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
  88 /*
  89  * an IO range starts from a meta data block and end at the next meta data
  90  * block. The io unit's the meta data block tracks data/parity followed it. io
  91  * unit is written to log disk with normal write, as we always flush log disk
  92  * first and then start move data to raid disks, there is no requirement to
  93  * write io unit with FLUSH/FUA
  94  */
  95 struct r5l_io_unit {
  96         struct r5l_log *log;
  97
  98         struct page *meta_page; /* store meta block */
  99         int meta_offset;        /* current offset in meta_page */
 100
 101         struct bio_list bios;
 102         atomic_t pending_io;    /* pending bios not written to log yet */
 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 };
 114
 115 /* r5l_io_unit state */
 116 enum r5l_io_unit_state {
 117         IO_UNIT_RUNNING = 0,    /* accepting new IO */
 118         IO_UNIT_IO_START = 1,   /* io_unit bio start writing to log,
 119                                  * don't accepting new bio */
 120         IO_UNIT_IO_END = 2,     /* io_unit bio finish writing to log */
 121         IO_UNIT_STRIPE_END = 3, /* stripes data finished writing to raid */
 122 };
 123
 124 static sector_t r5l_ring_add(struct r5l_log *log, sector_t start, sector_t inc)
 125 {
 126         start += inc;
 127         if (start >= log->device_size)
 128                 start = start - log->device_size;
 129         return start;
 130 }
 131
 132 static sector_t r5l_ring_distance(struct r5l_log *log, sector_t start,
 133                                   sector_t end)
 134 {
 135         if (end >= start)
 136                 return end - start;
 137         else
 138                 return end + log->device_size - start;
 139 }
 140
 141 static bool r5l_has_free_space(struct r5l_log *log, sector_t size)
 142 {
 143         sector_t used_size;
 144
 145         used_size = r5l_ring_distance(log, log->last_checkpoint,
 146                                         log->log_start);
 147
 148         return log->device_size > used_size + size;
 149 }
 150
 151 static struct r5l_io_unit *r5l_alloc_io_unit(struct r5l_log *log)
 152 {
 153         struct r5l_io_unit *io;
 154         /* We can't handle memory allocate failure so far */
 155         gfp_t gfp = GFP_NOIO | __GFP_NOFAIL;
 156
 157         io = kmem_cache_zalloc(log->io_kc, gfp);
 158         io->log = log;
 159         io->meta_page = alloc_page(gfp | __GFP_ZERO);
 160
 161         bio_list_init(&io->bios);
 162         INIT_LIST_HEAD(&io->log_sibling);
 163         INIT_LIST_HEAD(&io->stripe_list);
 164         io->state = IO_UNIT_RUNNING;
 165         return io;
 166 }
 167
 168 static void r5l_free_io_unit(struct r5l_log *log, struct r5l_io_unit *io)
 169 {
 170         __free_page(io->meta_page);
 171         kmem_cache_free(log->io_kc, io);
 172 }
 173
 174 static void r5l_move_io_unit_list(struct list_head *from, struct list_head *to,
 175                                   enum r5l_io_unit_state state)
 176 {
 177         struct r5l_io_unit *io;
 178
 179         while (!list_empty(from)) {
 180                 io = list_first_entry(from, struct r5l_io_unit, log_sibling);
 181                 /* don't change list order */
 182                 if (io->state >= state)
 183                         list_move_tail(&io->log_sibling, to);
 184                 else
 185                         break;
 186         }
 187 }
 188
 189 static void __r5l_set_io_unit_state(struct r5l_io_unit *io,
 190                                     enum r5l_io_unit_state state)
 191 {
 192         if (WARN_ON(io->state >= state))
 193                 return;
 194         io->state = state;
 195 }
 196
 197 static void r5l_io_run_stripes(struct r5l_io_unit *io)
 198 {
 199         struct stripe_head *sh, *next;
 200
 201         list_for_each_entry_safe(sh, next, &io->stripe_list, log_list) {
 202                 list_del_init(&sh->log_list);
 203                 set_bit(STRIPE_HANDLE, &sh->state);
 204                 raid5_release_stripe(sh);
 205         }
 206 }
 207
 208 /* XXX: totally ignores I/O errors */
 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         if (!atomic_dec_and_test(&io->pending_io))
 218                 return;
 219
 220         spin_lock_irqsave(&log->io_list_lock, flags);
 221         __r5l_set_io_unit_state(io, IO_UNIT_IO_END);
 222         r5l_move_io_unit_list(&log->running_ios, &log->io_end_ios,
 223                         IO_UNIT_IO_END);
 224         spin_unlock_irqrestore(&log->io_list_lock, flags);
 225
 226         md_wakeup_thread(log->rdev->mddev->thread);
 227 }
 228
 229 static void r5l_submit_current_io(struct r5l_log *log)
 230 {
 231         struct r5l_io_unit *io = log->current_io;
 232         struct r5l_meta_block *block;
 233         struct bio *bio;
 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         while ((bio = bio_list_pop(&io->bios))) {
 251                 /* all IO must start from rdev->data_offset */
 252                 bio->bi_iter.bi_sector += log->rdev->data_offset;
 253                 submit_bio(WRITE, bio);
 254         }
 255 }
 256
 257 static struct r5l_io_unit *r5l_new_meta(struct r5l_log *log)
 258 {
 259         struct r5l_io_unit *io;
 260         struct r5l_meta_block *block;
 261         struct bio *bio;
 262
 263         io = r5l_alloc_io_unit(log);
 264
 265         block = page_address(io->meta_page);
 266         block->magic = cpu_to_le32(R5LOG_MAGIC);
 267         block->version = R5LOG_VERSION;
 268         block->seq = cpu_to_le64(log->seq);
 269         block->position = cpu_to_le64(log->log_start);
 270
 271         io->log_start = log->log_start;
 272         io->meta_offset = sizeof(struct r5l_meta_block);
 273         io->seq = log->seq;
 274
 275         bio = bio_kmalloc(GFP_NOIO | __GFP_NOFAIL, BIO_MAX_PAGES);
 276         io->current_bio = bio;
 277         bio->bi_rw = WRITE;
 278         bio->bi_bdev = log->rdev->bdev;
 279         bio->bi_iter.bi_sector = log->log_start;
 280         bio_add_page(bio, io->meta_page, PAGE_SIZE, 0);
 281         bio->bi_end_io = r5l_log_endio;
 282         bio->bi_private = io;
 283
 284         bio_list_add(&io->bios, bio);
 285         atomic_inc(&io->pending_io);
 286
 287         log->seq++;
 288         log->log_start = r5l_ring_add(log, log->log_start, BLOCK_SECTORS);
 289         io->log_end = log->log_start;
 290         /* current bio hit disk end */
 291         if (log->log_start == 0)
 292                 io->current_bio = NULL;
 293
 294         spin_lock_irq(&log->io_list_lock);
 295         list_add_tail(&io->log_sibling, &log->running_ios);
 296         spin_unlock_irq(&log->io_list_lock);
 297
 298         return io;
 299 }
 300
 301 static int r5l_get_meta(struct r5l_log *log, unsigned int payload_size)
 302 {
 303         struct r5l_io_unit *io;
 304
 305         io = log->current_io;
 306         if (io && io->meta_offset + payload_size > PAGE_SIZE)
 307                 r5l_submit_current_io(log);
 308         io = log->current_io;
 309         if (io)
 310                 return 0;
 311
 312         log->current_io = r5l_new_meta(log);
 313         return 0;
 314 }
 315
 316 static void r5l_append_payload_meta(struct r5l_log *log, u16 type,
 317                                     sector_t location,
 318                                     u32 checksum1, u32 checksum2,
 319                                     bool checksum2_valid)
 320 {
 321         struct r5l_io_unit *io = log->current_io;
 322         struct r5l_payload_data_parity *payload;
 323
 324         payload = page_address(io->meta_page) + io->meta_offset;
 325         payload->header.type = cpu_to_le16(type);
 326         payload->header.flags = cpu_to_le16(0);
 327         payload->size = cpu_to_le32((1 + !!checksum2_valid) <<
 328                                     (PAGE_SHIFT - 9));
 329         payload->location = cpu_to_le64(location);
 330         payload->checksum[0] = cpu_to_le32(checksum1);
 331         if (checksum2_valid)
 332                 payload->checksum[1] = cpu_to_le32(checksum2);
 333
 334         io->meta_offset += sizeof(struct r5l_payload_data_parity) +
 335                 sizeof(__le32) * (1 + !!checksum2_valid);
 336 }
 337
 338 static void r5l_append_payload_page(struct r5l_log *log, struct page *page)
 339 {
 340         struct r5l_io_unit *io = log->current_io;
 341
 342 alloc_bio:
 343         if (!io->current_bio) {
 344                 struct bio *bio;
 345
 346                 bio = bio_kmalloc(GFP_NOIO | __GFP_NOFAIL, BIO_MAX_PAGES);
 347                 bio->bi_rw = WRITE;
 348                 bio->bi_bdev = log->rdev->bdev;
 349                 bio->bi_iter.bi_sector = log->log_start;
 350                 bio->bi_end_io = r5l_log_endio;
 351                 bio->bi_private = io;
 352                 bio_list_add(&io->bios, bio);
 353                 atomic_inc(&io->pending_io);
 354                 io->current_bio = bio;
 355         }
 356         if (!bio_add_page(io->current_bio, page, PAGE_SIZE, 0)) {
 357                 io->current_bio = NULL;
 358                 goto alloc_bio;
 359         }
 360         log->log_start = r5l_ring_add(log, log->log_start,
 361                                       BLOCK_SECTORS);
 362         /* current bio hit disk end */
 363         if (log->log_start == 0)
 364                 io->current_bio = NULL;
 365
 366         io->log_end = log->log_start;
 367 }
 368
 369 static void r5l_log_stripe(struct r5l_log *log, struct stripe_head *sh,
 370                            int data_pages, int parity_pages)
 371 {
 372         int i;
 373         int meta_size;
 374         struct r5l_io_unit *io;
 375
 376         meta_size =
 377                 ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32))
 378                  * data_pages) +
 379                 sizeof(struct r5l_payload_data_parity) +
 380                 sizeof(__le32) * parity_pages;
 381
 382         r5l_get_meta(log, meta_size);
 383         io = log->current_io;
 384
 385         for (i = 0; i < sh->disks; i++) {
 386                 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags))
 387                         continue;
 388                 if (i == sh->pd_idx || i == sh->qd_idx)
 389                         continue;
 390                 r5l_append_payload_meta(log, R5LOG_PAYLOAD_DATA,
 391                                         raid5_compute_blocknr(sh, i, 0),
 392                                         sh->dev[i].log_checksum, 0, false);
 393                 r5l_append_payload_page(log, sh->dev[i].page);
 394         }
 395
 396         if (sh->qd_idx >= 0) {
 397                 r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
 398                                         sh->sector, sh->dev[sh->pd_idx].log_checksum,
 399                                         sh->dev[sh->qd_idx].log_checksum, true);
 400                 r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
 401                 r5l_append_payload_page(log, sh->dev[sh->qd_idx].page);
 402         } else {
 403                 r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
 404                                         sh->sector, sh->dev[sh->pd_idx].log_checksum,
 405                                         0, false);
 406                 r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
 407         }
 408
 409         list_add_tail(&sh->log_list, &io->stripe_list);
 410         atomic_inc(&io->pending_stripe);
 411         sh->log_io = io;
 412 }
 413
 414 static void r5l_wake_reclaim(struct r5l_log *log, sector_t space);
 415 /*
 416  * running in raid5d, where reclaim could wait for raid5d too (when it flushes
 417  * data from log to raid disks), so we shouldn't wait for reclaim here
 418  */
 419 int r5l_write_stripe(struct r5l_log *log, struct stripe_head *sh)
 420 {
 421         int write_disks = 0;
 422         int data_pages, parity_pages;
 423         int meta_size;
 424         int reserve;
 425         int i;
 426
 427         if (!log)
 428                 return -EAGAIN;
 429         /* Don't support stripe batch */
 430         if (sh->log_io || !test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags) ||
 431             test_bit(STRIPE_SYNCING, &sh->state)) {
 432                 /* the stripe is written to log, we start writing it to raid */
 433                 clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
 434                 return -EAGAIN;
 435         }
 436
 437         for (i = 0; i < sh->disks; i++) {
 438                 void *addr;
 439
 440                 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags))
 441                         continue;
 442                 write_disks++;
 443                 /* checksum is already calculated in last run */
 444                 if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
 445                         continue;
 446                 addr = kmap_atomic(sh->dev[i].page);
 447                 sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum,
 448                                                     addr, PAGE_SIZE);
 449                 kunmap_atomic(addr);
 450         }
 451         parity_pages = 1 + !!(sh->qd_idx >= 0);
 452         data_pages = write_disks - parity_pages;
 453
 454         meta_size =
 455                 ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32))
 456                  * data_pages) +
 457                 sizeof(struct r5l_payload_data_parity) +
 458                 sizeof(__le32) * parity_pages;
 459         /* Doesn't work with very big raid array */
 460         if (meta_size + sizeof(struct r5l_meta_block) > PAGE_SIZE)
 461                 return -EINVAL;
 462
 463         set_bit(STRIPE_LOG_TRAPPED, &sh->state);
 464         /*
 465          * The stripe must enter state machine again to finish the write, so
 466          * don't delay.
 467          */
 468         clear_bit(STRIPE_DELAYED, &sh->state);
 469         atomic_inc(&sh->count);
 470
 471         mutex_lock(&log->io_mutex);
 472         /* meta + data */
 473         reserve = (1 + write_disks) << (PAGE_SHIFT - 9);
 474         if (r5l_has_free_space(log, reserve))
 475                 r5l_log_stripe(log, sh, data_pages, parity_pages);
 476         else {
 477                 spin_lock(&log->no_space_stripes_lock);
 478                 list_add_tail(&sh->log_list, &log->no_space_stripes);
 479                 spin_unlock(&log->no_space_stripes_lock);
 480
 481                 r5l_wake_reclaim(log, reserve);
 482         }
 483         mutex_unlock(&log->io_mutex);
 484
 485         return 0;
 486 }
 487
 488 void r5l_write_stripe_run(struct r5l_log *log)
 489 {
 490         if (!log)
 491                 return;
 492         mutex_lock(&log->io_mutex);
 493         r5l_submit_current_io(log);
 494         mutex_unlock(&log->io_mutex);
 495 }
 496
 497 int r5l_handle_flush_request(struct r5l_log *log, struct bio *bio)
 498 {
 499         if (!log)
 500                 return -ENODEV;
 501         /*
 502          * we flush log disk cache first, then write stripe data to raid disks.
 503          * So if bio is finished, the log disk cache is flushed already. The
 504          * recovery guarantees we can recovery the bio from log disk, so we
 505          * don't need to flush again
 506          */
 507         if (bio->bi_iter.bi_size == 0) {
 508                 bio_endio(bio);
 509                 return 0;
 510         }
 511         bio->bi_rw &= ~REQ_FLUSH;
 512         return -EAGAIN;
 513 }
 514
 515 /* This will run after log space is reclaimed */
 516 static void r5l_run_no_space_stripes(struct r5l_log *log)
 517 {
 518         struct stripe_head *sh;
 519
 520         spin_lock(&log->no_space_stripes_lock);
 521         while (!list_empty(&log->no_space_stripes)) {
 522                 sh = list_first_entry(&log->no_space_stripes,
 523                                       struct stripe_head, log_list);
 524                 list_del_init(&sh->log_list);
 525                 set_bit(STRIPE_HANDLE, &sh->state);
 526                 raid5_release_stripe(sh);
 527         }
 528         spin_unlock(&log->no_space_stripes_lock);
 529 }
 530
 531 static sector_t r5l_reclaimable_space(struct r5l_log *log)
 532 {
 533         return r5l_ring_distance(log, log->last_checkpoint,
 534                                  log->next_checkpoint);
 535 }
 536
 537 static bool r5l_complete_finished_ios(struct r5l_log *log)
 538 {
 539         struct r5l_io_unit *io, *next;
 540         bool found = false;
 541
 542         assert_spin_locked(&log->io_list_lock);
 543
 544         list_for_each_entry_safe(io, next, &log->finished_ios, log_sibling) {
 545                 /* don't change list order */
 546                 if (io->state < IO_UNIT_STRIPE_END)
 547                         break;
 548
 549                 log->next_checkpoint = io->log_start;
 550                 log->next_cp_seq = io->seq;
 551
 552                 list_del(&io->log_sibling);
 553                 r5l_free_io_unit(log, io);
 554
 555                 found = true;
 556         }
 557
 558         return found;
 559 }
 560
 561 static void __r5l_stripe_write_finished(struct r5l_io_unit *io)
 562 {
 563         struct r5l_log *log = io->log;
 564         unsigned long flags;
 565
 566         spin_lock_irqsave(&log->io_list_lock, flags);
 567         __r5l_set_io_unit_state(io, IO_UNIT_STRIPE_END);
 568
 569         if (!r5l_complete_finished_ios(log)) {
 570                 spin_unlock_irqrestore(&log->io_list_lock, flags);
 571                 return;
 572         }
 573
 574         if (r5l_reclaimable_space(log) > log->max_free_space)
 575                 r5l_wake_reclaim(log, 0);
 576
 577         spin_unlock_irqrestore(&log->io_list_lock, flags);
 578         wake_up(&log->iounit_wait);
 579 }
 580
 581 void r5l_stripe_write_finished(struct stripe_head *sh)
 582 {
 583         struct r5l_io_unit *io;
 584
 585         io = sh->log_io;
 586         sh->log_io = NULL;
 587
 588         if (io && atomic_dec_and_test(&io->pending_stripe))
 589                 __r5l_stripe_write_finished(io);
 590 }
 591
 592 static void r5l_log_flush_endio(struct bio *bio)
 593 {
 594         struct r5l_log *log = container_of(bio, struct r5l_log,
 595                 flush_bio);
 596         unsigned long flags;
 597         struct r5l_io_unit *io;
 598
 599         spin_lock_irqsave(&log->io_list_lock, flags);
 600         list_for_each_entry(io, &log->flushing_ios, log_sibling)
 601                 r5l_io_run_stripes(io);
 602         list_splice_tail_init(&log->flushing_ios, &log->finished_ios);
 603         spin_unlock_irqrestore(&log->io_list_lock, flags);
 604 }
 605
 606 /*
 607  * Starting dispatch IO to raid.
 608  * io_unit(meta) consists of a log. There is one situation we want to avoid. A
 609  * broken meta in the middle of a log causes recovery can't find meta at the
 610  * head of log. If operations require meta at the head persistent in log, we
 611  * must make sure meta before it persistent in log too. A case is:
 612  *
 613  * stripe data/parity is in log, we start write stripe to raid disks. stripe
 614  * data/parity must be persistent in log before we do the write to raid disks.
 615  *
 616  * The solution is we restrictly maintain io_unit list order. In this case, we
 617  * only write stripes of an io_unit to raid disks till the io_unit is the first
 618  * one whose data/parity is in log.
 619  */
 620 void r5l_flush_stripe_to_raid(struct r5l_log *log)
 621 {
 622         bool do_flush;
 623         if (!log)
 624                 return;
 625
 626         spin_lock_irq(&log->io_list_lock);
 627         /* flush bio is running */
 628         if (!list_empty(&log->flushing_ios)) {
 629                 spin_unlock_irq(&log->io_list_lock);
 630                 return;
 631         }
 632         list_splice_tail_init(&log->io_end_ios, &log->flushing_ios);
 633         do_flush = !list_empty(&log->flushing_ios);
 634         spin_unlock_irq(&log->io_list_lock);
 635
 636         if (!do_flush)
 637                 return;
 638         bio_reset(&log->flush_bio);
 639         log->flush_bio.bi_bdev = log->rdev->bdev;
 640         log->flush_bio.bi_end_io = r5l_log_flush_endio;
 641         submit_bio(WRITE_FLUSH, &log->flush_bio);
 642 }
 643
 644 static void r5l_write_super(struct r5l_log *log, sector_t cp);
 645 static void r5l_do_reclaim(struct r5l_log *log)
 646 {
 647         sector_t reclaim_target = xchg(&log->reclaim_target, 0);
 648         sector_t reclaimable;
 649         sector_t next_checkpoint;
 650         u64 next_cp_seq;
 651
 652         spin_lock_irq(&log->io_list_lock);
 653         /*
 654          * move proper io_unit to reclaim list. We should not change the order.
 655          * reclaimable/unreclaimable io_unit can be mixed in the list, we
 656          * shouldn't reuse space of an unreclaimable io_unit
 657          */
 658         while (1) {
 659                 reclaimable = r5l_reclaimable_space(log);
 660                 if (reclaimable >= reclaim_target ||
 661                     (list_empty(&log->running_ios) &&
 662                      list_empty(&log->io_end_ios) &&
 663                      list_empty(&log->flushing_ios) &&
 664                      list_empty(&log->finished_ios)))
 665                         break;
 666
 667                 md_wakeup_thread(log->rdev->mddev->thread);
 668                 wait_event_lock_irq(log->iounit_wait,
 669                                     r5l_reclaimable_space(log) > reclaimable,
 670                                     log->io_list_lock);
 671         }
 672
 673         next_checkpoint = log->next_checkpoint;
 674         next_cp_seq = log->next_cp_seq;
 675         spin_unlock_irq(&log->io_list_lock);
 676
 677         BUG_ON(reclaimable < 0);
 678         if (reclaimable == 0)
 679                 return;
 680
 681         /*
 682          * write_super will flush cache of each raid disk. We must write super
 683          * here, because the log area might be reused soon and we don't want to
 684          * confuse recovery
 685          */
 686         r5l_write_super(log, next_checkpoint);
 687
 688         mutex_lock(&log->io_mutex);
 689         log->last_checkpoint = next_checkpoint;
 690         log->last_cp_seq = next_cp_seq;
 691         mutex_unlock(&log->io_mutex);
 692
 693         r5l_run_no_space_stripes(log);
 694 }
 695
 696 static void r5l_reclaim_thread(struct md_thread *thread)
 697 {
 698         struct mddev *mddev = thread->mddev;
 699         struct r5conf *conf = mddev->private;
 700         struct r5l_log *log = conf->log;
 701
 702         if (!log)
 703                 return;
 704         r5l_do_reclaim(log);
 705 }
 706
 707 static void r5l_wake_reclaim(struct r5l_log *log, sector_t space)
 708 {
 709         unsigned long target;
 710         unsigned long new = (unsigned long)space; /* overflow in theory */
 711
 712         do {
 713                 target = log->reclaim_target;
 714                 if (new < target)
 715                         return;
 716         } while (cmpxchg(&log->reclaim_target, target, new) != target);
 717         md_wakeup_thread(log->reclaim_thread);
 718 }
 719
 720 void r5l_quiesce(struct r5l_log *log, int state)
 721 {
 722         if (!log || state == 2)
 723                 return;
 724         if (state == 0) {
 725                 log->reclaim_thread = md_register_thread(r5l_reclaim_thread,
 726                                         log->rdev->mddev, "reclaim");
 727         } else if (state == 1) {
 728                 /*
 729                  * at this point all stripes are finished, so io_unit is at
 730                  * least in STRIPE_END state
 731                  */
 732                 r5l_wake_reclaim(log, -1L);
 733                 md_unregister_thread(&log->reclaim_thread);
 734                 r5l_do_reclaim(log);
 735         }
 736 }
 737
 738 struct r5l_recovery_ctx {
 739         struct page *meta_page;         /* current meta */
 740         sector_t meta_total_blocks;     /* total size of current meta and data */
 741         sector_t pos;                   /* recovery position */
 742         u64 seq;                        /* recovery position seq */
 743 };
 744
 745 static int r5l_read_meta_block(struct r5l_log *log,
 746                                struct r5l_recovery_ctx *ctx)
 747 {
 748         struct page *page = ctx->meta_page;
 749         struct r5l_meta_block *mb;
 750         u32 crc, stored_crc;
 751
 752         if (!sync_page_io(log->rdev, ctx->pos, PAGE_SIZE, page, READ, false))
 753                 return -EIO;
 754
 755         mb = page_address(page);
 756         stored_crc = le32_to_cpu(mb->checksum);
 757         mb->checksum = 0;
 758
 759         if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
 760             le64_to_cpu(mb->seq) != ctx->seq ||
 761             mb->version != R5LOG_VERSION ||
 762             le64_to_cpu(mb->position) != ctx->pos)
 763                 return -EINVAL;
 764
 765         crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
 766         if (stored_crc != crc)
 767                 return -EINVAL;
 768
 769         if (le32_to_cpu(mb->meta_size) > PAGE_SIZE)
 770                 return -EINVAL;
 771
 772         ctx->meta_total_blocks = BLOCK_SECTORS;
 773
 774         return 0;
 775 }
 776
 777 static int r5l_recovery_flush_one_stripe(struct r5l_log *log,
 778                                          struct r5l_recovery_ctx *ctx,
 779                                          sector_t stripe_sect,
 780                                          int *offset, sector_t *log_offset)
 781 {
 782         struct r5conf *conf = log->rdev->mddev->private;
 783         struct stripe_head *sh;
 784         struct r5l_payload_data_parity *payload;
 785         int disk_index;
 786
 787         sh = raid5_get_active_stripe(conf, stripe_sect, 0, 0, 0);
 788         while (1) {
 789                 payload = page_address(ctx->meta_page) + *offset;
 790
 791                 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) {
 792                         raid5_compute_sector(conf,
 793                                              le64_to_cpu(payload->location), 0,
 794                                              &disk_index, sh);
 795
 796                         sync_page_io(log->rdev, *log_offset, PAGE_SIZE,
 797                                      sh->dev[disk_index].page, READ, false);
 798                         sh->dev[disk_index].log_checksum =
 799                                 le32_to_cpu(payload->checksum[0]);
 800                         set_bit(R5_Wantwrite, &sh->dev[disk_index].flags);
 801                         ctx->meta_total_blocks += BLOCK_SECTORS;
 802                 } else {
 803                         disk_index = sh->pd_idx;
 804                         sync_page_io(log->rdev, *log_offset, PAGE_SIZE,
 805                                      sh->dev[disk_index].page, READ, false);
 806                         sh->dev[disk_index].log_checksum =
 807                                 le32_to_cpu(payload->checksum[0]);
 808                         set_bit(R5_Wantwrite, &sh->dev[disk_index].flags);
 809
 810                         if (sh->qd_idx >= 0) {
 811                                 disk_index = sh->qd_idx;
 812                                 sync_page_io(log->rdev,
 813                                              r5l_ring_add(log, *log_offset, BLOCK_SECTORS),
 814                                              PAGE_SIZE, sh->dev[disk_index].page,
 815                                              READ, false);
 816                                 sh->dev[disk_index].log_checksum =
 817                                         le32_to_cpu(payload->checksum[1]);
 818                                 set_bit(R5_Wantwrite,
 819                                         &sh->dev[disk_index].flags);
 820                         }
 821                         ctx->meta_total_blocks += BLOCK_SECTORS * conf->max_degraded;
 822                 }
 823
 824                 *log_offset = r5l_ring_add(log, *log_offset,
 825                                            le32_to_cpu(payload->size));
 826                 *offset += sizeof(struct r5l_payload_data_parity) +
 827                         sizeof(__le32) *
 828                         (le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
 829                 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_PARITY)
 830                         break;
 831         }
 832
 833         for (disk_index = 0; disk_index < sh->disks; disk_index++) {
 834                 void *addr;
 835                 u32 checksum;
 836
 837                 if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
 838                         continue;
 839                 addr = kmap_atomic(sh->dev[disk_index].page);
 840                 checksum = crc32c_le(log->uuid_checksum, addr, PAGE_SIZE);
 841                 kunmap_atomic(addr);
 842                 if (checksum != sh->dev[disk_index].log_checksum)
 843                         goto error;
 844         }
 845
 846         for (disk_index = 0; disk_index < sh->disks; disk_index++) {
 847                 struct md_rdev *rdev, *rrdev;
 848
 849                 if (!test_and_clear_bit(R5_Wantwrite,
 850                                         &sh->dev[disk_index].flags))
 851                         continue;
 852
 853                 /* in case device is broken */
 854                 rdev = rcu_dereference(conf->disks[disk_index].rdev);
 855                 if (rdev)
 856                         sync_page_io(rdev, stripe_sect, PAGE_SIZE,
 857                                      sh->dev[disk_index].page, WRITE, false);
 858                 rrdev = rcu_dereference(conf->disks[disk_index].replacement);
 859                 if (rrdev)
 860                         sync_page_io(rrdev, stripe_sect, PAGE_SIZE,
 861                                      sh->dev[disk_index].page, WRITE, false);
 862         }
 863         raid5_release_stripe(sh);
 864         return 0;
 865
 866 error:
 867         for (disk_index = 0; disk_index < sh->disks; disk_index++)
 868                 sh->dev[disk_index].flags = 0;
 869         raid5_release_stripe(sh);
 870         return -EINVAL;
 871 }
 872
 873 static int r5l_recovery_flush_one_meta(struct r5l_log *log,
 874                                        struct r5l_recovery_ctx *ctx)
 875 {
 876         struct r5conf *conf = log->rdev->mddev->private;
 877         struct r5l_payload_data_parity *payload;
 878         struct r5l_meta_block *mb;
 879         int offset;
 880         sector_t log_offset;
 881         sector_t stripe_sector;
 882
 883         mb = page_address(ctx->meta_page);
 884         offset = sizeof(struct r5l_meta_block);
 885         log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
 886
 887         while (offset < le32_to_cpu(mb->meta_size)) {
 888                 int dd;
 889
 890                 payload = (void *)mb + offset;
 891                 stripe_sector = raid5_compute_sector(conf,
 892                                                      le64_to_cpu(payload->location), 0, &dd, NULL);
 893                 if (r5l_recovery_flush_one_stripe(log, ctx, stripe_sector,
 894                                                   &offset, &log_offset))
 895                         return -EINVAL;
 896         }
 897         return 0;
 898 }
 899
 900 /* copy data/parity from log to raid disks */
 901 static void r5l_recovery_flush_log(struct r5l_log *log,
 902                                    struct r5l_recovery_ctx *ctx)
 903 {
 904         while (1) {
 905                 if (r5l_read_meta_block(log, ctx))
 906                         return;
 907                 if (r5l_recovery_flush_one_meta(log, ctx))
 908                         return;
 909                 ctx->seq++;
 910                 ctx->pos = r5l_ring_add(log, ctx->pos, ctx->meta_total_blocks);
 911         }
 912 }
 913
 914 static int r5l_log_write_empty_meta_block(struct r5l_log *log, sector_t pos,
 915                                           u64 seq)
 916 {
 917         struct page *page;
 918         struct r5l_meta_block *mb;
 919         u32 crc;
 920
 921         page = alloc_page(GFP_KERNEL | __GFP_ZERO);
 922         if (!page)
 923                 return -ENOMEM;
 924         mb = page_address(page);
 925         mb->magic = cpu_to_le32(R5LOG_MAGIC);
 926         mb->version = R5LOG_VERSION;
 927         mb->meta_size = cpu_to_le32(sizeof(struct r5l_meta_block));
 928         mb->seq = cpu_to_le64(seq);
 929         mb->position = cpu_to_le64(pos);
 930         crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
 931         mb->checksum = cpu_to_le32(crc);
 932
 933         if (!sync_page_io(log->rdev, pos, PAGE_SIZE, page, WRITE_FUA, false)) {
 934                 __free_page(page);
 935                 return -EIO;
 936         }
 937         __free_page(page);
 938         return 0;
 939 }
 940
 941 static int r5l_recovery_log(struct r5l_log *log)
 942 {
 943         struct r5l_recovery_ctx ctx;
 944
 945         ctx.pos = log->last_checkpoint;
 946         ctx.seq = log->last_cp_seq;
 947         ctx.meta_page = alloc_page(GFP_KERNEL);
 948         if (!ctx.meta_page)
 949                 return -ENOMEM;
 950
 951         r5l_recovery_flush_log(log, &ctx);
 952         __free_page(ctx.meta_page);
 953
 954         /*
 955          * we did a recovery. Now ctx.pos points to an invalid meta block. New
 956          * log will start here. but we can't let superblock point to last valid
 957          * meta block. The log might looks like:
 958          * | meta 1| meta 2| meta 3|
 959          * meta 1 is valid, meta 2 is invalid. meta 3 could be valid. If
 960          * superblock points to meta 1, we write a new valid meta 2n.  if crash
 961          * happens again, new recovery will start from meta 1. Since meta 2n is
 962          * valid now, recovery will think meta 3 is valid, which is wrong.
 963          * The solution is we create a new meta in meta2 with its seq == meta
 964          * 1's seq + 10 and let superblock points to meta2. The same recovery will
 965          * not think meta 3 is a valid meta, because its seq doesn't match
 966          */
 967         if (ctx.seq > log->last_cp_seq + 1) {
 968                 int ret;
 969
 970                 ret = r5l_log_write_empty_meta_block(log, ctx.pos, ctx.seq + 10);
 971                 if (ret)
 972                         return ret;
 973                 log->seq = ctx.seq + 11;
 974                 log->log_start = r5l_ring_add(log, ctx.pos, BLOCK_SECTORS);
 975                 r5l_write_super(log, ctx.pos);
 976         } else {
 977                 log->log_start = ctx.pos;
 978                 log->seq = ctx.seq;
 979         }
 980         return 0;
 981 }
 982
 983 static void r5l_write_super(struct r5l_log *log, sector_t cp)
 984 {
 985         struct mddev *mddev = log->rdev->mddev;
 986
 987         log->rdev->journal_tail = cp;
 988         set_bit(MD_CHANGE_DEVS, &mddev->flags);
 989 }
 990
 991 static int r5l_load_log(struct r5l_log *log)
 992 {
 993         struct md_rdev *rdev = log->rdev;
 994         struct page *page;
 995         struct r5l_meta_block *mb;
 996         sector_t cp = log->rdev->journal_tail;
 997         u32 stored_crc, expected_crc;
 998         bool create_super = false;
 999         int ret;
1000
1001         /* Make sure it's valid */
1002         if (cp >= rdev->sectors || round_down(cp, BLOCK_SECTORS) != cp)
1003                 cp = 0;
1004         page = alloc_page(GFP_KERNEL);
1005         if (!page)
1006                 return -ENOMEM;
1007
1008         if (!sync_page_io(rdev, cp, PAGE_SIZE, page, READ, false)) {
1009                 ret = -EIO;
1010                 goto ioerr;
1011         }
1012         mb = page_address(page);
1013
1014         if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
1015             mb->version != R5LOG_VERSION) {
1016                 create_super = true;
1017                 goto create;
1018         }
1019         stored_crc = le32_to_cpu(mb->checksum);
1020         mb->checksum = 0;
1021         expected_crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
1022         if (stored_crc != expected_crc) {
1023                 create_super = true;
1024                 goto create;
1025         }
1026         if (le64_to_cpu(mb->position) != cp) {
1027                 create_super = true;
1028                 goto create;
1029         }
1030 create:
1031         if (create_super) {
1032                 log->last_cp_seq = prandom_u32();
1033                 cp = 0;
1034                 /*
1035                  * Make sure super points to correct address. Log might have
1036                  * data very soon. If super hasn't correct log tail address,
1037                  * recovery can't find the log
1038                  */
1039                 r5l_write_super(log, cp);
1040         } else
1041                 log->last_cp_seq = le64_to_cpu(mb->seq);
1042
1043         log->device_size = round_down(rdev->sectors, BLOCK_SECTORS);
1044         log->max_free_space = log->device_size >> RECLAIM_MAX_FREE_SPACE_SHIFT;
1045         if (log->max_free_space > RECLAIM_MAX_FREE_SPACE)
1046                 log->max_free_space = RECLAIM_MAX_FREE_SPACE;
1047         log->last_checkpoint = cp;
1048
1049         __free_page(page);
1050
1051         return r5l_recovery_log(log);
1052 ioerr:
1053         __free_page(page);
1054         return ret;
1055 }
1056
1057 int r5l_init_log(struct r5conf *conf, struct md_rdev *rdev)
1058 {
1059         struct r5l_log *log;
1060
1061         if (PAGE_SIZE != 4096)
1062                 return -EINVAL;
1063         log = kzalloc(sizeof(*log), GFP_KERNEL);
1064         if (!log)
1065                 return -ENOMEM;
1066         log->rdev = rdev;
1067
1068         log->uuid_checksum = crc32c_le(~0, rdev->mddev->uuid,
1069                                        sizeof(rdev->mddev->uuid));
1070
1071         mutex_init(&log->io_mutex);
1072
1073         spin_lock_init(&log->io_list_lock);
1074         INIT_LIST_HEAD(&log->running_ios);
1075         INIT_LIST_HEAD(&log->io_end_ios);
1076         INIT_LIST_HEAD(&log->flushing_ios);
1077         INIT_LIST_HEAD(&log->finished_ios);
1078         bio_init(&log->flush_bio);
1079
1080         log->io_kc = KMEM_CACHE(r5l_io_unit, 0);
1081         if (!log->io_kc)
1082                 goto io_kc;
1083
1084         log->reclaim_thread = md_register_thread(r5l_reclaim_thread,
1085                                                  log->rdev->mddev, "reclaim");
1086         if (!log->reclaim_thread)
1087                 goto reclaim_thread;
1088         init_waitqueue_head(&log->iounit_wait);
1089
1090         INIT_LIST_HEAD(&log->no_space_stripes);
1091         spin_lock_init(&log->no_space_stripes_lock);
1092
1093         if (r5l_load_log(log))
1094                 goto error;
1095
1096         conf->log = log;
1097         return 0;
1098 error:
1099         md_unregister_thread(&log->reclaim_thread);
1100 reclaim_thread:
1101         kmem_cache_destroy(log->io_kc);
1102 io_kc:
1103         kfree(log);
1104         return -EINVAL;
1105 }
1106
1107 void r5l_exit_log(struct r5l_log *log)
1108 {
1109         md_unregister_thread(&log->reclaim_thread);
1110         kmem_cache_destroy(log->io_kc);
1111         kfree(log);
1112 }