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         /*
 483          * The stripe must enter state machine again to finish the write, so
 484          * don't delay.
 485          */
 486         clear_bit(STRIPE_DELAYED, &sh->state);
 487         atomic_inc(&sh->count);
 488
 489         mutex_lock(&log->io_mutex);
 490         /* meta + data */
 491         reserve = (1 + write_disks) << (PAGE_SHIFT - 9);
 492         if (r5l_has_free_space(log, reserve))
 493                 r5l_log_stripe(log, sh, data_pages, parity_pages);
 494         else {
 495                 spin_lock(&log->no_space_stripes_lock);
 496                 list_add_tail(&sh->log_list, &log->no_space_stripes);
 497                 spin_unlock(&log->no_space_stripes_lock);
 498
 499                 r5l_wake_reclaim(log, reserve);
 500         }
 501         mutex_unlock(&log->io_mutex);
 502
 503         return 0;
 504 }
 505
 506 void r5l_write_stripe_run(struct r5l_log *log)
 507 {
 508         if (!log)
 509                 return;
 510         mutex_lock(&log->io_mutex);
 511         r5l_submit_current_io(log);
 512         mutex_unlock(&log->io_mutex);
 513 }
 514
 515 int r5l_handle_flush_request(struct r5l_log *log, struct bio *bio)
 516 {
 517         if (!log)
 518                 return -ENODEV;
 519         /*
 520          * we flush log disk cache first, then write stripe data to raid disks.
 521          * So if bio is finished, the log disk cache is flushed already. The
 522          * recovery guarantees we can recovery the bio from log disk, so we
 523          * don't need to flush again
 524          */
 525         if (bio->bi_iter.bi_size == 0) {
 526                 bio_endio(bio);
 527                 return 0;
 528         }
 529         bio->bi_rw &= ~REQ_FLUSH;
 530         return -EAGAIN;
 531 }
 532
 533 /* This will run after log space is reclaimed */
 534 static void r5l_run_no_space_stripes(struct r5l_log *log)
 535 {
 536         struct stripe_head *sh;
 537
 538         spin_lock(&log->no_space_stripes_lock);
 539         while (!list_empty(&log->no_space_stripes)) {
 540                 sh = list_first_entry(&log->no_space_stripes,
 541                                       struct stripe_head, log_list);
 542                 list_del_init(&sh->log_list);
 543                 set_bit(STRIPE_HANDLE, &sh->state);
 544                 raid5_release_stripe(sh);
 545         }
 546         spin_unlock(&log->no_space_stripes_lock);
 547 }
 548
 549 static void __r5l_stripe_write_finished(struct r5l_io_unit *io)
 550 {
 551         struct r5l_log *log = io->log;
 552         struct r5l_io_unit *last;
 553         sector_t reclaimable_space;
 554         unsigned long flags;
 555
 556         spin_lock_irqsave(&log->io_list_lock, flags);
 557         __r5l_set_io_unit_state(io, IO_UNIT_STRIPE_END);
 558         /* might move 0 entry */
 559         r5l_move_io_unit_list(&log->flushed_ios, &log->stripe_end_ios,
 560                               IO_UNIT_STRIPE_END);
 561         if (list_empty(&log->stripe_end_ios)) {
 562                 spin_unlock_irqrestore(&log->io_list_lock, flags);
 563                 return;
 564         }
 565
 566         last = list_last_entry(&log->stripe_end_ios,
 567                                struct r5l_io_unit, log_sibling);
 568         reclaimable_space = r5l_ring_distance(log, log->last_checkpoint,
 569                                               last->log_end);
 570         if (reclaimable_space >= log->max_free_space)
 571                 r5l_wake_reclaim(log, 0);
 572
 573         r5l_compress_stripe_end_list(log);
 574         spin_unlock_irqrestore(&log->io_list_lock, flags);
 575         wake_up(&log->iounit_wait);
 576 }
 577
 578 void r5l_stripe_write_finished(struct stripe_head *sh)
 579 {
 580         struct r5l_io_unit *io;
 581
 582         io = sh->log_io;
 583         sh->log_io = NULL;
 584
 585         if (io && atomic_dec_and_test(&io->pending_stripe))
 586                 __r5l_stripe_write_finished(io);
 587 }
 588
 589 static void r5l_log_flush_endio(struct bio *bio)
 590 {
 591         struct r5l_log *log = container_of(bio, struct r5l_log,
 592                 flush_bio);
 593         unsigned long flags;
 594         struct r5l_io_unit *io;
 595         struct stripe_head *sh;
 596
 597         spin_lock_irqsave(&log->io_list_lock, flags);
 598         list_for_each_entry(io, &log->flushing_ios, log_sibling) {
 599                 while (!list_empty(&io->stripe_list)) {
 600                         sh = list_first_entry(&io->stripe_list,
 601                                 struct stripe_head, log_list);
 602                         list_del_init(&sh->log_list);
 603                         set_bit(STRIPE_HANDLE, &sh->state);
 604                         raid5_release_stripe(sh);
 605                 }
 606         }
 607         list_splice_tail_init(&log->flushing_ios, &log->flushed_ios);
 608         spin_unlock_irqrestore(&log->io_list_lock, flags);
 609 }
 610
 611 /*
 612  * Starting dispatch IO to raid.
 613  * io_unit(meta) consists of a log. There is one situation we want to avoid. A
 614  * broken meta in the middle of a log causes recovery can't find meta at the
 615  * head of log. If operations require meta at the head persistent in log, we
 616  * must make sure meta before it persistent in log too. A case is:
 617  *
 618  * stripe data/parity is in log, we start write stripe to raid disks. stripe
 619  * data/parity must be persistent in log before we do the write to raid disks.
 620  *
 621  * The solution is we restrictly maintain io_unit list order. In this case, we
 622  * only write stripes of an io_unit to raid disks till the io_unit is the first
 623  * one whose data/parity is in log.
 624  */
 625 void r5l_flush_stripe_to_raid(struct r5l_log *log)
 626 {
 627         bool do_flush;
 628         if (!log)
 629                 return;
 630
 631         spin_lock_irq(&log->io_list_lock);
 632         /* flush bio is running */
 633         if (!list_empty(&log->flushing_ios)) {
 634                 spin_unlock_irq(&log->io_list_lock);
 635                 return;
 636         }
 637         list_splice_tail_init(&log->io_end_ios, &log->flushing_ios);
 638         do_flush = !list_empty(&log->flushing_ios);
 639         spin_unlock_irq(&log->io_list_lock);
 640
 641         if (!do_flush)
 642                 return;
 643         bio_reset(&log->flush_bio);
 644         log->flush_bio.bi_bdev = log->rdev->bdev;
 645         log->flush_bio.bi_end_io = r5l_log_flush_endio;
 646         submit_bio(WRITE_FLUSH, &log->flush_bio);
 647 }
 648
 649 static void r5l_kick_io_unit(struct r5l_log *log)
 650 {
 651         md_wakeup_thread(log->rdev->mddev->thread);
 652         wait_event_lock_irq(log->iounit_wait, !list_empty(&log->stripe_end_ios),
 653                             log->io_list_lock);
 654 }
 655
 656 static void r5l_write_super(struct r5l_log *log, sector_t cp);
 657 static void r5l_do_reclaim(struct r5l_log *log)
 658 {
 659         struct r5l_io_unit *io, *last;
 660         LIST_HEAD(list);
 661         sector_t free = 0;
 662         sector_t reclaim_target = xchg(&log->reclaim_target, 0);
 663
 664         spin_lock_irq(&log->io_list_lock);
 665         /*
 666          * move proper io_unit to reclaim list. We should not change the order.
 667          * reclaimable/unreclaimable io_unit can be mixed in the list, we
 668          * shouldn't reuse space of an unreclaimable io_unit
 669          */
 670         while (1) {
 671                 struct list_head *target_list = NULL;
 672
 673                 while (!list_empty(&log->stripe_end_ios)) {
 674                         io = list_first_entry(&log->stripe_end_ios,
 675                                               struct r5l_io_unit, log_sibling);
 676                         list_move_tail(&io->log_sibling, &list);
 677                         free += r5l_ring_distance(log, io->log_start,
 678                                                   io->log_end);
 679                 }
 680
 681                 if (free >= reclaim_target ||
 682                     (list_empty(&log->running_ios) &&
 683                      list_empty(&log->io_end_ios) &&
 684                      list_empty(&log->flushing_ios) &&
 685                      list_empty(&log->flushed_ios)))
 686                         break;
 687
 688                 /* Below waiting mostly happens when we shutdown the raid */
 689                 if (!list_empty(&log->flushed_ios))
 690                         target_list = &log->flushed_ios;
 691                 else if (!list_empty(&log->flushing_ios))
 692                         target_list = &log->flushing_ios;
 693                 else if (!list_empty(&log->io_end_ios))
 694                         target_list = &log->io_end_ios;
 695                 else if (!list_empty(&log->running_ios))
 696                         target_list = &log->running_ios;
 697
 698                 r5l_kick_io_unit(log);
 699         }
 700         spin_unlock_irq(&log->io_list_lock);
 701
 702         if (list_empty(&list))
 703                 return;
 704
 705         /* super always point to last valid meta */
 706         last = list_last_entry(&list, struct r5l_io_unit, log_sibling);
 707         /*
 708          * write_super will flush cache of each raid disk. We must write super
 709          * here, because the log area might be reused soon and we don't want to
 710          * confuse recovery
 711          */
 712         r5l_write_super(log, last->log_start);
 713
 714         mutex_lock(&log->io_mutex);
 715         log->last_checkpoint = last->log_start;
 716         log->last_cp_seq = last->seq;
 717         mutex_unlock(&log->io_mutex);
 718         r5l_run_no_space_stripes(log);
 719
 720         while (!list_empty(&list)) {
 721                 io = list_first_entry(&list, struct r5l_io_unit, log_sibling);
 722                 list_del(&io->log_sibling);
 723                 r5l_free_io_unit(log, io);
 724         }
 725 }
 726
 727 static void r5l_reclaim_thread(struct md_thread *thread)
 728 {
 729         struct mddev *mddev = thread->mddev;
 730         struct r5conf *conf = mddev->private;
 731         struct r5l_log *log = conf->log;
 732
 733         if (!log)
 734                 return;
 735         r5l_do_reclaim(log);
 736 }
 737
 738 static void r5l_wake_reclaim(struct r5l_log *log, sector_t space)
 739 {
 740         unsigned long target;
 741         unsigned long new = (unsigned long)space; /* overflow in theory */
 742
 743         do {
 744                 target = log->reclaim_target;
 745                 if (new < target)
 746                         return;
 747         } while (cmpxchg(&log->reclaim_target, target, new) != target);
 748         md_wakeup_thread(log->reclaim_thread);
 749 }
 750
 751 struct r5l_recovery_ctx {
 752         struct page *meta_page;         /* current meta */
 753         sector_t meta_total_blocks;     /* total size of current meta and data */
 754         sector_t pos;                   /* recovery position */
 755         u64 seq;                        /* recovery position seq */
 756 };
 757
 758 static int r5l_read_meta_block(struct r5l_log *log,
 759                                struct r5l_recovery_ctx *ctx)
 760 {
 761         struct page *page = ctx->meta_page;
 762         struct r5l_meta_block *mb;
 763         u32 crc, stored_crc;
 764
 765         if (!sync_page_io(log->rdev, ctx->pos, PAGE_SIZE, page, READ, false))
 766                 return -EIO;
 767
 768         mb = page_address(page);
 769         stored_crc = le32_to_cpu(mb->checksum);
 770         mb->checksum = 0;
 771
 772         if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
 773             le64_to_cpu(mb->seq) != ctx->seq ||
 774             mb->version != R5LOG_VERSION ||
 775             le64_to_cpu(mb->position) != ctx->pos)
 776                 return -EINVAL;
 777
 778         crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
 779         if (stored_crc != crc)
 780                 return -EINVAL;
 781
 782         if (le32_to_cpu(mb->meta_size) > PAGE_SIZE)
 783                 return -EINVAL;
 784
 785         ctx->meta_total_blocks = BLOCK_SECTORS;
 786
 787         return 0;
 788 }
 789
 790 static int r5l_recovery_flush_one_stripe(struct r5l_log *log,
 791                                          struct r5l_recovery_ctx *ctx,
 792                                          sector_t stripe_sect,
 793                                          int *offset, sector_t *log_offset)
 794 {
 795         struct r5conf *conf = log->rdev->mddev->private;
 796         struct stripe_head *sh;
 797         struct r5l_payload_data_parity *payload;
 798         int disk_index;
 799
 800         sh = raid5_get_active_stripe(conf, stripe_sect, 0, 0, 0);
 801         while (1) {
 802                 payload = page_address(ctx->meta_page) + *offset;
 803
 804                 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) {
 805                         raid5_compute_sector(conf,
 806                                              le64_to_cpu(payload->location), 0,
 807                                              &disk_index, sh);
 808
 809                         sync_page_io(log->rdev, *log_offset, PAGE_SIZE,
 810                                      sh->dev[disk_index].page, READ, false);
 811                         sh->dev[disk_index].log_checksum =
 812                                 le32_to_cpu(payload->checksum[0]);
 813                         set_bit(R5_Wantwrite, &sh->dev[disk_index].flags);
 814                         ctx->meta_total_blocks += BLOCK_SECTORS;
 815                 } else {
 816                         disk_index = sh->pd_idx;
 817                         sync_page_io(log->rdev, *log_offset, PAGE_SIZE,
 818                                      sh->dev[disk_index].page, READ, false);
 819                         sh->dev[disk_index].log_checksum =
 820                                 le32_to_cpu(payload->checksum[0]);
 821                         set_bit(R5_Wantwrite, &sh->dev[disk_index].flags);
 822
 823                         if (sh->qd_idx >= 0) {
 824                                 disk_index = sh->qd_idx;
 825                                 sync_page_io(log->rdev,
 826                                              r5l_ring_add(log, *log_offset, BLOCK_SECTORS),
 827                                              PAGE_SIZE, sh->dev[disk_index].page,
 828                                              READ, false);
 829                                 sh->dev[disk_index].log_checksum =
 830                                         le32_to_cpu(payload->checksum[1]);
 831                                 set_bit(R5_Wantwrite,
 832                                         &sh->dev[disk_index].flags);
 833                         }
 834                         ctx->meta_total_blocks += BLOCK_SECTORS * conf->max_degraded;
 835                 }
 836
 837                 *log_offset = r5l_ring_add(log, *log_offset,
 838                                            le32_to_cpu(payload->size));
 839                 *offset += sizeof(struct r5l_payload_data_parity) +
 840                         sizeof(__le32) *
 841                         (le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
 842                 if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_PARITY)
 843                         break;
 844         }
 845
 846         for (disk_index = 0; disk_index < sh->disks; disk_index++) {
 847                 void *addr;
 848                 u32 checksum;
 849
 850                 if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
 851                         continue;
 852                 addr = kmap_atomic(sh->dev[disk_index].page);
 853                 checksum = crc32c_le(log->uuid_checksum, addr, PAGE_SIZE);
 854                 kunmap_atomic(addr);
 855                 if (checksum != sh->dev[disk_index].log_checksum)
 856                         goto error;
 857         }
 858
 859         for (disk_index = 0; disk_index < sh->disks; disk_index++) {
 860                 struct md_rdev *rdev, *rrdev;
 861
 862                 if (!test_and_clear_bit(R5_Wantwrite,
 863                                         &sh->dev[disk_index].flags))
 864                         continue;
 865
 866                 /* in case device is broken */
 867                 rdev = rcu_dereference(conf->disks[disk_index].rdev);
 868                 if (rdev)
 869                         sync_page_io(rdev, stripe_sect, PAGE_SIZE,
 870                                      sh->dev[disk_index].page, WRITE, false);
 871                 rrdev = rcu_dereference(conf->disks[disk_index].replacement);
 872                 if (rrdev)
 873                         sync_page_io(rrdev, stripe_sect, PAGE_SIZE,
 874                                      sh->dev[disk_index].page, WRITE, false);
 875         }
 876         raid5_release_stripe(sh);
 877         return 0;
 878
 879 error:
 880         for (disk_index = 0; disk_index < sh->disks; disk_index++)
 881                 sh->dev[disk_index].flags = 0;
 882         raid5_release_stripe(sh);
 883         return -EINVAL;
 884 }
 885
 886 static int r5l_recovery_flush_one_meta(struct r5l_log *log,
 887                                        struct r5l_recovery_ctx *ctx)
 888 {
 889         struct r5conf *conf = log->rdev->mddev->private;
 890         struct r5l_payload_data_parity *payload;
 891         struct r5l_meta_block *mb;
 892         int offset;
 893         sector_t log_offset;
 894         sector_t stripe_sector;
 895
 896         mb = page_address(ctx->meta_page);
 897         offset = sizeof(struct r5l_meta_block);
 898         log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
 899
 900         while (offset < le32_to_cpu(mb->meta_size)) {
 901                 int dd;
 902
 903                 payload = (void *)mb + offset;
 904                 stripe_sector = raid5_compute_sector(conf,
 905                                                      le64_to_cpu(payload->location), 0, &dd, NULL);
 906                 if (r5l_recovery_flush_one_stripe(log, ctx, stripe_sector,
 907                                                   &offset, &log_offset))
 908                         return -EINVAL;
 909         }
 910         return 0;
 911 }
 912
 913 /* copy data/parity from log to raid disks */
 914 static void r5l_recovery_flush_log(struct r5l_log *log,
 915                                    struct r5l_recovery_ctx *ctx)
 916 {
 917         while (1) {
 918                 if (r5l_read_meta_block(log, ctx))
 919                         return;
 920                 if (r5l_recovery_flush_one_meta(log, ctx))
 921                         return;
 922                 ctx->seq++;
 923                 ctx->pos = r5l_ring_add(log, ctx->pos, ctx->meta_total_blocks);
 924         }
 925 }
 926
 927 static int r5l_log_write_empty_meta_block(struct r5l_log *log, sector_t pos,
 928                                           u64 seq)
 929 {
 930         struct page *page;
 931         struct r5l_meta_block *mb;
 932         u32 crc;
 933
 934         page = alloc_page(GFP_KERNEL | __GFP_ZERO);
 935         if (!page)
 936                 return -ENOMEM;
 937         mb = page_address(page);
 938         mb->magic = cpu_to_le32(R5LOG_MAGIC);
 939         mb->version = R5LOG_VERSION;
 940         mb->meta_size = cpu_to_le32(sizeof(struct r5l_meta_block));
 941         mb->seq = cpu_to_le64(seq);
 942         mb->position = cpu_to_le64(pos);
 943         crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
 944         mb->checksum = cpu_to_le32(crc);
 945
 946         if (!sync_page_io(log->rdev, pos, PAGE_SIZE, page, WRITE_FUA, false)) {
 947                 __free_page(page);
 948                 return -EIO;
 949         }
 950         __free_page(page);
 951         return 0;
 952 }
 953
 954 static int r5l_recovery_log(struct r5l_log *log)
 955 {
 956         struct r5l_recovery_ctx ctx;
 957
 958         ctx.pos = log->last_checkpoint;
 959         ctx.seq = log->last_cp_seq;
 960         ctx.meta_page = alloc_page(GFP_KERNEL);
 961         if (!ctx.meta_page)
 962                 return -ENOMEM;
 963
 964         r5l_recovery_flush_log(log, &ctx);
 965         __free_page(ctx.meta_page);
 966
 967         /*
 968          * we did a recovery. Now ctx.pos points to an invalid meta block. New
 969          * log will start here. but we can't let superblock point to last valid
 970          * meta block. The log might looks like:
 971          * | meta 1| meta 2| meta 3|
 972          * meta 1 is valid, meta 2 is invalid. meta 3 could be valid. If
 973          * superblock points to meta 1, we write a new valid meta 2n.  if crash
 974          * happens again, new recovery will start from meta 1. Since meta 2n is
 975          * valid now, recovery will think meta 3 is valid, which is wrong.
 976          * The solution is we create a new meta in meta2 with its seq == meta
 977          * 1's seq + 10 and let superblock points to meta2. The same recovery will
 978          * not think meta 3 is a valid meta, because its seq doesn't match
 979          */
 980         if (ctx.seq > log->last_cp_seq + 1) {
 981                 int ret;
 982
 983                 ret = r5l_log_write_empty_meta_block(log, ctx.pos, ctx.seq + 10);
 984                 if (ret)
 985                         return ret;
 986                 log->seq = ctx.seq + 11;
 987                 log->log_start = r5l_ring_add(log, ctx.pos, BLOCK_SECTORS);
 988                 r5l_write_super(log, ctx.pos);
 989         } else {
 990                 log->log_start = ctx.pos;
 991                 log->seq = ctx.seq;
 992         }
 993         return 0;
 994 }
 995
 996 static void r5l_write_super(struct r5l_log *log, sector_t cp)
 997 {
 998         struct mddev *mddev = log->rdev->mddev;
 999
1000         log->rdev->journal_tail = cp;
1001         set_bit(MD_CHANGE_DEVS, &mddev->flags);
1002 }
1003
1004 static int r5l_load_log(struct r5l_log *log)
1005 {
1006         struct md_rdev *rdev = log->rdev;
1007         struct page *page;
1008         struct r5l_meta_block *mb;
1009         sector_t cp = log->rdev->journal_tail;
1010         u32 stored_crc, expected_crc;
1011         bool create_super = false;
1012         int ret;
1013
1014         /* Make sure it's valid */
1015         if (cp >= rdev->sectors || round_down(cp, BLOCK_SECTORS) != cp)
1016                 cp = 0;
1017         page = alloc_page(GFP_KERNEL);
1018         if (!page)
1019                 return -ENOMEM;
1020
1021         if (!sync_page_io(rdev, cp, PAGE_SIZE, page, READ, false)) {
1022                 ret = -EIO;
1023                 goto ioerr;
1024         }
1025         mb = page_address(page);
1026
1027         if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
1028             mb->version != R5LOG_VERSION) {
1029                 create_super = true;
1030                 goto create;
1031         }
1032         stored_crc = le32_to_cpu(mb->checksum);
1033         mb->checksum = 0;
1034         expected_crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
1035         if (stored_crc != expected_crc) {
1036                 create_super = true;
1037                 goto create;
1038         }
1039         if (le64_to_cpu(mb->position) != cp) {
1040                 create_super = true;
1041                 goto create;
1042         }
1043 create:
1044         if (create_super) {
1045                 log->last_cp_seq = prandom_u32();
1046                 cp = 0;
1047                 /*
1048                  * Make sure super points to correct address. Log might have
1049                  * data very soon. If super hasn't correct log tail address,
1050                  * recovery can't find the log
1051                  */
1052                 r5l_write_super(log, cp);
1053         } else
1054                 log->last_cp_seq = le64_to_cpu(mb->seq);
1055
1056         log->device_size = round_down(rdev->sectors, BLOCK_SECTORS);
1057         log->max_free_space = log->device_size >> RECLAIM_MAX_FREE_SPACE_SHIFT;
1058         if (log->max_free_space > RECLAIM_MAX_FREE_SPACE)
1059                 log->max_free_space = RECLAIM_MAX_FREE_SPACE;
1060         log->last_checkpoint = cp;
1061
1062         __free_page(page);
1063
1064         return r5l_recovery_log(log);
1065 ioerr:
1066         __free_page(page);
1067         return ret;
1068 }
1069
1070 int r5l_init_log(struct r5conf *conf, struct md_rdev *rdev)
1071 {
1072         struct r5l_log *log;
1073
1074         if (PAGE_SIZE != 4096)
1075                 return -EINVAL;
1076         log = kzalloc(sizeof(*log), GFP_KERNEL);
1077         if (!log)
1078                 return -ENOMEM;
1079         log->rdev = rdev;
1080
1081         log->uuid_checksum = crc32c_le(~0, rdev->mddev->uuid,
1082                                        sizeof(rdev->mddev->uuid));
1083
1084         mutex_init(&log->io_mutex);
1085
1086         spin_lock_init(&log->io_list_lock);
1087         INIT_LIST_HEAD(&log->running_ios);
1088         INIT_LIST_HEAD(&log->io_end_ios);
1089         INIT_LIST_HEAD(&log->stripe_end_ios);
1090         INIT_LIST_HEAD(&log->flushing_ios);
1091         INIT_LIST_HEAD(&log->flushed_ios);
1092         bio_init(&log->flush_bio);
1093
1094         log->io_kc = KMEM_CACHE(r5l_io_unit, 0);
1095         if (!log->io_kc)
1096                 goto io_kc;
1097
1098         log->reclaim_thread = md_register_thread(r5l_reclaim_thread,
1099                                                  log->rdev->mddev, "reclaim");
1100         if (!log->reclaim_thread)
1101                 goto reclaim_thread;
1102         init_waitqueue_head(&log->iounit_wait);
1103
1104         INIT_LIST_HEAD(&log->no_space_stripes);
1105         spin_lock_init(&log->no_space_stripes_lock);
1106
1107         if (r5l_load_log(log))
1108                 goto error;
1109
1110         conf->log = log;
1111         return 0;
1112 error:
1113         md_unregister_thread(&log->reclaim_thread);
1114 reclaim_thread:
1115         kmem_cache_destroy(log->io_kc);
1116 io_kc:
1117         kfree(log);
1118         return -EINVAL;
1119 }
1120
1121 void r5l_exit_log(struct r5l_log *log)
1122 {
1123         /*
1124          * at this point all stripes are finished, so io_unit is at least in
1125          * STRIPE_END state
1126          */
1127         r5l_wake_reclaim(log, -1L);
1128         md_unregister_thread(&log->reclaim_thread);
1129         r5l_do_reclaim(log);
1130         /*
1131          * force a super update, r5l_do_reclaim might updated the super.
1132          * mddev->thread is already stopped
1133          */
1134         md_update_sb(log->rdev->mddev, 1);
1135
1136         kmem_cache_destroy(log->io_kc);
1137         kfree(log);
1138 }