tests/linux/md-resync/orig

   1 /*
   2  * raid1.c : Multiple Devices driver for Linux
   3  *
   4  * Copyright (C) 1999, 2000 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  * This program is free software; you can redistribute it and/or modify
  16  * it under the terms of the GNU General Public License as published by
  17  * the Free Software Foundation; either version 2, or (at your option)
  18  * any later version.
  19  *
  20  * You should have received a copy of the GNU General Public License
  21  * (for example /usr/src/linux/COPYING); if not, write to the Free
  22  * Software Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  23  */
  24
  25 #include <linux/module.h>
  26 #include <linux/config.h>
  27 #include <linux/slab.h>
  28 #include <linux/raid/raid1.h>
  29 #include <asm/atomic.h>
  30
  31 #define MAJOR_NR MD_MAJOR
  32 #define MD_DRIVER
  33 #define MD_PERSONALITY
  34
  35 #define MAX_WORK_PER_DISK 128
  36
  37 #define NR_RESERVED_BUFS        32
  38
  39
  40 /*
  41  * The following can be used to debug the driver
  42  */
  43 #define RAID1_DEBUG     0
  44
  45 #if RAID1_DEBUG
  46 #define PRINTK(x...)   printk(x)
  47 #define inline
  48 #define __inline__
  49 #else
  50 #define PRINTK(x...)  do { } while (0)
  51 #endif
  52
  53
  54 static mdk_personality_t raid1_personality;
  55 static md_spinlock_t retry_list_lock = MD_SPIN_LOCK_UNLOCKED;
  56 struct raid1_bh *raid1_retry_list = NULL, **raid1_retry_tail;
  57
  58 static struct buffer_head *raid1_alloc_bh(raid1_conf_t *conf, int cnt)
  59 {
  60         /* return a linked list of "cnt" struct buffer_heads.
  61          * don't take any off the free list unless we know we can
  62          * get all we need, otherwise we could deadlock
  63          */
  64         struct buffer_head *bh=NULL;
  65
  66         while(cnt) {
  67                 struct buffer_head *t;
  68                 md_spin_lock_irq(&conf->device_lock);
  69                 if (!conf->freebh_blocked && conf->freebh_cnt >= cnt)
  70                         while (cnt) {
  71                                 t = conf->freebh;
  72                                 conf->freebh = t->b_next;
  73                                 t->b_next = bh;
  74                                 bh = t;
  75                                 t->b_state = 0;
  76                                 conf->freebh_cnt--;
  77                                 cnt--;
  78                         }
  79                 md_spin_unlock_irq(&conf->device_lock);
  80                 if (cnt == 0)
  81                         break;
  82                 t = kmem_cache_alloc(bh_cachep, SLAB_NOIO);
  83                 if (t) {
  84                         t->b_next = bh;
  85                         bh = t;
  86                         cnt--;
  87                 } else {
  88                         PRINTK("raid1: waiting for %d bh\n", cnt);
  89                         conf->freebh_blocked = 1;
  90                         wait_disk_event(conf->wait_buffer,
  91                                         !conf->freebh_blocked ||
  92                                         conf->freebh_cnt > conf->raid_disks * NR_RESERVED_BUFS/2);
  93                         conf->freebh_blocked = 0;
  94                 }
  95         }
  96         return bh;
  97 }
  98
  99 static inline void raid1_free_bh(raid1_conf_t *conf, struct buffer_head *bh)
 100 {
 101         unsigned long flags;
 102         spin_lock_irqsave(&conf->device_lock, flags);
 103         while (bh) {
 104                 struct buffer_head *t = bh;
 105                 bh=bh->b_next;
 106                 if (t->b_pprev == NULL)
 107                         kmem_cache_free(bh_cachep, t);
 108                 else {
 109                         t->b_next= conf->freebh;
 110                         conf->freebh = t;
 111                         conf->freebh_cnt++;
 112                 }
 113         }
 114         spin_unlock_irqrestore(&conf->device_lock, flags);
 115         wake_up(&conf->wait_buffer);
 116 }
 117
 118 static int raid1_grow_bh(raid1_conf_t *conf, int cnt)
 119 {
 120         /* allocate cnt buffer_heads, possibly less if kmalloc fails */
 121         int i = 0;
 122
 123         while (i < cnt) {
 124                 struct buffer_head *bh;
 125                 bh = kmem_cache_alloc(bh_cachep, SLAB_KERNEL);
 126                 if (!bh) break;
 127
 128                 md_spin_lock_irq(&conf->device_lock);
 129                 bh->b_pprev = &conf->freebh;
 130                 bh->b_next = conf->freebh;
 131                 conf->freebh = bh;
 132                 conf->freebh_cnt++;
 133                 md_spin_unlock_irq(&conf->device_lock);
 134
 135                 i++;
 136         }
 137         return i;
 138 }
 139
 140 static void raid1_shrink_bh(raid1_conf_t *conf)
 141 {
 142         /* discard all buffer_heads */
 143
 144         md_spin_lock_irq(&conf->device_lock);
 145         while (conf->freebh) {
 146                 struct buffer_head *bh = conf->freebh;
 147                 conf->freebh = bh->b_next;
 148                 kmem_cache_free(bh_cachep, bh);
 149                 conf->freebh_cnt--;
 150         }
 151         md_spin_unlock_irq(&conf->device_lock);
 152 }
 153
 154
 155 static struct raid1_bh *raid1_alloc_r1bh(raid1_conf_t *conf)
 156 {
 157         struct raid1_bh *r1_bh = NULL;
 158
 159         do {
 160                 md_spin_lock_irq(&conf->device_lock);
 161                 if (!conf->freer1_blocked && conf->freer1) {
 162                         r1_bh = conf->freer1;
 163                         conf->freer1 = r1_bh->next_r1;
 164                         conf->freer1_cnt--;
 165                         r1_bh->next_r1 = NULL;
 166                         r1_bh->state = (1 << R1BH_PreAlloc);
 167                         r1_bh->bh_req.b_state = 0;
 168                 }
 169                 md_spin_unlock_irq(&conf->device_lock);
 170                 if (r1_bh)
 171                         return r1_bh;
 172                 r1_bh = (struct raid1_bh *) kmalloc(sizeof(struct raid1_bh), GFP_NOIO);
 173                 if (r1_bh) {
 174                         memset(r1_bh, 0, sizeof(*r1_bh));
 175                         return r1_bh;
 176                 }
 177                 conf->freer1_blocked = 1;
 178                 wait_disk_event(conf->wait_buffer,
 179                                 !conf->freer1_blocked ||
 180                                 conf->freer1_cnt > NR_RESERVED_BUFS/2
 181                         );
 182                 conf->freer1_blocked = 0;
 183         } while (1);
 184 }
 185
 186 static inline void raid1_free_r1bh(struct raid1_bh *r1_bh)
 187 {
 188         struct buffer_head *bh = r1_bh->mirror_bh_list;
 189         raid1_conf_t *conf = mddev_to_conf(r1_bh->mddev);
 190
 191         r1_bh->mirror_bh_list = NULL;
 192
 193         if (test_bit(R1BH_PreAlloc, &r1_bh->state)) {
 194                 unsigned long flags;
 195                 spin_lock_irqsave(&conf->device_lock, flags);
 196                 r1_bh->next_r1 = conf->freer1;
 197                 conf->freer1 = r1_bh;
 198                 conf->freer1_cnt++;
 199                 spin_unlock_irqrestore(&conf->device_lock, flags);
 200                 /* don't need to wakeup wait_buffer because
 201                  *  raid1_free_bh below will do that
 202                  */
 203         } else {
 204                 kfree(r1_bh);
 205         }
 206         raid1_free_bh(conf, bh);
 207 }
 208
 209 static int raid1_grow_r1bh (raid1_conf_t *conf, int cnt)
 210 {
 211         int i = 0;
 212
 213         while (i < cnt) {
 214                 struct raid1_bh *r1_bh;
 215                 r1_bh = (struct raid1_bh*)kmalloc(sizeof(*r1_bh), GFP_KERNEL);
 216                 if (!r1_bh)
 217                         break;
 218                 memset(r1_bh, 0, sizeof(*r1_bh));
 219                 set_bit(R1BH_PreAlloc, &r1_bh->state);
 220                 r1_bh->mddev = conf->mddev;
 221
 222                 raid1_free_r1bh(r1_bh);
 223                 i++;
 224         }
 225         return i;
 226 }
 227
 228 static void raid1_shrink_r1bh(raid1_conf_t *conf)
 229 {
 230         md_spin_lock_irq(&conf->device_lock);
 231         while (conf->freer1) {
 232                 struct raid1_bh *r1_bh = conf->freer1;
 233                 conf->freer1 = r1_bh->next_r1;
 234                 conf->freer1_cnt--;
 235                 kfree(r1_bh);
 236         }
 237         md_spin_unlock_irq(&conf->device_lock);
 238 }
 239
 240
 241
 242 static inline void raid1_free_buf(struct raid1_bh *r1_bh)
 243 {
 244         unsigned long flags;
 245         struct buffer_head *bh = r1_bh->mirror_bh_list;
 246         raid1_conf_t *conf = mddev_to_conf(r1_bh->mddev);
 247         r1_bh->mirror_bh_list = NULL;
 248
 249         spin_lock_irqsave(&conf->device_lock, flags);
 250         r1_bh->next_r1 = conf->freebuf;
 251         conf->freebuf = r1_bh;
 252         spin_unlock_irqrestore(&conf->device_lock, flags);
 253         raid1_free_bh(conf, bh);
 254 }
 255
 256 static struct raid1_bh *raid1_alloc_buf(raid1_conf_t *conf)
 257 {
 258         struct raid1_bh *r1_bh;
 259
 260         md_spin_lock_irq(&conf->device_lock);
 261         wait_event_lock_irq(conf->wait_buffer, conf->freebuf, conf->device_lock);
 262         r1_bh = conf->freebuf;
 263         conf->freebuf = r1_bh->next_r1;
 264         r1_bh->next_r1= NULL;
 265         md_spin_unlock_irq(&conf->device_lock);
 266
 267         return r1_bh;
 268 }
 269
 270 static int raid1_grow_buffers (raid1_conf_t *conf, int cnt)
 271 {
 272         int i = 0;
 273         struct raid1_bh *head = NULL, **tail;
 274         tail = &head;
 275
 276         while (i < cnt) {
 277                 struct raid1_bh *r1_bh;
 278                 struct page *page;
 279
 280                 page = alloc_page(GFP_KERNEL);
 281                 if (!page)
 282                         break;
 283
 284                 r1_bh = (struct raid1_bh *) kmalloc(sizeof(*r1_bh), GFP_KERNEL);
 285                 if (!r1_bh) {
 286                         __free_page(page);
 287                         break;
 288                 }
 289                 memset(r1_bh, 0, sizeof(*r1_bh));
 290                 r1_bh->bh_req.b_page = page;
 291                 r1_bh->bh_req.b_data = page_address(page);
 292                 *tail = r1_bh;
 293                 r1_bh->next_r1 = NULL;
 294                 tail = & r1_bh->next_r1;
 295                 i++;
 296         }
 297         /* this lock probably isn't needed, as at the time when
 298          * we are allocating buffers, nobody else will be touching the
 299          * freebuf list.  But it doesn't hurt....
 300          */
 301         md_spin_lock_irq(&conf->device_lock);
 302         *tail = conf->freebuf;
 303         conf->freebuf = head;
 304         md_spin_unlock_irq(&conf->device_lock);
 305         return i;
 306 }
 307
 308 static void raid1_shrink_buffers (raid1_conf_t *conf)
 309 {
 310         struct raid1_bh *head;
 311         md_spin_lock_irq(&conf->device_lock);
 312         head = conf->freebuf;
 313         conf->freebuf = NULL;
 314         md_spin_unlock_irq(&conf->device_lock);
 315
 316         while (head) {
 317                 struct raid1_bh *r1_bh = head;
 318                 head = r1_bh->next_r1;
 319                 __free_page(r1_bh->bh_req.b_page);
 320                 kfree(r1_bh);
 321         }
 322 }
 323
 324 static int raid1_map (mddev_t *mddev, kdev_t *rdev)
 325 {
 326         raid1_conf_t *conf = mddev_to_conf(mddev);
 327         int i, disks = MD_SB_DISKS;
 328
 329         /*
 330          * Later we do read balancing on the read side
 331          * now we use the first available disk.
 332          */
 333
 334         for (i = 0; i < disks; i++) {
 335                 if (conf->mirrors[i].operational) {
 336                         *rdev = conf->mirrors[i].dev;
 337                         return (0);
 338                 }
 339         }
 340
 341         printk (KERN_ERR "raid1_map(): huh, no more operational devices?\n");
 342         return (-1);
 343 }
 344
 345 static void raid1_reschedule_retry (struct raid1_bh *r1_bh)
 346 {
 347         unsigned long flags;
 348         mddev_t *mddev = r1_bh->mddev;
 349         raid1_conf_t *conf = mddev_to_conf(mddev);
 350
 351         md_spin_lock_irqsave(&retry_list_lock, flags);
 352         if (raid1_retry_list == NULL)
 353                 raid1_retry_tail = &raid1_retry_list;
 354         *raid1_retry_tail = r1_bh;
 355         raid1_retry_tail = &r1_bh->next_r1;
 356         r1_bh->next_r1 = NULL;
 357         md_spin_unlock_irqrestore(&retry_list_lock, flags);
 358         md_wakeup_thread(conf->thread);
 359 }
 360
 361
 362 static void inline io_request_done(unsigned long sector, raid1_conf_t *conf, int phase)
 363 {
 364         unsigned long flags;
 365         spin_lock_irqsave(&conf->segment_lock, flags);
 366         if (sector < conf->start_active)
 367                 conf->cnt_done--;
 368         else if (sector >= conf->start_future && conf->phase == phase)
 369                 conf->cnt_future--;
 370         else if (!--conf->cnt_pending)
 371                 wake_up(&conf->wait_ready);
 372
 373         spin_unlock_irqrestore(&conf->segment_lock, flags);
 374 }
 375
 376 static void inline sync_request_done (unsigned long sector, raid1_conf_t *conf)
 377 {
 378         unsigned long flags;
 379         spin_lock_irqsave(&conf->segment_lock, flags);
 380         if (sector >= conf->start_ready)
 381                 --conf->cnt_ready;
 382         else if (sector >= conf->start_active) {
 383                 if (!--conf->cnt_active) {
 384                         conf->start_active = conf->start_ready;
 385                         wake_up(&conf->wait_done);
 386                 }
 387         }
 388         spin_unlock_irqrestore(&conf->segment_lock, flags);
 389 }
 390
 391 /*
 392  * raid1_end_bh_io() is called when we have finished servicing a mirrored
 393  * operation and are ready to return a success/failure code to the buffer
 394  * cache layer.
 395  */
 396 static void raid1_end_bh_io (struct raid1_bh *r1_bh, int uptodate)
 397 {
 398         struct buffer_head *bh = r1_bh->master_bh;
 399
 400         io_request_done(bh->b_rsector, mddev_to_conf(r1_bh->mddev),
 401                         test_bit(R1BH_SyncPhase, &r1_bh->state));
 402
 403         bh->b_end_io(bh, uptodate);
 404         raid1_free_r1bh(r1_bh);
 405 }
 406 void raid1_end_request (struct buffer_head *bh, int uptodate)
 407 {
 408         struct raid1_bh * r1_bh = (struct raid1_bh *)(bh->b_private);
 409
 410         /*
 411          * this branch is our 'one mirror IO has finished' event handler:
 412          */
 413         if (!uptodate)
 414                 md_error (r1_bh->mddev, bh->b_dev);
 415         else
 416                 /*
 417                  * Set R1BH_Uptodate in our master buffer_head, so that
 418                  * we will return a good error code for to the higher
 419                  * levels even if IO on some other mirrored buffer fails.
 420                  *
 421                  * The 'master' represents the complex operation to
 422                  * user-side. So if something waits for IO, then it will
 423                  * wait for the 'master' buffer_head.
 424                  */
 425                 set_bit (R1BH_Uptodate, &r1_bh->state);
 426
 427         /*
 428          * We split up the read and write side, imho they are
 429          * conceptually different.
 430          */
 431
 432         if ( (r1_bh->cmd == READ) || (r1_bh->cmd == READA) ) {
 433                 /*
 434                  * we have only one buffer_head on the read side
 435                  */
 436
 437                 if (uptodate) {
 438                         raid1_end_bh_io(r1_bh, uptodate);
 439                         return;
 440                 }
 441                 /*
 442                  * oops, read error:
 443                  */
 444                 printk(KERN_ERR "raid1: %s: rescheduling block %lu\n",
 445                          partition_name(bh->b_dev), bh->b_blocknr);
 446                 raid1_reschedule_retry(r1_bh);
 447                 return;
 448         }
 449
 450         /*
 451          * WRITE:
 452          *
 453          * Let's see if all mirrored write operations have finished
 454          * already.
 455          */
 456
 457         if (atomic_dec_and_test(&r1_bh->remaining))
 458                 raid1_end_bh_io(r1_bh, test_bit(R1BH_Uptodate, &r1_bh->state));
 459 }
 460
 461 /*
 462  * This routine returns the disk from which the requested read should
 463  * be done. It bookkeeps the last read position for every disk
 464  * in array and when new read requests come, the disk which last
 465  * position is nearest to the request, is chosen.
 466  *
 467  * TODO: now if there are 2 mirrors in the same 2 devices, performance
 468  * degrades dramatically because position is mirror, not device based.
 469  * This should be changed to be device based. Also atomic sequential
 470  * reads should be somehow balanced.
 471  */
 472
 473 static int raid1_read_balance (raid1_conf_t *conf, struct buffer_head *bh)
 474 {
 475         int new_disk = conf->last_used;
 476         const int sectors = bh->b_size >> 9;
 477         const unsigned long this_sector = bh->b_rsector;
 478         int disk = new_disk;
 479         unsigned long new_distance;
 480         unsigned long current_distance;
 481
 482         /*
 483          * Check if it is sane at all to balance
 484          */
 485
 486         if (conf->resync_mirrors)
 487                 goto rb_out;
 488
 489
 490         /* make sure that disk is operational */
 491         while( !conf->mirrors[new_disk].operational) {
 492                 if (new_disk <= 0) new_disk = conf->raid_disks;
 493                 new_disk--;
 494                 if (new_disk == disk) {
 495                         /*
 496                          * This means no working disk was found
 497                          * Nothing much to do, lets not change anything
 498                          * and hope for the best...
 499                          */
 500
 501                         new_disk = conf->last_used;
 502
 503                         goto rb_out;
 504                 }
 505         }
 506         disk = new_disk;
 507         /* now disk == new_disk == starting point for search */
 508
 509         /*
 510          * Don't touch anything for sequential reads.
 511          */
 512
 513         if (this_sector == conf->mirrors[new_disk].head_position)
 514                 goto rb_out;
 515
 516         /*
 517          * If reads have been done only on a single disk
 518          * for a time, lets give another disk a change.
 519          * This is for kicking those idling disks so that
 520          * they would find work near some hotspot.
 521          */
 522
 523         if (conf->sect_count >= conf->mirrors[new_disk].sect_limit) {
 524                 conf->sect_count = 0;
 525
 526 #if defined(CONFIG_SPARC64) && (__GNUC__ == 2) && (__GNUC_MINOR__ == 92)
 527                 /* Work around a compiler bug in egcs-2.92.11 19980921 */
 528                 new_disk = *(volatile int *)&new_disk;
 529 #endif
 530                 do {
 531                         if (new_disk<=0)
 532                                 new_disk = conf->raid_disks;
 533                         new_disk--;
 534                         if (new_disk == disk)
 535                                 break;
 536                 } while ((conf->mirrors[new_disk].write_only) ||
 537                          (!conf->mirrors[new_disk].operational));
 538
 539                 goto rb_out;
 540         }
 541
 542         current_distance = abs(this_sector -
 543                                 conf->mirrors[disk].head_position);
 544
 545         /* Find the disk which is closest */
 546
 547         do {
 548                 if (disk <= 0)
 549                         disk = conf->raid_disks;
 550                 disk--;
 551
 552                 if ((conf->mirrors[disk].write_only) ||
 553                                 (!conf->mirrors[disk].operational))
 554                         continue;
 555
 556                 new_distance = abs(this_sector -
 557                                         conf->mirrors[disk].head_position);
 558
 559                 if (new_distance < current_distance) {
 560                         conf->sect_count = 0;
 561                         current_distance = new_distance;
 562                         new_disk = disk;
 563                 }
 564         } while (disk != conf->last_used);
 565
 566 rb_out:
 567         conf->mirrors[new_disk].head_position = this_sector + sectors;
 568
 569         conf->last_used = new_disk;
 570         conf->sect_count += sectors;
 571
 572         return new_disk;
 573 }
 574
 575 static int raid1_make_request (request_queue_t *q,
 576                                struct buffer_head * bh)
 577 {
 578         mddev_t *mddev = q->queuedata;
 579         raid1_conf_t *conf = mddev_to_conf(mddev);
 580         struct buffer_head *bh_req, *bhl;
 581         struct raid1_bh * r1_bh;
 582         int disks = MD_SB_DISKS;
 583         int i, sum_bhs = 0;
 584         struct mirror_info *mirror;
 585
 586         if (!buffer_locked(bh))
 587                 BUG();
 588
 589 /*
 590  * make_request() can abort the operation when READA is being
 591  * used and no empty request is available.
 592  *
 593  * Currently, just replace the command with READ/WRITE.
 594  */
 595         r1_bh = raid1_alloc_r1bh (conf);
 596
 597         spin_lock_irq(&conf->segment_lock);
 598         wait_event_lock_irq(conf->wait_done,
 599                         bh->b_rsector < conf->start_active ||
 600                         bh->b_rsector >= conf->start_future,
 601                         conf->segment_lock);
 602         if (bh->b_rsector < conf->start_active)
 603                 conf->cnt_done++;
 604         else {
 605                 conf->cnt_future++;
 606                 if (conf->phase)
 607                         set_bit(R1BH_SyncPhase, &r1_bh->state);
 608         }
 609         spin_unlock_irq(&conf->segment_lock);
 610
 611         /*
 612          * i think the read and write branch should be separated completely,
 613          * since we want to do read balancing on the read side for example.
 614          * Alternative implementations? :) --mingo
 615          */
 616
 617         r1_bh->master_bh = bh;
 618         r1_bh->mddev = mddev;
 619         r1_bh->cmd = rw;
 620
 621         if (rw == READ) {
 622                 /*
 623                  * read balancing logic:
 624                  */
 625                 mirror = conf->mirrors + raid1_read_balance(conf, bh);
 626
 627                 bh_req = &r1_bh->bh_req;
 628                 memcpy(bh_req, bh, sizeof(*bh));
 629                 bh_req->b_blocknr = bh->b_rsector;
 630                 bh_req->b_dev = mirror->dev;
 631                 bh_req->b_rdev = mirror->dev;
 632         /*      bh_req->b_rsector = bh->n_rsector; */
 633                 bh_req->b_end_io = raid1_end_request;
 634                 bh_req->b_private = r1_bh;
 635                 generic_make_request (rw, bh_req);
 636                 return 0;
 637         }
 638
 639         /*
 640          * WRITE:
 641          */
 642
 643         bhl = raid1_alloc_bh(conf, conf->raid_disks);
 644         for (i = 0; i < disks; i++) {
 645                 struct buffer_head *mbh;
 646                 if (!conf->mirrors[i].operational)
 647                         continue;
 648
 649         /*
 650          * We should use a private pool (size depending on NR_REQUEST),
 651          * to avoid writes filling up the memory with bhs
 652          *
 653          * Such pools are much faster than kmalloc anyways (so we waste
 654          * almost nothing by not using the master bh when writing and
 655          * win alot of cleanness) but for now we are cool enough. --mingo
 656          *
 657          * It's safe to sleep here, buffer heads cannot be used in a shared
 658          * manner in the write branch. Look how we lock the buffer at the
 659          * beginning of this function to grok the difference ;)
 660          */
 661                 mbh = bhl;
 662                 if (mbh == NULL) {
 663                         MD_BUG();
 664                         break;
 665                 }
 666                 bhl = mbh->b_next;
 667                 mbh->b_next = NULL;
 668                 mbh->b_this_page = (struct buffer_head *)1;
 669
 670         /*
 671          * prepare mirrored mbh (fields ordered for max mem throughput):
 672          */
 673                 mbh->b_blocknr    = bh->b_rsector;
 674                 mbh->b_dev        = conf->mirrors[i].dev;
 675                 mbh->b_rdev       = conf->mirrors[i].dev;
 676                 mbh->b_rsector    = bh->b_rsector;
 677                 mbh->b_state      = (1<<BH_Req) | (1<<BH_Dirty) |
 678                                                 (1<<BH_Mapped) | (1<<BH_Lock);
 679
 680                 atomic_set(&mbh->b_count, 1);
 681                 mbh->b_size       = bh->b_size;
 682                 mbh->b_page       = bh->b_page;
 683                 mbh->b_data       = bh->b_data;
 684                 mbh->b_list       = BUF_LOCKED;
 685                 mbh->b_end_io     = raid1_end_request;
 686                 mbh->b_private    = r1_bh;
 687
 688                 mbh->b_next = r1_bh->mirror_bh_list;
 689                 r1_bh->mirror_bh_list = mbh;
 690                 sum_bhs++;
 691         }
 692         if (bhl) raid1_free_bh(conf,bhl);
 693         if (!sum_bhs) {
 694                 /* Gag - all mirrors non-operational.. */
 695                 raid1_end_bh_io(r1_bh, 0);
 696                 return 0;
 697         }
 698         md_atomic_set(&r1_bh->remaining, sum_bhs);
 699
 700         /*
 701          * We have to be a bit careful about the semaphore above, thats
 702          * why we start the requests separately. Since kmalloc() could
 703          * fail, sleep and make_request() can sleep too, this is the
 704          * safer solution. Imagine, end_request decreasing the semaphore
 705          * before we could have set it up ... We could play tricks with
 706          * the semaphore (presetting it and correcting at the end if
 707          * sum_bhs is not 'n' but we have to do end_request by hand if
 708          * all requests finish until we had a chance to set up the
 709          * semaphore correctly ... lots of races).
 710          */
 711         bh = r1_bh->mirror_bh_list;
 712         while(bh) {
 713                 struct buffer_head *bh2 = bh;
 714                 bh = bh->b_next;
 715                 generic_make_request(rw, bh2);
 716         }
 717         return (0);
 718 }
 719
 720 static void raid1_status(struct seq_file *seq, mddev_t *mddev)
 721 {
 722         raid1_conf_t *conf = mddev_to_conf(mddev);
 723         int i;
 724
 725         seq_printf(seq, " [%d/%d] [", conf->raid_disks,
 726                                                  conf->working_disks);
 727         for (i = 0; i < conf->raid_disks; i++)
 728                 seq_printf(seq, "%s",
 729                         conf->mirrors[i].operational ? "U" : "_");
 730         seq_printf(seq, "]");
 731 }
 732
 733 #define LAST_DISK KERN_ALERT \
 734 "raid1: only one disk left and IO error.\n"
 735
 736 #define NO_SPARE_DISK KERN_ALERT \
 737 "raid1: no spare disk left, degrading mirror level by one.\n"
 738
 739 #define DISK_FAILED KERN_ALERT \
 740 "raid1: Disk failure on %s, disabling device. \n" \
 741 "       Operation continuing on %d devices\n"
 742
 743 #define START_SYNCING KERN_ALERT \
 744 "raid1: start syncing spare disk.\n"
 745
 746 #define ALREADY_SYNCING KERN_INFO \
 747 "raid1: syncing already in progress.\n"
 748
 749 static void mark_disk_bad (mddev_t *mddev, int failed)
 750 {
 751         raid1_conf_t *conf = mddev_to_conf(mddev);
 752         struct mirror_info *mirror = conf->mirrors+failed;
 753         mdp_super_t *sb = mddev->sb;
 754
 755         mirror->operational = 0;
 756         mark_disk_faulty(sb->disks+mirror->number);
 757         mark_disk_nonsync(sb->disks+mirror->number);
 758         mark_disk_inactive(sb->disks+mirror->number);
 759         if (!mirror->write_only)
 760                 sb->active_disks--;
 761         sb->working_disks--;
 762         sb->failed_disks++;
 763         mddev->sb_dirty = 1;
 764         md_wakeup_thread(conf->thread);
 765         if (!mirror->write_only)
 766                 conf->working_disks--;
 767         printk (DISK_FAILED, partition_name (mirror->dev),
 768                                  conf->working_disks);
 769 }
 770
 771 static int raid1_error (mddev_t *mddev, kdev_t dev)
 772 {
 773         raid1_conf_t *conf = mddev_to_conf(mddev);
 774         struct mirror_info * mirrors = conf->mirrors;
 775         int disks = MD_SB_DISKS;
 776         int i;
 777
 778         /* Find the drive.
 779          * If it is not operational, then we have already marked it as dead
 780          * else if it is the last working disks, ignore the error, let the
 781          * next level up know.
 782          * else mark the drive as failed
 783          */
 784
 785         for (i = 0; i < disks; i++)
 786                 if (mirrors[i].dev==dev && mirrors[i].operational)
 787                         break;
 788         if (i == disks)
 789                 return 0;
 790
 791         if (i < conf->raid_disks && conf->working_disks == 1) {
 792                 /* Don't fail the drive, act as though we were just a
 793                  * normal single drive
 794                  */
 795
 796                 return 1;
 797         }
 798         mark_disk_bad(mddev, i);
 799         return 0;
 800 }
 801
 802 #undef LAST_DISK
 803 #undef NO_SPARE_DISK
 804 #undef DISK_FAILED
 805 #undef START_SYNCING
 806
 807
 808 static void print_raid1_conf (raid1_conf_t *conf)
 809 {
 810         int i;
 811         struct mirror_info *tmp;
 812
 813         printk("RAID1 conf printout:\n");
 814         if (!conf) {
 815                 printk("(conf==NULL)\n");
 816                 return;
 817         }
 818         printk(" --- wd:%d rd:%d nd:%d\n", conf->working_disks,
 819                          conf->raid_disks, conf->nr_disks);
 820
 821         for (i = 0; i < MD_SB_DISKS; i++) {
 822                 tmp = conf->mirrors + i;
 823                 printk(" disk %d, s:%d, o:%d, n:%d rd:%d us:%d dev:%s\n",
 824                         i, tmp->spare,tmp->operational,
 825                         tmp->number,tmp->raid_disk,tmp->used_slot,
 826                         partition_name(tmp->dev));
 827         }
 828 }
 829
 830 static void close_sync(raid1_conf_t *conf)
 831 {
 832         mddev_t *mddev = conf->mddev;
 833         /* If reconstruction was interrupted, we need to close the "active" and "pending"
 834          * holes.
 835          * we know that there are no active rebuild requests, os cnt_active == cnt_ready ==0
 836          */
 837         /* this is really needed when recovery stops too... */
 838         spin_lock_irq(&conf->segment_lock);
 839         conf->start_active = conf->start_pending;
 840         conf->start_ready = conf->start_pending;
 841         wait_event_lock_irq(conf->wait_ready, !conf->cnt_pending, conf->segment_lock);
 842         conf->start_active =conf->start_ready = conf->start_pending = conf->start_future;
 843         conf->start_future = (mddev->sb->size<<1)+1;
 844         conf->cnt_pending = conf->cnt_future;
 845         conf->cnt_future = 0;
 846         conf->phase = conf->phase ^1;
 847         wait_event_lock_irq(conf->wait_ready, !conf->cnt_pending, conf->segment_lock);
 848         conf->start_active = conf->start_ready = conf->start_pending = conf->start_future = 0;
 849         conf->phase = 0;
 850         conf->cnt_future = conf->cnt_done;;
 851         conf->cnt_done = 0;
 852         spin_unlock_irq(&conf->segment_lock);
 853         wake_up(&conf->wait_done);
 854 }
 855
 856 static int raid1_diskop(mddev_t *mddev, mdp_disk_t **d, int state)
 857 {
 858         int err = 0;
 859         int i, failed_disk=-1, spare_disk=-1, removed_disk=-1, added_disk=-1;
 860         raid1_conf_t *conf = mddev->private;
 861         struct mirror_info *tmp, *sdisk, *fdisk, *rdisk, *adisk;
 862         mdp_super_t *sb = mddev->sb;
 863         mdp_disk_t *failed_desc, *spare_desc, *added_desc;
 864         mdk_rdev_t *spare_rdev, *failed_rdev;
 865
 866         print_raid1_conf(conf);
 867
 868         switch (state) {
 869         case DISKOP_SPARE_ACTIVE:
 870         case DISKOP_SPARE_INACTIVE:
 871                 /* need to wait for pending sync io before locking device */
 872                 close_sync(conf);
 873         }
 874
 875         md_spin_lock_irq(&conf->device_lock);
 876         /*
 877          * find the disk ...
 878          */
 879         switch (state) {
 880
 881         case DISKOP_SPARE_ACTIVE:
 882
 883                 /*
 884                  * Find the failed disk within the RAID1 configuration ...
 885                  * (this can only be in the first conf->working_disks part)
 886                  */
 887                 for (i = 0; i < conf->raid_disks; i++) {
 888                         tmp = conf->mirrors + i;
 889                         if ((!tmp->operational && !tmp->spare) ||
 890                                         !tmp->used_slot) {
 891                                 failed_disk = i;
 892                                 break;
 893                         }
 894                 }
 895                 /*
 896                  * When we activate a spare disk we _must_ have a disk in
 897                  * the lower (active) part of the array to replace.
 898                  */
 899                 if ((failed_disk == -1) || (failed_disk >= conf->raid_disks)) {
 900                         MD_BUG();
 901                         err = 1;
 902                         goto abort;
 903                 }
 904                 /* fall through */
 905
 906         case DISKOP_SPARE_WRITE:
 907         case DISKOP_SPARE_INACTIVE:
 908
 909                 /*
 910                  * Find the spare disk ... (can only be in the 'high'
 911                  * area of the array)
 912                  */
 913                 for (i = conf->raid_disks; i < MD_SB_DISKS; i++) {
 914                         tmp = conf->mirrors + i;
 915                         if (tmp->spare && tmp->number == (*d)->number) {
 916                                 spare_disk = i;
 917                                 break;
 918                         }
 919                 }
 920                 if (spare_disk == -1) {
 921                         MD_BUG();
 922                         err = 1;
 923                         goto abort;
 924                 }
 925                 break;
 926
 927         case DISKOP_HOT_REMOVE_DISK:
 928
 929                 for (i = 0; i < MD_SB_DISKS; i++) {
 930                         tmp = conf->mirrors + i;
 931                         if (tmp->used_slot && (tmp->number == (*d)->number)) {
 932                                 if (tmp->operational) {
 933                                         err = -EBUSY;
 934                                         goto abort;
 935                                 }
 936                                 removed_disk = i;
 937                                 break;
 938                         }
 939                 }
 940                 if (removed_disk == -1) {
 941                         MD_BUG();
 942                         err = 1;
 943                         goto abort;
 944                 }
 945                 break;
 946
 947         case DISKOP_HOT_ADD_DISK:
 948
 949                 for (i = conf->raid_disks; i < MD_SB_DISKS; i++) {
 950                         tmp = conf->mirrors + i;
 951                         if (!tmp->used_slot) {
 952                                 added_disk = i;
 953                                 break;
 954                         }
 955                 }
 956                 if (added_disk == -1) {
 957                         MD_BUG();
 958                         err = 1;
 959                         goto abort;
 960                 }
 961                 break;
 962         }
 963
 964         switch (state) {
 965         /*
 966          * Switch the spare disk to write-only mode:
 967          */
 968         case DISKOP_SPARE_WRITE:
 969                 sdisk = conf->mirrors + spare_disk;
 970                 sdisk->operational = 1;
 971                 sdisk->write_only = 1;
 972                 break;
 973         /*
 974          * Deactivate a spare disk:
 975          */
 976         case DISKOP_SPARE_INACTIVE:
 977                 if (conf->start_future > 0) {
 978                         MD_BUG();
 979                         err = -EBUSY;
 980                         break;
 981                 }
 982                 sdisk = conf->mirrors + spare_disk;
 983                 sdisk->operational = 0;
 984                 sdisk->write_only = 0;
 985                 break;
 986         /*
 987          * Activate (mark read-write) the (now sync) spare disk,
 988          * which means we switch it's 'raid position' (->raid_disk)
 989          * with the failed disk. (only the first 'conf->nr_disks'
 990          * slots are used for 'real' disks and we must preserve this
 991          * property)
 992          */
 993         case DISKOP_SPARE_ACTIVE:
 994                 if (conf->start_future > 0) {
 995                         MD_BUG();
 996                         err = -EBUSY;
 997                         break;
 998                 }
 999                 sdisk = conf->mirrors + spare_disk;
1000                 fdisk = conf->mirrors + failed_disk;
1001
1002                 spare_desc = &sb->disks[sdisk->number];
1003                 failed_desc = &sb->disks[fdisk->number];
1004
1005                 if (spare_desc != *d) {
1006                         MD_BUG();
1007                         err = 1;
1008                         goto abort;
1009                 }
1010
1011                 if (spare_desc->raid_disk != sdisk->raid_disk) {
1012                         MD_BUG();
1013                         err = 1;
1014                         goto abort;
1015                 }
1016
1017                 if (sdisk->raid_disk != spare_disk) {
1018                         MD_BUG();
1019                         err = 1;
1020                         goto abort;
1021                 }
1022
1023                 if (failed_desc->raid_disk != fdisk->raid_disk) {
1024                         MD_BUG();
1025                         err = 1;
1026                         goto abort;
1027                 }
1028
1029                 if (fdisk->raid_disk != failed_disk) {
1030                         MD_BUG();
1031                         err = 1;
1032                         goto abort;
1033                 }
1034
1035                 /*
1036                  * do the switch finally
1037                  */
1038                 spare_rdev = find_rdev_nr(mddev, spare_desc->number);
1039                 failed_rdev = find_rdev_nr(mddev, failed_desc->number);
1040
1041                 /* There must be a spare_rdev, but there may not be a
1042                  * failed_rdev.  That slot might be empty...
1043                  */
1044                 spare_rdev->desc_nr = failed_desc->number;
1045                 if (failed_rdev)
1046                         failed_rdev->desc_nr = spare_desc->number;
1047
1048                 xchg_values(*spare_desc, *failed_desc);
1049                 xchg_values(*fdisk, *sdisk);
1050
1051                 /*
1052                  * (careful, 'failed' and 'spare' are switched from now on)
1053                  *
1054                  * we want to preserve linear numbering and we want to
1055                  * give the proper raid_disk number to the now activated
1056                  * disk. (this means we switch back these values)
1057                  */
1058
1059                 xchg_values(spare_desc->raid_disk, failed_desc->raid_disk);
1060                 xchg_values(sdisk->raid_disk, fdisk->raid_disk);
1061                 xchg_values(spare_desc->number, failed_desc->number);
1062                 xchg_values(sdisk->number, fdisk->number);
1063
1064                 *d = failed_desc;
1065
1066                 if (sdisk->dev == MKDEV(0,0))
1067                         sdisk->used_slot = 0;
1068                 /*
1069                  * this really activates the spare.
1070                  */
1071                 fdisk->spare = 0;
1072                 fdisk->write_only = 0;
1073
1074                 /*
1075                  * if we activate a spare, we definitely replace a
1076                  * non-operational disk slot in the 'low' area of
1077                  * the disk array.
1078                  */
1079
1080                 conf->working_disks++;
1081
1082                 break;
1083
1084         case DISKOP_HOT_REMOVE_DISK:
1085                 rdisk = conf->mirrors + removed_disk;
1086
1087                 if (rdisk->spare && (removed_disk < conf->raid_disks)) {
1088                         MD_BUG();
1089                         err = 1;
1090                         goto abort;
1091                 }
1092                 rdisk->dev = MKDEV(0,0);
1093                 rdisk->used_slot = 0;
1094                 conf->nr_disks--;
1095                 break;
1096
1097         case DISKOP_HOT_ADD_DISK:
1098                 adisk = conf->mirrors + added_disk;
1099                 added_desc = *d;
1100
1101                 if (added_disk != added_desc->number) {
1102                         MD_BUG();
1103                         err = 1;
1104                         goto abort;
1105                 }
1106
1107                 adisk->number = added_desc->number;
1108                 adisk->raid_disk = added_desc->raid_disk;
1109                 adisk->dev = MKDEV(added_desc->major,added_desc->minor);
1110
1111                 adisk->operational = 0;
1112                 adisk->write_only = 0;
1113                 adisk->spare = 1;
1114                 adisk->used_slot = 1;
1115                 adisk->head_position = 0;
1116                 conf->nr_disks++;
1117
1118                 break;
1119
1120         default:
1121                 MD_BUG();
1122                 err = 1;
1123                 goto abort;
1124         }
1125 abort:
1126         md_spin_unlock_irq(&conf->device_lock);
1127         if (state == DISKOP_SPARE_ACTIVE || state == DISKOP_SPARE_INACTIVE)
1128                 /* should move to "END_REBUILD" when such exists */
1129                 raid1_shrink_buffers(conf);
1130
1131         print_raid1_conf(conf);
1132         return err;
1133 }
1134
1135
1136 #define IO_ERROR KERN_ALERT \
1137 "raid1: %s: unrecoverable I/O read error for block %lu\n"
1138
1139 #define REDIRECT_SECTOR KERN_ERR \
1140 "raid1: %s: redirecting sector %lu to another mirror\n"
1141
1142 /*
1143  * This is a kernel thread which:
1144  *
1145  *      1.      Retries failed read operations on working mirrors.
1146  *      2.      Updates the raid superblock when problems encounter.
1147  *      3.      Performs writes following reads for array syncronising.
1148  */
1149 static void end_sync_write(struct buffer_head *bh, int uptodate);
1150 static void end_sync_read(struct buffer_head *bh, int uptodate);
1151
1152 static void raid1d (void *data)
1153 {
1154         struct raid1_bh *r1_bh;
1155         struct buffer_head *bh;
1156         unsigned long flags;
1157         raid1_conf_t *conf = data;
1158         mddev_t *mddev = conf->mddev;
1159         kdev_t dev;
1160
1161         if (mddev->sb_dirty)
1162                 md_update_sb(mddev);
1163
1164         for (;;) {
1165                 md_spin_lock_irqsave(&retry_list_lock, flags);
1166                 r1_bh = raid1_retry_list;
1167                 if (!r1_bh)
1168                         break;
1169                 raid1_retry_list = r1_bh->next_r1;
1170                 md_spin_unlock_irqrestore(&retry_list_lock, flags);
1171
1172                 mddev = r1_bh->mddev;
1173                 bh = &r1_bh->bh_req;
1174                 switch(r1_bh->cmd) {
1175                 case SPECIAL:
1176                         /* have to allocate lots of bh structures and
1177                          * schedule writes
1178                          */
1179                         if (test_bit(R1BH_Uptodate, &r1_bh->state)) {
1180                                 int i, sum_bhs = 0;
1181                                 int disks = MD_SB_DISKS;
1182                                 struct buffer_head *bhl, *mbh;
1183
1184                                 conf = mddev_to_conf(mddev);
1185                                 bhl = raid1_alloc_bh(conf, conf->raid_disks); /* don't really need this many */
1186                                 for (i = 0; i < disks ; i++) {
1187                                         if (!conf->mirrors[i].operational)
1188                                                 continue;
1189                                         if (i==conf->last_used)
1190                                                 /* we read from here, no need to write */
1191                                                 continue;
1192                                         if (i < conf->raid_disks
1193                                             && !conf->resync_mirrors)
1194                                                 /* don't need to write this,
1195                                                  * we are just rebuilding */
1196                                                 continue;
1197                                         mbh = bhl;
1198                                         if (!mbh) {
1199                                                 MD_BUG();
1200                                                 break;
1201                                         }
1202                                         bhl = mbh->b_next;
1203                                         mbh->b_this_page = (struct buffer_head *)1;
1204
1205
1206                                 /*
1207                                  * prepare mirrored bh (fields ordered for max mem throughput):
1208                                  */
1209                                         mbh->b_blocknr    = bh->b_blocknr;
1210                                         mbh->b_dev        = conf->mirrors[i].dev;
1211                                         mbh->b_rdev       = conf->mirrors[i].dev;
1212                                         mbh->b_rsector    = bh->b_blocknr;
1213                                         mbh->b_state      = (1<<BH_Req) | (1<<BH_Dirty) |
1214                                                 (1<<BH_Mapped) | (1<<BH_Lock);
1215                                         atomic_set(&mbh->b_count, 1);
1216                                         mbh->b_size       = bh->b_size;
1217                                         mbh->b_page       = bh->b_page;
1218                                         mbh->b_data       = bh->b_data;
1219                                         mbh->b_list       = BUF_LOCKED;
1220                                         mbh->b_end_io     = end_sync_write;
1221                                         mbh->b_private    = r1_bh;
1222
1223                                         mbh->b_next = r1_bh->mirror_bh_list;
1224                                         r1_bh->mirror_bh_list = mbh;
1225
1226                                         sum_bhs++;
1227                                 }
1228                                 md_atomic_set(&r1_bh->remaining, sum_bhs);
1229                                 if (bhl) raid1_free_bh(conf, bhl);
1230                                 mbh = r1_bh->mirror_bh_list;
1231
1232                                 if (!sum_bhs) {
1233                                         /* nowhere to write this too... I guess we
1234                                          * must be done
1235                                          */
1236                                         sync_request_done(bh->b_blocknr, conf);
1237                                         md_done_sync(mddev, bh->b_size>>9, 0);
1238                                         raid1_free_buf(r1_bh);
1239                                 } else
1240                                 while (mbh) {
1241                                         struct buffer_head *bh1 = mbh;
1242                                         mbh = mbh->b_next;
1243                                         generic_make_request(WRITE, bh1);
1244                                         md_sync_acct(bh1->b_dev, bh1->b_size/512);
1245                                 }
1246                         } else {
1247                                 /* There is no point trying a read-for-reconstruct
1248                                  * as reconstruct is about to be aborted
1249                                  */
1250
1251                                 printk (IO_ERROR, partition_name(bh->b_dev), bh->b_blocknr);
1252                                 md_done_sync(mddev, bh->b_size>>9, 0);
1253                         }
1254
1255                         break;
1256                 case READ:
1257                 case READA:
1258                         dev = bh->b_dev;
1259                         raid1_map (mddev, &bh->b_dev);
1260                         if (bh->b_dev == dev) {
1261                                 printk (IO_ERROR, partition_name(bh->b_dev), bh->b_blocknr);
1262                                 raid1_end_bh_io(r1_bh, 0);
1263                         } else {
1264                                 printk (REDIRECT_SECTOR,
1265                                         partition_name(bh->b_dev), bh->b_blocknr);
1266                                 bh->b_rdev = bh->b_dev;
1267                                 bh->b_rsector = bh->b_blocknr;
1268                                 generic_make_request (r1_bh->cmd, bh);
1269                         }
1270                         break;
1271                 }
1272         }
1273         md_spin_unlock_irqrestore(&retry_list_lock, flags);
1274 }
1275 #undef IO_ERROR
1276 #undef REDIRECT_SECTOR
1277
1278 /*
1279  * Private kernel thread to reconstruct mirrors after an unclean
1280  * shutdown.
1281  */
1282 static void raid1syncd (void *data)
1283 {
1284         raid1_conf_t *conf = data;
1285         mddev_t *mddev = conf->mddev;
1286
1287         if (!conf->resync_mirrors)
1288                 return;
1289         if (mddev->recovery_running != 2)
1290                 return;
1291         if (!md_do_sync(mddev, NULL)) {
1292                 /*
1293                  * Only if everything went Ok.
1294                  */
1295                 conf->resync_mirrors = 0;
1296         }
1297
1298         close_sync(conf);
1299
1300 }
1301
1302 /*
1303  * perform a "sync" on one "block"
1304  *
1305  * We need to make sure that no normal I/O request - particularly write
1306  * requests - conflict with active sync requests.
1307  * This is achieved by conceptually dividing the device space into a
1308  * number of sections:
1309  *  DONE: 0 .. a-1     These blocks are in-sync
1310  *  ACTIVE: a.. b-1    These blocks may have active sync requests, but
1311  *                     no normal IO requests
1312  *  READY: b .. c-1    These blocks have no normal IO requests - sync
1313  *                     request may be happening
1314  *  PENDING: c .. d-1  These blocks may have IO requests, but no new
1315  *                     ones will be added
1316  *  FUTURE:  d .. end  These blocks are not to be considered yet. IO may
1317  *                     be happening, but not sync
1318  *
1319  * We keep a
1320  *   phase    which flips (0 or 1) each time d moves and
1321  * a count of:
1322  *   z =  active io requests in FUTURE since d moved - marked with
1323  *        current phase
1324  *   y =  active io requests in FUTURE before d moved, or PENDING -
1325  *        marked with previous phase
1326  *   x =  active sync requests in READY
1327  *   w =  active sync requests in ACTIVE
1328  *   v =  active io requests in DONE
1329  *
1330  * Normally, a=b=c=d=0 and z= active io requests
1331  *   or a=b=c=d=END and v= active io requests
1332  * Allowed changes to a,b,c,d:
1333  * A:  c==d &&  y==0 -> d+=window, y=z, z=0, phase=!phase
1334  * B:  y==0 -> c=d
1335  * C:   b=c, w+=x, x=0
1336  * D:  w==0 -> a=b
1337  * E: a==b==c==d==end -> a=b=c=d=0, z=v, v=0
1338  *
1339  * At start of sync we apply A.
1340  * When y reaches 0, we apply B then A then being sync requests
1341  * When sync point reaches c-1, we wait for y==0, and W==0, and
1342  * then apply apply B then A then D then C.
1343  * Finally, we apply E
1344  *
1345  * The sync request simply issues a "read" against a working drive
1346  * This is marked so that on completion the raid1d thread is woken to
1347  * issue suitable write requests
1348  */
1349
1350 static int raid1_sync_request (mddev_t *mddev, unsigned long sector_nr)
1351 {
1352         raid1_conf_t *conf = mddev_to_conf(mddev);
1353         struct mirror_info *mirror;
1354         struct raid1_bh *r1_bh;
1355         struct buffer_head *bh;
1356         int bsize;
1357         int disk;
1358         int block_nr;
1359         int buffs;
1360
1361         if (!sector_nr) {
1362                 /* we want enough buffers to hold twice the window of 128*/
1363                 buffs = 128 *2 / (PAGE_SIZE>>9);
1364                 buffs = raid1_grow_buffers(conf, buffs);
1365                 if (buffs < 2)
1366                         goto nomem;
1367                 conf->window = buffs*(PAGE_SIZE>>9)/2;
1368         }
1369         spin_lock_irq(&conf->segment_lock);
1370         if (!sector_nr) {
1371                 /* initialize ...*/
1372                 conf->start_active = 0;
1373                 conf->start_ready = 0;
1374                 conf->start_pending = 0;
1375                 conf->start_future = 0;
1376                 conf->phase = 0;
1377
1378                 conf->cnt_future += conf->cnt_done+conf->cnt_pending;
1379                 conf->cnt_done = conf->cnt_pending = 0;
1380                 if (conf->cnt_ready || conf->cnt_active)
1381                         MD_BUG();
1382         }
1383         while (sector_nr >= conf->start_pending) {
1384                 PRINTK("wait .. sect=%lu start_active=%d ready=%d pending=%d future=%d, cnt_done=%d active=%d ready=%d pending=%d future=%d\n",
1385                         sector_nr, conf->start_active, conf->start_ready, conf->start_pending, conf->start_future,
1386                         conf->cnt_done, conf->cnt_active, conf->cnt_ready, conf->cnt_pending, conf->cnt_future);
1387                 wait_event_lock_irq(conf->wait_done,
1388                                         !conf->cnt_active,
1389                                         conf->segment_lock);
1390                 wait_event_lock_irq(conf->wait_ready,
1391                                         !conf->cnt_pending,
1392                                         conf->segment_lock);
1393                 conf->start_active = conf->start_ready;
1394                 conf->start_ready = conf->start_pending;
1395                 conf->start_pending = conf->start_future;
1396                 conf->start_future = conf->start_future+conf->window;
1397                 // Note: falling off the end is not a problem
1398                 conf->phase = conf->phase ^1;
1399                 conf->cnt_active = conf->cnt_ready;
1400                 conf->cnt_ready = 0;
1401                 conf->cnt_pending = conf->cnt_future;
1402                 conf->cnt_future = 0;
1403                 wake_up(&conf->wait_done);
1404         }
1405         conf->cnt_ready++;
1406         spin_unlock_irq(&conf->segment_lock);
1407
1408
1409         /* If reconstructing, and >1 working disc,
1410          * could dedicate one to rebuild and others to
1411          * service read requests ..
1412          */
1413         disk = conf->last_used;
1414         /* make sure disk is operational */
1415         while (!conf->mirrors[disk].operational) {
1416                 if (disk <= 0) disk = conf->raid_disks;
1417                 disk--;
1418                 if (disk == conf->last_used)
1419                         break;
1420         }
1421         conf->last_used = disk;
1422
1423         mirror = conf->mirrors+conf->last_used;
1424
1425         r1_bh = raid1_alloc_buf (conf);
1426         r1_bh->master_bh = NULL;
1427         r1_bh->mddev = mddev;
1428         r1_bh->cmd = SPECIAL;
1429         bh = &r1_bh->bh_req;
1430
1431         block_nr = sector_nr;
1432         bsize = 512;
1433         while (!(block_nr & 1) && bsize < PAGE_SIZE
1434                         && (block_nr+2)*(bsize>>9) < (mddev->sb->size *2)) {
1435                 block_nr >>= 1;
1436                 bsize <<= 1;
1437         }
1438         bh->b_size = bsize;
1439         bh->b_list = BUF_LOCKED;
1440         bh->b_dev = mirror->dev;
1441         bh->b_rdev = mirror->dev;
1442         bh->b_state = (1<<BH_Req) | (1<<BH_Mapped) | (1<<BH_Lock);
1443         if (!bh->b_page)
1444                 BUG();
1445         if (!bh->b_data)
1446                 BUG();
1447         if (bh->b_data != page_address(bh->b_page))
1448                 BUG();
1449         bh->b_end_io = end_sync_read;
1450         bh->b_private = r1_bh;
1451         bh->b_blocknr = sector_nr;
1452         bh->b_rsector = sector_nr;
1453         init_waitqueue_head(&bh->b_wait);
1454
1455         generic_make_request(READ, bh);
1456         md_sync_acct(bh->b_dev, bh->b_size/512);
1457
1458         return (bsize >> 9);
1459
1460 nomem:
1461         raid1_shrink_buffers(conf);
1462         return -ENOMEM;
1463 }
1464
1465 static void end_sync_read(struct buffer_head *bh, int uptodate)
1466 {
1467         struct raid1_bh * r1_bh = (struct raid1_bh *)(bh->b_private);
1468
1469         /* we have read a block, now it needs to be re-written,
1470          * or re-read if the read failed.
1471          * We don't do much here, just schedule handling by raid1d
1472          */
1473         if (!uptodate)
1474                 md_error (r1_bh->mddev, bh->b_dev);
1475         else
1476                 set_bit(R1BH_Uptodate, &r1_bh->state);
1477         raid1_reschedule_retry(r1_bh);
1478 }
1479
1480 static void end_sync_write(struct buffer_head *bh, int uptodate)
1481 {
1482         struct raid1_bh * r1_bh = (struct raid1_bh *)(bh->b_private);
1483
1484         if (!uptodate)
1485                 md_error (r1_bh->mddev, bh->b_dev);
1486         if (atomic_dec_and_test(&r1_bh->remaining)) {
1487                 mddev_t *mddev = r1_bh->mddev;
1488                 unsigned long sect = bh->b_blocknr;
1489                 int size = bh->b_size;
1490                 raid1_free_buf(r1_bh);
1491                 sync_request_done(sect, mddev_to_conf(mddev));
1492                 md_done_sync(mddev,size>>9, uptodate);
1493         }
1494 }
1495
1496 #define INVALID_LEVEL KERN_WARNING \
1497 "raid1: md%d: raid level not set to mirroring (%d)\n"
1498
1499 #define NO_SB KERN_ERR \
1500 "raid1: disabled mirror %s (couldn't access raid superblock)\n"
1501
1502 #define ERRORS KERN_ERR \
1503 "raid1: disabled mirror %s (errors detected)\n"
1504
1505 #define NOT_IN_SYNC KERN_ERR \
1506 "raid1: disabled mirror %s (not in sync)\n"
1507
1508 #define INCONSISTENT KERN_ERR \
1509 "raid1: disabled mirror %s (inconsistent descriptor)\n"
1510
1511 #define ALREADY_RUNNING KERN_ERR \
1512 "raid1: disabled mirror %s (mirror %d already operational)\n"
1513
1514 #define OPERATIONAL KERN_INFO \
1515 "raid1: device %s operational as mirror %d\n"
1516
1517 #define MEM_ERROR KERN_ERR \
1518 "raid1: couldn't allocate memory for md%d\n"
1519
1520 #define SPARE KERN_INFO \
1521 "raid1: spare disk %s\n"
1522
1523 #define NONE_OPERATIONAL KERN_ERR \
1524 "raid1: no operational mirrors for md%d\n"
1525
1526 #define ARRAY_IS_ACTIVE KERN_INFO \
1527 "raid1: raid set md%d active with %d out of %d mirrors\n"
1528
1529 #define THREAD_ERROR KERN_ERR \
1530 "raid1: couldn't allocate thread for md%d\n"
1531
1532 #define START_RESYNC KERN_WARNING \
1533 "raid1: raid set md%d not clean; reconstructing mirrors\n"
1534
1535 static int raid1_run (mddev_t *mddev)
1536 {
1537         raid1_conf_t *conf;
1538         int i, j, disk_idx;
1539         struct mirror_info *disk;
1540         mdp_super_t *sb = mddev->sb;
1541         mdp_disk_t *descriptor;
1542         mdk_rdev_t *rdev;
1543         struct md_list_head *tmp;
1544         int start_recovery = 0;
1545
1546         MOD_INC_USE_COUNT;
1547
1548         if (sb->level != 1) {
1549                 printk(INVALID_LEVEL, mdidx(mddev), sb->level);
1550                 goto out;
1551         }
1552         /*
1553          * copy the already verified devices into our private RAID1
1554          * bookkeeping area. [whatever we allocate in raid1_run(),
1555          * should be freed in raid1_stop()]
1556          */
1557
1558         conf = kmalloc(sizeof(raid1_conf_t), GFP_KERNEL);
1559         mddev->private = conf;
1560         if (!conf) {
1561                 printk(MEM_ERROR, mdidx(mddev));
1562                 goto out;
1563         }
1564         memset(conf, 0, sizeof(*conf));
1565
1566         ITERATE_RDEV(mddev,rdev,tmp) {
1567                 if (rdev->faulty) {
1568                         printk(ERRORS, partition_name(rdev->dev));
1569                 } else {
1570                         if (!rdev->sb) {
1571                                 MD_BUG();
1572                                 continue;
1573                         }
1574                 }
1575                 if (rdev->desc_nr == -1) {
1576                         MD_BUG();
1577                         continue;
1578                 }
1579                 descriptor = &sb->disks[rdev->desc_nr];
1580                 disk_idx = descriptor->raid_disk;
1581                 disk = conf->mirrors + disk_idx;
1582
1583                 if (disk_faulty(descriptor)) {
1584                         disk->number = descriptor->number;
1585                         disk->raid_disk = disk_idx;
1586                         disk->dev = rdev->dev;
1587                         disk->sect_limit = MAX_WORK_PER_DISK;
1588                         disk->operational = 0;
1589                         disk->write_only = 0;
1590                         disk->spare = 0;
1591                         disk->used_slot = 1;
1592                         disk->head_position = 0;
1593                         continue;
1594                 }
1595                 if (disk_active(descriptor)) {
1596                         if (!disk_sync(descriptor)) {
1597                                 printk(NOT_IN_SYNC,
1598                                         partition_name(rdev->dev));
1599                                 continue;
1600                         }
1601                         if ((descriptor->number > MD_SB_DISKS) ||
1602                                          (disk_idx > sb->raid_disks)) {
1603
1604                                 printk(INCONSISTENT,
1605                                         partition_name(rdev->dev));
1606                                 continue;
1607                         }
1608                         if (disk->operational) {
1609                                 printk(ALREADY_RUNNING,
1610                                         partition_name(rdev->dev),
1611                                         disk_idx);
1612                                 continue;
1613                         }
1614                         printk(OPERATIONAL, partition_name(rdev->dev),
1615                                         disk_idx);
1616                         disk->number = descriptor->number;
1617                         disk->raid_disk = disk_idx;
1618                         disk->dev = rdev->dev;
1619                         disk->sect_limit = MAX_WORK_PER_DISK;
1620                         disk->operational = 1;
1621                         disk->write_only = 0;
1622                         disk->spare = 0;
1623                         disk->used_slot = 1;
1624                         disk->head_position = 0;
1625                         conf->working_disks++;
1626                 } else {
1627                 /*
1628                  * Must be a spare disk ..
1629                  */
1630                         printk(SPARE, partition_name(rdev->dev));
1631                         disk->number = descriptor->number;
1632                         disk->raid_disk = disk_idx;
1633                         disk->dev = rdev->dev;
1634                         disk->sect_limit = MAX_WORK_PER_DISK;
1635                         disk->operational = 0;
1636                         disk->write_only = 0;
1637                         disk->spare = 1;
1638                         disk->used_slot = 1;
1639                         disk->head_position = 0;
1640                 }
1641         }
1642         conf->raid_disks = sb->raid_disks;
1643         conf->nr_disks = sb->nr_disks;
1644         conf->mddev = mddev;
1645         conf->device_lock = MD_SPIN_LOCK_UNLOCKED;
1646
1647         conf->segment_lock = MD_SPIN_LOCK_UNLOCKED;
1648         init_waitqueue_head(&conf->wait_buffer);
1649         init_waitqueue_head(&conf->wait_done);
1650         init_waitqueue_head(&conf->wait_ready);
1651
1652         if (!conf->working_disks) {
1653                 printk(NONE_OPERATIONAL, mdidx(mddev));
1654                 goto out_free_conf;
1655         }
1656
1657
1658         /* pre-allocate some buffer_head structures.
1659          * As a minimum, 1 r1bh and raid_disks buffer_heads
1660          * would probably get us by in tight memory situations,
1661          * but a few more is probably a good idea.
1662          * For now, try NR_RESERVED_BUFS r1bh and
1663          * NR_RESERVED_BUFS*raid_disks bufferheads
1664          * This will allow at least NR_RESERVED_BUFS concurrent
1665          * reads or writes even if kmalloc starts failing
1666          */
1667         if (raid1_grow_r1bh(conf, NR_RESERVED_BUFS) < NR_RESERVED_BUFS ||
1668             raid1_grow_bh(conf, NR_RESERVED_BUFS*conf->raid_disks)
1669                               < NR_RESERVED_BUFS*conf->raid_disks) {
1670                 printk(MEM_ERROR, mdidx(mddev));
1671                 goto out_free_conf;
1672         }
1673
1674         for (i = 0; i < MD_SB_DISKS; i++) {
1675
1676                 descriptor = sb->disks+i;
1677                 disk_idx = descriptor->raid_disk;
1678                 disk = conf->mirrors + disk_idx;
1679
1680                 if (disk_faulty(descriptor) && (disk_idx < conf->raid_disks) &&
1681                                 !disk->used_slot) {
1682
1683                         disk->number = descriptor->number;
1684                         disk->raid_disk = disk_idx;
1685                         disk->dev = MKDEV(0,0);
1686
1687                         disk->operational = 0;
1688                         disk->write_only = 0;
1689                         disk->spare = 0;
1690                         disk->used_slot = 1;
1691                         disk->head_position = 0;
1692                 }
1693         }
1694
1695         /*
1696          * find the first working one and use it as a starting point
1697          * to read balancing.
1698          */
1699         for (j = 0; !conf->mirrors[j].operational && j < MD_SB_DISKS; j++)
1700                 /* nothing */;
1701         conf->last_used = j;
1702
1703
1704
1705         {
1706                 const char * name = "raid1d";
1707
1708                 conf->thread = md_register_thread(raid1d, conf, name);
1709                 if (!conf->thread) {
1710                         printk(THREAD_ERROR, mdidx(mddev));
1711                         goto out_free_conf;
1712                 }
1713         }
1714
1715         if (!start_recovery && !(sb->state & (1 << MD_SB_CLEAN)) &&
1716             (conf->working_disks > 1)) {
1717                 const char * name = "raid1syncd";
1718
1719                 conf->resync_thread = md_register_thread(raid1syncd, conf,name);
1720                 if (!conf->resync_thread) {
1721                         printk(THREAD_ERROR, mdidx(mddev));
1722                         goto out_free_conf;
1723                 }
1724
1725                 printk(START_RESYNC, mdidx(mddev));
1726                 conf->resync_mirrors = 1;
1727                 mddev->recovery_running = 2;
1728                 md_wakeup_thread(conf->resync_thread);
1729         }
1730
1731         /*
1732          * Regenerate the "device is in sync with the raid set" bit for
1733          * each device.
1734          */
1735         for (i = 0; i < MD_SB_DISKS; i++) {
1736                 mark_disk_nonsync(sb->disks+i);
1737                 for (j = 0; j < sb->raid_disks; j++) {
1738                         if (!conf->mirrors[j].operational)
1739                                 continue;
1740                         if (sb->disks[i].number == conf->mirrors[j].number)
1741                                 mark_disk_sync(sb->disks+i);
1742                 }
1743         }
1744         sb->active_disks = conf->working_disks;
1745
1746         printk(ARRAY_IS_ACTIVE, mdidx(mddev), sb->active_disks, sb->raid_disks);
1747         /*
1748          * Ok, everything is just fine now
1749          */
1750         return 0;
1751
1752 out_free_conf:
1753         raid1_shrink_r1bh(conf);
1754         raid1_shrink_bh(conf);
1755         raid1_shrink_buffers(conf);
1756         kfree(conf);
1757         mddev->private = NULL;
1758 out:
1759         MOD_DEC_USE_COUNT;
1760         return -EIO;
1761 }
1762
1763 #undef INVALID_LEVEL
1764 #undef NO_SB
1765 #undef ERRORS
1766 #undef NOT_IN_SYNC
1767 #undef INCONSISTENT
1768 #undef ALREADY_RUNNING
1769 #undef OPERATIONAL
1770 #undef SPARE
1771 #undef NONE_OPERATIONAL
1772 #undef ARRAY_IS_ACTIVE
1773
1774 static int raid1_stop_resync (mddev_t *mddev)
1775 {
1776         raid1_conf_t *conf = mddev_to_conf(mddev);
1777
1778         if (conf->resync_thread) {
1779                 if (conf->resync_mirrors) {
1780                         md_interrupt_thread(conf->resync_thread);
1781
1782                         printk(KERN_INFO "raid1: mirror resync was not fully finished, restarting next time.\n");
1783                         return 1;
1784                 }
1785                 return 0;
1786         }
1787         return 0;
1788 }
1789
1790 static int raid1_restart_resync (mddev_t *mddev)
1791 {
1792         raid1_conf_t *conf = mddev_to_conf(mddev);
1793
1794         if (conf->resync_mirrors) {
1795                 if (!conf->resync_thread) {
1796                         MD_BUG();
1797                         return 0;
1798                 }
1799                 mddev->recovery_running = 2;
1800                 md_wakeup_thread(conf->resync_thread);
1801                 return 1;
1802         }
1803         return 0;
1804 }
1805
1806 static int raid1_stop (mddev_t *mddev)
1807 {
1808         raid1_conf_t *conf = mddev_to_conf(mddev);
1809
1810         md_unregister_thread(conf->thread);
1811         if (conf->resync_thread)
1812                 md_unregister_thread(conf->resync_thread);
1813         raid1_shrink_r1bh(conf);
1814         raid1_shrink_bh(conf);
1815         raid1_shrink_buffers(conf);
1816         kfree(conf);
1817         mddev->private = NULL;
1818         MOD_DEC_USE_COUNT;
1819         return 0;
1820 }
1821
1822 static mdk_personality_t raid1_personality=
1823 {
1824         name:           "raid1",
1825         make_request:   raid1_make_request,
1826         run:            raid1_run,
1827         stop:           raid1_stop,
1828         status:         raid1_status,
1829         error_handler:  raid1_error,
1830         diskop:         raid1_diskop,
1831         stop_resync:    raid1_stop_resync,
1832         restart_resync: raid1_restart_resync,
1833         sync_request:   raid1_sync_request
1834 };
1835
1836 static int md__init raid1_init (void)
1837 {
1838         return register_md_personality (RAID1, &raid1_personality);
1839 }
1840
1841 static void raid1_exit (void)
1842 {
1843         unregister_md_personality (RAID1);
1844 }
1845
1846 module_init(raid1_init);
1847 module_exit(raid1_exit);
1848 MODULE_LICENSE("GPL");