drivers/md/raid5-cache.c

   1 /*
   2  * Copyright (C) 2015 Shaohua Li <shli@fb.com>
   3  *
   4  * This program is free software; you can redistribute it and/or modify it
   5  * under the terms and conditions of the GNU General Public License,
   6  * version 2, as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope it will be useful, but WITHOUT
   9  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  10  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  11  * more details.
  12  *
  13  */
  14 #include <linux/kernel.h>
  15 #include <linux/wait.h>
  16 #include <linux/blkdev.h>
  17 #include <linux/slab.h>
  18 #include <linux/raid/md_p.h>
  19 #include <linux/crc32c.h>
  20 #include <linux/random.h>
  21 #include "md.h"
  22 #include "raid5.h"
  23
  24 /*
  25  * metadata/data stored in disk with 4k size unit (a block) regardless
  26  * underneath hardware sector size. only works with PAGE_SIZE == 4096
  27  */
  28 #define BLOCK_SECTORS (8)
  29
  30 /*
  31  * reclaim runs every 1/4 disk size or 10G reclaimable space. This can prevent
  32  * recovery scans a very long log
  33  */
  34 #define RECLAIM_MAX_FREE_SPACE (10 * 1024 * 1024 * 2) /* sector */
  35 #define RECLAIM_MAX_FREE_SPACE_SHIFT (2)
  36
  37 struct r5l_log {
  38         struct md_rdev *rdev;
  39
  40         u32 uuid_checksum;
  41
  42         sector_t device_size;           /* log device size, round to
  43                                          * BLOCK_SECTORS */
  44         sector_t max_free_space;        /* reclaim run if free space is at
  45                                          * this size */
  46
  47         sector_t last_checkpoint;       /* log tail. where recovery scan
  48                                          * starts from */
  49         u64 last_cp_seq;                /* log tail sequence */
  50
  51         sector_t log_start;             /* log head. where new data appends */
  52         u64 seq;                        /* log head sequence */
  53
  54         sector_t next_checkpoint;
  55         u64 next_cp_seq;
  56
  57         struct mutex io_mutex;
  58         struct r5l_io_unit *current_io; /* current io_unit accepting new data */
  59
  60         spinlock_t io_list_lock;
  61         struct list_head running_ios;   /* io_units which are still running,
  62                                          * and have not yet been completely
  63                                          * written to the log */
  64         struct list_head io_end_ios;    /* io_units which have been completely
  65                                          * written to the log but not yet written
  66                                          * to the RAID */
  67         struct list_head flushing_ios;  /* io_units which are waiting for log
  68                                          * cache flush */
  69         struct list_head finished_ios;  /* io_units which settle down in log disk */
  70         struct bio flush_bio;
  71
  72         struct kmem_cache *io_kc;
  73
  74         struct md_thread *reclaim_thread;
  75         unsigned long reclaim_target;   /* number of space that need to be
  76                                          * reclaimed.  if it's 0, reclaim spaces
  77                                          * used by io_units which are in
  78                                          * IO_UNIT_STRIPE_END state (eg, reclaim
  79                                          * dones't wait for specific io_unit
  80                                          * switching to IO_UNIT_STRIPE_END
  81                                          * state) */
  82         wait_queue_head_t iounit_wait;
  83
  84         struct list_head no_space_stripes; /* pending stripes, log has no space */
  85         spinlock_t no_space_stripes_lock;
  86
  87         bool need_cache_flush;
  88 };
  89
  90 /*
  91  * an IO range starts from a meta data block and end at the next meta data
  92  * block. The io unit's the meta data block tracks data/parity followed it. io
  93  * unit is written to log disk with normal write, as we always flush log disk
  94  * first and then start move data to raid disks, there is no requirement to
  95  * write io unit with FLUSH/FUA
  96  */
  97 struct r5l_io_unit {
  98         struct r5l_log *log;
  99
 100         struct page *meta_page; /* store meta block */
 101         int meta_offset;        /* current offset in meta_page */
 102
 103         struct bio_list bios;
 104         atomic_t pending_io;    /* pending bios not written to log yet */
 105         struct bio *current_bio;/* current_bio accepting new data */
 106
 107         atomic_t pending_stripe;/* how many stripes not flushed to raid */
 108         u64 seq;                /* seq number of the metablock */
 109         sector_t log_start;     /* where the io_unit starts */
 110         sector_t log_end;       /* where the io_unit ends */
 111         struct list_head log_sibling; /* log->running_ios */
 112         struct list_head stripe_list; /* stripes added to the io_unit */
 113
 114         int state;
 115 };
 116
 117 /* r5l_io_unit state */
 118 enum r5l_io_unit_state {
 119         IO_UNIT_RUNNING = 0,    /* accepting new IO */
 120         IO_UNIT_IO_START = 1,   /* io_unit bio start writing to log,
 121                                  * don't accepting new bio */
 122         IO_UNIT_IO_END = 2,     /* io_unit bio finish writing to log */
 123         IO_UNIT_STRIPE_END = 3, /* stripes data finished writing to raid */
 124 };
 125
 126 static sector_t r5l_ring_add(struct r5l_log *log, sector_t start, sector_t inc)
 127 {
 128         start += inc;
 129         if (start >= log->device_size)
 130                 start = start - log->device_size;
 131         return start;
 132 }
 133
 134 static sector_t r5l_ring_distance(struct r5l_log *log, sector_t start,
 135                                   sector_t end)
 136 {
 137         if (end >= start)
 138                 return end - start;
 139         else
 140                 return end + log->device_size - start;
 141 }
 142
 143 static bool r5l_has_free_space(struct r5l_log *log, sector_t size)
 144 {
 145         sector_t used_size;
 146
 147         used_size = r5l_ring_distance(log, log->last_checkpoint,
 148                                         log->log_start);
 149
 150         return log->device_size > used_size + size;
 151 }
 152
 153 static void r5l_free_io_unit(struct r5l_log *log, struct r5l_io_unit *io)
 154 {
 155         __free_page(io->meta_page);
 156         kmem_cache_free(log->io_kc, io);
 157 }
 158
 159 static void r5l_move_io_unit_list(struct list_head *from, struct list_head *to,
 160                                   enum r5l_io_unit_state state)
 161 {
 162         struct r5l_io_unit *io;
 163
 164         while (!list_empty(from)) {
 165                 io = list_first_entry(from, struct r5l_io_unit, log_sibling);
 166                 /* don't change list order */
 167                 if (io->state >= state)
 168                         list_move_tail(&io->log_sibling, to);
 169                 else
 170                         break;
 171         }
 172 }
 173
 174 static void __r5l_set_io_unit_state(struct r5l_io_unit *io,
 175                                     enum r5l_io_unit_state state)
 176 {
 177         if (WARN_ON(io->state >= state))
 178                 return;
 179         io->state = state;
 180 }
 181
 182 static void r5l_io_run_stripes(struct r5l_io_unit *io)
 183 {
 184         struct stripe_head *sh, *next;
 185
 186         list_for_each_entry_safe(sh, next, &io->stripe_list, log_list) {
 187                 list_del_init(&sh->log_list);
 188                 set_bit(STRIPE_HANDLE, &sh->state);
 189                 raid5_release_stripe(sh);
 190         }
 191 }
 192
 193 /* XXX: totally ignores I/O errors */
 194 static void r5l_log_run_stripes(struct r5l_log *log)
 195 {
 196         struct r5l_io_unit *io, *next;
 197
 198         assert_spin_locked(&log->io_list_lock);
 199
 200         list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) {
 201                 /* don't change list order */
 202                 if (io->state < IO_UNIT_IO_END)
 203                         break;
 204
 205                 list_move_tail(&io->log_sibling, &log->finished_ios);
 206                 r5l_io_run_stripes(io);
 207         }
 208 }
 209
 210 static void r5l_log_endio(struct bio *bio)
 211 {
 212         struct r5l_io_unit *io = bio->bi_private;
 213         struct r5l_log *log = io->log;
 214         unsigned long flags;
 215
 216         bio_put(bio);
 217
 218         if (!atomic_dec_and_test(&io->pending_io))
 219                 return;
 220
 221         spin_lock_irqsave(&log->io_list_lock, flags);
 222         __r5l_set_io_unit_state(io, IO_UNIT_IO_END);
 223         if (log->need_cache_flush)
 224                 r5l_move_io_unit_list(&log->running_ios, &log->io_end_ios,
 225                                       IO_UNIT_IO_END);
 226         else
 227                 r5l_log_run_stripes(log);
 228         spin_unlock_irqrestore(&log->io_list_lock, flags);
 229
 230         if (log->need_cache_flush)
 231                 md_wakeup_thread(log->rdev->mddev->thread);
 232 }
 233
 234 static void r5l_submit_current_io(struct r5l_log *log)
 235 {
 236         struct r5l_io_unit *io = log->current_io;
 237         struct r5l_meta_block *block;
 238         struct bio *bio;
 239         unsigned long flags;
 240         u32 crc;
 241
 242         if (!io)
 243                 return;
 244
 245         block = page_address(io->meta_page);
 246         block->meta_size = cpu_to_le32(io->meta_offset);
 247         crc = crc32c_le(log->uuid_checksum, block, PAGE_SIZE);
 248         block->checksum = cpu_to_le32(crc);
 249
 250         log->current_io = NULL;
 251         spin_lock_irqsave(&log->io_list_lock, flags);
 252         __r5l_set_io_unit_state(io, IO_UNIT_IO_START);
 253         spin_unlock_irqrestore(&log->io_list_lock, flags);
 254
 255         while ((bio = bio_list_pop(&io->bios)))
 256                 submit_bio(WRITE, bio);
 257 }
 258
 259 static struct bio *r5l_bio_alloc(struct r5l_log *log, struct r5l_io_unit *io)
 260 {
 261         struct bio *bio = bio_kmalloc(GFP_NOIO | __GFP_NOFAIL, BIO_MAX_PAGES);
 262
 263         bio->bi_rw = WRITE;
 264         bio->bi_bdev = log->rdev->bdev;
 265         bio->bi_iter.bi_sector = log->rdev->data_offset + log->log_start;
 266         bio->bi_end_io = r5l_log_endio;
 267         bio->bi_private = io;
 268
 269         bio_list_add(&io->bios, bio);
 270         atomic_inc(&io->pending_io);
 271         return bio;
 272 }
 273
 274 static void r5_reserve_log_entry(struct r5l_log *log, struct r5l_io_unit *io)
 275 {
 276         log->log_start = r5l_ring_add(log, log->log_start, BLOCK_SECTORS);
 277
 278         /*
 279          * If we filled up the log device start from the beginning again,
 280          * which will require a new bio.
 281          *
 282          * Note: for this to work properly the log size needs to me a multiple
 283          * of BLOCK_SECTORS.
 284          */
 285         if (log->log_start == 0)
 286                 io->current_bio = NULL;
 287
 288         io->log_end = log->log_start;
 289 }
 290
 291 static struct r5l_io_unit *r5l_new_meta(struct r5l_log *log)
 292 {
 293         struct r5l_io_unit *io;
 294         struct r5l_meta_block *block;
 295
 296         /* We can't handle memory allocate failure so far */
 297         io = kmem_cache_zalloc(log->io_kc, GFP_NOIO | __GFP_NOFAIL);
 298         io->log = log;
 299         bio_list_init(&io->bios);
 300         INIT_LIST_HEAD(&io->log_sibling);
 301         INIT_LIST_HEAD(&io->stripe_list);
 302         io->state = IO_UNIT_RUNNING;
 303
 304         io->meta_page = alloc_page(GFP_NOIO | __GFP_NOFAIL | __GFP_ZERO);
 305         block = page_address(io->meta_page);
 306         block->magic = cpu_to_le32(R5LOG_MAGIC);
 307         block->version = R5LOG_VERSION;
 308         block->seq = cpu_to_le64(log->seq);
 309         block->position = cpu_to_le64(log->log_start);
 310
 311         io->log_start = log->log_start;
 312         io->meta_offset = sizeof(struct r5l_meta_block);
 313         io->seq = log->seq++;
 314
 315         io->current_bio = r5l_bio_alloc(log, io);
 316         bio_add_page(io->current_bio, io->meta_page, PAGE_SIZE, 0);
 317
 318         r5_reserve_log_entry(log, io);
 319
 320         spin_lock_irq(&log->io_list_lock);
 321         list_add_tail(&io->log_sibling, &log->running_ios);
 322         spin_unlock_irq(&log->io_list_lock);
 323
 324         return io;
 325 }
 326
 327 static int r5l_get_meta(struct r5l_log *log, unsigned int payload_size)
 328 {
 329         if (log->current_io &&
 330             log->current_io->meta_offset + payload_size > PAGE_SIZE)
 331                 r5l_submit_current_io(log);
 332
 333         if (!log->current_io)
 334                 log->current_io = r5l_new_meta(log);
 335         return 0;
 336 }
 337
 338 static void r5l_append_payload_meta(struct r5l_log *log, u16 type,
 339                                     sector_t location,
 340                                     u32 checksum1, u32 checksum2,
 341                                     bool checksum2_valid)
 342 {
 343         struct r5l_io_unit *io = log->current_io;
 344         struct r5l_payload_data_parity *payload;
 345
 346         payload = page_address(io->meta_page) + io->meta_offset;
 347         payload->header.type = cpu_to_le16(type);
 348         payload->header.flags = cpu_to_le16(0);
 349         payload->size = cpu_to_le32((1 + !!checksum2_valid) <<
 350                                     (PAGE_SHIFT - 9));
 351         payload->location = cpu_to_le64(location);
 352         payload->checksum[0] = cpu_to_le32(checksum1);
 353         if (checksum2_valid)
 354                 payload->checksum[1] = cpu_to_le32(checksum2);
 355
 356         io->meta_offset += sizeof(struct r5l_payload_data_parity) +
 357                 sizeof(__le32) * (1 + !!checksum2_valid);
 358 }
 359
 360 static void r5l_append_payload_page(struct r5l_log *log, struct page *page)
 361 {
 362         struct r5l_io_unit *io = log->current_io;
 363
 364 alloc_bio:
 365         if (!io->current_bio)
 366                 io->current_bio = r5l_bio_alloc(log, io);
 367
 368         if (!bio_add_page(io->current_bio, page, PAGE_SIZE, 0)) {
 369                 io->current_bio = NULL;
 370                 goto alloc_bio;
 371         }
 372
 373         r5_reserve_log_entry(log, io);
 374 }
 375
 376 static void r5l_log_stripe(struct r5l_log *log, struct stripe_head *sh,
 377                            int data_pages, int parity_pages)
 378 {
 379         int i;
 380         int meta_size;
 381         struct r5l_io_unit *io;
 382
 383         meta_size =
 384                 ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32))
 385                  * data_pages) +
 386                 sizeof(struct r5l_payload_data_parity) +
 387                 sizeof(__le32) * parity_pages;
 388
 389         r5l_get_meta(log, meta_size);
 390         io = log->current_io;
 391
 392         for (i = 0; i < sh->disks; i++) {
 393                 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags))
 394                         continue;
 395                 if (i == sh->pd_idx || i == sh->qd_idx)
 396                         continue;
 397                 r5l_append_payload_meta(log, R5LOG_PAYLOAD_DATA,
 398                                         raid5_compute_blocknr(sh, i, 0),
 399                                         sh->dev[i].log_checksum, 0, false);
 400                 r5l_append_payload_page(log, sh->dev[i].page);
 401         }
 402
 403         if (sh->qd_idx >= 0) {
 404                 r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
 405                                         sh->sector, sh->dev[sh->pd_idx].log_checksum,
 406                                         sh->dev[sh->qd_idx].log_checksum, true);
 407                 r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
 408                 r5l_append_payload_page(log, sh->dev[sh->qd_idx].page);
 409         } else {
 410                 r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
 411                                         sh->sector, sh->dev[sh->pd_idx].log_checksum,
 412                                         0, false);
 413                 r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
 414         }
 415
 416         list_add_tail(&sh->log_list, &io->stripe_list);
 417         atomic_inc(&io->pending_stripe);
 418         sh->log_io = io;
 419 }
 420
 421 static void r5l_wake_reclaim(struct r5l_log *log, sector_t space);
 422 /*
 423  * running in raid5d, where reclaim could wait for raid5d too (when it flushes
 424  * data from log to raid disks), so we shouldn't wait for reclaim here
 425  */
 426 int r5l_write_stripe(struct r5l_log *log, struct stripe_head *sh)
 427 {
 428         int write_disks = 0;
 429         int data_pages, parity_pages;
 430         int meta_size;
 431         int reserve;
 432         int i;
 433
 434         if (!log)
 435                 return -EAGAIN;
 436         /* Don't support stripe batch */
 437         if (sh->log_io || !test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags) ||
 438             test_bit(STRIPE_SYNCING, &sh->state)) {
 439                 /* the stripe is written to log, we start writing it to raid */
 440                 clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
 441                 return -EAGAIN;
 442         }
 443
 444         for (i = 0; i < sh->disks; i++) {
 445                 void *addr;
 446
 447                 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags))
 448                         continue;
 449                 write_disks++;
 450                 /* checksum is already calculated in last run */
 451                 if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
 452                         continue;
 453                 addr = kmap_atomic(sh->dev[i].page);
 454                 sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum,
 455                                                     addr, PAGE_SIZE);
 456                 kunmap_atomic(addr);
 457         }
 458         parity_pages = 1 + !!(sh->qd_idx >= 0);
 459         data_pages = write_disks - parity_pages;
 460
 461         meta_size =
 462                 ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32))
 463                  * data_pages) +
 464                 sizeof(struct r5l_payload_data_parity) +
 465                 sizeof(__le32) * parity_pages;
 466         /* Doesn't work with very big raid array */
 467         if (meta_size + sizeof(struct r5l_meta_block) > PAGE_SIZE)
 468                 return -EINVAL;
 469
 470         set_bit(STRIPE_LOG_TRAPPED, &sh->state);
 471         /*
 472          * The stripe must enter state machine again to finish the write, so
 473          * don't delay.
 474          */
 475         clear_bit(STRIPE_DELAYED, &sh->state);
 476         atomic_inc(&sh->count);
 477
 478         mutex_lock(&log->io_mutex);
 479         /* meta + data */
 480         reserve = (1 + write_disks) << (PAGE_SHIFT - 9);
 481         if (r5l_has_free_space(log, reserve))
 482                 r5l_log_stripe(log, sh, data_pages, parity_pages);
 483         else {
 484                 spin_lock(&log->no_space_stripes_lock);
 485                 list_add_tail(&sh->log_list, &log->no_space_stripes);
 486                 spin_unlock(&log->no_space_stripes_lock);
 487
 488                 r5l_wake_reclaim(log, reserve);
 489         }
 490         mutex_unlock(&log->io_mutex);
 491
 492         return 0;
 493 }
 494
 495 void r5l_write_stripe_run(struct r5l_log *log)
 496 {
 497         if (!log)
 498                 return;
 499         mutex_lock(&log->io_mutex);
 500         r5l_submit_current_io(log);
 501         mutex_unlock(&log->io_mutex);
 502 }
 503
 504 int r5l_handle_flush_request(struct r5l_log *log, struct bio *bio)
 505 {
 506         if (!log)
 507                 return -ENODEV;
 508         /*
 509          * we flush log disk cache first, then write stripe data to raid disks.
 510          * So if bio is finished, the log disk cache is flushed already. The
 511          * recovery guarantees we can recovery the bio from log disk, so we
 512          * don't need to flush again
 513          */
 514         if (bio->bi_iter.bi_size == 0) {
 515                 bio_endio(bio);
 516                 return 0;
 517         }
 518         bio->bi_rw &= ~REQ_FLUSH;
 519         return -EAGAIN;
 520 }
 521
 522 /* This will run after log space is reclaimed */
 523 static void r5l_run_no_space_stripes(struct r5l_log *log)
 524 {
 525         struct stripe_head *sh;
 526
 527         spin_lock(&log->no_space_stripes_lock);
 528         while (!list_empty(&log->no_space_stripes)) {
 529                 sh = list_first_entry(&log->no_space_stripes,
 530                                       struct stripe_head, log_list);
 531                 list_del_init(&sh->log_list);
 532                 set_bit(STRIPE_HANDLE, &sh->state);
 533                 raid5_release_stripe(sh);
 534         }
 535         spin_unlock(&log->no_space_stripes_lock);
 536 }
 537
 538 static sector_t r5l_reclaimable_space(struct r5l_log *log)
 539 {
 540         return r5l_ring_distance(log, log->last_checkpoint,
 541                                  log->next_checkpoint);
 542 }
 543
 544 static bool r5l_complete_finished_ios(struct r5l_log *log)
 545 {
 546         struct r5l_io_unit *io, *next;
 547         bool found = false;
 548
 549         assert_spin_locked(&log->io_list_lock);
 550
 551         list_for_each_entry_safe(io, next, &log->finished_ios, log_sibling) {
 552                 /* don't change list order */
 553                 if (io->state < IO_UNIT_STRIPE_END)
 554                         break;
 555
 556                 log->next_checkpoint = io->log_start;
 557                 log->next_cp_seq = io->seq;
 558
 559                 list_del(&io->log_sibling);
 560                 r5l_free_io_unit(log, io);
 561
 562                 found = true;
 563         }
 564
 565         return found;
 566 }
 567
 568 static void __r5l_stripe_write_finished(struct r5l_io_unit *io)
 569 {
 570         struct r5l_log *log = io->log;
 571         unsigned long flags;
 572
 573         spin_lock_irqsave(&log->io_list_lock, flags);
 574         __r5l_set_io_unit_state(io, IO_UNIT_STRIPE_END);
 575
 576         if (!r5l_complete_finished_ios(log)) {
 577                 spin_unlock_irqrestore(&log->io_list_lock, flags);
 578                 return;
 579         }
 580
 581         if (r5l_reclaimable_space(log) > log->max_free_space)
 582                 r5l_wake_reclaim(log, 0);
 583
 584         spin_unlock_irqrestore(&log->io_list_lock, flags);
 585         wake_up(&log->iounit_wait);
 586 }
 587
 588 void r5l_stripe_write_finished(struct stripe_head *sh)
 589 {
 590         struct r5l_io_unit *io;
 591
 592         io = sh->log_io;
 593         sh->log_io = NULL;
 594
 595         if (io && atomic_dec_and_test(&io->pending_stripe))
 596                 __r5l_stripe_write_finished(io);
 597 }
 598
 599 static void r5l_log_flush_endio(struct bio *bio)
 600 {
 601         struct r5l_log *log = container_of(bio, struct r5l_log,
 602                 flush_bio);
 603         unsigned long flags;
 604         struct r5l_io_unit *io;
 605
 606         spin_lock_irqsave(&log->io_list_lock, flags);
 607         list_for_each_entry(io, &log->flushing_ios, log_sibling)
 608                 r5l_io_run_stripes(io);
 609         list_splice_tail_init(&log->flushing_ios, &log->finished_ios);
 610         spin_unlock_irqrestore(&log->io_list_lock, flags);
 611 }
 612
 613 /*
 614  * Starting dispatch IO to raid.
 615  * io_unit(meta) consists of a log. There is one situation we want to avoid. A
 616  * broken meta in the middle of a log causes recovery can't find meta at the
 617  * head of log. If operations require meta at the head persistent in log, we
 618  * must make sure meta before it persistent in log too. A case is:
 619  *
 620  * stripe data/parity is in log, we start write stripe to raid disks. stripe
 621  * data/parity must be persistent in log before we do the write to raid disks.
 622  *
 623  * The solution is we restrictly maintain io_unit list order. In this case, we
 624  * only write stripes of an io_unit to raid disks till the io_unit is the first
 625  * one whose data/parity is in log.
 626  */
 627 void r5l_flush_stripe_to_raid(struct r5l_log *log)
 628 {
 629         bool do_flush;
 630
 631         if (!log || !log->need_cache_flush)
 632                 return;
 633
 634         spin_lock_irq(&log->io_list_lock);
 635         /* flush bio is running */
 636         if (!list_empty(&log->flushing_ios)) {
 637                 spin_unlock_irq(&log->io_list_lock);
 638                 return;
 639         }
 640         list_splice_tail_init(&log->io_end_ios, &log->flushing_ios);
 641         do_flush = !list_empty(&log->flushing_ios);
 642         spin_unlock_irq(&log->io_list_lock);
 643
 644         if (!do_flush)
 645                 return;
 646         bio_reset(&log->flush_bio);
 647         log->flush_bio.bi_bdev = log->rdev->bdev;
 648         log->flush_bio.bi_end_io = r5l_log_flush_endio;
 649         submit_bio(WRITE_FLUSH, &log->flush_bio);
 650 }
 651
 652 static void r5l_write_super(struct r5l_log *log, sector_t cp);
 653 static void r5l_do_reclaim(struct r5l_log *log)
 654 {
 655         sector_t reclaim_target = xchg(&log->reclaim_target, 0);
 656         sector_t reclaimable;
 657         sector_t next_checkpoint;
 658         u64 next_cp_seq;
 659
 660         spin_lock_irq(&log->io_list_lock);
 661         /*
 662          * move proper io_unit to reclaim list. We should not change the order.
 663          * reclaimable/unreclaimable io_unit can be mixed in the list, we
 664          * shouldn't reuse space of an unreclaimable io_unit
 665          */
 666         while (1) {
 667                 reclaimable = r5l_reclaimable_space(log);
 668                 if (reclaimable >= reclaim_target ||
 669                     (list_empty(&log->running_ios) &&
 670                      list_empty(&log->io_end_ios) &&
 671                      list_empty(&log->flushing_ios) &&
 672                      list_empty(&log->finished_ios)))
 673                         break;
 674
 675                 md_wakeup_thread(log->rdev->mddev->thread);
 676                 wait_event_lock_irq(log->iounit_wait,
 677                                     r5l_reclaimable_space(log) > reclaimable,
 678                                     log->io_list_lock);
 679         }
 680
 681         next_checkpoint = log->next_checkpoint;
 682         next_cp_seq = log->next_cp_seq;
 683         spin_unlock_irq(&log->io_list_lock);
 684
 685         BUG_ON(reclaimable < 0);
 686         if (reclaimable == 0)
 687                 return;
 688
 689         /*
 690          * write_super will flush cache of each raid disk. We must write super
 691          * here, because the log area might be reused soon and we don't want to
 692          * confuse recovery
 693          */
 694         r5l_write_super(log, next_checkpoint);
 695
 696         mutex_lock(&log->io_mutex);
 697         log->last_checkpoint = next_checkpoint;
 698         log->last_cp_seq = next_cp_seq;
 699         mutex_unlock(&log->io_mutex);
 700
 701         r5l_run_no_space_stripes(log);
 702 }
 703
 704 static void r5l_reclaim_thread(struct md_thread *thread)
 705 {
 706         struct mddev *mddev = thread->mddev;
 707         struct r5conf *conf = mddev->private;
 708         struct r5l_log *log = conf->log;
 709
 710         if (!log)
 711                 return;
 712         r5l_do_reclaim(log);
 713 }
 714
 715 static void r5l_wake_reclaim(struct r5l_log *log, sector_t space)
 716 {
 717         unsigned long target;
 718         unsigned long new = (unsigned long)space; /* overflow in theory */
 719
 720         do {
 721                 target = log->reclaim_target;
 722                 if (new < target)
 723                         return;
 724         } while (cmpxchg(&log->reclaim_target, target, new) != target);
 725         md_wakeup_thread(log->reclaim_thread);
 726 }
 727
 728 void r5l_quiesce(struct r5l_log *log, int state)
 729 {
 730         if (!log || state == 2)
 731                 return;
 732         if (state == 0) {
 733                 log->reclaim_thread = md_register_thread(r5l_reclaim_thread,
 734                                         log->rdev->mddev, "reclaim");
 735         } else if (state == 1) {
 736                 /*
 737                  * at this point all stripes are finished, so io_unit is at
 738                  * least in STRIPE_END state
 739                  */
 740                 r5l_wake_reclaim(log, -1L);
 741                 md_unregister_thread(&log->reclaim_thread);
 742                 r5l_do_reclaim(log);
 743         }
 744 }
 745
 746 struct r5l_recovery_ctx {
 747         struct page *meta_page;         /* current meta */
 748         sector_t meta_total_blocks;     /* total size of current meta and data */
 749         sector_t pos;                   /* recovery position */
 750         u64 seq;                        /* recovery position seq */
 751 };
 752
 753 static int r5l_read_meta_block(struct r5l_log *log,
 754                                struct r5l_recovery_ctx *ctx)
 755 {
 756         struct page *page = ctx->meta_page;
 757         struct r5l_meta_block *mb;
 758         u32 crc, stored_crc;
 759
 760         if (!sync_page_io(log->rdev, ctx->pos, PAGE_SIZE, page, READ, false))
 761                 return -EIO;
 762
 763         mb = page_address(page);
 764         stored_crc = le32_to_cpu(mb->checksum);
 765         mb->checksum = 0;
 766
 767         if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
 768             le64_to_cpu(mb->seq) != ctx->seq ||
 769             mb->version != R5LOG_VERSION ||
 770             le64_to_cpu(mb->position) != ctx->pos)
 771                 return -EINVAL;
 772
 773         crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
 774         if (stored_crc != crc)
 775                 return -EINVAL;
 776
 777         if (le32_to_cpu(mb->meta_size) > PAGE_SIZE)
 778                 return -EINVAL;
 779
 780         ctx->meta_total_blocks = BLOCK_SECTORS;
 781
 782         return 0;
 783 }
 784
 785 static int r5l_recovery_flush_one_stripe(struct r5l_log *log,
 786                                          struct r5l_recovery_ctx *ctx,
 787                                          sector_t stripe_sect,
 788                                          int *offset, sector_t *log_offset)
 789 {
 790         struct r5conf *conf = log->rdev->mddev->private;
 791         struct stripe_head *sh;
 792         struct r5l_payload_data_parity *payload;
 793         int disk_index;
 794
 795         sh = raid5_get_active_stripe(conf, stripe_sect, 0, 0, 0);
 796         while (1) {
 797                 payload = page_address(ctx->meta_page) + *offset;
 798
 799                 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) {
 800                         raid5_compute_sector(conf,
 801                                              le64_to_cpu(payload->location), 0,
 802                                              &disk_index, sh);
 803
 804                         sync_page_io(log->rdev, *log_offset, PAGE_SIZE,
 805                                      sh->dev[disk_index].page, READ, false);
 806                         sh->dev[disk_index].log_checksum =
 807                                 le32_to_cpu(payload->checksum[0]);
 808                         set_bit(R5_Wantwrite, &sh->dev[disk_index].flags);
 809                         ctx->meta_total_blocks += BLOCK_SECTORS;
 810                 } else {
 811                         disk_index = sh->pd_idx;
 812                         sync_page_io(log->rdev, *log_offset, PAGE_SIZE,
 813                                      sh->dev[disk_index].page, READ, false);
 814                         sh->dev[disk_index].log_checksum =
 815                                 le32_to_cpu(payload->checksum[0]);
 816                         set_bit(R5_Wantwrite, &sh->dev[disk_index].flags);
 817
 818                         if (sh->qd_idx >= 0) {
 819                                 disk_index = sh->qd_idx;
 820                                 sync_page_io(log->rdev,
 821                                              r5l_ring_add(log, *log_offset, BLOCK_SECTORS),
 822                                              PAGE_SIZE, sh->dev[disk_index].page,
 823                                              READ, false);
 824                                 sh->dev[disk_index].log_checksum =
 825                                         le32_to_cpu(payload->checksum[1]);
 826                                 set_bit(R5_Wantwrite,
 827                                         &sh->dev[disk_index].flags);
 828                         }
 829                         ctx->meta_total_blocks += BLOCK_SECTORS * conf->max_degraded;
 830                 }
 831
 832                 *log_offset = r5l_ring_add(log, *log_offset,
 833                                            le32_to_cpu(payload->size));
 834                 *offset += sizeof(struct r5l_payload_data_parity) +
 835                         sizeof(__le32) *
 836                         (le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
 837                 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_PARITY)
 838                         break;
 839         }
 840
 841         for (disk_index = 0; disk_index < sh->disks; disk_index++) {
 842                 void *addr;
 843                 u32 checksum;
 844
 845                 if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
 846                         continue;
 847                 addr = kmap_atomic(sh->dev[disk_index].page);
 848                 checksum = crc32c_le(log->uuid_checksum, addr, PAGE_SIZE);
 849                 kunmap_atomic(addr);
 850                 if (checksum != sh->dev[disk_index].log_checksum)
 851                         goto error;
 852         }
 853
 854         for (disk_index = 0; disk_index < sh->disks; disk_index++) {
 855                 struct md_rdev *rdev, *rrdev;
 856
 857                 if (!test_and_clear_bit(R5_Wantwrite,
 858                                         &sh->dev[disk_index].flags))
 859                         continue;
 860
 861                 /* in case device is broken */
 862                 rdev = rcu_dereference(conf->disks[disk_index].rdev);
 863                 if (rdev)
 864                         sync_page_io(rdev, stripe_sect, PAGE_SIZE,
 865                                      sh->dev[disk_index].page, WRITE, false);
 866                 rrdev = rcu_dereference(conf->disks[disk_index].replacement);
 867                 if (rrdev)
 868                         sync_page_io(rrdev, stripe_sect, PAGE_SIZE,
 869                                      sh->dev[disk_index].page, WRITE, false);
 870         }
 871         raid5_release_stripe(sh);
 872         return 0;
 873
 874 error:
 875         for (disk_index = 0; disk_index < sh->disks; disk_index++)
 876                 sh->dev[disk_index].flags = 0;
 877         raid5_release_stripe(sh);
 878         return -EINVAL;
 879 }
 880
 881 static int r5l_recovery_flush_one_meta(struct r5l_log *log,
 882                                        struct r5l_recovery_ctx *ctx)
 883 {
 884         struct r5conf *conf = log->rdev->mddev->private;
 885         struct r5l_payload_data_parity *payload;
 886         struct r5l_meta_block *mb;
 887         int offset;
 888         sector_t log_offset;
 889         sector_t stripe_sector;
 890
 891         mb = page_address(ctx->meta_page);
 892         offset = sizeof(struct r5l_meta_block);
 893         log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
 894
 895         while (offset < le32_to_cpu(mb->meta_size)) {
 896                 int dd;
 897
 898                 payload = (void *)mb + offset;
 899                 stripe_sector = raid5_compute_sector(conf,
 900                                                      le64_to_cpu(payload->location), 0, &dd, NULL);
 901                 if (r5l_recovery_flush_one_stripe(log, ctx, stripe_sector,
 902                                                   &offset, &log_offset))
 903                         return -EINVAL;
 904         }
 905         return 0;
 906 }
 907
 908 /* copy data/parity from log to raid disks */
 909 static void r5l_recovery_flush_log(struct r5l_log *log,
 910                                    struct r5l_recovery_ctx *ctx)
 911 {
 912         while (1) {
 913                 if (r5l_read_meta_block(log, ctx))
 914                         return;
 915                 if (r5l_recovery_flush_one_meta(log, ctx))
 916                         return;
 917                 ctx->seq++;
 918                 ctx->pos = r5l_ring_add(log, ctx->pos, ctx->meta_total_blocks);
 919         }
 920 }
 921
 922 static int r5l_log_write_empty_meta_block(struct r5l_log *log, sector_t pos,
 923                                           u64 seq)
 924 {
 925         struct page *page;
 926         struct r5l_meta_block *mb;
 927         u32 crc;
 928
 929         page = alloc_page(GFP_KERNEL | __GFP_ZERO);
 930         if (!page)
 931                 return -ENOMEM;
 932         mb = page_address(page);
 933         mb->magic = cpu_to_le32(R5LOG_MAGIC);
 934         mb->version = R5LOG_VERSION;
 935         mb->meta_size = cpu_to_le32(sizeof(struct r5l_meta_block));
 936         mb->seq = cpu_to_le64(seq);
 937         mb->position = cpu_to_le64(pos);
 938         crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
 939         mb->checksum = cpu_to_le32(crc);
 940
 941         if (!sync_page_io(log->rdev, pos, PAGE_SIZE, page, WRITE_FUA, false)) {
 942                 __free_page(page);
 943                 return -EIO;
 944         }
 945         __free_page(page);
 946         return 0;
 947 }
 948
 949 static int r5l_recovery_log(struct r5l_log *log)
 950 {
 951         struct r5l_recovery_ctx ctx;
 952
 953         ctx.pos = log->last_checkpoint;
 954         ctx.seq = log->last_cp_seq;
 955         ctx.meta_page = alloc_page(GFP_KERNEL);
 956         if (!ctx.meta_page)
 957                 return -ENOMEM;
 958
 959         r5l_recovery_flush_log(log, &ctx);
 960         __free_page(ctx.meta_page);
 961
 962         /*
 963          * we did a recovery. Now ctx.pos points to an invalid meta block. New
 964          * log will start here. but we can't let superblock point to last valid
 965          * meta block. The log might looks like:
 966          * | meta 1| meta 2| meta 3|
 967          * meta 1 is valid, meta 2 is invalid. meta 3 could be valid. If
 968          * superblock points to meta 1, we write a new valid meta 2n.  if crash
 969          * happens again, new recovery will start from meta 1. Since meta 2n is
 970          * valid now, recovery will think meta 3 is valid, which is wrong.
 971          * The solution is we create a new meta in meta2 with its seq == meta
 972          * 1's seq + 10 and let superblock points to meta2. The same recovery will
 973          * not think meta 3 is a valid meta, because its seq doesn't match
 974          */
 975         if (ctx.seq > log->last_cp_seq + 1) {
 976                 int ret;
 977
 978                 ret = r5l_log_write_empty_meta_block(log, ctx.pos, ctx.seq + 10);
 979                 if (ret)
 980                         return ret;
 981                 log->seq = ctx.seq + 11;
 982                 log->log_start = r5l_ring_add(log, ctx.pos, BLOCK_SECTORS);
 983                 r5l_write_super(log, ctx.pos);
 984         } else {
 985                 log->log_start = ctx.pos;
 986                 log->seq = ctx.seq;
 987         }
 988         return 0;
 989 }
 990
 991 static void r5l_write_super(struct r5l_log *log, sector_t cp)
 992 {
 993         struct mddev *mddev = log->rdev->mddev;
 994
 995         log->rdev->journal_tail = cp;
 996         set_bit(MD_CHANGE_DEVS, &mddev->flags);
 997 }
 998
 999 static int r5l_load_log(struct r5l_log *log)
1000 {
1001         struct md_rdev *rdev = log->rdev;
1002         struct page *page;
1003         struct r5l_meta_block *mb;
1004         sector_t cp = log->rdev->journal_tail;
1005         u32 stored_crc, expected_crc;
1006         bool create_super = false;
1007         int ret;
1008
1009         /* Make sure it's valid */
1010         if (cp >= rdev->sectors || round_down(cp, BLOCK_SECTORS) != cp)
1011                 cp = 0;
1012         page = alloc_page(GFP_KERNEL);
1013         if (!page)
1014                 return -ENOMEM;
1015
1016         if (!sync_page_io(rdev, cp, PAGE_SIZE, page, READ, false)) {
1017                 ret = -EIO;
1018                 goto ioerr;
1019         }
1020         mb = page_address(page);
1021
1022         if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
1023             mb->version != R5LOG_VERSION) {
1024                 create_super = true;
1025                 goto create;
1026         }
1027         stored_crc = le32_to_cpu(mb->checksum);
1028         mb->checksum = 0;
1029         expected_crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
1030         if (stored_crc != expected_crc) {
1031                 create_super = true;
1032                 goto create;
1033         }
1034         if (le64_to_cpu(mb->position) != cp) {
1035                 create_super = true;
1036                 goto create;
1037         }
1038 create:
1039         if (create_super) {
1040                 log->last_cp_seq = prandom_u32();
1041                 cp = 0;
1042                 /*
1043                  * Make sure super points to correct address. Log might have
1044                  * data very soon. If super hasn't correct log tail address,
1045                  * recovery can't find the log
1046                  */
1047                 r5l_write_super(log, cp);
1048         } else
1049                 log->last_cp_seq = le64_to_cpu(mb->seq);
1050
1051         log->device_size = round_down(rdev->sectors, BLOCK_SECTORS);
1052         log->max_free_space = log->device_size >> RECLAIM_MAX_FREE_SPACE_SHIFT;
1053         if (log->max_free_space > RECLAIM_MAX_FREE_SPACE)
1054                 log->max_free_space = RECLAIM_MAX_FREE_SPACE;
1055         log->last_checkpoint = cp;
1056
1057         __free_page(page);
1058
1059         return r5l_recovery_log(log);
1060 ioerr:
1061         __free_page(page);
1062         return ret;
1063 }
1064
1065 int r5l_init_log(struct r5conf *conf, struct md_rdev *rdev)
1066 {
1067         struct r5l_log *log;
1068
1069         if (PAGE_SIZE != 4096)
1070                 return -EINVAL;
1071         log = kzalloc(sizeof(*log), GFP_KERNEL);
1072         if (!log)
1073                 return -ENOMEM;
1074         log->rdev = rdev;
1075
1076         log->need_cache_flush = (rdev->bdev->bd_disk->queue->flush_flags != 0);
1077
1078         log->uuid_checksum = crc32c_le(~0, rdev->mddev->uuid,
1079                                        sizeof(rdev->mddev->uuid));
1080
1081         mutex_init(&log->io_mutex);
1082
1083         spin_lock_init(&log->io_list_lock);
1084         INIT_LIST_HEAD(&log->running_ios);
1085         INIT_LIST_HEAD(&log->io_end_ios);
1086         INIT_LIST_HEAD(&log->flushing_ios);
1087         INIT_LIST_HEAD(&log->finished_ios);
1088         bio_init(&log->flush_bio);
1089
1090         log->io_kc = KMEM_CACHE(r5l_io_unit, 0);
1091         if (!log->io_kc)
1092                 goto io_kc;
1093
1094         log->reclaim_thread = md_register_thread(r5l_reclaim_thread,
1095                                                  log->rdev->mddev, "reclaim");
1096         if (!log->reclaim_thread)
1097                 goto reclaim_thread;
1098         init_waitqueue_head(&log->iounit_wait);
1099
1100         INIT_LIST_HEAD(&log->no_space_stripes);
1101         spin_lock_init(&log->no_space_stripes_lock);
1102
1103         if (r5l_load_log(log))
1104                 goto error;
1105
1106         conf->log = log;
1107         return 0;
1108 error:
1109         md_unregister_thread(&log->reclaim_thread);
1110 reclaim_thread:
1111         kmem_cache_destroy(log->io_kc);
1112 io_kc:
1113         kfree(log);
1114         return -EINVAL;
1115 }
1116
1117 void r5l_exit_log(struct r5l_log *log)
1118 {
1119         md_unregister_thread(&log->reclaim_thread);
1120         kmem_cache_destroy(log->io_kc);
1121         kfree(log);
1122 }