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 �stergaard" <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 54static mdk_personality_t raid1_personality; 55static md_spinlock_t retry_list_lock = MD_SPIN_LOCK_UNLOCKED; 56struct raid1_bh *raid1_retry_list = NULL, **raid1_retry_tail; 57 58static 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 99static 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 118static 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 140static 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 155static 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 186static 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 209static 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 228static 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 242static 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 256static 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 270static 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 308static 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 324static 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 345static 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 362static 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 376static 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 */ 396static 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} 406void 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 473static 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 566rb_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 575static int raid1_make_request (mddev_t *mddev, int rw, 576 struct buffer_head * bh) 577{ 578 raid1_conf_t *conf = mddev_to_conf(mddev); 579 struct buffer_head *bh_req, *bhl; 580 struct raid1_bh * r1_bh; 581 int disks = MD_SB_DISKS; 582 int i, sum_bhs = 0; 583 struct mirror_info *mirror; 584 585 if (!buffer_locked(bh)) 586 BUG(); 587 588/* 589 * make_request() can abort the operation when READA is being 590 * used and no empty request is available. 591 * 592 * Currently, just replace the command with READ/WRITE. 593 */ 594 if (rw == READA) 595 rw = READ; 596 597 r1_bh = raid1_alloc_r1bh (conf); 598 599 spin_lock_irq(&conf->segment_lock); 600 wait_event_lock_irq(conf->wait_done, 601 bh->b_rsector < conf->start_active || 602 bh->b_rsector >= conf->start_future, 603 conf->segment_lock); 604 if (bh->b_rsector < conf->start_active) 605 conf->cnt_done++; 606 else { 607 conf->cnt_future++; 608 if (conf->phase) 609 set_bit(R1BH_SyncPhase, &r1_bh->state); 610 } 611 spin_unlock_irq(&conf->segment_lock); 612 613 /* 614 * i think the read and write branch should be separated completely, 615 * since we want to do read balancing on the read side for example. 616 * Alternative implementations? :) --mingo 617 */ 618 619 r1_bh->master_bh = bh; 620 r1_bh->mddev = mddev; 621 r1_bh->cmd = rw; 622 623 if (rw == READ) { 624 /* 625 * read balancing logic: 626 */ 627 mirror = conf->mirrors + raid1_read_balance(conf, bh); 628 629 bh_req = &r1_bh->bh_req; 630 memcpy(bh_req, bh, sizeof(*bh)); 631 bh_req->b_blocknr = bh->b_rsector; 632 bh_req->b_dev = mirror->dev; 633 bh_req->b_rdev = mirror->dev; 634 /* bh_req->b_rsector = bh->n_rsector; */ 635 bh_req->b_end_io = raid1_end_request; 636 bh_req->b_private = r1_bh; 637 generic_make_request (rw, bh_req); 638 return 0; 639 } 640 641 /* 642 * WRITE: 643 */ 644 645 bhl = raid1_alloc_bh(conf, conf->raid_disks); 646 for (i = 0; i < disks; i++) { 647 struct buffer_head *mbh; 648 if (!conf->mirrors[i].operational) 649 continue; 650 651 /* 652 * We should use a private pool (size depending on NR_REQUEST), 653 * to avoid writes filling up the memory with bhs 654 * 655 * Such pools are much faster than kmalloc anyways (so we waste 656 * almost nothing by not using the master bh when writing and 657 * win alot of cleanness) but for now we are cool enough. --mingo 658 * 659 * It's safe to sleep here, buffer heads cannot be used in a shared 660 * manner in the write branch. Look how we lock the buffer at the 661 * beginning of this function to grok the difference ;) 662 */ 663 mbh = bhl; 664 if (mbh == NULL) { 665 MD_BUG(); 666 break; 667 } 668 bhl = mbh->b_next; 669 mbh->b_next = NULL; 670 mbh->b_this_page = (struct buffer_head *)1; 671 672 /* 673 * prepare mirrored mbh (fields ordered for max mem throughput): 674 */ 675 mbh->b_blocknr = bh->b_rsector; 676 mbh->b_dev = conf->mirrors[i].dev; 677 mbh->b_rdev = conf->mirrors[i].dev; 678 mbh->b_rsector = bh->b_rsector; 679 mbh->b_state = (1<<BH_Req) | (1<<BH_Dirty) | 680 (1<<BH_Mapped) | (1<<BH_Lock); 681 682 atomic_set(&mbh->b_count, 1); 683 mbh->b_size = bh->b_size; 684 mbh->b_page = bh->b_page; 685 mbh->b_data = bh->b_data; 686 mbh->b_list = BUF_LOCKED; 687 mbh->b_end_io = raid1_end_request; 688 mbh->b_private = r1_bh; 689 690 mbh->b_next = r1_bh->mirror_bh_list; 691 r1_bh->mirror_bh_list = mbh; 692 sum_bhs++; 693 } 694 if (bhl) raid1_free_bh(conf,bhl); 695 if (!sum_bhs) { 696 /* Gag - all mirrors non-operational.. */ 697 raid1_end_bh_io(r1_bh, 0); 698 return 0; 699 } 700 md_atomic_set(&r1_bh->remaining, sum_bhs); 701 702 /* 703 * We have to be a bit careful about the semaphore above, thats 704 * why we start the requests separately. Since kmalloc() could 705 * fail, sleep and make_request() can sleep too, this is the 706 * safer solution. Imagine, end_request decreasing the semaphore 707 * before we could have set it up ... We could play tricks with 708 * the semaphore (presetting it and correcting at the end if 709 * sum_bhs is not 'n' but we have to do end_request by hand if 710 * all requests finish until we had a chance to set up the 711 * semaphore correctly ... lots of races). 712 */ 713 bh = r1_bh->mirror_bh_list; 714 while(bh) { 715 struct buffer_head *bh2 = bh; 716 bh = bh->b_next; 717 generic_make_request(rw, bh2); 718 } 719 return (0); 720} 721 722static int raid1_status (char *page, mddev_t *mddev) 723{ 724 raid1_conf_t *conf = mddev_to_conf(mddev); 725 int sz = 0, i; 726 727 sz += sprintf (page+sz, " [%d/%d] [", conf->raid_disks, 728 conf->working_disks); 729 for (i = 0; i < conf->raid_disks; i++) 730 sz += sprintf (page+sz, "%s", 731 conf->mirrors[i].operational ? "U" : "_"); 732 sz += sprintf (page+sz, "]"); 733 return sz; 734} 735 736#define LAST_DISK KERN_ALERT \ 737"raid1: only one disk left and IO error.\n" 738 739#define NO_SPARE_DISK KERN_ALERT \ 740"raid1: no spare disk left, degrading mirror level by one.\n" 741 742#define DISK_FAILED KERN_ALERT \ 743"raid1: Disk failure on %s, disabling device. \n" \ 744" Operation continuing on %d devices\n" 745 746#define START_SYNCING KERN_ALERT \ 747"raid1: start syncing spare disk.\n" 748 749#define ALREADY_SYNCING KERN_INFO \ 750"raid1: syncing already in progress.\n" 751 752static void mark_disk_bad (mddev_t *mddev, int failed) 753{ 754 raid1_conf_t *conf = mddev_to_conf(mddev); 755 struct mirror_info *mirror = conf->mirrors+failed; 756 mdp_super_t *sb = mddev->sb; 757 758 mirror->operational = 0; 759 mark_disk_faulty(sb->disks+mirror->number); 760 mark_disk_nonsync(sb->disks+mirror->number); 761 mark_disk_inactive(sb->disks+mirror->number); 762 if (!mirror->write_only) 763 sb->active_disks--; 764 sb->working_disks--; 765 sb->failed_disks++; 766 mddev->sb_dirty = 1; 767 md_wakeup_thread(conf->thread); 768 if (!mirror->write_only) 769 conf->working_disks--; 770 printk (DISK_FAILED, partition_name (mirror->dev), 771 conf->working_disks); 772} 773 774static int raid1_error (mddev_t *mddev, kdev_t dev) 775{ 776 raid1_conf_t *conf = mddev_to_conf(mddev); 777 struct mirror_info * mirrors = conf->mirrors; 778 int disks = MD_SB_DISKS; 779 int i; 780 781 /* Find the drive. 782 * If it is not operational, then we have already marked it as dead 783 * else if it is the last working disks, ignore the error, let the 784 * next level up know. 785 * else mark the drive as failed 786 */ 787 788 for (i = 0; i < disks; i++) 789 if (mirrors[i].dev==dev && mirrors[i].operational) 790 break; 791 if (i == disks) 792 return 0; 793 794 if (i < conf->raid_disks && conf->working_disks == 1) { 795 /* Don't fail the drive, act as though we were just a 796 * normal single drive 797 */ 798 799 return 1; 800 } 801 mark_disk_bad(mddev, i); 802 return 0; 803} 804 805#undef LAST_DISK 806#undef NO_SPARE_DISK 807#undef DISK_FAILED 808#undef START_SYNCING 809 810 811static void print_raid1_conf (raid1_conf_t *conf) 812{ 813 int i; 814 struct mirror_info *tmp; 815 816 printk("RAID1 conf printout:\n"); 817 if (!conf) { 818 printk("(conf==NULL)\n"); 819 return; 820 } 821 printk(" --- wd:%d rd:%d nd:%d\n", conf->working_disks, 822 conf->raid_disks, conf->nr_disks); 823 824 for (i = 0; i < MD_SB_DISKS; i++) { 825 tmp = conf->mirrors + i; 826 printk(" disk %d, s:%d, o:%d, n:%d rd:%d us:%d dev:%s\n", 827 i, tmp->spare,tmp->operational, 828 tmp->number,tmp->raid_disk,tmp->used_slot, 829 partition_name(tmp->dev)); 830 } 831} 832 833static void close_sync(raid1_conf_t *conf) 834{ 835 mddev_t *mddev = conf->mddev; 836 /* If reconstruction was interrupted, we need to close the "active" and "pending" 837 * holes. 838 * we know that there are no active rebuild requests, os cnt_active == cnt_ready ==0 839 */ 840 /* this is really needed when recovery stops too... */ 841 spin_lock_irq(&conf->segment_lock); 842 conf->start_active = conf->start_pending; 843 conf->start_ready = conf->start_pending; 844 wait_event_lock_irq(conf->wait_ready, !conf->cnt_pending, conf->segment_lock); 845 conf->start_active =conf->start_ready = conf->start_pending = conf->start_future; 846 conf->start_future = (mddev->sb->size<<1)+1; 847 conf->cnt_pending = conf->cnt_future; 848 conf->cnt_future = 0; 849 conf->phase = conf->phase ^1; 850 wait_event_lock_irq(conf->wait_ready, !conf->cnt_pending, conf->segment_lock); 851 conf->start_active = conf->start_ready = conf->start_pending = conf->start_future = 0; 852 conf->phase = 0; 853 conf->cnt_future = conf->cnt_done;; 854 conf->cnt_done = 0; 855 spin_unlock_irq(&conf->segment_lock); 856 wake_up(&conf->wait_done); 857} 858 859static int raid1_diskop(mddev_t *mddev, mdp_disk_t **d, int state) 860{ 861 int err = 0; 862 int i, failed_disk=-1, spare_disk=-1, removed_disk=-1, added_disk=-1; 863 raid1_conf_t *conf = mddev->private; 864 struct mirror_info *tmp, *sdisk, *fdisk, *rdisk, *adisk; 865 mdp_super_t *sb = mddev->sb; 866 mdp_disk_t *failed_desc, *spare_desc, *added_desc; 867 mdk_rdev_t *spare_rdev, *failed_rdev; 868 869 print_raid1_conf(conf); 870 871 switch (state) { 872 case DISKOP_SPARE_ACTIVE: 873 case DISKOP_SPARE_INACTIVE: 874 /* need to wait for pending sync io before locking device */ 875 close_sync(conf); 876 } 877 878 md_spin_lock_irq(&conf->device_lock); 879 /* 880 * find the disk ... 881 */ 882 switch (state) { 883 884 case DISKOP_SPARE_ACTIVE: 885 886 /* 887 * Find the failed disk within the RAID1 configuration ... 888 * (this can only be in the first conf->working_disks part) 889 */ 890 for (i = 0; i < conf->raid_disks; i++) { 891 tmp = conf->mirrors + i; 892 if ((!tmp->operational && !tmp->spare) || 893 !tmp->used_slot) { 894 failed_disk = i; 895 break; 896 } 897 } 898 /* 899 * When we activate a spare disk we _must_ have a disk in 900 * the lower (active) part of the array to replace. 901 */ 902 if ((failed_disk == -1) || (failed_disk >= conf->raid_disks)) { 903 MD_BUG(); 904 err = 1; 905 goto abort; 906 } 907 /* fall through */ 908 909 case DISKOP_SPARE_WRITE: 910 case DISKOP_SPARE_INACTIVE: 911 912 /* 913 * Find the spare disk ... (can only be in the 'high' 914 * area of the array) 915 */ 916 for (i = conf->raid_disks; i < MD_SB_DISKS; i++) { 917 tmp = conf->mirrors + i; 918 if (tmp->spare && tmp->number == (*d)->number) { 919 spare_disk = i; 920 break; 921 } 922 } 923 if (spare_disk == -1) { 924 MD_BUG(); 925 err = 1; 926 goto abort; 927 } 928 break; 929 930 case DISKOP_HOT_REMOVE_DISK: 931 932 for (i = 0; i < MD_SB_DISKS; i++) { 933 tmp = conf->mirrors + i; 934 if (tmp->used_slot && (tmp->number == (*d)->number)) { 935 if (tmp->operational) { 936 err = -EBUSY; 937 goto abort; 938 } 939 removed_disk = i; 940 break; 941 } 942 } 943 if (removed_disk == -1) { 944 MD_BUG(); 945 err = 1; 946 goto abort; 947 } 948 break; 949 950 case DISKOP_HOT_ADD_DISK: 951 952 for (i = conf->raid_disks; i < MD_SB_DISKS; i++) { 953 tmp = conf->mirrors + i; 954 if (!tmp->used_slot) { 955 added_disk = i; 956 break; 957 } 958 } 959 if (added_disk == -1) { 960 MD_BUG(); 961 err = 1; 962 goto abort; 963 } 964 break; 965 } 966 967 switch (state) { 968 /* 969 * Switch the spare disk to write-only mode: 970 */ 971 case DISKOP_SPARE_WRITE: 972 sdisk = conf->mirrors + spare_disk; 973 sdisk->operational = 1; 974 sdisk->write_only = 1; 975 break; 976 /* 977 * Deactivate a spare disk: 978 */ 979 case DISKOP_SPARE_INACTIVE: 980 if (conf->start_future > 0) { 981 MD_BUG(); 982 err = -EBUSY; 983 break; 984 } 985 sdisk = conf->mirrors + spare_disk; 986 sdisk->operational = 0; 987 sdisk->write_only = 0; 988 break; 989 /* 990 * Activate (mark read-write) the (now sync) spare disk, 991 * which means we switch it's 'raid position' (->raid_disk) 992 * with the failed disk. (only the first 'conf->nr_disks' 993 * slots are used for 'real' disks and we must preserve this 994 * property) 995 */ 996 case DISKOP_SPARE_ACTIVE: 997 if (conf->start_future > 0) { 998 MD_BUG(); 999 err = -EBUSY; 1000 break; 1001 } 1002 sdisk = conf->mirrors + spare_disk; 1003 fdisk = conf->mirrors + failed_disk; 1004 1005 spare_desc = &sb->disks[sdisk->number]; 1006 failed_desc = &sb->disks[fdisk->number]; 1007 1008 if (spare_desc != *d) { 1009 MD_BUG(); 1010 err = 1; 1011 goto abort; 1012 } 1013 1014 if (spare_desc->raid_disk != sdisk->raid_disk) { 1015 MD_BUG(); 1016 err = 1; 1017 goto abort; 1018 } 1019 1020 if (sdisk->raid_disk != spare_disk) { 1021 MD_BUG(); 1022 err = 1; 1023 goto abort; 1024 } 1025 1026 if (failed_desc->raid_disk != fdisk->raid_disk) { 1027 MD_BUG(); 1028 err = 1; 1029 goto abort; 1030 } 1031 1032 if (fdisk->raid_disk != failed_disk) { 1033 MD_BUG(); 1034 err = 1; 1035 goto abort; 1036 } 1037 1038 /* 1039 * do the switch finally 1040 */ 1041 spare_rdev = find_rdev_nr(mddev, spare_desc->number); 1042 failed_rdev = find_rdev_nr(mddev, failed_desc->number); 1043 1044 /* There must be a spare_rdev, but there may not be a 1045 * failed_rdev. That slot might be empty... 1046 */ 1047 spare_rdev->desc_nr = failed_desc->number; 1048 if (failed_rdev) 1049 failed_rdev->desc_nr = spare_desc->number; 1050 1051 xchg_values(*spare_desc, *failed_desc); 1052 xchg_values(*fdisk, *sdisk); 1053 1054 /* 1055 * (careful, 'failed' and 'spare' are switched from now on) 1056 * 1057 * we want to preserve linear numbering and we want to 1058 * give the proper raid_disk number to the now activated 1059 * disk. (this means we switch back these values) 1060 */ 1061 1062 xchg_values(spare_desc->raid_disk, failed_desc->raid_disk); 1063 xchg_values(sdisk->raid_disk, fdisk->raid_disk); 1064 xchg_values(spare_desc->number, failed_desc->number); 1065 xchg_values(sdisk->number, fdisk->number); 1066 1067 *d = failed_desc; 1068 1069 if (sdisk->dev == MKDEV(0,0)) 1070 sdisk->used_slot = 0; 1071 /* 1072 * this really activates the spare. 1073 */ 1074 fdisk->spare = 0; 1075 fdisk->write_only = 0; 1076 1077 /* 1078 * if we activate a spare, we definitely replace a 1079 * non-operational disk slot in the 'low' area of 1080 * the disk array. 1081 */ 1082 1083 conf->working_disks++; 1084 1085 break; 1086 1087 case DISKOP_HOT_REMOVE_DISK: 1088 rdisk = conf->mirrors + removed_disk; 1089 1090 if (rdisk->spare && (removed_disk < conf->raid_disks)) { 1091 MD_BUG(); 1092 err = 1; 1093 goto abort; 1094 } 1095 rdisk->dev = MKDEV(0,0); 1096 rdisk->used_slot = 0; 1097 conf->nr_disks--; 1098 break; 1099 1100 case DISKOP_HOT_ADD_DISK: 1101 adisk = conf->mirrors + added_disk; 1102 added_desc = *d; 1103 1104 if (added_disk != added_desc->number) { 1105 MD_BUG(); 1106 err = 1; 1107 goto abort; 1108 } 1109 1110 adisk->number = added_desc->number; 1111 adisk->raid_disk = added_desc->raid_disk; 1112 adisk->dev = MKDEV(added_desc->major,added_desc->minor); 1113 1114 adisk->operational = 0; 1115 adisk->write_only = 0; 1116 adisk->spare = 1; 1117 adisk->used_slot = 1; 1118 adisk->head_position = 0; 1119 conf->nr_disks++; 1120 1121 break; 1122 1123 default: 1124 MD_BUG(); 1125 err = 1; 1126 goto abort; 1127 } 1128abort: 1129 md_spin_unlock_irq(&conf->device_lock); 1130 if (state == DISKOP_SPARE_ACTIVE || state == DISKOP_SPARE_INACTIVE) 1131 /* should move to "END_REBUILD" when such exists */ 1132 raid1_shrink_buffers(conf); 1133 1134 print_raid1_conf(conf); 1135 return err; 1136} 1137 1138 1139#define IO_ERROR KERN_ALERT \ 1140"raid1: %s: unrecoverable I/O read error for block %lu\n" 1141 1142#define REDIRECT_SECTOR KERN_ERR \ 1143"raid1: %s: redirecting sector %lu to another mirror\n" 1144 1145/* 1146 * This is a kernel thread which: 1147 * 1148 * 1. Retries failed read operations on working mirrors. 1149 * 2. Updates the raid superblock when problems encounter. 1150 * 3. Performs writes following reads for array syncronising. 1151 */ 1152static void end_sync_write(struct buffer_head *bh, int uptodate); 1153static void end_sync_read(struct buffer_head *bh, int uptodate); 1154 1155static void raid1d (void *data) 1156{ 1157 struct raid1_bh *r1_bh; 1158 struct buffer_head *bh; 1159 unsigned long flags; 1160 raid1_conf_t *conf = data; 1161 mddev_t *mddev = conf->mddev; 1162 kdev_t dev; 1163 1164 if (mddev->sb_dirty) 1165 md_update_sb(mddev); 1166 1167 for (;;) { 1168 md_spin_lock_irqsave(&retry_list_lock, flags); 1169 r1_bh = raid1_retry_list; 1170 if (!r1_bh) 1171 break; 1172 raid1_retry_list = r1_bh->next_r1; 1173 md_spin_unlock_irqrestore(&retry_list_lock, flags); 1174 1175 mddev = r1_bh->mddev; 1176 bh = &r1_bh->bh_req; 1177 switch(r1_bh->cmd) { 1178 case SPECIAL: 1179 /* have to allocate lots of bh structures and 1180 * schedule writes 1181 */ 1182 if (test_bit(R1BH_Uptodate, &r1_bh->state)) { 1183 int i, sum_bhs = 0; 1184 int disks = MD_SB_DISKS; 1185 struct buffer_head *bhl, *mbh; 1186 1187 conf = mddev_to_conf(mddev); 1188 bhl = raid1_alloc_bh(conf, conf->raid_disks); /* don't really need this many */ 1189 for (i = 0; i < disks ; i++) { 1190 if (!conf->mirrors[i].operational) 1191 continue; 1192 if (i==conf->last_used) 1193 /* we read from here, no need to write */ 1194 continue; 1195 if (i < conf->raid_disks 1196 && !conf->resync_mirrors) 1197 /* don't need to write this, 1198 * we are just rebuilding */ 1199 continue; 1200 mbh = bhl; 1201 if (!mbh) { 1202 MD_BUG(); 1203 break; 1204 } 1205 bhl = mbh->b_next; 1206 mbh->b_this_page = (struct buffer_head *)1; 1207 1208 1209 /* 1210 * prepare mirrored bh (fields ordered for max mem throughput): 1211 */ 1212 mbh->b_blocknr = bh->b_blocknr; 1213 mbh->b_dev = conf->mirrors[i].dev; 1214 mbh->b_rdev = conf->mirrors[i].dev; 1215 mbh->b_rsector = bh->b_blocknr; 1216 mbh->b_state = (1<<BH_Req) | (1<<BH_Dirty) | 1217 (1<<BH_Mapped) | (1<<BH_Lock); 1218 atomic_set(&mbh->b_count, 1); 1219 mbh->b_size = bh->b_size; 1220 mbh->b_page = bh->b_page; 1221 mbh->b_data = bh->b_data; 1222 mbh->b_list = BUF_LOCKED; 1223 mbh->b_end_io = end_sync_write; 1224 mbh->b_private = r1_bh; 1225 1226 mbh->b_next = r1_bh->mirror_bh_list; 1227 r1_bh->mirror_bh_list = mbh; 1228 1229 sum_bhs++; 1230 } 1231 md_atomic_set(&r1_bh->remaining, sum_bhs); 1232 if (bhl) raid1_free_bh(conf, bhl); 1233 mbh = r1_bh->mirror_bh_list; 1234 1235 if (!sum_bhs) { 1236 /* nowhere to write this too... I guess we 1237 * must be done 1238 */ 1239 sync_request_done(bh->b_blocknr, conf); 1240 md_done_sync(mddev, bh->b_size>>9, 0); 1241 raid1_free_buf(r1_bh); 1242 } else 1243 while (mbh) { 1244 struct buffer_head *bh1 = mbh; 1245 mbh = mbh->b_next; 1246 generic_make_request(WRITE, bh1); 1247 md_sync_acct(bh1->b_dev, bh1->b_size/512); 1248 } 1249 } else { 1250 /* There is no point trying a read-for-reconstruct 1251 * as reconstruct is about to be aborted 1252 */ 1253 1254 printk (IO_ERROR, partition_name(bh->b_dev), bh->b_blocknr); 1255 md_done_sync(mddev, bh->b_size>>9, 0); 1256 } 1257 1258 break; 1259 case READ: 1260 case READA: 1261 dev = bh->b_dev; 1262 raid1_map (mddev, &bh->b_dev); 1263 if (bh->b_dev == dev) { 1264 printk (IO_ERROR, partition_name(bh->b_dev), bh->b_blocknr); 1265 raid1_end_bh_io(r1_bh, 0); 1266 } else { 1267 printk (REDIRECT_SECTOR, 1268 partition_name(bh->b_dev), bh->b_blocknr); 1269 bh->b_rdev = bh->b_dev; 1270 bh->b_rsector = bh->b_blocknr; 1271 generic_make_request (r1_bh->cmd, bh); 1272 } 1273 break; 1274 } 1275 } 1276 md_spin_unlock_irqrestore(&retry_list_lock, flags); 1277} 1278#undef IO_ERROR 1279#undef REDIRECT_SECTOR 1280 1281/* 1282 * Private kernel thread to reconstruct mirrors after an unclean 1283 * shutdown. 1284 */ 1285static void raid1syncd (void *data) 1286{ 1287 raid1_conf_t *conf = data; 1288 mddev_t *mddev = conf->mddev; 1289 1290 if (!conf->resync_mirrors) 1291 return; 1292 if (conf->resync_mirrors == 2) 1293 return; 1294 down(&mddev->recovery_sem); 1295 if (!md_do_sync(mddev, NULL)) { 1296 /* 1297 * Only if everything went Ok. 1298 */ 1299 conf->resync_mirrors = 0; 1300 } 1301 1302 close_sync(conf); 1303 1304 up(&mddev->recovery_sem); 1305 raid1_shrink_buffers(conf); 1306} 1307 1308/* 1309 * perform a "sync" on one "block" 1310 * 1311 * We need to make sure that no normal I/O request - particularly write 1312 * requests - conflict with active sync requests. 1313 * This is achieved by conceptually dividing the device space into a 1314 * number of sections: 1315 * DONE: 0 .. a-1 These blocks are in-sync 1316 * ACTIVE: a.. b-1 These blocks may have active sync requests, but 1317 * no normal IO requests 1318 * READY: b .. c-1 These blocks have no normal IO requests - sync 1319 * request may be happening 1320 * PENDING: c .. d-1 These blocks may have IO requests, but no new 1321 * ones will be added 1322 * FUTURE: d .. end These blocks are not to be considered yet. IO may 1323 * be happening, but not sync 1324 * 1325 * We keep a 1326 * phase which flips (0 or 1) each time d moves and 1327 * a count of: 1328 * z = active io requests in FUTURE since d moved - marked with 1329 * current phase 1330 * y = active io requests in FUTURE before d moved, or PENDING - 1331 * marked with previous phase 1332 * x = active sync requests in READY 1333 * w = active sync requests in ACTIVE 1334 * v = active io requests in DONE 1335 * 1336 * Normally, a=b=c=d=0 and z= active io requests 1337 * or a=b=c=d=END and v= active io requests 1338 * Allowed changes to a,b,c,d: 1339 * A: c==d && y==0 -> d+=window, y=z, z=0, phase=!phase 1340 * B: y==0 -> c=d 1341 * C: b=c, w+=x, x=0 1342 * D: w==0 -> a=b 1343 * E: a==b==c==d==end -> a=b=c=d=0, z=v, v=0 1344 * 1345 * At start of sync we apply A. 1346 * When y reaches 0, we apply B then A then being sync requests 1347 * When sync point reaches c-1, we wait for y==0, and W==0, and 1348 * then apply apply B then A then D then C. 1349 * Finally, we apply E 1350 * 1351 * The sync request simply issues a "read" against a working drive 1352 * This is marked so that on completion the raid1d thread is woken to 1353 * issue suitable write requests 1354 */ 1355 1356static int raid1_sync_request (mddev_t *mddev, unsigned long sector_nr) 1357{ 1358 raid1_conf_t *conf = mddev_to_conf(mddev); 1359 struct mirror_info *mirror; 1360 struct raid1_bh *r1_bh; 1361 struct buffer_head *bh; 1362 int bsize; 1363 int disk; 1364 int block_nr; 1365 int buffs; 1366 1367 if (!sector_nr) { 1368 /* we want enough buffers to hold twice the window of 128*/ 1369 buffs = 128 *2 / (PAGE_SIZE>>9); 1370 buffs = raid1_grow_buffers(conf, buffs); 1371 if (buffs < 2) 1372 goto nomem; 1373 conf->window = buffs*(PAGE_SIZE>>9)/2; 1374 } 1375 spin_lock_irq(&conf->segment_lock); 1376 if (!sector_nr) { 1377 /* initialize ...*/ 1378 conf->start_active = 0; 1379 conf->start_ready = 0; 1380 conf->start_pending = 0; 1381 conf->start_future = 0; 1382 conf->phase = 0; 1383 1384 conf->cnt_future += conf->cnt_done+conf->cnt_pending; 1385 conf->cnt_done = conf->cnt_pending = 0; 1386 if (conf->cnt_ready || conf->cnt_active) 1387 MD_BUG(); 1388 } 1389 while (sector_nr >= conf->start_pending) { 1390 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", 1391 sector_nr, conf->start_active, conf->start_ready, conf->start_pending, conf->start_future, 1392 conf->cnt_done, conf->cnt_active, conf->cnt_ready, conf->cnt_pending, conf->cnt_future); 1393 wait_event_lock_irq(conf->wait_done, 1394 !conf->cnt_active, 1395 conf->segment_lock); 1396 wait_event_lock_irq(conf->wait_ready, 1397 !conf->cnt_pending, 1398 conf->segment_lock); 1399 conf->start_active = conf->start_ready; 1400 conf->start_ready = conf->start_pending; 1401 conf->start_pending = conf->start_future; 1402 conf->start_future = conf->start_future+conf->window; 1403 // Note: falling off the end is not a problem 1404 conf->phase = conf->phase ^1; 1405 conf->cnt_active = conf->cnt_ready; 1406 conf->cnt_ready = 0; 1407 conf->cnt_pending = conf->cnt_future; 1408 conf->cnt_future = 0; 1409 wake_up(&conf->wait_done); 1410 } 1411 conf->cnt_ready++; 1412 spin_unlock_irq(&conf->segment_lock); 1413 1414 1415 /* If reconstructing, and >1 working disc, 1416 * could dedicate one to rebuild and others to 1417 * service read requests .. 1418 */ 1419 disk = conf->last_used; 1420 /* make sure disk is operational */ 1421 while (!conf->mirrors[disk].operational) { 1422 if (disk <= 0) disk = conf->raid_disks; 1423 disk--; 1424 if (disk == conf->last_used) 1425 break; 1426 } 1427 conf->last_used = disk; 1428 1429 mirror = conf->mirrors+conf->last_used; 1430 1431 r1_bh = raid1_alloc_buf (conf); 1432 r1_bh->master_bh = NULL; 1433 r1_bh->mddev = mddev; 1434 r1_bh->cmd = SPECIAL; 1435 bh = &r1_bh->bh_req; 1436 1437 block_nr = sector_nr; 1438 bsize = 512; 1439 while (!(block_nr & 1) && bsize < PAGE_SIZE 1440 && (block_nr+2)*(bsize>>9) < (mddev->sb->size *2)) { 1441 block_nr >>= 1; 1442 bsize <<= 1; 1443 } 1444 bh->b_size = bsize; 1445 bh->b_list = BUF_LOCKED; 1446 bh->b_dev = mirror->dev; 1447 bh->b_rdev = mirror->dev; 1448 bh->b_state = (1<<BH_Req) | (1<<BH_Mapped) | (1<<BH_Lock); 1449 if (!bh->b_page) 1450 BUG(); 1451 if (!bh->b_data) 1452 BUG(); 1453 if (bh->b_data != page_address(bh->b_page)) 1454 BUG(); 1455 bh->b_end_io = end_sync_read; 1456 bh->b_private = r1_bh; 1457 bh->b_blocknr = sector_nr; 1458 bh->b_rsector = sector_nr; 1459 init_waitqueue_head(&bh->b_wait); 1460 1461 generic_make_request(READ, bh); 1462 md_sync_acct(bh->b_dev, bh->b_size/512); 1463 1464 return (bsize >> 9); 1465 1466nomem: 1467 raid1_shrink_buffers(conf); 1468 return -ENOMEM; 1469} 1470 1471static void end_sync_read(struct buffer_head *bh, int uptodate) 1472{ 1473 struct raid1_bh * r1_bh = (struct raid1_bh *)(bh->b_private); 1474 1475 /* we have read a block, now it needs to be re-written, 1476 * or re-read if the read failed. 1477 * We don't do much here, just schedule handling by raid1d 1478 */ 1479 if (!uptodate) 1480 md_error (r1_bh->mddev, bh->b_dev); 1481 else 1482 set_bit(R1BH_Uptodate, &r1_bh->state); 1483 raid1_reschedule_retry(r1_bh); 1484} 1485 1486static void end_sync_write(struct buffer_head *bh, int uptodate) 1487{ 1488 struct raid1_bh * r1_bh = (struct raid1_bh *)(bh->b_private); 1489 1490 if (!uptodate) 1491 md_error (r1_bh->mddev, bh->b_dev); 1492 if (atomic_dec_and_test(&r1_bh->remaining)) { 1493 mddev_t *mddev = r1_bh->mddev; 1494 unsigned long sect = bh->b_blocknr; 1495 int size = bh->b_size; 1496 raid1_free_buf(r1_bh); 1497 sync_request_done(sect, mddev_to_conf(mddev)); 1498 md_done_sync(mddev,size>>9, uptodate); 1499 } 1500} 1501 1502#define INVALID_LEVEL KERN_WARNING \ 1503"raid1: md%d: raid level not set to mirroring (%d)\n" 1504 1505#define NO_SB KERN_ERR \ 1506"raid1: disabled mirror %s (couldn't access raid superblock)\n" 1507 1508#define ERRORS KERN_ERR \ 1509"raid1: disabled mirror %s (errors detected)\n" 1510 1511#define NOT_IN_SYNC KERN_ERR \ 1512"raid1: disabled mirror %s (not in sync)\n" 1513 1514#define INCONSISTENT KERN_ERR \ 1515"raid1: disabled mirror %s (inconsistent descriptor)\n" 1516 1517#define ALREADY_RUNNING KERN_ERR \ 1518"raid1: disabled mirror %s (mirror %d already operational)\n" 1519 1520#define OPERATIONAL KERN_INFO \ 1521"raid1: device %s operational as mirror %d\n" 1522 1523#define MEM_ERROR KERN_ERR \ 1524"raid1: couldn't allocate memory for md%d\n" 1525 1526#define SPARE KERN_INFO \ 1527"raid1: spare disk %s\n" 1528 1529#define NONE_OPERATIONAL KERN_ERR \ 1530"raid1: no operational mirrors for md%d\n" 1531 1532#define ARRAY_IS_ACTIVE KERN_INFO \ 1533"raid1: raid set md%d active with %d out of %d mirrors\n" 1534 1535#define THREAD_ERROR KERN_ERR \ 1536"raid1: couldn't allocate thread for md%d\n" 1537 1538#define START_RESYNC KERN_WARNING \ 1539"raid1: raid set md%d not clean; reconstructing mirrors\n" 1540 1541static int raid1_run (mddev_t *mddev) 1542{ 1543 raid1_conf_t *conf; 1544 int i, j, disk_idx; 1545 struct mirror_info *disk; 1546 mdp_super_t *sb = mddev->sb; 1547 mdp_disk_t *descriptor; 1548 mdk_rdev_t *rdev; 1549 struct md_list_head *tmp; 1550 int start_recovery = 0; 1551 1552 MOD_INC_USE_COUNT; 1553 1554 if (sb->level != 1) { 1555 printk(INVALID_LEVEL, mdidx(mddev), sb->level); 1556 goto out; 1557 } 1558 /* 1559 * copy the already verified devices into our private RAID1 1560 * bookkeeping area. [whatever we allocate in raid1_run(), 1561 * should be freed in raid1_stop()] 1562 */ 1563 1564 conf = kmalloc(sizeof(raid1_conf_t), GFP_KERNEL); 1565 mddev->private = conf; 1566 if (!conf) { 1567 printk(MEM_ERROR, mdidx(mddev)); 1568 goto out; 1569 } 1570 memset(conf, 0, sizeof(*conf)); 1571 1572 ITERATE_RDEV(mddev,rdev,tmp) { 1573 if (rdev->faulty) { 1574 printk(ERRORS, partition_name(rdev->dev)); 1575 } else { 1576 if (!rdev->sb) { 1577 MD_BUG(); 1578 continue; 1579 } 1580 } 1581 if (rdev->desc_nr == -1) { 1582 MD_BUG(); 1583 continue; 1584 } 1585 descriptor = &sb->disks[rdev->desc_nr]; 1586 disk_idx = descriptor->raid_disk; 1587 disk = conf->mirrors + disk_idx; 1588 1589 if (disk_faulty(descriptor)) { 1590 disk->number = descriptor->number; 1591 disk->raid_disk = disk_idx; 1592 disk->dev = rdev->dev; 1593 disk->sect_limit = MAX_WORK_PER_DISK; 1594 disk->operational = 0; 1595 disk->write_only = 0; 1596 disk->spare = 0; 1597 disk->used_slot = 1; 1598 disk->head_position = 0; 1599 continue; 1600 } 1601 if (disk_active(descriptor)) { 1602 if (!disk_sync(descriptor)) { 1603 printk(NOT_IN_SYNC, 1604 partition_name(rdev->dev)); 1605 continue; 1606 } 1607 if ((descriptor->number > MD_SB_DISKS) || 1608 (disk_idx > sb->raid_disks)) { 1609 1610 printk(INCONSISTENT, 1611 partition_name(rdev->dev)); 1612 continue; 1613 } 1614 if (disk->operational) { 1615 printk(ALREADY_RUNNING, 1616 partition_name(rdev->dev), 1617 disk_idx); 1618 continue; 1619 } 1620 printk(OPERATIONAL, partition_name(rdev->dev), 1621 disk_idx); 1622 disk->number = descriptor->number; 1623 disk->raid_disk = disk_idx; 1624 disk->dev = rdev->dev; 1625 disk->sect_limit = MAX_WORK_PER_DISK; 1626 disk->operational = 1; 1627 disk->write_only = 0; 1628 disk->spare = 0; 1629 disk->used_slot = 1; 1630 disk->head_position = 0; 1631 conf->working_disks++; 1632 } else { 1633 /* 1634 * Must be a spare disk .. 1635 */ 1636 printk(SPARE, partition_name(rdev->dev)); 1637 disk->number = descriptor->number; 1638 disk->raid_disk = disk_idx; 1639 disk->dev = rdev->dev; 1640 disk->sect_limit = MAX_WORK_PER_DISK; 1641 disk->operational = 0; 1642 disk->write_only = 0; 1643 disk->spare = 1; 1644 disk->used_slot = 1; 1645 disk->head_position = 0; 1646 } 1647 } 1648 conf->raid_disks = sb->raid_disks; 1649 conf->nr_disks = sb->nr_disks; 1650 conf->mddev = mddev; 1651 conf->device_lock = MD_SPIN_LOCK_UNLOCKED; 1652 1653 conf->segment_lock = MD_SPIN_LOCK_UNLOCKED; 1654 init_waitqueue_head(&conf->wait_buffer); 1655 init_waitqueue_head(&conf->wait_done); 1656 init_waitqueue_head(&conf->wait_ready); 1657 1658 if (!conf->working_disks) { 1659 printk(NONE_OPERATIONAL, mdidx(mddev)); 1660 goto out_free_conf; 1661 } 1662 1663 1664 /* pre-allocate some buffer_head structures. 1665 * As a minimum, 1 r1bh and raid_disks buffer_heads 1666 * would probably get us by in tight memory situations, 1667 * but a few more is probably a good idea. 1668 * For now, try NR_RESERVED_BUFS r1bh and 1669 * NR_RESERVED_BUFS*raid_disks bufferheads 1670 * This will allow at least NR_RESERVED_BUFS concurrent 1671 * reads or writes even if kmalloc starts failing 1672 */ 1673 if (raid1_grow_r1bh(conf, NR_RESERVED_BUFS) < NR_RESERVED_BUFS || 1674 raid1_grow_bh(conf, NR_RESERVED_BUFS*conf->raid_disks) 1675 < NR_RESERVED_BUFS*conf->raid_disks) { 1676 printk(MEM_ERROR, mdidx(mddev)); 1677 goto out_free_conf; 1678 } 1679 1680 for (i = 0; i < MD_SB_DISKS; i++) { 1681 1682 descriptor = sb->disks+i; 1683 disk_idx = descriptor->raid_disk; 1684 disk = conf->mirrors + disk_idx; 1685 1686 if (disk_faulty(descriptor) && (disk_idx < conf->raid_disks) && 1687 !disk->used_slot) { 1688 1689 disk->number = descriptor->number; 1690 disk->raid_disk = disk_idx; 1691 disk->dev = MKDEV(0,0); 1692 1693 disk->operational = 0; 1694 disk->write_only = 0; 1695 disk->spare = 0; 1696 disk->used_slot = 1; 1697 disk->head_position = 0; 1698 } 1699 } 1700 1701 /* 1702 * find the first working one and use it as a starting point 1703 * to read balancing. 1704 */ 1705 for (j = 0; !conf->mirrors[j].operational && j < MD_SB_DISKS; j++) 1706 /* nothing */; 1707 conf->last_used = j; 1708 1709 1710 if (conf->working_disks != sb->raid_disks) { 1711 printk(KERN_ALERT "raid1: md%d, not all disks are operational -- trying to recover array\n", mdidx(mddev)); 1712 start_recovery = 1; 1713 } 1714 1715 { 1716 const char * name = "raid1d"; 1717 1718 conf->thread = md_register_thread(raid1d, conf, name); 1719 if (!conf->thread) { 1720 printk(THREAD_ERROR, mdidx(mddev)); 1721 goto out_free_conf; 1722 } 1723 } 1724 1725 if (!start_recovery && !(sb->state & (1 << MD_SB_CLEAN)) && 1726 (conf->working_disks > 1)) { 1727 const char * name = "raid1syncd"; 1728 1729 conf->resync_thread = md_register_thread(raid1syncd, conf,name); 1730 if (!conf->resync_thread) { 1731 printk(THREAD_ERROR, mdidx(mddev)); 1732 goto out_free_conf; 1733 } 1734 1735 printk(START_RESYNC, mdidx(mddev)); 1736 conf->resync_mirrors = 1; 1737 md_wakeup_thread(conf->resync_thread); 1738 } 1739 1740 /* 1741 * Regenerate the "device is in sync with the raid set" bit for 1742 * each device. 1743 */ 1744 for (i = 0; i < MD_SB_DISKS; i++) { 1745 mark_disk_nonsync(sb->disks+i); 1746 for (j = 0; j < sb->raid_disks; j++) { 1747 if (!conf->mirrors[j].operational) 1748 continue; 1749 if (sb->disks[i].number == conf->mirrors[j].number) 1750 mark_disk_sync(sb->disks+i); 1751 } 1752 } 1753 sb->active_disks = conf->working_disks; 1754 1755 if (start_recovery) 1756 md_recover_arrays(); 1757 1758 1759 printk(ARRAY_IS_ACTIVE, mdidx(mddev), sb->active_disks, sb->raid_disks); 1760 /* 1761 * Ok, everything is just fine now 1762 */ 1763 return 0; 1764 1765out_free_conf: 1766 raid1_shrink_r1bh(conf); 1767 raid1_shrink_bh(conf); 1768 raid1_shrink_buffers(conf); 1769 kfree(conf); 1770 mddev->private = NULL; 1771out: 1772 MOD_DEC_USE_COUNT; 1773 return -EIO; 1774} 1775 1776#undef INVALID_LEVEL 1777#undef NO_SB 1778#undef ERRORS 1779#undef NOT_IN_SYNC 1780#undef INCONSISTENT 1781#undef ALREADY_RUNNING 1782#undef OPERATIONAL 1783#undef SPARE 1784#undef NONE_OPERATIONAL 1785#undef ARRAY_IS_ACTIVE 1786 1787static int raid1_stop_resync (mddev_t *mddev) 1788{ 1789 raid1_conf_t *conf = mddev_to_conf(mddev); 1790 1791 if (conf->resync_thread) { 1792 if (conf->resync_mirrors) { 1793 conf->resync_mirrors = 2; 1794 md_interrupt_thread(conf->resync_thread); 1795 1796 printk(KERN_INFO "raid1: mirror resync was not fully finished, restarting next time.\n"); 1797 return 1; 1798 } 1799 return 0; 1800 } 1801 return 0; 1802} 1803 1804static int raid1_restart_resync (mddev_t *mddev) 1805{ 1806 raid1_conf_t *conf = mddev_to_conf(mddev); 1807 1808 if (conf->resync_mirrors) { 1809 if (!conf->resync_thread) { 1810 MD_BUG(); 1811 return 0; 1812 } 1813 conf->resync_mirrors = 1; 1814 md_wakeup_thread(conf->resync_thread); 1815 return 1; 1816 } 1817 return 0; 1818} 1819 1820static int raid1_stop (mddev_t *mddev) 1821{ 1822 raid1_conf_t *conf = mddev_to_conf(mddev); 1823 1824 md_unregister_thread(conf->thread); 1825 if (conf->resync_thread) 1826 md_unregister_thread(conf->resync_thread); 1827 raid1_shrink_r1bh(conf); 1828 raid1_shrink_bh(conf); 1829 raid1_shrink_buffers(conf); 1830 kfree(conf); 1831 mddev->private = NULL; 1832 MOD_DEC_USE_COUNT; 1833 return 0; 1834} 1835 1836static mdk_personality_t raid1_personality= 1837{ 1838 name: "raid1", 1839 make_request: raid1_make_request, 1840 run: raid1_run, 1841 stop: raid1_stop, 1842 status: raid1_status, 1843 error_handler: raid1_error, 1844 diskop: raid1_diskop, 1845 stop_resync: raid1_stop_resync, 1846 restart_resync: raid1_restart_resync, 1847 sync_request: raid1_sync_request 1848}; 1849 1850static int md__init raid1_init (void) 1851{ 1852 return register_md_personality (RAID1, &raid1_personality); 1853} 1854 1855static void raid1_exit (void) 1856{ 1857 unregister_md_personality (RAID1); 1858} 1859 1860module_init(raid1_init); 1861module_exit(raid1_exit); 1862MODULE_LICENSE("GPL"); 1863