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