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         bool in_teardown;
  89 };
  90
  91 /*
  92  * an IO range starts from a meta data block and end at the next meta data
  93  * block. The io unit's the meta data block tracks data/parity followed it. io
  94  * unit is written to log disk with normal write, as we always flush log disk
  95  * first and then start move data to raid disks, there is no requirement to
  96  * write io unit with FLUSH/FUA
  97  */
  98 struct r5l_io_unit {
  99         struct r5l_log *log;
 100
 101         struct page *meta_page; /* store meta block */
 102         int meta_offset;        /* current offset in meta_page */
 103
 104         struct bio *current_bio;/* current_bio accepting new data */
 105
 106         atomic_t pending_stripe;/* how many stripes not flushed to raid */
 107         u64 seq;                /* seq number of the metablock */
 108         sector_t log_start;     /* where the io_unit starts */
 109         sector_t log_end;       /* where the io_unit ends */
 110         struct list_head log_sibling; /* log->running_ios */
 111         struct list_head stripe_list; /* stripes added to the io_unit */
 112
 113         int state;
 114         bool need_split_bio;
 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 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         if (bio->bi_error)
 216                 md_error(log->rdev->mddev, log->rdev);
 217
 218         bio_put(bio);
 219
 220         spin_lock_irqsave(&log->io_list_lock, flags);
 221         __r5l_set_io_unit_state(io, IO_UNIT_IO_END);
 222         if (log->need_cache_flush)
 223                 r5l_move_io_unit_list(&log->running_ios, &log->io_end_ios,
 224                                       IO_UNIT_IO_END);
 225         else
 226                 r5l_log_run_stripes(log);
 227         spin_unlock_irqrestore(&log->io_list_lock, flags);
 228
 229         if (log->need_cache_flush)
 230                 md_wakeup_thread(log->rdev->mddev->thread);
 231 }
 232
 233 static void r5l_submit_current_io(struct r5l_log *log)
 234 {
 235         struct r5l_io_unit *io = log->current_io;
 236         struct r5l_meta_block *block;
 237         unsigned long flags;
 238         u32 crc;
 239
 240         if (!io)
 241                 return;
 242
 243         block = page_address(io->meta_page);
 244         block->meta_size = cpu_to_le32(io->meta_offset);
 245         crc = crc32c_le(log->uuid_checksum, block, PAGE_SIZE);
 246         block->checksum = cpu_to_le32(crc);
 247
 248         log->current_io = NULL;
 249         spin_lock_irqsave(&log->io_list_lock, flags);
 250         __r5l_set_io_unit_state(io, IO_UNIT_IO_START);
 251         spin_unlock_irqrestore(&log->io_list_lock, flags);
 252
 253         submit_bio(WRITE, io->current_bio);
 254 }
 255
 256 static struct bio *r5l_bio_alloc(struct r5l_log *log)
 257 {
 258         struct bio *bio = bio_kmalloc(GFP_NOIO | __GFP_NOFAIL, BIO_MAX_PAGES);
 259
 260         bio->bi_rw = WRITE;
 261         bio->bi_bdev = log->rdev->bdev;
 262         bio->bi_iter.bi_sector = log->rdev->data_offset + log->log_start;
 263
 264         return bio;
 265 }
 266
 267 static void r5_reserve_log_entry(struct r5l_log *log, struct r5l_io_unit *io)
 268 {
 269         log->log_start = r5l_ring_add(log, log->log_start, BLOCK_SECTORS);
 270
 271         /*
 272          * If we filled up the log device start from the beginning again,
 273          * which will require a new bio.
 274          *
 275          * Note: for this to work properly the log size needs to me a multiple
 276          * of BLOCK_SECTORS.
 277          */
 278         if (log->log_start == 0)
 279                 io->need_split_bio = true;
 280
 281         io->log_end = log->log_start;
 282 }
 283
 284 static struct r5l_io_unit *r5l_new_meta(struct r5l_log *log)
 285 {
 286         struct r5l_io_unit *io;
 287         struct r5l_meta_block *block;
 288
 289         /* We can't handle memory allocate failure so far */
 290         io = kmem_cache_zalloc(log->io_kc, GFP_NOIO | __GFP_NOFAIL);
 291         io->log = log;
 292         INIT_LIST_HEAD(&io->log_sibling);
 293         INIT_LIST_HEAD(&io->stripe_list);
 294         io->state = IO_UNIT_RUNNING;
 295
 296         io->meta_page = alloc_page(GFP_NOIO | __GFP_NOFAIL | __GFP_ZERO);
 297         block = page_address(io->meta_page);
 298         block->magic = cpu_to_le32(R5LOG_MAGIC);
 299         block->version = R5LOG_VERSION;
 300         block->seq = cpu_to_le64(log->seq);
 301         block->position = cpu_to_le64(log->log_start);
 302
 303         io->log_start = log->log_start;
 304         io->meta_offset = sizeof(struct r5l_meta_block);
 305         io->seq = log->seq++;
 306
 307         io->current_bio = r5l_bio_alloc(log);
 308         io->current_bio->bi_end_io = r5l_log_endio;
 309         io->current_bio->bi_private = io;
 310         bio_add_page(io->current_bio, io->meta_page, PAGE_SIZE, 0);
 311
 312         r5_reserve_log_entry(log, io);
 313
 314         spin_lock_irq(&log->io_list_lock);
 315         list_add_tail(&io->log_sibling, &log->running_ios);
 316         spin_unlock_irq(&log->io_list_lock);
 317
 318         return io;
 319 }
 320
 321 static int r5l_get_meta(struct r5l_log *log, unsigned int payload_size)
 322 {
 323         if (log->current_io &&
 324             log->current_io->meta_offset + payload_size > PAGE_SIZE)
 325                 r5l_submit_current_io(log);
 326
 327         if (!log->current_io)
 328                 log->current_io = r5l_new_meta(log);
 329         return 0;
 330 }
 331
 332 static void r5l_append_payload_meta(struct r5l_log *log, u16 type,
 333                                     sector_t location,
 334                                     u32 checksum1, u32 checksum2,
 335                                     bool checksum2_valid)
 336 {
 337         struct r5l_io_unit *io = log->current_io;
 338         struct r5l_payload_data_parity *payload;
 339
 340         payload = page_address(io->meta_page) + io->meta_offset;
 341         payload->header.type = cpu_to_le16(type);
 342         payload->header.flags = cpu_to_le16(0);
 343         payload->size = cpu_to_le32((1 + !!checksum2_valid) <<
 344                                     (PAGE_SHIFT - 9));
 345         payload->location = cpu_to_le64(location);
 346         payload->checksum[0] = cpu_to_le32(checksum1);
 347         if (checksum2_valid)
 348                 payload->checksum[1] = cpu_to_le32(checksum2);
 349
 350         io->meta_offset += sizeof(struct r5l_payload_data_parity) +
 351                 sizeof(__le32) * (1 + !!checksum2_valid);
 352 }
 353
 354 static void r5l_append_payload_page(struct r5l_log *log, struct page *page)
 355 {
 356         struct r5l_io_unit *io = log->current_io;
 357
 358         if (io->need_split_bio) {
 359                 struct bio *prev = io->current_bio;
 360
 361                 io->current_bio = r5l_bio_alloc(log);
 362                 bio_chain(io->current_bio, prev);
 363
 364                 submit_bio(WRITE, prev);
 365         }
 366
 367         if (!bio_add_page(io->current_bio, page, PAGE_SIZE, 0))
 368                 BUG();
 369
 370         r5_reserve_log_entry(log, io);
 371 }
 372
 373 static void r5l_log_stripe(struct r5l_log *log, struct stripe_head *sh,
 374                            int data_pages, int parity_pages)
 375 {
 376         int i;
 377         int meta_size;
 378         struct r5l_io_unit *io;
 379
 380         meta_size =
 381                 ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32))
 382                  * data_pages) +
 383                 sizeof(struct r5l_payload_data_parity) +
 384                 sizeof(__le32) * parity_pages;
 385
 386         r5l_get_meta(log, meta_size);
 387         io = log->current_io;
 388
 389         for (i = 0; i < sh->disks; i++) {
 390                 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags))
 391                         continue;
 392                 if (i == sh->pd_idx || i == sh->qd_idx)
 393                         continue;
 394                 r5l_append_payload_meta(log, R5LOG_PAYLOAD_DATA,
 395                                         raid5_compute_blocknr(sh, i, 0),
 396                                         sh->dev[i].log_checksum, 0, false);
 397                 r5l_append_payload_page(log, sh->dev[i].page);
 398         }
 399
 400         if (sh->qd_idx >= 0) {
 401                 r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
 402                                         sh->sector, sh->dev[sh->pd_idx].log_checksum,
 403                                         sh->dev[sh->qd_idx].log_checksum, true);
 404                 r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
 405                 r5l_append_payload_page(log, sh->dev[sh->qd_idx].page);
 406         } else {
 407                 r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
 408                                         sh->sector, sh->dev[sh->pd_idx].log_checksum,
 409                                         0, false);
 410                 r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
 411         }
 412
 413         list_add_tail(&sh->log_list, &io->stripe_list);
 414         atomic_inc(&io->pending_stripe);
 415         sh->log_io = io;
 416 }
 417
 418 static void r5l_wake_reclaim(struct r5l_log *log, sector_t space);
 419 /*
 420  * running in raid5d, where reclaim could wait for raid5d too (when it flushes
 421  * data from log to raid disks), so we shouldn't wait for reclaim here
 422  */
 423 int r5l_write_stripe(struct r5l_log *log, struct stripe_head *sh)
 424 {
 425         int write_disks = 0;
 426         int data_pages, parity_pages;
 427         int meta_size;
 428         int reserve;
 429         int i;
 430
 431         if (!log)
 432                 return -EAGAIN;
 433         /* Don't support stripe batch */
 434         if (sh->log_io || !test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags) ||
 435             test_bit(STRIPE_SYNCING, &sh->state)) {
 436                 /* the stripe is written to log, we start writing it to raid */
 437                 clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
 438                 return -EAGAIN;
 439         }
 440
 441         for (i = 0; i < sh->disks; i++) {
 442                 void *addr;
 443
 444                 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags))
 445                         continue;
 446                 write_disks++;
 447                 /* checksum is already calculated in last run */
 448                 if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
 449                         continue;
 450                 addr = kmap_atomic(sh->dev[i].page);
 451                 sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum,
 452                                                     addr, PAGE_SIZE);
 453                 kunmap_atomic(addr);
 454         }
 455         parity_pages = 1 + !!(sh->qd_idx >= 0);
 456         data_pages = write_disks - parity_pages;
 457
 458         meta_size =
 459                 ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32))
 460                  * data_pages) +
 461                 sizeof(struct r5l_payload_data_parity) +
 462                 sizeof(__le32) * parity_pages;
 463         /* Doesn't work with very big raid array */
 464         if (meta_size + sizeof(struct r5l_meta_block) > PAGE_SIZE)
 465                 return -EINVAL;
 466
 467         set_bit(STRIPE_LOG_TRAPPED, &sh->state);
 468         /*
 469          * The stripe must enter state machine again to finish the write, so
 470          * don't delay.
 471          */
 472         clear_bit(STRIPE_DELAYED, &sh->state);
 473         atomic_inc(&sh->count);
 474
 475         mutex_lock(&log->io_mutex);
 476         /* meta + data */
 477         reserve = (1 + write_disks) << (PAGE_SHIFT - 9);
 478         if (r5l_has_free_space(log, reserve))
 479                 r5l_log_stripe(log, sh, data_pages, parity_pages);
 480         else {
 481                 spin_lock(&log->no_space_stripes_lock);
 482                 list_add_tail(&sh->log_list, &log->no_space_stripes);
 483                 spin_unlock(&log->no_space_stripes_lock);
 484
 485                 r5l_wake_reclaim(log, reserve);
 486         }
 487         mutex_unlock(&log->io_mutex);
 488
 489         return 0;
 490 }
 491
 492 void r5l_write_stripe_run(struct r5l_log *log)
 493 {
 494         if (!log)
 495                 return;
 496         mutex_lock(&log->io_mutex);
 497         r5l_submit_current_io(log);
 498         mutex_unlock(&log->io_mutex);
 499 }
 500
 501 int r5l_handle_flush_request(struct r5l_log *log, struct bio *bio)
 502 {
 503         if (!log)
 504                 return -ENODEV;
 505         /*
 506          * we flush log disk cache first, then write stripe data to raid disks.
 507          * So if bio is finished, the log disk cache is flushed already. The
 508          * recovery guarantees we can recovery the bio from log disk, so we
 509          * don't need to flush again
 510          */
 511         if (bio->bi_iter.bi_size == 0) {
 512                 bio_endio(bio);
 513                 return 0;
 514         }
 515         bio->bi_rw &= ~REQ_FLUSH;
 516         return -EAGAIN;
 517 }
 518
 519 /* This will run after log space is reclaimed */
 520 static void r5l_run_no_space_stripes(struct r5l_log *log)
 521 {
 522         struct stripe_head *sh;
 523
 524         spin_lock(&log->no_space_stripes_lock);
 525         while (!list_empty(&log->no_space_stripes)) {
 526                 sh = list_first_entry(&log->no_space_stripes,
 527                                       struct stripe_head, log_list);
 528                 list_del_init(&sh->log_list);
 529                 set_bit(STRIPE_HANDLE, &sh->state);
 530                 raid5_release_stripe(sh);
 531         }
 532         spin_unlock(&log->no_space_stripes_lock);
 533 }
 534
 535 static sector_t r5l_reclaimable_space(struct r5l_log *log)
 536 {
 537         return r5l_ring_distance(log, log->last_checkpoint,
 538                                  log->next_checkpoint);
 539 }
 540
 541 static bool r5l_complete_finished_ios(struct r5l_log *log)
 542 {
 543         struct r5l_io_unit *io, *next;
 544         bool found = false;
 545
 546         assert_spin_locked(&log->io_list_lock);
 547
 548         list_for_each_entry_safe(io, next, &log->finished_ios, log_sibling) {
 549                 /* don't change list order */
 550                 if (io->state < IO_UNIT_STRIPE_END)
 551                         break;
 552
 553                 log->next_checkpoint = io->log_start;
 554                 log->next_cp_seq = io->seq;
 555
 556                 list_del(&io->log_sibling);
 557                 r5l_free_io_unit(log, io);
 558
 559                 found = true;
 560         }
 561
 562         return found;
 563 }
 564
 565 static void __r5l_stripe_write_finished(struct r5l_io_unit *io)
 566 {
 567         struct r5l_log *log = io->log;
 568         unsigned long flags;
 569
 570         spin_lock_irqsave(&log->io_list_lock, flags);
 571         __r5l_set_io_unit_state(io, IO_UNIT_STRIPE_END);
 572
 573         if (!r5l_complete_finished_ios(log)) {
 574                 spin_unlock_irqrestore(&log->io_list_lock, flags);
 575                 return;
 576         }
 577
 578         if (r5l_reclaimable_space(log) > log->max_free_space)
 579                 r5l_wake_reclaim(log, 0);
 580
 581         spin_unlock_irqrestore(&log->io_list_lock, flags);
 582         wake_up(&log->iounit_wait);
 583 }
 584
 585 void r5l_stripe_write_finished(struct stripe_head *sh)
 586 {
 587         struct r5l_io_unit *io;
 588
 589         io = sh->log_io;
 590         sh->log_io = NULL;
 591
 592         if (io && atomic_dec_and_test(&io->pending_stripe))
 593                 __r5l_stripe_write_finished(io);
 594 }
 595
 596 static void r5l_log_flush_endio(struct bio *bio)
 597 {
 598         struct r5l_log *log = container_of(bio, struct r5l_log,
 599                 flush_bio);
 600         unsigned long flags;
 601         struct r5l_io_unit *io;
 602
 603         if (bio->bi_error)
 604                 md_error(log->rdev->mddev, log->rdev);
 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_write_super_and_discard_space(struct r5l_log *log,
 654         sector_t end)
 655 {
 656         struct block_device *bdev = log->rdev->bdev;
 657         struct mddev *mddev;
 658
 659         r5l_write_super(log, end);
 660
 661         if (!blk_queue_discard(bdev_get_queue(bdev)))
 662                 return;
 663
 664         mddev = log->rdev->mddev;
 665         /*
 666          * This is to avoid a deadlock. r5l_quiesce holds reconfig_mutex and
 667          * wait for this thread to finish. This thread waits for
 668          * MD_CHANGE_PENDING clear, which is supposed to be done in
 669          * md_check_recovery(). md_check_recovery() tries to get
 670          * reconfig_mutex. Since r5l_quiesce already holds the mutex,
 671          * md_check_recovery() fails, so the PENDING never get cleared. The
 672          * in_teardown check workaround this issue.
 673          */
 674         if (!log->in_teardown) {
 675                 set_bit(MD_CHANGE_DEVS, &mddev->flags);
 676                 set_bit(MD_CHANGE_PENDING, &mddev->flags);
 677                 md_wakeup_thread(mddev->thread);
 678                 wait_event(mddev->sb_wait,
 679                         !test_bit(MD_CHANGE_PENDING, &mddev->flags) ||
 680                         log->in_teardown);
 681                 /*
 682                  * r5l_quiesce could run after in_teardown check and hold
 683                  * mutex first. Superblock might get updated twice.
 684                  */
 685                 if (log->in_teardown)
 686                         md_update_sb(mddev, 1);
 687         } else {
 688                 WARN_ON(!mddev_is_locked(mddev));
 689                 md_update_sb(mddev, 1);
 690         }
 691
 692         /* discard IO error really doesn't matter, ignore it */
 693         if (log->last_checkpoint < end) {
 694                 blkdev_issue_discard(bdev,
 695                                 log->last_checkpoint + log->rdev->data_offset,
 696                                 end - log->last_checkpoint, GFP_NOIO, 0);
 697         } else {
 698                 blkdev_issue_discard(bdev,
 699                                 log->last_checkpoint + log->rdev->data_offset,
 700                                 log->device_size - log->last_checkpoint,
 701                                 GFP_NOIO, 0);
 702                 blkdev_issue_discard(bdev, log->rdev->data_offset, end,
 703                                 GFP_NOIO, 0);
 704         }
 705 }
 706
 707
 708 static void r5l_do_reclaim(struct r5l_log *log)
 709 {
 710         sector_t reclaim_target = xchg(&log->reclaim_target, 0);
 711         sector_t reclaimable;
 712         sector_t next_checkpoint;
 713         u64 next_cp_seq;
 714
 715         spin_lock_irq(&log->io_list_lock);
 716         /*
 717          * move proper io_unit to reclaim list. We should not change the order.
 718          * reclaimable/unreclaimable io_unit can be mixed in the list, we
 719          * shouldn't reuse space of an unreclaimable io_unit
 720          */
 721         while (1) {
 722                 reclaimable = r5l_reclaimable_space(log);
 723                 if (reclaimable >= reclaim_target ||
 724                     (list_empty(&log->running_ios) &&
 725                      list_empty(&log->io_end_ios) &&
 726                      list_empty(&log->flushing_ios) &&
 727                      list_empty(&log->finished_ios)))
 728                         break;
 729
 730                 md_wakeup_thread(log->rdev->mddev->thread);
 731                 wait_event_lock_irq(log->iounit_wait,
 732                                     r5l_reclaimable_space(log) > reclaimable,
 733                                     log->io_list_lock);
 734         }
 735
 736         next_checkpoint = log->next_checkpoint;
 737         next_cp_seq = log->next_cp_seq;
 738         spin_unlock_irq(&log->io_list_lock);
 739
 740         BUG_ON(reclaimable < 0);
 741         if (reclaimable == 0)
 742                 return;
 743
 744         /*
 745          * write_super will flush cache of each raid disk. We must write super
 746          * here, because the log area might be reused soon and we don't want to
 747          * confuse recovery
 748          */
 749         r5l_write_super_and_discard_space(log, next_checkpoint);
 750
 751         mutex_lock(&log->io_mutex);
 752         log->last_checkpoint = next_checkpoint;
 753         log->last_cp_seq = next_cp_seq;
 754         mutex_unlock(&log->io_mutex);
 755
 756         r5l_run_no_space_stripes(log);
 757 }
 758
 759 static void r5l_reclaim_thread(struct md_thread *thread)
 760 {
 761         struct mddev *mddev = thread->mddev;
 762         struct r5conf *conf = mddev->private;
 763         struct r5l_log *log = conf->log;
 764
 765         if (!log)
 766                 return;
 767         r5l_do_reclaim(log);
 768 }
 769
 770 static void r5l_wake_reclaim(struct r5l_log *log, sector_t space)
 771 {
 772         unsigned long target;
 773         unsigned long new = (unsigned long)space; /* overflow in theory */
 774
 775         do {
 776                 target = log->reclaim_target;
 777                 if (new < target)
 778                         return;
 779         } while (cmpxchg(&log->reclaim_target, target, new) != target);
 780         md_wakeup_thread(log->reclaim_thread);
 781 }
 782
 783 void r5l_quiesce(struct r5l_log *log, int state)
 784 {
 785         struct mddev *mddev;
 786         if (!log || state == 2)
 787                 return;
 788         if (state == 0) {
 789                 log->in_teardown = 0;
 790                 log->reclaim_thread = md_register_thread(r5l_reclaim_thread,
 791                                         log->rdev->mddev, "reclaim");
 792         } else if (state == 1) {
 793                 /*
 794                  * at this point all stripes are finished, so io_unit is at
 795                  * least in STRIPE_END state
 796                  */
 797                 log->in_teardown = 1;
 798                 /* make sure r5l_write_super_and_discard_space exits */
 799                 mddev = log->rdev->mddev;
 800                 wake_up(&mddev->sb_wait);
 801                 r5l_wake_reclaim(log, -1L);
 802                 md_unregister_thread(&log->reclaim_thread);
 803                 r5l_do_reclaim(log);
 804         }
 805 }
 806
 807 bool r5l_log_disk_error(struct r5conf *conf)
 808 {
 809         /* don't allow write if journal disk is missing */
 810         if (!conf->log)
 811                 return test_bit(MD_HAS_JOURNAL, &conf->mddev->flags);
 812         return test_bit(Faulty, &conf->log->rdev->flags);
 813 }
 814
 815 struct r5l_recovery_ctx {
 816         struct page *meta_page;         /* current meta */
 817         sector_t meta_total_blocks;     /* total size of current meta and data */
 818         sector_t pos;                   /* recovery position */
 819         u64 seq;                        /* recovery position seq */
 820 };
 821
 822 static int r5l_read_meta_block(struct r5l_log *log,
 823                                struct r5l_recovery_ctx *ctx)
 824 {
 825         struct page *page = ctx->meta_page;
 826         struct r5l_meta_block *mb;
 827         u32 crc, stored_crc;
 828
 829         if (!sync_page_io(log->rdev, ctx->pos, PAGE_SIZE, page, READ, false))
 830                 return -EIO;
 831
 832         mb = page_address(page);
 833         stored_crc = le32_to_cpu(mb->checksum);
 834         mb->checksum = 0;
 835
 836         if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
 837             le64_to_cpu(mb->seq) != ctx->seq ||
 838             mb->version != R5LOG_VERSION ||
 839             le64_to_cpu(mb->position) != ctx->pos)
 840                 return -EINVAL;
 841
 842         crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
 843         if (stored_crc != crc)
 844                 return -EINVAL;
 845
 846         if (le32_to_cpu(mb->meta_size) > PAGE_SIZE)
 847                 return -EINVAL;
 848
 849         ctx->meta_total_blocks = BLOCK_SECTORS;
 850
 851         return 0;
 852 }
 853
 854 static int r5l_recovery_flush_one_stripe(struct r5l_log *log,
 855                                          struct r5l_recovery_ctx *ctx,
 856                                          sector_t stripe_sect,
 857                                          int *offset, sector_t *log_offset)
 858 {
 859         struct r5conf *conf = log->rdev->mddev->private;
 860         struct stripe_head *sh;
 861         struct r5l_payload_data_parity *payload;
 862         int disk_index;
 863
 864         sh = raid5_get_active_stripe(conf, stripe_sect, 0, 0, 0);
 865         while (1) {
 866                 payload = page_address(ctx->meta_page) + *offset;
 867
 868                 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) {
 869                         raid5_compute_sector(conf,
 870                                              le64_to_cpu(payload->location), 0,
 871                                              &disk_index, sh);
 872
 873                         sync_page_io(log->rdev, *log_offset, PAGE_SIZE,
 874                                      sh->dev[disk_index].page, READ, false);
 875                         sh->dev[disk_index].log_checksum =
 876                                 le32_to_cpu(payload->checksum[0]);
 877                         set_bit(R5_Wantwrite, &sh->dev[disk_index].flags);
 878                         ctx->meta_total_blocks += BLOCK_SECTORS;
 879                 } else {
 880                         disk_index = sh->pd_idx;
 881                         sync_page_io(log->rdev, *log_offset, PAGE_SIZE,
 882                                      sh->dev[disk_index].page, READ, false);
 883                         sh->dev[disk_index].log_checksum =
 884                                 le32_to_cpu(payload->checksum[0]);
 885                         set_bit(R5_Wantwrite, &sh->dev[disk_index].flags);
 886
 887                         if (sh->qd_idx >= 0) {
 888                                 disk_index = sh->qd_idx;
 889                                 sync_page_io(log->rdev,
 890                                              r5l_ring_add(log, *log_offset, BLOCK_SECTORS),
 891                                              PAGE_SIZE, sh->dev[disk_index].page,
 892                                              READ, false);
 893                                 sh->dev[disk_index].log_checksum =
 894                                         le32_to_cpu(payload->checksum[1]);
 895                                 set_bit(R5_Wantwrite,
 896                                         &sh->dev[disk_index].flags);
 897                         }
 898                         ctx->meta_total_blocks += BLOCK_SECTORS * conf->max_degraded;
 899                 }
 900
 901                 *log_offset = r5l_ring_add(log, *log_offset,
 902                                            le32_to_cpu(payload->size));
 903                 *offset += sizeof(struct r5l_payload_data_parity) +
 904                         sizeof(__le32) *
 905                         (le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
 906                 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_PARITY)
 907                         break;
 908         }
 909
 910         for (disk_index = 0; disk_index < sh->disks; disk_index++) {
 911                 void *addr;
 912                 u32 checksum;
 913
 914                 if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
 915                         continue;
 916                 addr = kmap_atomic(sh->dev[disk_index].page);
 917                 checksum = crc32c_le(log->uuid_checksum, addr, PAGE_SIZE);
 918                 kunmap_atomic(addr);
 919                 if (checksum != sh->dev[disk_index].log_checksum)
 920                         goto error;
 921         }
 922
 923         for (disk_index = 0; disk_index < sh->disks; disk_index++) {
 924                 struct md_rdev *rdev, *rrdev;
 925
 926                 if (!test_and_clear_bit(R5_Wantwrite,
 927                                         &sh->dev[disk_index].flags))
 928                         continue;
 929
 930                 /* in case device is broken */
 931                 rdev = rcu_dereference(conf->disks[disk_index].rdev);
 932                 if (rdev)
 933                         sync_page_io(rdev, stripe_sect, PAGE_SIZE,
 934                                      sh->dev[disk_index].page, WRITE, false);
 935                 rrdev = rcu_dereference(conf->disks[disk_index].replacement);
 936                 if (rrdev)
 937                         sync_page_io(rrdev, stripe_sect, PAGE_SIZE,
 938                                      sh->dev[disk_index].page, WRITE, false);
 939         }
 940         raid5_release_stripe(sh);
 941         return 0;
 942
 943 error:
 944         for (disk_index = 0; disk_index < sh->disks; disk_index++)
 945                 sh->dev[disk_index].flags = 0;
 946         raid5_release_stripe(sh);
 947         return -EINVAL;
 948 }
 949
 950 static int r5l_recovery_flush_one_meta(struct r5l_log *log,
 951                                        struct r5l_recovery_ctx *ctx)
 952 {
 953         struct r5conf *conf = log->rdev->mddev->private;
 954         struct r5l_payload_data_parity *payload;
 955         struct r5l_meta_block *mb;
 956         int offset;
 957         sector_t log_offset;
 958         sector_t stripe_sector;
 959
 960         mb = page_address(ctx->meta_page);
 961         offset = sizeof(struct r5l_meta_block);
 962         log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
 963
 964         while (offset < le32_to_cpu(mb->meta_size)) {
 965                 int dd;
 966
 967                 payload = (void *)mb + offset;
 968                 stripe_sector = raid5_compute_sector(conf,
 969                                                      le64_to_cpu(payload->location), 0, &dd, NULL);
 970                 if (r5l_recovery_flush_one_stripe(log, ctx, stripe_sector,
 971                                                   &offset, &log_offset))
 972                         return -EINVAL;
 973         }
 974         return 0;
 975 }
 976
 977 /* copy data/parity from log to raid disks */
 978 static void r5l_recovery_flush_log(struct r5l_log *log,
 979                                    struct r5l_recovery_ctx *ctx)
 980 {
 981         while (1) {
 982                 if (r5l_read_meta_block(log, ctx))
 983                         return;
 984                 if (r5l_recovery_flush_one_meta(log, ctx))
 985                         return;
 986                 ctx->seq++;
 987                 ctx->pos = r5l_ring_add(log, ctx->pos, ctx->meta_total_blocks);
 988         }
 989 }
 990
 991 static int r5l_log_write_empty_meta_block(struct r5l_log *log, sector_t pos,
 992                                           u64 seq)
 993 {
 994         struct page *page;
 995         struct r5l_meta_block *mb;
 996         u32 crc;
 997
 998         page = alloc_page(GFP_KERNEL | __GFP_ZERO);
 999         if (!page)
1000                 return -ENOMEM;
1001         mb = page_address(page);
1002         mb->magic = cpu_to_le32(R5LOG_MAGIC);
1003         mb->version = R5LOG_VERSION;
1004         mb->meta_size = cpu_to_le32(sizeof(struct r5l_meta_block));
1005         mb->seq = cpu_to_le64(seq);
1006         mb->position = cpu_to_le64(pos);
1007         crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
1008         mb->checksum = cpu_to_le32(crc);
1009
1010         if (!sync_page_io(log->rdev, pos, PAGE_SIZE, page, WRITE_FUA, false)) {
1011                 __free_page(page);
1012                 return -EIO;
1013         }
1014         __free_page(page);
1015         return 0;
1016 }
1017
1018 static int r5l_recovery_log(struct r5l_log *log)
1019 {
1020         struct r5l_recovery_ctx ctx;
1021
1022         ctx.pos = log->last_checkpoint;
1023         ctx.seq = log->last_cp_seq;
1024         ctx.meta_page = alloc_page(GFP_KERNEL);
1025         if (!ctx.meta_page)
1026                 return -ENOMEM;
1027
1028         r5l_recovery_flush_log(log, &ctx);
1029         __free_page(ctx.meta_page);
1030
1031         /*
1032          * we did a recovery. Now ctx.pos points to an invalid meta block. New
1033          * log will start here. but we can't let superblock point to last valid
1034          * meta block. The log might looks like:
1035          * | meta 1| meta 2| meta 3|
1036          * meta 1 is valid, meta 2 is invalid. meta 3 could be valid. If
1037          * superblock points to meta 1, we write a new valid meta 2n.  if crash
1038          * happens again, new recovery will start from meta 1. Since meta 2n is
1039          * valid now, recovery will think meta 3 is valid, which is wrong.
1040          * The solution is we create a new meta in meta2 with its seq == meta
1041          * 1's seq + 10 and let superblock points to meta2. The same recovery will
1042          * not think meta 3 is a valid meta, because its seq doesn't match
1043          */
1044         if (ctx.seq > log->last_cp_seq + 1) {
1045                 int ret;
1046
1047                 ret = r5l_log_write_empty_meta_block(log, ctx.pos, ctx.seq + 10);
1048                 if (ret)
1049                         return ret;
1050                 log->seq = ctx.seq + 11;
1051                 log->log_start = r5l_ring_add(log, ctx.pos, BLOCK_SECTORS);
1052                 r5l_write_super(log, ctx.pos);
1053         } else {
1054                 log->log_start = ctx.pos;
1055                 log->seq = ctx.seq;
1056         }
1057         return 0;
1058 }
1059
1060 static void r5l_write_super(struct r5l_log *log, sector_t cp)
1061 {
1062         struct mddev *mddev = log->rdev->mddev;
1063
1064         log->rdev->journal_tail = cp;
1065         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1066 }
1067
1068 static int r5l_load_log(struct r5l_log *log)
1069 {
1070         struct md_rdev *rdev = log->rdev;
1071         struct page *page;
1072         struct r5l_meta_block *mb;
1073         sector_t cp = log->rdev->journal_tail;
1074         u32 stored_crc, expected_crc;
1075         bool create_super = false;
1076         int ret;
1077
1078         /* Make sure it's valid */
1079         if (cp >= rdev->sectors || round_down(cp, BLOCK_SECTORS) != cp)
1080                 cp = 0;
1081         page = alloc_page(GFP_KERNEL);
1082         if (!page)
1083                 return -ENOMEM;
1084
1085         if (!sync_page_io(rdev, cp, PAGE_SIZE, page, READ, false)) {
1086                 ret = -EIO;
1087                 goto ioerr;
1088         }
1089         mb = page_address(page);
1090
1091         if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
1092             mb->version != R5LOG_VERSION) {
1093                 create_super = true;
1094                 goto create;
1095         }
1096         stored_crc = le32_to_cpu(mb->checksum);
1097         mb->checksum = 0;
1098         expected_crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
1099         if (stored_crc != expected_crc) {
1100                 create_super = true;
1101                 goto create;
1102         }
1103         if (le64_to_cpu(mb->position) != cp) {
1104                 create_super = true;
1105                 goto create;
1106         }
1107 create:
1108         if (create_super) {
1109                 log->last_cp_seq = prandom_u32();
1110                 cp = 0;
1111                 /*
1112                  * Make sure super points to correct address. Log might have
1113                  * data very soon. If super hasn't correct log tail address,
1114                  * recovery can't find the log
1115                  */
1116                 r5l_write_super(log, cp);
1117         } else
1118                 log->last_cp_seq = le64_to_cpu(mb->seq);
1119
1120         log->device_size = round_down(rdev->sectors, BLOCK_SECTORS);
1121         log->max_free_space = log->device_size >> RECLAIM_MAX_FREE_SPACE_SHIFT;
1122         if (log->max_free_space > RECLAIM_MAX_FREE_SPACE)
1123                 log->max_free_space = RECLAIM_MAX_FREE_SPACE;
1124         log->last_checkpoint = cp;
1125
1126         __free_page(page);
1127
1128         return r5l_recovery_log(log);
1129 ioerr:
1130         __free_page(page);
1131         return ret;
1132 }
1133
1134 int r5l_init_log(struct r5conf *conf, struct md_rdev *rdev)
1135 {
1136         struct r5l_log *log;
1137
1138         if (PAGE_SIZE != 4096)
1139                 return -EINVAL;
1140         log = kzalloc(sizeof(*log), GFP_KERNEL);
1141         if (!log)
1142                 return -ENOMEM;
1143         log->rdev = rdev;
1144
1145         log->need_cache_flush = (rdev->bdev->bd_disk->queue->flush_flags != 0);
1146
1147         log->uuid_checksum = crc32c_le(~0, rdev->mddev->uuid,
1148                                        sizeof(rdev->mddev->uuid));
1149
1150         mutex_init(&log->io_mutex);
1151
1152         spin_lock_init(&log->io_list_lock);
1153         INIT_LIST_HEAD(&log->running_ios);
1154         INIT_LIST_HEAD(&log->io_end_ios);
1155         INIT_LIST_HEAD(&log->flushing_ios);
1156         INIT_LIST_HEAD(&log->finished_ios);
1157         bio_init(&log->flush_bio);
1158
1159         log->io_kc = KMEM_CACHE(r5l_io_unit, 0);
1160         if (!log->io_kc)
1161                 goto io_kc;
1162
1163         log->reclaim_thread = md_register_thread(r5l_reclaim_thread,
1164                                                  log->rdev->mddev, "reclaim");
1165         if (!log->reclaim_thread)
1166                 goto reclaim_thread;
1167         init_waitqueue_head(&log->iounit_wait);
1168
1169         INIT_LIST_HEAD(&log->no_space_stripes);
1170         spin_lock_init(&log->no_space_stripes_lock);
1171
1172         if (r5l_load_log(log))
1173                 goto error;
1174
1175         conf->log = log;
1176         return 0;
1177 error:
1178         md_unregister_thread(&log->reclaim_thread);
1179 reclaim_thread:
1180         kmem_cache_destroy(log->io_kc);
1181 io_kc:
1182         kfree(log);
1183         return -EINVAL;
1184 }
1185
1186 void r5l_exit_log(struct r5l_log *log)
1187 {
1188         md_unregister_thread(&log->reclaim_thread);
1189         kmem_cache_destroy(log->io_kc);
1190         kfree(log);
1191 }