned long long)r1_bio->sector);
1426                                 md_done_sync(mddev, r1_bio->sectors, 0);
1427                                 put_buf(r1_bio);
1428                                 return;
1429                         }
1430                         sectors -= s;
1431                         sect += s;
1432                         idx ++;
1433                 }
1434         }
1435
1436         /*
1437          * schedule writes
1438          */
1439         atomic_set(&r1_bio->remaining, 1);
1440         for (i = 0; i < disks ; i++) {
1441                 wbio = r1_bio->bios[i];
1442                 if (wbio->bi_end_io == NULL ||
1443                     (wbio->bi_end_io == end_sync_read &&
1444                      (i == r1_bio->read_disk ||
1445                       !test_bit(MD_RECOVERY_SYNC, &mddev->recovery))))
1446                         continue;
1447
1448                 wbio->bi_rw = WRITE;
1449                 wbio->bi_end_io = end_sync_write;
1450                 atomic_inc(&r1_bio->remaining);
1451                 md_sync_acct(conf->mirrors[i].rdev->bdev, wbio->bi_size >> 9);
1452
1453                 generic_make_request(wbio);
1454         }
1455
1456         if (atomic_dec_and_test(&r1_bio->remaining)) {
1457                 /* if we're here, all write(s) have completed, so clean up */
1458                 md_done_sync(mddev, r1_bio->sectors, 1);
1459                 put_buf(r1_bio);
1460         }
1461 }
1462
1463 /*
1464  * This is a kernel thread which:
1465  *
1466  *      1.      Retries failed read operations on working mirrors.
1467  *      2.      Updates the raid superblock when problems encounter.
1468  *      3.      Performs writes following reads for array syncronising.
1469  */
1470
1471 static void fix_read_error(conf_t *conf, int read_disk,
1472                            sector_t sect, int sectors)
1473 {
1474         mddev_t *mddev = conf->mddev;
1475         while(sectors) {
1476                 int s = sectors;
1477                 int d = read_disk;
1478                 int success = 0;
1479                 int start;
1480                 mdk_rdev_t *rdev;
1481
1482                 if (s > (PAGE_SIZE>>9))
1483                         s = PAGE_SIZE >> 9;
1484
1485                 do {
1486                         /* Note: no rcu protection needed here
1487                          * as this is synchronous in the raid1d thread
1488                          * which is the thread that might remove
1489                          * a device.  If raid1d ever becomes multi-threaded....
1490                          */
1491                         rdev = conf->mirrors[d].rdev;
1492                         if (rdev &&
1493                             test_bit(In_sync, &rdev->flags) &&
1494                             sync_page_io(rdev->bdev,
1495                                          sect + rdev->data_offset,
1496                                          s<<9,
1497                                          conf->tmppage, READ))
1498                                 success = 1;
1499                         else {
1500                                 d++;
1501                                 if (d == conf->raid_disks)
1502                                         d = 0;
1503                         }
1504                 } while (!success && d != read_disk);
1505
1506                 if (!success) {
1507                         /* Cannot read from anywhere -- bye bye array */
1508                         md_error(mddev, conf->mirrors[read_disk].rdev);
1509                         break;
1510                 }
1511                 /* write it back and re-read */
1512                 start = d;
1513                 while (d != read_disk) {
1514                         if (d==0)
1515                                 d = conf->raid_disks;
1516                         d--;
1517                         rdev = conf->mirrors[d].rdev;
1518                         if (rdev &&
1519                             test_bit(In_sync, &rdev->flags)) {
1520                                 if (sync_page_io(rdev->bdev,
1521                                                  sect + rdev->data_offset,
1522                                                  s<<9, conf->tmppage, WRITE)
1523                                     == 0)
1524                                         /* Well, this device is dead */
1525                                         md_error(mddev, rdev);
1526                         }
1527                 }
1528                 d = start;
1529                 while (d != read_disk) {
1530                         char b[BDEVNAME_SIZE];
1531                         if (d==0)
1532                                 d = conf->raid_disks;
1533                         d--;
1534                         rdev = conf->mirrors[d].rdev;
1535                         if (rdev &&
1536                             test_bit(In_sync, &rdev->flags)) {
1537                                 if (sync_page_io(rdev->bdev,
1538                                                  sect + rdev->data_offset,
1539                                                  s<<9, conf->tmppage, READ)
1540                                     == 0)
1541                                         /* Well, this device is dead */
1542                                         md_error(mddev, rdev);
1543                                 else {
1544                                         atomic_add(s, &rdev->corrected_errors);
1545                                         printk(KERN_INFO
1546                                                "raid1:%s: read error corrected "
1547                                                "(%d sectors at %llu on %s)\n",
1548                                                mdname(mddev), s,
1549                                                (unsigned long long)(sect +
1550                                                    rdev->data_offset),
1551                                                bdevname(rdev->bdev, b));
1552                                 }
1553                         }
1554                 }
1555                 sectors -= s;
1556                 sect += s;
1557         }
1558 }
1559
1560 static void raid1d(mddev_t *mddev)
1561 {
1562         r1bio_t *r1_bio;
1563         struct bio *bio;
1564         unsigned long flags;
1565         conf_t *conf = mddev->private;
1566         struct list_head *head = &conf->retry_list;
1567         int unplug=0;
1568         mdk_rdev_t *rdev;
1569
1570         md_check_recovery(mddev);
1571
1572         for (;;) {
1573                 char b[BDEVNAME_SIZE];
1574
1575                 unplug += flush_pending_writes(conf);
1576
1577                 spin_lock_irqsave(&conf->device_lock, flags);
1578                 if (list_empty(head)) {
1579                         spin_unlock_irqrestore(&conf->device_lock, flags);
1580                         break;
1581                 }
1582                 r1_bio = list_entry(head->prev, r1bio_t, retry_list);
1583                 list_del(head->prev);
1584                 conf->nr_queued--;
1585                 spin_unlock_irqrestore(&conf->device_lock, flags);
1586
1587                 mddev = r1_bio->mddev;
1588                 conf = mddev->private;
1589                 if (test_bit(R1BIO_IsSync, &r1_bio->state)) {
1590                         sync_request_write(mddev, r1_bio);
1591                         unplug = 1;
1592                 } else if (test_bit(R1BIO_BarrierRetry, &r1_bio->state)) {
1593                         /* some requests in the r1bio were BIO_RW_BARRIER
1594                          * requests which failed with -EOPNOTSUPP.  Hohumm..
1595                          * Better resubmit without the barrier.
1596                          * We know which devices to resubmit for, because
1597                          * all others have had their bios[] entry cleared.
1598                          * We already have a nr_pending reference on these rdevs.
1599                          */
1600                         int i;
1601                         const int do_sync = bio_sync(r1_bio->master_bio);
1602                         clear_bit(R1BIO_BarrierRetry, &r1_bio->state);
1603                         clear_bit(R1BIO_Barrier, &r1_bio->state);
1604                         for (i=0; i < conf->raid_disks; i++)
1605                                 if (r1_bio->bios[i])
1606                                         atomic_inc(&r1_bio->remaining);
1607                         for (i=0; i < conf->raid_disks; i++)
1608                                 if (r1_bio->bios[i]) {
1609                                         struct bio_vec *bvec;
1610                                         int j;
1611
1612                                         bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1613                                         /* copy pages from the failed bio, as
1614                                          * this might be a write-behind device */
1615                                         __bio_for_each_segment(bvec, bio, j, 0)
1616                                                 bvec->bv_page = bio_iovec_idx(r1_bio->bios[i], j)->bv_page;
1617                                         bio_put(r1_bio->bios[i]);
1618                                         bio->bi_sector = r1_bio->sector +
1619                                                 conf->mirrors[i].rdev->data_offset;
1620                                         bio->bi_bdev = conf->mirrors[i].rdev->bdev;
1621                                         bio->bi_end_io = raid1_end_write_request;
1622                                         bio->bi_rw = WRITE | do_sync;
1623                                         bio->bi_private = r1_bio;
1624                                         r1_bio->bios[i] = bio;
1625                                         generic_make_request(bio);
1626                                 }
1627                 } else {
1628                         int disk;
1629
1630                         /* we got a read error. Maybe the drive is bad.  Maybe just
1631                          * the block and we can fix it.
1632                          * We freeze all other IO, and try reading the block from
1633                          * other devices.  When we find one, we re-write
1634                          * and check it that fixes the read error.
1635                          * This is all done synchronously while the array is
1636                          * frozen
1637                          */
1638                         if (mddev->ro == 0) {
1639                                 freeze_array(conf);
1640                                 fix_read_error(conf, r1_bio->read_disk,
1641                                                r1_bio->sector,
1642                                                r1_bio->sectors);
1643                                 unfreeze_array(conf);
1644                         }
1645
1646                         bio = r1_bio->bios[r1_bio->read_disk];
1647                         if ((disk=read_balance(conf, r1_bio)) == -1 ||
1648                             disk == r1_bio->read_disk) {
1649                                 printk(KERN_ALERT "raid1: %s: unrecoverable I/O"
1650                                        " read error for block %llu\n",
1651                                        bdevname(bio->bi_bdev,b),
1652                                        (unsigned long long)r1_bio->sector);
1653                                 raid_end_bio_io(r1_bio);
1654                         } else {
1655                                 const int do_sync = bio_sync(r1_bio->master_bio);
1656                                 r1_bio->bios[r1_bio->read_disk] =
1657                                         mddev->ro ? IO_BLOCKED : NULL;
1658                                 r1_bio->read_disk = disk;
1659                                 bio_put(bio);
1660                                 bio = bio_clone(r1_bio->master_bio, GFP_NOIO);
1661                                 r1_bio->bios[r1_bio->read_disk] = bio;
1662                                 rdev = conf->mirrors[disk].rdev;
1663                                 if (printk_ratelimit())
1664                                         printk(KERN_ERR "raid1: %s: redirecting sector %llu to"
1665                                                " another mirror\n",
1666                                                bdevname(rdev->bdev,b),
1667                                                (unsigned long long)r1_bio->sector);
1668                                 bio->bi_sector = r1_bio->sector + rdev->data_offset;
1669                                 bio->bi_bdev = rdev->bdev;
1670                                 bio->bi_end_io = raid1_end_read_request;
1671                                 bio->bi_rw = READ | do_sync;
1672                                 bio->bi_private = r1_bio;
1673                                 unplug = 1;
1674                                 generic_make_request(bio);
1675                         }
1676                 }
1677         }
1678         if (unplug)
1679                 unplug_slaves(mddev);
1680 }
1681
1682
1683 static int init_resync(conf_t *conf)
1684 {
1685         int buffs;
1686
1687         buffs = RESYNC_WINDOW / RESYNC_BLOCK_SIZE;
1688         BUG_ON(conf->r1buf_pool);
1689         conf->r1buf_pool = mempool_create(buffs, r1buf_pool_alloc, r1buf_pool_free,
1690                                           conf->poolinfo);
1691         if (!conf->r1buf_pool)
1692                 return -ENOMEM;
1693         conf->next_resync = 0;
1694         return 0;
1695 }
1696
1697 /*
1698  * perform a "sync" on one "block"
1699  *
1700  * We need to make sure that no normal I/O request - particularly write
1701  * requests - conflict with active sync requests.
1702  *
1703  * This is achieved by tracking pending requests and a 'barrier' concept
1704  * that can be installed to exclude normal IO requests.
1705  */
1706
1707 static sector_t sync_request(mddev_t *mddev, sector_t sector_nr, int *skipped, int go_faster)
1708 {
1709         conf_t *conf = mddev->private;
1710         r1bio_t *r1_bio;
1711         struct bio *bio;
1712         sector_t max_sector, nr_sectors;
1713         int disk = -1;
1714         int i;
1715         int wonly = -1;
1716         int write_targets = 0, read_targets = 0;
1717         int sync_blocks;
1718         int still_degraded = 0;
1719
1720         if (!conf->r1buf_pool)
1721         {
1722 /*
1723                 printk("sync start - bitmap %p\n", mddev->bitmap);
1724 */
1725                 if (init_resync(conf))
1726                         return 0;
1727         }
1728
1729         max_sector = mddev->dev_sectors;
1730         if (sector_nr >= max_sector) {
1731                 /* If we aborted, we need to abort the
1732                  * sync on the 'current' bitmap chunk (there will
1733                  * only be one in raid1 resync.
1734                  * We can find the current addess in mddev->curr_resync
1735                  */
1736                 if (mddev->curr_resync < max_sector) /* aborted */
1737                         bitmap_end_sync(mddev->bitmap, mddev->curr_resync,
1738                                                 &sync_blocks, 1);
1739                 else /* completed sync */
1740                         conf->fullsync = 0;
1741
1742                 bitmap_close_sync(mddev->bitmap);
1743                 close_sync(conf);
1744                 return 0;
1745         }
1746
1747         if (mddev->bitmap == NULL &&
1748             mddev->recovery_cp == MaxSector &&
1749             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery) &&
1750             conf->fullsync == 0) {
1751                 *skipped = 1;
1752                 return max_sector - sector_nr;
1753         }
1754         /* before building a request, check if we can skip these blocks..
1755          * This call the bitmap_start_sync doesn't actually record anything
1756          */
1757         if (!bitmap_start_sync(mddev->bitmap, sector_nr, &sync_blocks, 1) &&
1758             !conf->fullsync && !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1759                 /* We can skip this block, and probably several more */
1760                 *skipped = 1;
1761                 return sync_blocks;
1762         }
1763         /*
1764          * If there is non-resync activity waiting for a turn,
1765          * and resync is going fast enough,
1766          * then let it though before starting on this new sync request.
1767          */
1768         if (!go_faster && conf->nr_waiting)
1769                 msleep_interruptible(1000);
1770
1771         bitmap_cond_end_sync(mddev->bitmap, sector_nr);
1772         raise_barrier(conf);
1773
1774         conf->next_resync = sector_nr;
1775
1776         r1_bio = mempool_alloc(conf->r1buf_pool, GFP_NOIO);
1777         rcu_read_lock();
1778         /*
1779          * If we get a correctably read error during resync or recovery,
1780          * we might want to read from a different device.  So we
1781          * flag all drives that could conceivably be read from for READ,
1782          * and any others (which will be non-In_sync devices) for WRITE.
1783          * If a read fails, we try reading from something else for which READ
1784          * is OK.
1785          */
1786
1787         r1_bio->mddev = mddev;
1788         r1_bio->sector = sector_nr;
1789         r1_bio->state = 0;
1790         set_bit(R1BIO_IsSync, &r1_bio->state);
1791
1792         for (i=0; i < conf->raid_disks; i++) {
1793                 mdk_rdev_t *rdev;
1794                 bio = r1_bio->bios[i];
1795
1796                 /* take from bio_init */
1797                 bio->bi_next = NULL;
1798                 bio->bi_flags |= 1 << BIO_UPTODATE;
1799                 bio->bi_rw = READ;
1800                 bio->bi_vcnt = 0;
1801                 bio->bi_idx = 0;
1802                 bio->bi_phys_segments = 0;
1803                 bio->bi_size = 0;
1804                 bio->bi_end_io = NULL;
1805                 bio->bi_private = NULL;
1806
1807                 rdev = rcu_dereference(conf->mirrors[i].rdev);
1808                 if (rdev == NULL ||
1809                            test_bit(Faulty, &rdev->flags)) {
1810                         still_degraded = 1;
1811                         continue;
1812                 } else if (!test_bit(In_sync, &rdev->flags)) {
1813                         bio->bi_rw = WRITE;
1814                         bio->bi_end_io = end_sync_write;
1815                         write_targets ++;
1816                 } else {
1817                         /* may need to read from here */
1818                         bio->bi_rw = READ;
1819                         bio->bi_end_io = end_sync_read;
1820                         if (test_bit(WriteMostly, &rdev->flags)) {
1821                                 if (wonly < 0)
1822                                         wonly = i;
1823                         } else {
1824                                 if (disk < 0)
1825                                         disk = i;
1826                         }
1827                         read_targets++;
1828                 }
1829                 atomic_inc(&rdev->nr_pending);
1830                 bio->bi_sector = sector_nr + rdev->data_offset;
1831                 bio->bi_bdev = rdev->bdev;
1832                 bio->bi_private = r1_bio;
1833         }
1834         rcu_read_unlock();
1835         if (disk < 0)
1836                 disk = wonly;
1837         r1_bio->read_disk = disk;
1838
1839         if (test_bit(MD_RECOVERY_SYNC, &mddev->recovery) && read_targets > 0)
1840                 /* extra read targets are also write targets */
1841                 write_targets += read_targets-1;
1842
1843         if (write_targets == 0 || read_targets == 0) {
1844                 /* There is nowhere to write, so all non-sync
1845                  * drives must be failed - so we are finished
1846                  */
1847                 sector_t rv = max_sector - sector_nr;
1848                 *skipped = 1;
1849                 put_buf(r1_bio);
1850                 return rv;
1851         }
1852
1853         if (max_sector > mddev->resync_max)
1854                 max_sector = mddev->resync_max; /* Don't do IO beyond here */
1855         nr_sectors = 0;
1856         sync_blocks = 0;
1857         do {
1858                 struct page *page;
1859                 int len = PAGE_SIZE;
1860                 if (sector_nr + (len>>9) > max_sector)
1861                         len = (max_sector - sector_nr) << 9;
1862                 if (len == 0)
1863                         break;
1864                 if (sync_blocks == 0) {
1865                         if (!bitmap_start_sync(mddev->bitmap, sector_nr,
1866                                                &sync_blocks, still_degraded) &&
1867                             !conf->fullsync &&
1868                             !test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery))
1869                                 break;
1870                         BUG_ON(sync_blocks < (PAGE_SIZE>>9));
1871                         if (len > (sync_blocks<<9))
1872                                 len = sync_blocks<<9;
1873                 }
1874
1875                 for (i=0 ; i < conf->raid_disks; i++) {
1876                         bio = r1_bio->bios[i];
1877                         if (bio->bi_end_io) {
1878                                 page = bio->bi_io_vec[bio->bi_vcnt].bv_page;
1879                                 if (bio_add_page(bio, page, len, 0) == 0) {
1880                                         /* stop here */
1881                                         bio->bi_io_vec[bio->bi_vcnt].bv_page = page;
1882                                         while (i > 0) {
1883                                                 i--;
1884                                                 bio = r1_bio->bios[i];
1885                                                 if (bio->bi_end_io==NULL)
1886                                                         continue;
1887                                                 /* remove last page from this bio */
1888                                                 bio->bi_vcnt--;
1889                                                 bio->bi_size -= len;
1890                                                 bio->bi_flags &= ~(1<< BIO_SEG_VALID);
1891                                         }
1892                                         goto bio_full;
1893                                 }
1894                         }
1895                 }
1896                 nr_sectors += len>>9;
1897                 sector_nr += len>>9;
1898                 sync_blocks -= (len>>9);
1899         } while (r1_bio->bios[disk]->bi_vcnt < RESYNC_PAGES);
1900  bio_full:
1901         r1_bio->sectors = nr_sectors;
1902
1903         /* For a user-requested sync, we read all readable devices and do a
1904          * compare
1905          */
1906         if (test_bit(MD_RECOVERY_REQUESTED, &mddev->recovery)) {
1907                 atomic_set(&r1_bio->remaining, read_targets);
1908                 for (i=0; i<conf->raid_disks; i++) {
1909                         bio = r1_bio->bios[i];
1910                         if (bio->bi_end_io == end_sync_read) {
1911                                 md_sync_acct(bio->bi_bdev, nr_sectors);
1912                                 generic_make_request(bio);
1913                         }
1914                 }
1915         } else {
1916                 atomic_set(&r1_bio->remaining, 1);
1917                 bio = r1_bio->bios[r1_bio->read_disk];
1918                 md_sync_acct(bio->bi_bdev, nr_sectors);
1919                 generic_make_request(bio);
1920
1921         }
1922         return nr_sectors;
1923 }
1924
1925 static sector_t raid1_size(mddev_t *mddev, sector_t sectors, int raid_disks)
1926 {
1927         if (sectors)
1928                 return sectors;
1929
1930         return mddev->dev_sectors;
1931 }
1932
1933 static int run(mddev_t *mddev)
1934 {
1935         conf_t *conf;
1936         int i, j, disk_idx;
1937         mirror_info_t *disk;
1938         mdk_rdev_t *rdev;
1939
1940         if (mddev->level != 1) {
1941                 printk("raid1: %s: raid level not set to mirroring (%d)\n",
1942                        mdname(mddev), mddev->level);
1943                 goto out;
1944         }
1945         if (mddev->reshape_position != MaxSector) {
1946                 printk("raid1: %s: reshape_position set but not supported\n",
1947                        mdname(mddev));
1948                 goto out;
1949         }
1950         /*
1951          * copy the already verified devices into our private RAID1
1952          * bookkeeping area. [whatever we allocate in run(),
1953          * should be freed in stop()]
1954          */
1955         conf = kzalloc(sizeof(conf_t), GFP_KERNEL);
1956         mddev->private = conf;
1957         if (!conf)
1958                 goto out_no_mem;
1959
1960         conf->mirrors = kzalloc(sizeof(struct mirror_info)*mddev->raid_disks,
1961                                  GFP_KERNEL);
1962         if (!conf->mirrors)
1963                 goto out_no_mem;
1964
1965         conf->tmppage = alloc_page(GFP_KERNEL);
1966         if (!conf->tmppage)
1967                 goto out_no_mem;
1968
1969         conf->poolinfo = kmalloc(sizeof(*conf->poolinfo), GFP_KERNEL);
1970         if (!conf->poolinfo)
1971                 goto out_no_mem;
1972         conf->poolinfo->mddev = mddev;
1973         conf->poolinfo->raid_disks = mddev->raid_disks;
1974         conf->r1bio_pool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
1975                                           r1bio_pool_free,
1976                                           conf->poolinfo);
1977         if (!conf->r1bio_pool)
1978                 goto out_no_mem;
1979
1980         spin_lock_init(&conf->device_lock);
1981         mddev->queue->queue_lock = &conf->device_lock;
1982
1983         list_for_each_entry(rdev, &mddev->disks, same_set) {
1984                 disk_idx = rdev->raid_disk;
1985                 if (disk_idx >= mddev->raid_disks
1986                     || disk_idx < 0)
1987                         continue;
1988                 disk = conf->mirrors + disk_idx;
1989
1990                 disk->rdev = rdev;
1991
1992                 blk_queue_stack_limits(mddev->queue,
1993                                        rdev->bdev->bd_disk->queue);
1994                 /* as we don't honour merge_bvec_fn, we must never risk
1995                  * violating it, so limit ->max_sector to one PAGE, as
1996                  * a one page request is never in violation.
1997                  */
1998                 if (rdev->bdev->bd_disk->queue->merge_bvec_fn &&
1999                     queue_max_sectors(mddev->queue) > (PAGE_SIZE>>9))
2000                         blk_queue_max_sectors(mddev->queue, PAGE_SIZE>>9);
2001
2002                 disk->head_position = 0;
2003         }
2004         conf->raid_disks = mddev->raid_disks;
2005         conf->mddev = mddev;
2006         INIT_LIST_HEAD(&conf->retry_list);
2007
2008         spin_lock_init(&conf->resync_lock);
2009         init_waitqueue_head(&conf->wait_barrier);
2010
2011         bio_list_init(&conf->pending_bio_list);
2012         bio_list_init(&conf->flushing_bio_list);
2013
2014
2015         mddev->degraded = 0;
2016         for (i = 0; i < conf->raid_disks; i++) {
2017
2018                 disk = conf->mirrors + i;
2019
2020                 if (!disk->rdev ||
2021                     !test_bit(In_sync, &disk->rdev->flags)) {
2022                         disk->head_position = 0;
2023                         mddev->degraded++;
2024                         if (disk->rdev)
2025                                 conf->fullsync = 1;
2026                 }
2027         }
2028         if (mddev->degraded == conf->raid_disks) {
2029                 printk(KERN_ERR "raid1: no operational mirrors for %s\n",
2030                         mdname(mddev));
2031                 goto out_free_conf;
2032         }
2033         if (conf->raid_disks - mddev->degraded == 1)
2034                 mddev->recovery_cp = MaxSector;
2035
2036         /*
2037          * find the first working one and use it as a starting point
2038          * to read balancing.
2039          */
2040         for (j = 0; j < conf->raid_disks &&
2041                      (!conf->mirrors[j].rdev ||
2042                       !test_bit(In_sync, &conf->mirrors[j].rdev->flags)) ; j++)
2043                 /* nothing */;
2044         conf->last_used = j;
2045
2046
2047         mddev->thread = md_register_thread(raid1d, mddev, "%s_raid1");
2048         if (!mddev->thread) {
2049                 printk(KERN_ERR
2050                        "raid1: couldn't allocate thread for %s\n",
2051                        mdname(mddev));
2052                 goto out_free_conf;
2053         }
2054
2055         if (mddev->recovery_cp != MaxSector)
2056                 printk(KERN_NOTICE "raid1: %s is not clean"
2057                        " -- starting background reconstruction\n",
2058                        mdname(mddev));
2059         printk(KERN_INFO
2060                 "raid1: raid set %s active with %d out of %d mirrors\n",
2061                 mdname(mddev), mddev->raid_disks - mddev->degraded,
2062                 mddev->raid_disks);
2063         /*
2064          * Ok, everything is just fine now
2065          */
2066         md_set_array_sectors(mddev, raid1_size(mddev, 0, 0));
2067
2068         mddev->queue->unplug_fn = raid1_unplug;
2069         mddev->queue->backing_dev_info.congested_fn = raid1_congested;
2070         mddev->queue->backing_dev_info.congested_data = mddev;
2071
2072         return 0;
2073
2074 out_no_mem:
2075         printk(KERN_ERR "raid1: couldn't allocate memory for %s\n",
2076                mdname(mddev));
2077
2078 out_free_conf:
2079         if (conf) {
2080                 if (conf->r1bio_pool)
2081                         mempool_destroy(conf->r1bio_pool);
2082                 kfree(conf->mirrors);
2083                 safe_put_page(conf->tmppage);
2084                 kfree(conf->poolinfo);
2085                 kfree(conf);
2086                 mddev->private = NULL;
2087         }
2088 out:
2089         return -EIO;
2090 }
2091
2092 static int stop(mddev_t *mddev)
2093 {
2094         conf_t *conf = mddev->private;
2095         struct bitmap *bitmap = mddev->bitmap;
2096         int behind_wait = 0;
2097
2098         /* wait for behind writes to complete */
2099         while (bitmap && atomic_read(&bitmap->behind_writes) > 0) {
2100                 behind_wait++;
2101                 printk(KERN_INFO "raid1: behind writes in progress on device %s, waiting to stop (%d)\n", mdname(mddev), behind_wait);
2102                 set_current_state(TASK_UNINTERRUPTIBLE);
2103                 schedule_timeout(HZ); /* wait a second */
2104                 /* need to kick something here to make sure I/O goes? */
2105         }
2106
2107         raise_barrier(conf);
2108         lower_barrier(conf);
2109
2110         md_unregister_thread(mddev->thread);
2111         mddev->thread = NULL;
2112         blk_sync_queue(mddev->queue); /* the unplug fn references 'conf'*/
2113         if (conf->r1bio_pool)
2114                 mempool_destroy(conf->r1bio_pool);
2115         kfree(conf->mirrors);
2116         kfree(conf->poolinfo);
2117         kfree(conf);
2118         mddev->private = NULL;
2119         return 0;
2120 }
2121
2122 static int raid1_resize(mddev_t *mddev, sector_t sectors)
2123 {
2124         /* no resync is happening, and there is enough space
2125          * on all devices, so we can resize.
2126          * We need to make sure resync covers any new space.
2127          * If the array is shrinking we should possibly wait until
2128          * any io in the removed space completes, but it hardly seems
2129          * worth it.
2130          */
2131         md_set_array_sectors(mddev, raid1_size(mddev, sectors, 0));
2132         if (mddev->array_sectors > raid1_size(mddev, sectors, 0))
2133                 return -EINVAL;
2134         set_capacity(mddev->gendisk, mddev->array_sectors);
2135         mddev->changed = 1;
2136         if (sectors > mddev->dev_sectors &&
2137             mddev->recovery_cp == MaxSector) {
2138                 mddev->recovery_cp = mddev->dev_sectors;
2139                 set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2140         }
2141         mddev->dev_sectors = sectors;
2142         mddev->resync_max_sectors = sectors;
2143         return 0;
2144 }
2145
2146 static int raid1_reshape(mddev_t *mddev)
2147 {
2148         /* We need to:
2149          * 1/ resize the r1bio_pool
2150          * 2/ resize conf->mirrors
2151          *
2152          * We allocate a new r1bio_pool if we can.
2153          * Then raise a device barrier and wait until all IO stops.
2154          * Then resize conf->mirrors and swap in the new r1bio pool.
2155          *
2156          * At the same time, we "pack" the devices so that all the missing
2157          * devices have the higher raid_disk numbers.
2158          */
2159         mempool_t *newpool, *oldpool;
2160         struct pool_info *newpoolinfo;
2161         mirror_info_t *newmirrors;
2162         conf_t *conf = mddev->private;
2163         int cnt, raid_disks;
2164         unsigned long flags;
2165         int d, d2, err;
2166
2167         /* Cannot change chunk_size, layout, or level */
2168         if (mddev->chunk_sectors != mddev->new_chunk_sectors ||
2169             mddev->layout != mddev->new_layout ||
2170             mddev->level != mddev->new_level) {
2171                 mddev->new_chunk_sectors = mddev->chunk_sectors;
2172                 mddev->new_layout = mddev->layout;
2173                 mddev->new_level = mddev->level;
2174                 return -EINVAL;
2175         }
2176
2177         err = md_allow_write(mddev);
2178         if (err)
2179                 return err;
2180
2181         raid_disks = mddev->raid_disks + mddev->delta_disks;
2182
2183         if (raid_disks < conf->raid_disks) {
2184                 cnt=0;
2185                 for (d= 0; d < conf->raid_disks; d++)
2186                         if (conf->mirrors[d].rdev)
2187                                 cnt++;
2188                 if (cnt > raid_disks)
2189                         return -EBUSY;
2190         }
2191
2192         newpoolinfo = kmalloc(sizeof(*newpoolinfo), GFP_KERNEL);
2193         if (!newpoolinfo)
2194                 return -ENOMEM;
2195         newpoolinfo->mddev = mddev;
2196         newpoolinfo->raid_disks = raid_disks;
2197
2198         newpool = mempool_create(NR_RAID1_BIOS, r1bio_pool_alloc,
2199                                  r1bio_pool_free, newpoolinfo);
2200         if (!newpool) {
2201                 kfree(newpoolinfo);
2202                 return -ENOMEM;
2203         }
2204         newmirrors = kzalloc(sizeof(struct mirror_info) * raid_disks, GFP_KERNEL);
2205         if (!newmirrors) {
2206                 kfree(newpoolinfo);
2207                 mempool_destroy(newpool);
2208                 return -ENOMEM;
2209         }
2210
2211         raise_barrier(conf);
2212
2213         /* ok, everything is stopped */
2214         oldpool = conf->r1bio_pool;
2215         conf->r1bio_pool = newpool;
2216
2217         for (d = d2 = 0; d < conf->raid_disks; d++) {
2218                 mdk_rdev_t *rdev = conf->mirrors[d].rdev;
2219                 if (rdev && rdev->raid_disk != d2) {
2220                         char nm[20];
2221                         sprintf(nm, "rd%d", rdev->raid_disk);
2222                         sysfs_remove_link(&mddev->kobj, nm);
2223                         rdev->raid_disk = d2;
2224                         sprintf(nm, "rd%d", rdev->raid_disk);
2225                         sysfs_remove_link(&mddev->kobj, nm);
2226                         if (sysfs_create_link(&mddev->kobj,
2227                                               &rdev->kobj, nm))
2228                                 printk(KERN_WARNING
2229                                        "md/raid1: cannot register "
2230                                        "%s for %s\n",
2231                                        nm, mdname(mddev));
2232                 }
2233                 if (rdev)
2234                         newmirrors[d2++].rdev = rdev;
2235         }
2236         kfree(conf->mirrors);
2237         conf->mirrors = newmirrors;
2238         kfree(conf->poolinfo);
2239         conf->poolinfo = newpoolinfo;
2240
2241         spin_lock_irqsave(&conf->device_lock, flags);
2242         mddev->degraded += (raid_disks - conf->raid_disks);
2243         spin_unlock_irqrestore(&conf->device_lock, flags);
2244         conf->raid_disks = mddev->raid_disks = raid_disks;
2245         mddev->delta_disks = 0;
2246
2247         conf->last_used = 0; /* just make sure it is in-range */
2248         lower_barrier(conf);
2249
2250         set_bit(MD_RECOVERY_NEEDED, &mddev->recovery);
2251         md_wakeup_thread(mddev->thread);
2252
2253         mempool_destroy(oldpool);
2254         return 0;
2255 }
2256
2257 static void raid1_quiesce(mddev_t *mddev, int state)
2258 {
2259         conf_t *conf = mddev->private;
2260
2261         switch(state) {
2262         case 1:
2263                 raise_barrier(conf);
2264                 break;
2265         case 0:
2266                 lower_barrier(conf);
2267                 break;
2268         }
2269 }
2270
2271
2272 static struct mdk_personality raid1_personality =
2273 {
2274         .name           = "raid1",
2275         .level          = 1,
2276         .owner          = THIS_MODULE,
2277         .make_request   = make_request,
2278         .run            = run,
2279         .stop           = stop,
2280         .status         = status,
2281         .error_handler  = error,
2282         .hot_add_disk   = raid1_add_disk,
2283         .hot_remove_disk= raid1_remove_disk,
2284         .spare_active   = raid1_spare_active,
2285         .sync_request   = sync_request,
2286         .resize         = raid1_resize,
2287         .size           = raid1_size,
2288         .check_reshape  = raid1_reshape,
2289         .quiesce        = raid1_quiesce,
2290 };
2291
2292 static int __init raid_init(void)
2293 {
2294         return register_md_personality(&raid1_personality);
2295 }
2296
2297 static void raid_exit(void)
2298 {
2299         unregister_md_personality(&raid1_personality);
2300 }
2301
2302 module_init(raid_init);
2303 module_exit(raid_exit);
2304 MODULE_LICENSE("GPL");
2305 MODULE_ALIAS("md-personality-3"); /* RAID1 */
2306 MODULE_ALIAS("md-raid1");
2307 MODULE_ALIAS("md-level-1");