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         struct mutex io_mutex;
  55         struct r5l_io_unit *current_io; /* current io_unit accepting new data */
  56
  57         spinlock_t io_list_lock;
  58         struct list_head running_ios;   /* io_units which are still running,
  59                                          * and have not yet been completely
  60                                          * written to the log */
  61         struct list_head io_end_ios;    /* io_units which have been completely
  62                                          * written to the log but not yet written
  63                                          * to the RAID */
  64         struct list_head flushing_ios;  /* io_units which are waiting for log
  65                                          * cache flush */
  66         struct list_head flushed_ios;   /* io_units which settle down in log disk */
  67         struct bio flush_bio;
  68         struct list_head stripe_end_ios;/* io_units which have been completely
  69                                          * written to the RAID but have not yet
  70                                          * been considered for updating super */
  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 /*
 190  * We don't want too many io_units reside in stripe_end_ios list, which will
 191  * waste a lot of memory. So we try to remove some. But we must keep at least 2
 192  * io_units. The superblock must point to a valid meta, if it's the last meta,
 193  * recovery can scan less
 194  */
 195 static void r5l_compress_stripe_end_list(struct r5l_log *log)
 196 {
 197         struct r5l_io_unit *first, *last, *io;
 198
 199         first = list_first_entry(&log->stripe_end_ios,
 200                                  struct r5l_io_unit, log_sibling);
 201         last = list_last_entry(&log->stripe_end_ios,
 202                                struct r5l_io_unit, log_sibling);
 203         if (first == last)
 204                 return;
 205         list_del(&first->log_sibling);
 206         list_del(&last->log_sibling);
 207         while (!list_empty(&log->stripe_end_ios)) {
 208                 io = list_first_entry(&log->stripe_end_ios,
 209                                       struct r5l_io_unit, log_sibling);
 210                 list_del(&io->log_sibling);
 211                 first->log_end = io->log_end;
 212                 r5l_free_io_unit(log, io);
 213         }
 214         list_add_tail(&first->log_sibling, &log->stripe_end_ios);
 215         list_add_tail(&last->log_sibling, &log->stripe_end_ios);
 216 }
 217
 218 static void __r5l_set_io_unit_state(struct r5l_io_unit *io,
 219                                     enum r5l_io_unit_state state)
 220 {
 221         if (WARN_ON(io->state >= state))
 222                 return;
 223         io->state = state;
 224 }
 225
 226 /* XXX: totally ignores I/O errors */
 227 static void r5l_log_endio(struct bio *bio)
 228 {
 229         struct r5l_io_unit *io = bio->bi_private;
 230         struct r5l_log *log = io->log;
 231         unsigned long flags;
 232
 233         bio_put(bio);
 234
 235         if (!atomic_dec_and_test(&io->pending_io))
 236                 return;
 237
 238         spin_lock_irqsave(&log->io_list_lock, flags);
 239         __r5l_set_io_unit_state(io, IO_UNIT_IO_END);
 240         r5l_move_io_unit_list(&log->running_ios, &log->io_end_ios,
 241                         IO_UNIT_IO_END);
 242         spin_unlock_irqrestore(&log->io_list_lock, flags);
 243
 244         md_wakeup_thread(log->rdev->mddev->thread);
 245 }
 246
 247 static void r5l_submit_current_io(struct r5l_log *log)
 248 {
 249         struct r5l_io_unit *io = log->current_io;
 250         struct r5l_meta_block *block;
 251         struct bio *bio;
 252         unsigned long flags;
 253         u32 crc;
 254
 255         if (!io)
 256                 return;
 257
 258         block = page_address(io->meta_page);
 259         block->meta_size = cpu_to_le32(io->meta_offset);
 260         crc = crc32c_le(log->uuid_checksum, block, PAGE_SIZE);
 261         block->checksum = cpu_to_le32(crc);
 262
 263         log->current_io = NULL;
 264         spin_lock_irqsave(&log->io_list_lock, flags);
 265         __r5l_set_io_unit_state(io, IO_UNIT_IO_START);
 266         spin_unlock_irqrestore(&log->io_list_lock, flags);
 267
 268         while ((bio = bio_list_pop(&io->bios))) {
 269                 /* all IO must start from rdev->data_offset */
 270                 bio->bi_iter.bi_sector += log->rdev->data_offset;
 271                 submit_bio(WRITE, bio);
 272         }
 273 }
 274
 275 static struct r5l_io_unit *r5l_new_meta(struct r5l_log *log)
 276 {
 277         struct r5l_io_unit *io;
 278         struct r5l_meta_block *block;
 279         struct bio *bio;
 280
 281         io = r5l_alloc_io_unit(log);
 282
 283         block = page_address(io->meta_page);
 284         block->magic = cpu_to_le32(R5LOG_MAGIC);
 285         block->version = R5LOG_VERSION;
 286         block->seq = cpu_to_le64(log->seq);
 287         block->position = cpu_to_le64(log->log_start);
 288
 289         io->log_start = log->log_start;
 290         io->meta_offset = sizeof(struct r5l_meta_block);
 291         io->seq = log->seq;
 292
 293         bio = bio_kmalloc(GFP_NOIO | __GFP_NOFAIL, BIO_MAX_PAGES);
 294         io->current_bio = bio;
 295         bio->bi_rw = WRITE;
 296         bio->bi_bdev = log->rdev->bdev;
 297         bio->bi_iter.bi_sector = log->log_start;
 298         bio_add_page(bio, io->meta_page, PAGE_SIZE, 0);
 299         bio->bi_end_io = r5l_log_endio;
 300         bio->bi_private = io;
 301
 302         bio_list_add(&io->bios, bio);
 303         atomic_inc(&io->pending_io);
 304
 305         log->seq++;
 306         log->log_start = r5l_ring_add(log, log->log_start, BLOCK_SECTORS);
 307         io->log_end = log->log_start;
 308         /* current bio hit disk end */
 309         if (log->log_start == 0)
 310                 io->current_bio = NULL;
 311
 312         spin_lock_irq(&log->io_list_lock);
 313         list_add_tail(&io->log_sibling, &log->running_ios);
 314         spin_unlock_irq(&log->io_list_lock);
 315
 316         return io;
 317 }
 318
 319 static int r5l_get_meta(struct r5l_log *log, unsigned int payload_size)
 320 {
 321         struct r5l_io_unit *io;
 322
 323         io = log->current_io;
 324         if (io && io->meta_offset + payload_size > PAGE_SIZE)
 325                 r5l_submit_current_io(log);
 326         io = log->current_io;
 327         if (io)
 328                 return 0;
 329
 330         log->current_io = r5l_new_meta(log);
 331         return 0;
 332 }
 333
 334 static void r5l_append_payload_meta(struct r5l_log *log, u16 type,
 335                                     sector_t location,
 336                                     u32 checksum1, u32 checksum2,
 337                                     bool checksum2_valid)
 338 {
 339         struct r5l_io_unit *io = log->current_io;
 340         struct r5l_payload_data_parity *payload;
 341
 342         payload = page_address(io->meta_page) + io->meta_offset;
 343         payload->header.type = cpu_to_le16(type);
 344         payload->header.flags = cpu_to_le16(0);
 345         payload->size = cpu_to_le32((1 + !!checksum2_valid) <<
 346                                     (PAGE_SHIFT - 9));
 347         payload->location = cpu_to_le64(location);
 348         payload->checksum[0] = cpu_to_le32(checksum1);
 349         if (checksum2_valid)
 350                 payload->checksum[1] = cpu_to_le32(checksum2);
 351
 352         io->meta_offset += sizeof(struct r5l_payload_data_parity) +
 353                 sizeof(__le32) * (1 + !!checksum2_valid);
 354 }
 355
 356 static void r5l_append_payload_page(struct r5l_log *log, struct page *page)
 357 {
 358         struct r5l_io_unit *io = log->current_io;
 359
 360 alloc_bio:
 361         if (!io->current_bio) {
 362                 struct bio *bio;
 363
 364                 bio = bio_kmalloc(GFP_NOIO | __GFP_NOFAIL, BIO_MAX_PAGES);
 365                 bio->bi_rw = WRITE;
 366                 bio->bi_bdev = log->rdev->bdev;
 367                 bio->bi_iter.bi_sector = log->log_start;
 368                 bio->bi_end_io = r5l_log_endio;
 369                 bio->bi_private = io;
 370                 bio_list_add(&io->bios, bio);
 371                 atomic_inc(&io->pending_io);
 372                 io->current_bio = bio;
 373         }
 374         if (!bio_add_page(io->current_bio, page, PAGE_SIZE, 0)) {
 375                 io->current_bio = NULL;
 376                 goto alloc_bio;
 377         }
 378         log->log_start = r5l_ring_add(log, log->log_start,
 379                                       BLOCK_SECTORS);
 380         /* current bio hit disk end */
 381         if (log->log_start == 0)
 382                 io->current_bio = NULL;
 383
 384         io->log_end = log->log_start;
 385 }
 386
 387 static void r5l_log_stripe(struct r5l_log *log, struct stripe_head *sh,
 388                            int data_pages, int parity_pages)
 389 {
 390         int i;
 391         int meta_size;
 392         struct r5l_io_unit *io;
 393
 394         meta_size =
 395                 ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32))
 396                  * data_pages) +
 397                 sizeof(struct r5l_payload_data_parity) +
 398                 sizeof(__le32) * parity_pages;
 399
 400         r5l_get_meta(log, meta_size);
 401         io = log->current_io;
 402
 403         for (i = 0; i < sh->disks; i++) {
 404                 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags))
 405                         continue;
 406                 if (i == sh->pd_idx || i == sh->qd_idx)
 407                         continue;
 408                 r5l_append_payload_meta(log, R5LOG_PAYLOAD_DATA,
 409                                         raid5_compute_blocknr(sh, i, 0),
 410                                         sh->dev[i].log_checksum, 0, false);
 411                 r5l_append_payload_page(log, sh->dev[i].page);
 412         }
 413
 414         if (sh->qd_idx >= 0) {
 415                 r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
 416                                         sh->sector, sh->dev[sh->pd_idx].log_checksum,
 417                                         sh->dev[sh->qd_idx].log_checksum, true);
 418                 r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
 419                 r5l_append_payload_page(log, sh->dev[sh->qd_idx].page);
 420         } else {
 421                 r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
 422                                         sh->sector, sh->dev[sh->pd_idx].log_checksum,
 423                                         0, false);
 424                 r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
 425         }
 426
 427         list_add_tail(&sh->log_list, &io->stripe_list);
 428         atomic_inc(&io->pending_stripe);
 429         sh->log_io = io;
 430 }
 431
 432 static void r5l_wake_reclaim(struct r5l_log *log, sector_t space);
 433 /*
 434  * running in raid5d, where reclaim could wait for raid5d too (when it flushes
 435  * data from log to raid disks), so we shouldn't wait for reclaim here
 436  */
 437 int r5l_write_stripe(struct r5l_log *log, struct stripe_head *sh)
 438 {
 439         int write_disks = 0;
 440         int data_pages, parity_pages;
 441         int meta_size;
 442         int reserve;
 443         int i;
 444
 445         if (!log)
 446                 return -EAGAIN;
 447         /* Don't support stripe batch */
 448         if (sh->log_io || !test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags) ||
 449             test_bit(STRIPE_SYNCING, &sh->state)) {
 450                 /* the stripe is written to log, we start writing it to raid */
 451                 clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
 452                 return -EAGAIN;
 453         }
 454
 455         for (i = 0; i < sh->disks; i++) {
 456                 void *addr;
 457
 458                 if (!test_bit(R5_Wantwrite, &sh->dev[i].flags))
 459                         continue;
 460                 write_disks++;
 461                 /* checksum is already calculated in last run */
 462                 if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
 463                         continue;
 464                 addr = kmap_atomic(sh->dev[i].page);
 465                 sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum,
 466                                                     addr, PAGE_SIZE);
 467                 kunmap_atomic(addr);
 468         }
 469         parity_pages = 1 + !!(sh->qd_idx >= 0);
 470         data_pages = write_disks - parity_pages;
 471
 472         meta_size =
 473                 ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32))
 474                  * data_pages) +
 475                 sizeof(struct r5l_payload_data_parity) +
 476                 sizeof(__le32) * parity_pages;
 477         /* Doesn't work with very big raid array */
 478         if (meta_size + sizeof(struct r5l_meta_block) > PAGE_SIZE)
 479                 return -EINVAL;
 480
 481         set_bit(STRIPE_LOG_TRAPPED, &sh->state);
 482         atomic_inc(&sh->count);
 483
 484         mutex_lock(&log->io_mutex);
 485         /* meta + data */
 486         reserve = (1 + write_disks) << (PAGE_SHIFT - 9);
 487         if (r5l_has_free_space(log, reserve))
 488                 r5l_log_stripe(log, sh, data_pages, parity_pages);
 489         else {
 490                 spin_lock(&log->no_space_stripes_lock);
 491                 list_add_tail(&sh->log_list, &log->no_space_stripes);
 492                 spin_unlock(&log->no_space_stripes_lock);
 493
 494                 r5l_wake_reclaim(log, reserve);
 495         }
 496         mutex_unlock(&log->io_mutex);
 497
 498         return 0;
 499 }
 500
 501 void r5l_write_stripe_run(struct r5l_log *log)
 502 {
 503         if (!log)
 504                 return;
 505         mutex_lock(&log->io_mutex);
 506         r5l_submit_current_io(log);
 507         mutex_unlock(&log->io_mutex);
 508 }
 509
 510 int r5l_handle_flush_request(struct r5l_log *log, struct bio *bio)
 511 {
 512         if (!log)
 513                 return -ENODEV;
 514         /*
 515          * we flush log disk cache first, then write stripe data to raid disks.
 516          * So if bio is finished, the log disk cache is flushed already. The
 517          * recovery guarantees we can recovery the bio from log disk, so we
 518          * don't need to flush again
 519          */
 520         if (bio->bi_iter.bi_size == 0) {
 521                 bio_endio(bio);
 522                 return 0;
 523         }
 524         bio->bi_rw &= ~REQ_FLUSH;
 525         return -EAGAIN;
 526 }
 527
 528 /* This will run after log space is reclaimed */
 529 static void r5l_run_no_space_stripes(struct r5l_log *log)
 530 {
 531         struct stripe_head *sh;
 532
 533         spin_lock(&log->no_space_stripes_lock);
 534         while (!list_empty(&log->no_space_stripes)) {
 535                 sh = list_first_entry(&log->no_space_stripes,
 536                                       struct stripe_head, log_list);
 537                 list_del_init(&sh->log_list);
 538                 set_bit(STRIPE_HANDLE, &sh->state);
 539                 raid5_release_stripe(sh);
 540         }
 541         spin_unlock(&log->no_space_stripes_lock);
 542 }
 543
 544 static void __r5l_stripe_write_finished(struct r5l_io_unit *io)
 545 {
 546         struct r5l_log *log = io->log;
 547         struct r5l_io_unit *last;
 548         sector_t reclaimable_space;
 549         unsigned long flags;
 550
 551         spin_lock_irqsave(&log->io_list_lock, flags);
 552         __r5l_set_io_unit_state(io, IO_UNIT_STRIPE_END);
 553         /* might move 0 entry */
 554         r5l_move_io_unit_list(&log->flushed_ios, &log->stripe_end_ios,
 555                               IO_UNIT_STRIPE_END);
 556         if (list_empty(&log->stripe_end_ios)) {
 557                 spin_unlock_irqrestore(&log->io_list_lock, flags);
 558                 return;
 559         }
 560
 561         last = list_last_entry(&log->stripe_end_ios,
 562                                struct r5l_io_unit, log_sibling);
 563         reclaimable_space = r5l_ring_distance(log, log->last_checkpoint,
 564                                               last->log_end);
 565         if (reclaimable_space >= log->max_free_space)
 566                 r5l_wake_reclaim(log, 0);
 567
 568         r5l_compress_stripe_end_list(log);
 569         spin_unlock_irqrestore(&log->io_list_lock, flags);
 570         wake_up(&log->iounit_wait);
 571 }
 572
 573 void r5l_stripe_write_finished(struct stripe_head *sh)
 574 {
 575         struct r5l_io_unit *io;
 576
 577         io = sh->log_io;
 578         sh->log_io = NULL;
 579
 580         if (io && atomic_dec_and_test(&io->pending_stripe))
 581                 __r5l_stripe_write_finished(io);
 582 }
 583
 584 static void r5l_log_flush_endio(struct bio *bio)
 585 {
 586         struct r5l_log *log = container_of(bio, struct r5l_log,
 587                 flush_bio);
 588         unsigned long flags;
 589         struct r5l_io_unit *io;
 590         struct stripe_head *sh;
 591
 592         spin_lock_irqsave(&log->io_list_lock, flags);
 593         list_for_each_entry(io, &log->flushing_ios, log_sibling) {
 594                 while (!list_empty(&io->stripe_list)) {
 595                         sh = list_first_entry(&io->stripe_list,
 596                                 struct stripe_head, log_list);
 597                         list_del_init(&sh->log_list);
 598                         set_bit(STRIPE_HANDLE, &sh->state);
 599                         raid5_release_stripe(sh);
 600                 }
 601         }
 602         list_splice_tail_init(&log->flushing_ios, &log->flushed_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_kick_io_unit(struct r5l_log *log)
 645 {
 646         md_wakeup_thread(log->rdev->mddev->thread);
 647         wait_event_lock_irq(log->iounit_wait, !list_empty(&log->stripe_end_ios),
 648                             log->io_list_lock);
 649 }
 650
 651 static void r5l_write_super(struct r5l_log *log, sector_t cp);
 652 static void r5l_do_reclaim(struct r5l_log *log)
 653 {
 654         struct r5l_io_unit *io, *last;
 655         LIST_HEAD(list);
 656         sector_t free = 0;
 657         sector_t reclaim_target = xchg(&log->reclaim_target, 0);
 658
 659         spin_lock_irq(&log->io_list_lock);
 660         /*
 661          * move proper io_unit to reclaim list. We should not change the order.
 662          * reclaimable/unreclaimable io_unit can be mixed in the list, we
 663          * shouldn't reuse space of an unreclaimable io_unit
 664          */
 665         while (1) {
 666                 struct list_head *target_list = NULL;
 667
 668                 while (!list_empty(&log->stripe_end_ios)) {
 669                         io = list_first_entry(&log->stripe_end_ios,
 670                                               struct r5l_io_unit, log_sibling);
 671                         list_move_tail(&io->log_sibling, &list);
 672                         free += r5l_ring_distance(log, io->log_start,
 673                                                   io->log_end);
 674                 }
 675
 676                 if (free >= reclaim_target ||
 677                     (list_empty(&log->running_ios) &&
 678                      list_empty(&log->io_end_ios) &&
 679                      list_empty(&log->flushing_ios) &&
 680                      list_empty(&log->flushed_ios)))
 681                         break;
 682
 683                 /* Below waiting mostly happens when we shutdown the raid */
 684                 if (!list_empty(&log->flushed_ios))
 685                         target_list = &log->flushed_ios;
 686                 else if (!list_empty(&log->flushing_ios))
 687                         target_list = &log->flushing_ios;
 688                 else if (!list_empty(&log->io_end_ios))
 689                         target_list = &log->io_end_ios;
 690                 else if (!list_empty(&log->running_ios))
 691                         target_list = &log->running_ios;
 692
 693                 r5l_kick_io_unit(log);
 694         }
 695         spin_unlock_irq(&log->io_list_lock);
 696
 697         if (list_empty(&list))
 698                 return;
 699
 700         /* super always point to last valid meta */
 701         last = list_last_entry(&list, struct r5l_io_unit, log_sibling);
 702         /*
 703          * write_super will flush cache of each raid disk. We must write super
 704          * here, because the log area might be reused soon and we don't want to
 705          * confuse recovery
 706          */
 707         r5l_write_super(log, last->log_start);
 708
 709         mutex_lock(&log->io_mutex);
 710         log->last_checkpoint = last->log_start;
 711         log->last_cp_seq = last->seq;
 712         mutex_unlock(&log->io_mutex);
 713         r5l_run_no_space_stripes(log);
 714
 715         while (!list_empty(&list)) {
 716                 io = list_first_entry(&list, struct r5l_io_unit, log_sibling);
 717                 list_del(&io->log_sibling);
 718                 r5l_free_io_unit(log, io);
 719         }
 720 }
 721
 722 static void r5l_reclaim_thread(struct md_thread *thread)
 723 {
 724         struct mddev *mddev = thread->mddev;
 725         struct r5conf *conf = mddev->private;
 726         struct r5l_log *log = conf->log;
 727
 728         if (!log)
 729                 return;
 730         r5l_do_reclaim(log);
 731 }
 732
 733 static void r5l_wake_reclaim(struct r5l_log *log, sector_t space)
 734 {
 735         unsigned long target;
 736         unsigned long new = (unsigned long)space; /* overflow in theory */
 737
 738         do {
 739                 target = log->reclaim_target;
 740                 if (new < target)
 741                         return;
 742         } while (cmpxchg(&log->reclaim_target, target, new) != target);
 743         md_wakeup_thread(log->reclaim_thread);
 744 }
 745
 746 struct r5l_recovery_ctx {
 747         struct page *meta_page;         /* current meta */
 748         sector_t meta_total_blocks;     /* total size of current meta and data */
 749         sector_t pos;                   /* recovery position */
 750         u64 seq;                        /* recovery position seq */
 751 };
 752
 753 static int r5l_read_meta_block(struct r5l_log *log,
 754                                struct r5l_recovery_ctx *ctx)
 755 {
 756         struct page *page = ctx->meta_page;
 757         struct r5l_meta_block *mb;
 758         u32 crc, stored_crc;
 759
 760         if (!sync_page_io(log->rdev, ctx->pos, PAGE_SIZE, page, READ, false))
 761                 return -EIO;
 762
 763         mb = page_address(page);
 764         stored_crc = le32_to_cpu(mb->checksum);
 765         mb->checksum = 0;
 766
 767         if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
 768             le64_to_cpu(mb->seq) != ctx->seq ||
 769             mb->version != R5LOG_VERSION ||
 770             le64_to_cpu(mb->position) != ctx->pos)
 771                 return -EINVAL;
 772
 773         crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
 774         if (stored_crc != crc)
 775                 return -EINVAL;
 776
 777         if (le32_to_cpu(mb->meta_size) > PAGE_SIZE)
 778                 return -EINVAL;
 779
 780         ctx->meta_total_blocks = BLOCK_SECTORS;
 781
 782         return 0;
 783 }
 784
 785 static int r5l_recovery_flush_one_stripe(struct r5l_log *log,
 786                                          struct r5l_recovery_ctx *ctx,
 787                                          sector_t stripe_sect,
 788                                          int *offset, sector_t *log_offset)
 789 {
 790         struct r5conf *conf = log->rdev->mddev->private;
 791         struct stripe_head *sh;
 792         struct r5l_payload_data_parity *payload;
 793         int disk_index;
 794
 795         sh = raid5_get_active_stripe(conf, stripe_sect, 0, 0, 0);
 796         while (1) {
 797                 payload = page_address(ctx->meta_page) + *offset;
 798
 799                 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) {
 800                         raid5_compute_sector(conf,
 801                                              le64_to_cpu(payload->location), 0,
 802                                              &disk_index, sh);
 803
 804                         sync_page_io(log->rdev, *log_offset, PAGE_SIZE,
 805                                      sh->dev[disk_index].page, READ, false);
 806                         sh->dev[disk_index].log_checksum =
 807                                 le32_to_cpu(payload->checksum[0]);
 808                         set_bit(R5_Wantwrite, &sh->dev[disk_index].flags);
 809                         ctx->meta_total_blocks += BLOCK_SECTORS;
 810                 } else {
 811                         disk_index = sh->pd_idx;
 812                         sync_page_io(log->rdev, *log_offset, PAGE_SIZE,
 813                                      sh->dev[disk_index].page, READ, false);
 814                         sh->dev[disk_index].log_checksum =
 815                                 le32_to_cpu(payload->checksum[0]);
 816                         set_bit(R5_Wantwrite, &sh->dev[disk_index].flags);
 817
 818                         if (sh->qd_idx >= 0) {
 819                                 disk_index = sh->qd_idx;
 820                                 sync_page_io(log->rdev,
 821                                              r5l_ring_add(log, *log_offset, BLOCK_SECTORS),
 822                                              PAGE_SIZE, sh->dev[disk_index].page,
 823                                              READ, false);
 824                                 sh->dev[disk_index].log_checksum =
 825                                         le32_to_cpu(payload->checksum[1]);
 826                                 set_bit(R5_Wantwrite,
 827                                         &sh->dev[disk_index].flags);
 828                         }
 829                         ctx->meta_total_blocks += BLOCK_SECTORS * conf->max_degraded;
 830                 }
 831
 832                 *log_offset = r5l_ring_add(log, *log_offset,
 833                                            le32_to_cpu(payload->size));
 834                 *offset += sizeof(struct r5l_payload_data_parity) +
 835                         sizeof(__le32) *
 836                         (le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
 837                 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_PARITY)
 838                         break;
 839         }
 840
 841         for (disk_index = 0; disk_index < sh->disks; disk_index++) {
 842                 void *addr;
 843                 u32 checksum;
 844
 845                 if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
 846                         continue;
 847                 addr = kmap_atomic(sh->dev[disk_index].page);
 848                 checksum = crc32c_le(log->uuid_checksum, addr, PAGE_SIZE);
 849                 kunmap_atomic(addr);
 850                 if (checksum != sh->dev[disk_index].log_checksum)
 851                         goto error;
 852         }
 853
 854         for (disk_index = 0; disk_index < sh->disks; disk_index++) {
 855                 struct md_rdev *rdev, *rrdev;
 856
 857                 if (!test_and_clear_bit(R5_Wantwrite,
 858                                         &sh->dev[disk_index].flags))
 859                         continue;
 860
 861                 /* in case device is broken */
 862                 rdev = rcu_dereference(conf->disks[disk_index].rdev);
 863                 if (rdev)
 864                         sync_page_io(rdev, stripe_sect, PAGE_SIZE,
 865                                      sh->dev[disk_index].page, WRITE, false);
 866                 rrdev = rcu_dereference(conf->disks[disk_index].replacement);
 867                 if (rrdev)
 868                         sync_page_io(rrdev, stripe_sect, PAGE_SIZE,
 869                                      sh->dev[disk_index].page, WRITE, false);
 870         }
 871         raid5_release_stripe(sh);
 872         return 0;
 873
 874 error:
 875         for (disk_index = 0; disk_index < sh->disks; disk_index++)
 876                 sh->dev[disk_index].flags = 0;
 877         raid5_release_stripe(sh);
 878         return -EINVAL;
 879 }
 880
 881 static int r5l_recovery_flush_one_meta(struct r5l_log *log,
 882                                        struct r5l_recovery_ctx *ctx)
 883 {
 884         struct r5conf *conf = log->rdev->mddev->private;
 885         struct r5l_payload_data_parity *payload;
 886         struct r5l_meta_block *mb;
 887         int offset;
 888         sector_t log_offset;
 889         sector_t stripe_sector;
 890
 891         mb = page_address(ctx->meta_page);
 892         offset = sizeof(struct r5l_meta_block);
 893         log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
 894
 895         while (offset < le32_to_cpu(mb->meta_size)) {
 896                 int dd;
 897
 898                 payload = (void *)mb + offset;
 899                 stripe_sector = raid5_compute_sector(conf,
 900                                                      le64_to_cpu(payload->location), 0, &dd, NULL);
 901                 if (r5l_recovery_flush_one_stripe(log, ctx, stripe_sector,
 902                                                   &offset, &log_offset))
 903                         return -EINVAL;
 904         }
 905         return 0;
 906 }
 907
 908 /* copy data/parity from log to raid disks */
 909 static void r5l_recovery_flush_log(struct r5l_log *log,
 910                                    struct r5l_recovery_ctx *ctx)
 911 {
 912         while (1) {
 913                 if (r5l_read_meta_block(log, ctx))
 914                         return;
 915                 if (r5l_recovery_flush_one_meta(log, ctx))
 916                         return;
 917                 ctx->seq++;
 918                 ctx->pos = r5l_ring_add(log, ctx->pos, ctx->meta_total_blocks);
 919         }
 920 }
 921
 922 static int r5l_log_write_empty_meta_block(struct r5l_log *log, sector_t pos,
 923                                           u64 seq)
 924 {
 925         struct page *page;
 926         struct r5l_meta_block *mb;
 927         u32 crc;
 928
 929         page = alloc_page(GFP_KERNEL | __GFP_ZERO);
 930         if (!page)
 931                 return -ENOMEM;
 932         mb = page_address(page);
 933         mb->magic = cpu_to_le32(R5LOG_MAGIC);
 934         mb->version = R5LOG_VERSION;
 935         mb->meta_size = cpu_to_le32(sizeof(struct r5l_meta_block));
 936         mb->seq = cpu_to_le64(seq);
 937         mb->position = cpu_to_le64(pos);
 938         crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
 939         mb->checksum = cpu_to_le32(crc);
 940
 941         if (!sync_page_io(log->rdev, pos, PAGE_SIZE, page, WRITE_FUA, false)) {
 942                 __free_page(page);
 943                 return -EIO;
 944         }
 945         __free_page(page);
 946         return 0;
 947 }
 948
 949 static int r5l_recovery_log(struct r5l_log *log)
 950 {
 951         struct r5l_recovery_ctx ctx;
 952
 953         ctx.pos = log->last_checkpoint;
 954         ctx.seq = log->last_cp_seq;
 955         ctx.meta_page = alloc_page(GFP_KERNEL);
 956         if (!ctx.meta_page)
 957                 return -ENOMEM;
 958
 959         r5l_recovery_flush_log(log, &ctx);
 960         __free_page(ctx.meta_page);
 961
 962         /*
 963          * we did a recovery. Now ctx.pos points to an invalid meta block. New
 964          * log will start here. but we can't let superblock point to last valid
 965          * meta block. The log might looks like:
 966          * | meta 1| meta 2| meta 3|
 967          * meta 1 is valid, meta 2 is invalid. meta 3 could be valid. If
 968          * superblock points to meta 1, we write a new valid meta 2n.  if crash
 969          * happens again, new recovery will start from meta 1. Since meta 2n is
 970          * valid now, recovery will think meta 3 is valid, which is wrong.
 971          * The solution is we create a new meta in meta2 with its seq == meta
 972          * 1's seq + 10 and let superblock points to meta2. The same recovery will
 973          * not think meta 3 is a valid meta, because its seq doesn't match
 974          */
 975         if (ctx.seq > log->last_cp_seq + 1) {
 976                 int ret;
 977
 978                 ret = r5l_log_write_empty_meta_block(log, ctx.pos, ctx.seq + 10);
 979                 if (ret)
 980                         return ret;
 981                 log->seq = ctx.seq + 11;
 982                 log->log_start = r5l_ring_add(log, ctx.pos, BLOCK_SECTORS);
 983                 r5l_write_super(log, ctx.pos);
 984         } else {
 985                 log->log_start = ctx.pos;
 986                 log->seq = ctx.seq;
 987         }
 988         return 0;
 989 }
 990
 991 static void r5l_write_super(struct r5l_log *log, sector_t cp)
 992 {
 993         struct mddev *mddev = log->rdev->mddev;
 994
 995         log->rdev->journal_tail = cp;
 996         set_bit(MD_CHANGE_DEVS, &mddev->flags);
 997 }
 998
 999 static int r5l_load_log(struct r5l_log *log)
1000 {
1001         struct md_rdev *rdev = log->rdev;
1002         struct page *page;
1003         struct r5l_meta_block *mb;
1004         sector_t cp = log->rdev->journal_tail;
1005         u32 stored_crc, expected_crc;
1006         bool create_super = false;
1007         int ret;
1008
1009         /* Make sure it's valid */
1010         if (cp >= rdev->sectors || round_down(cp, BLOCK_SECTORS) != cp)
1011                 cp = 0;
1012         page = alloc_page(GFP_KERNEL);
1013         if (!page)
1014                 return -ENOMEM;
1015
1016         if (!sync_page_io(rdev, cp, PAGE_SIZE, page, READ, false)) {
1017                 ret = -EIO;
1018                 goto ioerr;
1019         }
1020         mb = page_address(page);
1021
1022         if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
1023             mb->version != R5LOG_VERSION) {
1024                 create_super = true;
1025                 goto create;
1026         }
1027         stored_crc = le32_to_cpu(mb->checksum);
1028         mb->checksum = 0;
1029         expected_crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
1030         if (stored_crc != expected_crc) {
1031                 create_super = true;
1032                 goto create;
1033         }
1034         if (le64_to_cpu(mb->position) != cp) {
1035                 create_super = true;
1036                 goto create;
1037         }
1038 create:
1039         if (create_super) {
1040                 log->last_cp_seq = prandom_u32();
1041                 cp = 0;
1042                 /*
1043                  * Make sure super points to correct address. Log might have
1044                  * data very soon. If super hasn't correct log tail address,
1045                  * recovery can't find the log
1046                  */
1047                 r5l_write_super(log, cp);
1048         } else
1049                 log->last_cp_seq = le64_to_cpu(mb->seq);
1050
1051         log->device_size = round_down(rdev->sectors, BLOCK_SECTORS);
1052         log->max_free_space = log->device_size >> RECLAIM_MAX_FREE_SPACE_SHIFT;
1053         if (log->max_free_space > RECLAIM_MAX_FREE_SPACE)
1054                 log->max_free_space = RECLAIM_MAX_FREE_SPACE;
1055         log->last_checkpoint = cp;
1056
1057         __free_page(page);
1058
1059         return r5l_recovery_log(log);
1060 ioerr:
1061         __free_page(page);
1062         return ret;
1063 }
1064
1065 int r5l_init_log(struct r5conf *conf, struct md_rdev *rdev)
1066 {
1067         struct r5l_log *log;
1068
1069         if (PAGE_SIZE != 4096)
1070                 return -EINVAL;
1071         log = kzalloc(sizeof(*log), GFP_KERNEL);
1072         if (!log)
1073                 return -ENOMEM;
1074         log->rdev = rdev;
1075
1076         log->uuid_checksum = crc32c_le(~0, rdev->mddev->uuid,
1077                                        sizeof(rdev->mddev->uuid));
1078
1079         mutex_init(&log->io_mutex);
1080
1081         spin_lock_init(&log->io_list_lock);
1082         INIT_LIST_HEAD(&log->running_ios);
1083         INIT_LIST_HEAD(&log->io_end_ios);
1084         INIT_LIST_HEAD(&log->stripe_end_ios);
1085         INIT_LIST_HEAD(&log->flushing_ios);
1086         INIT_LIST_HEAD(&log->flushed_ios);
1087         bio_init(&log->flush_bio);
1088
1089         log->io_kc = KMEM_CACHE(r5l_io_unit, 0);
1090         if (!log->io_kc)
1091                 goto io_kc;
1092
1093         log->reclaim_thread = md_register_thread(r5l_reclaim_thread,
1094                                                  log->rdev->mddev, "reclaim");
1095         if (!log->reclaim_thread)
1096                 goto reclaim_thread;
1097         init_waitqueue_head(&log->iounit_wait);
1098
1099         INIT_LIST_HEAD(&log->no_space_stripes);
1100         spin_lock_init(&log->no_space_stripes_lock);
1101
1102         if (r5l_load_log(log))
1103                 goto error;
1104
1105         conf->log = log;
1106         return 0;
1107 error:
1108         md_unregister_thread(&log->reclaim_thread);
1109 reclaim_thread:
1110         kmem_cache_destroy(log->io_kc);
1111 io_kc:
1112         kfree(log);
1113         return -EINVAL;
1114 }
1115
1116 void r5l_exit_log(struct r5l_log *log)
1117 {
1118         /*
1119          * at this point all stripes are finished, so io_unit is at least in
1120          * STRIPE_END state
1121          */
1122         r5l_wake_reclaim(log, -1L);
1123         md_unregister_thread(&log->reclaim_thread);
1124         r5l_do_reclaim(log);
1125         /*
1126          * force a super update, r5l_do_reclaim might updated the super.
1127          * mddev->thread is already stopped
1128          */
1129         md_update_sb(log->rdev->mddev, 1);
1130
1131         kmem_cache_destroy(log->io_kc);
1132         kfree(log);
1133 }