1f6bed0efSShaohua Li /* 2f6bed0efSShaohua Li * Copyright (C) 2015 Shaohua Li <shli@fb.com> 3b4c625c6SSong Liu * Copyright (C) 2016 Song Liu <songliubraving@fb.com> 4f6bed0efSShaohua Li * 5f6bed0efSShaohua Li * This program is free software; you can redistribute it and/or modify it 6f6bed0efSShaohua Li * under the terms and conditions of the GNU General Public License, 7f6bed0efSShaohua Li * version 2, as published by the Free Software Foundation. 8f6bed0efSShaohua Li * 9f6bed0efSShaohua Li * This program is distributed in the hope it will be useful, but WITHOUT 10f6bed0efSShaohua Li * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11f6bed0efSShaohua Li * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for 12f6bed0efSShaohua Li * more details. 13f6bed0efSShaohua Li * 14f6bed0efSShaohua Li */ 15f6bed0efSShaohua Li #include <linux/kernel.h> 16f6bed0efSShaohua Li #include <linux/wait.h> 17f6bed0efSShaohua Li #include <linux/blkdev.h> 18f6bed0efSShaohua Li #include <linux/slab.h> 19f6bed0efSShaohua Li #include <linux/raid/md_p.h> 205cb2fbd6SShaohua Li #include <linux/crc32c.h> 21f6bed0efSShaohua Li #include <linux/random.h> 22ce1ccd07SShaohua Li #include <linux/kthread.h> 23f6bed0efSShaohua Li #include "md.h" 24f6bed0efSShaohua Li #include "raid5.h" 251e6d690bSSong Liu #include "bitmap.h" 26f6bed0efSShaohua Li 27f6bed0efSShaohua Li /* 28f6bed0efSShaohua Li * metadata/data stored in disk with 4k size unit (a block) regardless 29f6bed0efSShaohua Li * underneath hardware sector size. only works with PAGE_SIZE == 4096 30f6bed0efSShaohua Li */ 31f6bed0efSShaohua Li #define BLOCK_SECTORS (8) 32f6bed0efSShaohua Li 330576b1c6SShaohua Li /* 34a39f7afdSSong Liu * log->max_free_space is min(1/4 disk size, 10G reclaimable space). 35a39f7afdSSong Liu * 36a39f7afdSSong Liu * In write through mode, the reclaim runs every log->max_free_space. 37a39f7afdSSong Liu * This can prevent the recovery scans for too long 380576b1c6SShaohua Li */ 390576b1c6SShaohua Li #define RECLAIM_MAX_FREE_SPACE (10 * 1024 * 1024 * 2) /* sector */ 400576b1c6SShaohua Li #define RECLAIM_MAX_FREE_SPACE_SHIFT (2) 410576b1c6SShaohua Li 42a39f7afdSSong Liu /* wake up reclaim thread periodically */ 43a39f7afdSSong Liu #define R5C_RECLAIM_WAKEUP_INTERVAL (30 * HZ) 44a39f7afdSSong Liu /* start flush with these full stripes */ 45a39f7afdSSong Liu #define R5C_FULL_STRIPE_FLUSH_BATCH 256 46a39f7afdSSong Liu /* reclaim stripes in groups */ 47a39f7afdSSong Liu #define R5C_RECLAIM_STRIPE_GROUP (NR_STRIPE_HASH_LOCKS * 2) 48a39f7afdSSong Liu 49c38d29b3SChristoph Hellwig /* 50c38d29b3SChristoph Hellwig * We only need 2 bios per I/O unit to make progress, but ensure we 51c38d29b3SChristoph Hellwig * have a few more available to not get too tight. 52c38d29b3SChristoph Hellwig */ 53c38d29b3SChristoph Hellwig #define R5L_POOL_SIZE 4 54c38d29b3SChristoph Hellwig 552ded3703SSong Liu /* 562ded3703SSong Liu * r5c journal modes of the array: write-back or write-through. 572ded3703SSong Liu * write-through mode has identical behavior as existing log only 582ded3703SSong Liu * implementation. 592ded3703SSong Liu */ 602ded3703SSong Liu enum r5c_journal_mode { 612ded3703SSong Liu R5C_JOURNAL_MODE_WRITE_THROUGH = 0, 622ded3703SSong Liu R5C_JOURNAL_MODE_WRITE_BACK = 1, 632ded3703SSong Liu }; 642ded3703SSong Liu 652c7da14bSSong Liu static char *r5c_journal_mode_str[] = {"write-through", 662c7da14bSSong Liu "write-back"}; 672ded3703SSong Liu /* 682ded3703SSong Liu * raid5 cache state machine 692ded3703SSong Liu * 702ded3703SSong Liu * With rhe RAID cache, each stripe works in two phases: 712ded3703SSong Liu * - caching phase 722ded3703SSong Liu * - writing-out phase 732ded3703SSong Liu * 742ded3703SSong Liu * These two phases are controlled by bit STRIPE_R5C_CACHING: 752ded3703SSong Liu * if STRIPE_R5C_CACHING == 0, the stripe is in writing-out phase 762ded3703SSong Liu * if STRIPE_R5C_CACHING == 1, the stripe is in caching phase 772ded3703SSong Liu * 782ded3703SSong Liu * When there is no journal, or the journal is in write-through mode, 792ded3703SSong Liu * the stripe is always in writing-out phase. 802ded3703SSong Liu * 812ded3703SSong Liu * For write-back journal, the stripe is sent to caching phase on write 822ded3703SSong Liu * (r5c_try_caching_write). r5c_make_stripe_write_out() kicks off 832ded3703SSong Liu * the write-out phase by clearing STRIPE_R5C_CACHING. 842ded3703SSong Liu * 852ded3703SSong Liu * Stripes in caching phase do not write the raid disks. Instead, all 862ded3703SSong Liu * writes are committed from the log device. Therefore, a stripe in 872ded3703SSong Liu * caching phase handles writes as: 882ded3703SSong Liu * - write to log device 892ded3703SSong Liu * - return IO 902ded3703SSong Liu * 912ded3703SSong Liu * Stripes in writing-out phase handle writes as: 922ded3703SSong Liu * - calculate parity 932ded3703SSong Liu * - write pending data and parity to journal 942ded3703SSong Liu * - write data and parity to raid disks 952ded3703SSong Liu * - return IO for pending writes 962ded3703SSong Liu */ 972ded3703SSong Liu 98f6bed0efSShaohua Li struct r5l_log { 99f6bed0efSShaohua Li struct md_rdev *rdev; 100f6bed0efSShaohua Li 101f6bed0efSShaohua Li u32 uuid_checksum; 102f6bed0efSShaohua Li 103f6bed0efSShaohua Li sector_t device_size; /* log device size, round to 104f6bed0efSShaohua Li * BLOCK_SECTORS */ 1050576b1c6SShaohua Li sector_t max_free_space; /* reclaim run if free space is at 1060576b1c6SShaohua Li * this size */ 107f6bed0efSShaohua Li 108f6bed0efSShaohua Li sector_t last_checkpoint; /* log tail. where recovery scan 109f6bed0efSShaohua Li * starts from */ 110f6bed0efSShaohua Li u64 last_cp_seq; /* log tail sequence */ 111f6bed0efSShaohua Li 112f6bed0efSShaohua Li sector_t log_start; /* log head. where new data appends */ 113f6bed0efSShaohua Li u64 seq; /* log head sequence */ 114f6bed0efSShaohua Li 11517036461SChristoph Hellwig sector_t next_checkpoint; 11617036461SChristoph Hellwig u64 next_cp_seq; 11717036461SChristoph Hellwig 118f6bed0efSShaohua Li struct mutex io_mutex; 119f6bed0efSShaohua Li struct r5l_io_unit *current_io; /* current io_unit accepting new data */ 120f6bed0efSShaohua Li 121f6bed0efSShaohua Li spinlock_t io_list_lock; 122f6bed0efSShaohua Li struct list_head running_ios; /* io_units which are still running, 123f6bed0efSShaohua Li * and have not yet been completely 124f6bed0efSShaohua Li * written to the log */ 125f6bed0efSShaohua Li struct list_head io_end_ios; /* io_units which have been completely 126f6bed0efSShaohua Li * written to the log but not yet written 127f6bed0efSShaohua Li * to the RAID */ 128a8c34f91SShaohua Li struct list_head flushing_ios; /* io_units which are waiting for log 129a8c34f91SShaohua Li * cache flush */ 13004732f74SChristoph Hellwig struct list_head finished_ios; /* io_units which settle down in log disk */ 131a8c34f91SShaohua Li struct bio flush_bio; 132f6bed0efSShaohua Li 1335036c390SChristoph Hellwig struct list_head no_mem_stripes; /* pending stripes, -ENOMEM */ 1345036c390SChristoph Hellwig 135f6bed0efSShaohua Li struct kmem_cache *io_kc; 1365036c390SChristoph Hellwig mempool_t *io_pool; 137c38d29b3SChristoph Hellwig struct bio_set *bs; 138e8deb638SChristoph Hellwig mempool_t *meta_pool; 139f6bed0efSShaohua Li 1400576b1c6SShaohua Li struct md_thread *reclaim_thread; 1410576b1c6SShaohua Li unsigned long reclaim_target; /* number of space that need to be 1420576b1c6SShaohua Li * reclaimed. if it's 0, reclaim spaces 1430576b1c6SShaohua Li * used by io_units which are in 1440576b1c6SShaohua Li * IO_UNIT_STRIPE_END state (eg, reclaim 1450576b1c6SShaohua Li * dones't wait for specific io_unit 1460576b1c6SShaohua Li * switching to IO_UNIT_STRIPE_END 1470576b1c6SShaohua Li * state) */ 1480fd22b45SShaohua Li wait_queue_head_t iounit_wait; 1490576b1c6SShaohua Li 150f6bed0efSShaohua Li struct list_head no_space_stripes; /* pending stripes, log has no space */ 151f6bed0efSShaohua Li spinlock_t no_space_stripes_lock; 15256fef7c6SChristoph Hellwig 15356fef7c6SChristoph Hellwig bool need_cache_flush; 1542ded3703SSong Liu 1552ded3703SSong Liu /* for r5c_cache */ 1562ded3703SSong Liu enum r5c_journal_mode r5c_journal_mode; 157a39f7afdSSong Liu 158a39f7afdSSong Liu /* all stripes in r5cache, in the order of seq at sh->log_start */ 159a39f7afdSSong Liu struct list_head stripe_in_journal_list; 160a39f7afdSSong Liu 161a39f7afdSSong Liu spinlock_t stripe_in_journal_lock; 162a39f7afdSSong Liu atomic_t stripe_in_journal_count; 1633bddb7f8SSong Liu 1643bddb7f8SSong Liu /* to submit async io_units, to fulfill ordering of flush */ 1653bddb7f8SSong Liu struct work_struct deferred_io_work; 166f6bed0efSShaohua Li }; 167f6bed0efSShaohua Li 168f6bed0efSShaohua Li /* 169f6bed0efSShaohua Li * an IO range starts from a meta data block and end at the next meta data 170f6bed0efSShaohua Li * block. The io unit's the meta data block tracks data/parity followed it. io 171f6bed0efSShaohua Li * unit is written to log disk with normal write, as we always flush log disk 172f6bed0efSShaohua Li * first and then start move data to raid disks, there is no requirement to 173f6bed0efSShaohua Li * write io unit with FLUSH/FUA 174f6bed0efSShaohua Li */ 175f6bed0efSShaohua Li struct r5l_io_unit { 176f6bed0efSShaohua Li struct r5l_log *log; 177f6bed0efSShaohua Li 178f6bed0efSShaohua Li struct page *meta_page; /* store meta block */ 179f6bed0efSShaohua Li int meta_offset; /* current offset in meta_page */ 180f6bed0efSShaohua Li 181f6bed0efSShaohua Li struct bio *current_bio;/* current_bio accepting new data */ 182f6bed0efSShaohua Li 183f6bed0efSShaohua Li atomic_t pending_stripe;/* how many stripes not flushed to raid */ 184f6bed0efSShaohua Li u64 seq; /* seq number of the metablock */ 185f6bed0efSShaohua Li sector_t log_start; /* where the io_unit starts */ 186f6bed0efSShaohua Li sector_t log_end; /* where the io_unit ends */ 187f6bed0efSShaohua Li struct list_head log_sibling; /* log->running_ios */ 188f6bed0efSShaohua Li struct list_head stripe_list; /* stripes added to the io_unit */ 189f6bed0efSShaohua Li 190f6bed0efSShaohua Li int state; 1916143e2ceSChristoph Hellwig bool need_split_bio; 1923bddb7f8SSong Liu struct bio *split_bio; 1933bddb7f8SSong Liu 1943bddb7f8SSong Liu unsigned int has_flush:1; /* include flush request */ 1953bddb7f8SSong Liu unsigned int has_fua:1; /* include fua request */ 1963bddb7f8SSong Liu unsigned int has_null_flush:1; /* include empty flush request */ 1973bddb7f8SSong Liu /* 1983bddb7f8SSong Liu * io isn't sent yet, flush/fua request can only be submitted till it's 1993bddb7f8SSong Liu * the first IO in running_ios list 2003bddb7f8SSong Liu */ 2013bddb7f8SSong Liu unsigned int io_deferred:1; 2023bddb7f8SSong Liu 2033bddb7f8SSong Liu struct bio_list flush_barriers; /* size == 0 flush bios */ 204f6bed0efSShaohua Li }; 205f6bed0efSShaohua Li 206f6bed0efSShaohua Li /* r5l_io_unit state */ 207f6bed0efSShaohua Li enum r5l_io_unit_state { 208f6bed0efSShaohua Li IO_UNIT_RUNNING = 0, /* accepting new IO */ 209f6bed0efSShaohua Li IO_UNIT_IO_START = 1, /* io_unit bio start writing to log, 210f6bed0efSShaohua Li * don't accepting new bio */ 211f6bed0efSShaohua Li IO_UNIT_IO_END = 2, /* io_unit bio finish writing to log */ 212a8c34f91SShaohua Li IO_UNIT_STRIPE_END = 3, /* stripes data finished writing to raid */ 213f6bed0efSShaohua Li }; 214f6bed0efSShaohua Li 2152ded3703SSong Liu bool r5c_is_writeback(struct r5l_log *log) 2162ded3703SSong Liu { 2172ded3703SSong Liu return (log != NULL && 2182ded3703SSong Liu log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK); 2192ded3703SSong Liu } 2202ded3703SSong Liu 221f6bed0efSShaohua Li static sector_t r5l_ring_add(struct r5l_log *log, sector_t start, sector_t inc) 222f6bed0efSShaohua Li { 223f6bed0efSShaohua Li start += inc; 224f6bed0efSShaohua Li if (start >= log->device_size) 225f6bed0efSShaohua Li start = start - log->device_size; 226f6bed0efSShaohua Li return start; 227f6bed0efSShaohua Li } 228f6bed0efSShaohua Li 229f6bed0efSShaohua Li static sector_t r5l_ring_distance(struct r5l_log *log, sector_t start, 230f6bed0efSShaohua Li sector_t end) 231f6bed0efSShaohua Li { 232f6bed0efSShaohua Li if (end >= start) 233f6bed0efSShaohua Li return end - start; 234f6bed0efSShaohua Li else 235f6bed0efSShaohua Li return end + log->device_size - start; 236f6bed0efSShaohua Li } 237f6bed0efSShaohua Li 238f6bed0efSShaohua Li static bool r5l_has_free_space(struct r5l_log *log, sector_t size) 239f6bed0efSShaohua Li { 240f6bed0efSShaohua Li sector_t used_size; 241f6bed0efSShaohua Li 242f6bed0efSShaohua Li used_size = r5l_ring_distance(log, log->last_checkpoint, 243f6bed0efSShaohua Li log->log_start); 244f6bed0efSShaohua Li 245f6bed0efSShaohua Li return log->device_size > used_size + size; 246f6bed0efSShaohua Li } 247f6bed0efSShaohua Li 248f6bed0efSShaohua Li static void __r5l_set_io_unit_state(struct r5l_io_unit *io, 249f6bed0efSShaohua Li enum r5l_io_unit_state state) 250f6bed0efSShaohua Li { 251f6bed0efSShaohua Li if (WARN_ON(io->state >= state)) 252f6bed0efSShaohua Li return; 253f6bed0efSShaohua Li io->state = state; 254f6bed0efSShaohua Li } 255f6bed0efSShaohua Li 2561e6d690bSSong Liu static void 2571e6d690bSSong Liu r5c_return_dev_pending_writes(struct r5conf *conf, struct r5dev *dev, 2581e6d690bSSong Liu struct bio_list *return_bi) 2591e6d690bSSong Liu { 2601e6d690bSSong Liu struct bio *wbi, *wbi2; 2611e6d690bSSong Liu 2621e6d690bSSong Liu wbi = dev->written; 2631e6d690bSSong Liu dev->written = NULL; 2641e6d690bSSong Liu while (wbi && wbi->bi_iter.bi_sector < 2651e6d690bSSong Liu dev->sector + STRIPE_SECTORS) { 2661e6d690bSSong Liu wbi2 = r5_next_bio(wbi, dev->sector); 2671e6d690bSSong Liu if (!raid5_dec_bi_active_stripes(wbi)) { 2681e6d690bSSong Liu md_write_end(conf->mddev); 2691e6d690bSSong Liu bio_list_add(return_bi, wbi); 2701e6d690bSSong Liu } 2711e6d690bSSong Liu wbi = wbi2; 2721e6d690bSSong Liu } 2731e6d690bSSong Liu } 2741e6d690bSSong Liu 2751e6d690bSSong Liu void r5c_handle_cached_data_endio(struct r5conf *conf, 2761e6d690bSSong Liu struct stripe_head *sh, int disks, struct bio_list *return_bi) 2771e6d690bSSong Liu { 2781e6d690bSSong Liu int i; 2791e6d690bSSong Liu 2801e6d690bSSong Liu for (i = sh->disks; i--; ) { 2811e6d690bSSong Liu if (sh->dev[i].written) { 2821e6d690bSSong Liu set_bit(R5_UPTODATE, &sh->dev[i].flags); 2831e6d690bSSong Liu r5c_return_dev_pending_writes(conf, &sh->dev[i], 2841e6d690bSSong Liu return_bi); 2851e6d690bSSong Liu bitmap_endwrite(conf->mddev->bitmap, sh->sector, 2861e6d690bSSong Liu STRIPE_SECTORS, 2871e6d690bSSong Liu !test_bit(STRIPE_DEGRADED, &sh->state), 2881e6d690bSSong Liu 0); 2891e6d690bSSong Liu } 2901e6d690bSSong Liu } 2911e6d690bSSong Liu } 2921e6d690bSSong Liu 293a39f7afdSSong Liu /* Check whether we should flush some stripes to free up stripe cache */ 294a39f7afdSSong Liu void r5c_check_stripe_cache_usage(struct r5conf *conf) 295a39f7afdSSong Liu { 296a39f7afdSSong Liu int total_cached; 297a39f7afdSSong Liu 298a39f7afdSSong Liu if (!r5c_is_writeback(conf->log)) 299a39f7afdSSong Liu return; 300a39f7afdSSong Liu 301a39f7afdSSong Liu total_cached = atomic_read(&conf->r5c_cached_partial_stripes) + 302a39f7afdSSong Liu atomic_read(&conf->r5c_cached_full_stripes); 303a39f7afdSSong Liu 304a39f7afdSSong Liu /* 305a39f7afdSSong Liu * The following condition is true for either of the following: 306a39f7afdSSong Liu * - stripe cache pressure high: 307a39f7afdSSong Liu * total_cached > 3/4 min_nr_stripes || 308a39f7afdSSong Liu * empty_inactive_list_nr > 0 309a39f7afdSSong Liu * - stripe cache pressure moderate: 310a39f7afdSSong Liu * total_cached > 1/2 min_nr_stripes 311a39f7afdSSong Liu */ 312a39f7afdSSong Liu if (total_cached > conf->min_nr_stripes * 1 / 2 || 313a39f7afdSSong Liu atomic_read(&conf->empty_inactive_list_nr) > 0) 314a39f7afdSSong Liu r5l_wake_reclaim(conf->log, 0); 315a39f7afdSSong Liu } 316a39f7afdSSong Liu 317a39f7afdSSong Liu /* 318a39f7afdSSong Liu * flush cache when there are R5C_FULL_STRIPE_FLUSH_BATCH or more full 319a39f7afdSSong Liu * stripes in the cache 320a39f7afdSSong Liu */ 321a39f7afdSSong Liu void r5c_check_cached_full_stripe(struct r5conf *conf) 322a39f7afdSSong Liu { 323a39f7afdSSong Liu if (!r5c_is_writeback(conf->log)) 324a39f7afdSSong Liu return; 325a39f7afdSSong Liu 326a39f7afdSSong Liu /* 327a39f7afdSSong Liu * wake up reclaim for R5C_FULL_STRIPE_FLUSH_BATCH cached stripes 328a39f7afdSSong Liu * or a full stripe (chunk size / 4k stripes). 329a39f7afdSSong Liu */ 330a39f7afdSSong Liu if (atomic_read(&conf->r5c_cached_full_stripes) >= 331a39f7afdSSong Liu min(R5C_FULL_STRIPE_FLUSH_BATCH, 332a39f7afdSSong Liu conf->chunk_sectors >> STRIPE_SHIFT)) 333a39f7afdSSong Liu r5l_wake_reclaim(conf->log, 0); 334a39f7afdSSong Liu } 335a39f7afdSSong Liu 336a39f7afdSSong Liu /* 337a39f7afdSSong Liu * Total log space (in sectors) needed to flush all data in cache 338a39f7afdSSong Liu * 339a39f7afdSSong Liu * Currently, writing-out phase automatically includes all pending writes 340a39f7afdSSong Liu * to the same sector. So the reclaim of each stripe takes up to 341a39f7afdSSong Liu * (conf->raid_disks + 1) pages of log space. 342a39f7afdSSong Liu * 343a39f7afdSSong Liu * To totally avoid deadlock due to log space, the code reserves 344a39f7afdSSong Liu * (conf->raid_disks + 1) pages for each stripe in cache, which is not 345a39f7afdSSong Liu * necessary in most cases. 346a39f7afdSSong Liu * 347a39f7afdSSong Liu * To improve this, we will need writing-out phase to be able to NOT include 348a39f7afdSSong Liu * pending writes, which will reduce the requirement to 349a39f7afdSSong Liu * (conf->max_degraded + 1) pages per stripe in cache. 350a39f7afdSSong Liu */ 351a39f7afdSSong Liu static sector_t r5c_log_required_to_flush_cache(struct r5conf *conf) 352a39f7afdSSong Liu { 353a39f7afdSSong Liu struct r5l_log *log = conf->log; 354a39f7afdSSong Liu 355a39f7afdSSong Liu if (!r5c_is_writeback(log)) 356a39f7afdSSong Liu return 0; 357a39f7afdSSong Liu 358a39f7afdSSong Liu return BLOCK_SECTORS * (conf->raid_disks + 1) * 359a39f7afdSSong Liu atomic_read(&log->stripe_in_journal_count); 360a39f7afdSSong Liu } 361a39f7afdSSong Liu 362a39f7afdSSong Liu /* 363a39f7afdSSong Liu * evaluate log space usage and update R5C_LOG_TIGHT and R5C_LOG_CRITICAL 364a39f7afdSSong Liu * 365a39f7afdSSong Liu * R5C_LOG_TIGHT is set when free space on the log device is less than 3x of 366a39f7afdSSong Liu * reclaim_required_space. R5C_LOG_CRITICAL is set when free space on the log 367a39f7afdSSong Liu * device is less than 2x of reclaim_required_space. 368a39f7afdSSong Liu */ 369a39f7afdSSong Liu static inline void r5c_update_log_state(struct r5l_log *log) 370a39f7afdSSong Liu { 371a39f7afdSSong Liu struct r5conf *conf = log->rdev->mddev->private; 372a39f7afdSSong Liu sector_t free_space; 373a39f7afdSSong Liu sector_t reclaim_space; 374a39f7afdSSong Liu 375a39f7afdSSong Liu if (!r5c_is_writeback(log)) 376a39f7afdSSong Liu return; 377a39f7afdSSong Liu 378a39f7afdSSong Liu free_space = r5l_ring_distance(log, log->log_start, 379a39f7afdSSong Liu log->last_checkpoint); 380a39f7afdSSong Liu reclaim_space = r5c_log_required_to_flush_cache(conf); 381a39f7afdSSong Liu if (free_space < 2 * reclaim_space) 382a39f7afdSSong Liu set_bit(R5C_LOG_CRITICAL, &conf->cache_state); 383a39f7afdSSong Liu else 384a39f7afdSSong Liu clear_bit(R5C_LOG_CRITICAL, &conf->cache_state); 385a39f7afdSSong Liu if (free_space < 3 * reclaim_space) 386a39f7afdSSong Liu set_bit(R5C_LOG_TIGHT, &conf->cache_state); 387a39f7afdSSong Liu else 388a39f7afdSSong Liu clear_bit(R5C_LOG_TIGHT, &conf->cache_state); 389a39f7afdSSong Liu } 390a39f7afdSSong Liu 3912ded3703SSong Liu /* 3922ded3703SSong Liu * Put the stripe into writing-out phase by clearing STRIPE_R5C_CACHING. 3932ded3703SSong Liu * This function should only be called in write-back mode. 3942ded3703SSong Liu */ 395a39f7afdSSong Liu void r5c_make_stripe_write_out(struct stripe_head *sh) 3962ded3703SSong Liu { 3972ded3703SSong Liu struct r5conf *conf = sh->raid_conf; 3982ded3703SSong Liu struct r5l_log *log = conf->log; 3992ded3703SSong Liu 4002ded3703SSong Liu BUG_ON(!r5c_is_writeback(log)); 4012ded3703SSong Liu 4022ded3703SSong Liu WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state)); 4032ded3703SSong Liu clear_bit(STRIPE_R5C_CACHING, &sh->state); 4041e6d690bSSong Liu 4051e6d690bSSong Liu if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) 4061e6d690bSSong Liu atomic_inc(&conf->preread_active_stripes); 4071e6d690bSSong Liu 4081e6d690bSSong Liu if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) { 4091e6d690bSSong Liu BUG_ON(atomic_read(&conf->r5c_cached_partial_stripes) == 0); 4101e6d690bSSong Liu atomic_dec(&conf->r5c_cached_partial_stripes); 4111e6d690bSSong Liu } 4121e6d690bSSong Liu 4131e6d690bSSong Liu if (test_and_clear_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) { 4141e6d690bSSong Liu BUG_ON(atomic_read(&conf->r5c_cached_full_stripes) == 0); 4151e6d690bSSong Liu atomic_dec(&conf->r5c_cached_full_stripes); 4161e6d690bSSong Liu } 4171e6d690bSSong Liu } 4181e6d690bSSong Liu 4191e6d690bSSong Liu static void r5c_handle_data_cached(struct stripe_head *sh) 4201e6d690bSSong Liu { 4211e6d690bSSong Liu int i; 4221e6d690bSSong Liu 4231e6d690bSSong Liu for (i = sh->disks; i--; ) 4241e6d690bSSong Liu if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) { 4251e6d690bSSong Liu set_bit(R5_InJournal, &sh->dev[i].flags); 4261e6d690bSSong Liu clear_bit(R5_LOCKED, &sh->dev[i].flags); 4271e6d690bSSong Liu } 4281e6d690bSSong Liu clear_bit(STRIPE_LOG_TRAPPED, &sh->state); 4291e6d690bSSong Liu } 4301e6d690bSSong Liu 4311e6d690bSSong Liu /* 4321e6d690bSSong Liu * this journal write must contain full parity, 4331e6d690bSSong Liu * it may also contain some data pages 4341e6d690bSSong Liu */ 4351e6d690bSSong Liu static void r5c_handle_parity_cached(struct stripe_head *sh) 4361e6d690bSSong Liu { 4371e6d690bSSong Liu int i; 4381e6d690bSSong Liu 4391e6d690bSSong Liu for (i = sh->disks; i--; ) 4401e6d690bSSong Liu if (test_bit(R5_InJournal, &sh->dev[i].flags)) 4411e6d690bSSong Liu set_bit(R5_Wantwrite, &sh->dev[i].flags); 4422ded3703SSong Liu } 4432ded3703SSong Liu 4442ded3703SSong Liu /* 4452ded3703SSong Liu * Setting proper flags after writing (or flushing) data and/or parity to the 4462ded3703SSong Liu * log device. This is called from r5l_log_endio() or r5l_log_flush_endio(). 4472ded3703SSong Liu */ 4482ded3703SSong Liu static void r5c_finish_cache_stripe(struct stripe_head *sh) 4492ded3703SSong Liu { 4502ded3703SSong Liu struct r5l_log *log = sh->raid_conf->log; 4512ded3703SSong Liu 4522ded3703SSong Liu if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) { 4532ded3703SSong Liu BUG_ON(test_bit(STRIPE_R5C_CACHING, &sh->state)); 4542ded3703SSong Liu /* 4552ded3703SSong Liu * Set R5_InJournal for parity dev[pd_idx]. This means 4562ded3703SSong Liu * all data AND parity in the journal. For RAID 6, it is 4572ded3703SSong Liu * NOT necessary to set the flag for dev[qd_idx], as the 4582ded3703SSong Liu * two parities are written out together. 4592ded3703SSong Liu */ 4602ded3703SSong Liu set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags); 4611e6d690bSSong Liu } else if (test_bit(STRIPE_R5C_CACHING, &sh->state)) { 4621e6d690bSSong Liu r5c_handle_data_cached(sh); 4631e6d690bSSong Liu } else { 4641e6d690bSSong Liu r5c_handle_parity_cached(sh); 4651e6d690bSSong Liu set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags); 4661e6d690bSSong Liu } 4672ded3703SSong Liu } 4682ded3703SSong Liu 469d8858f43SChristoph Hellwig static void r5l_io_run_stripes(struct r5l_io_unit *io) 470d8858f43SChristoph Hellwig { 471d8858f43SChristoph Hellwig struct stripe_head *sh, *next; 472d8858f43SChristoph Hellwig 473d8858f43SChristoph Hellwig list_for_each_entry_safe(sh, next, &io->stripe_list, log_list) { 474d8858f43SChristoph Hellwig list_del_init(&sh->log_list); 4752ded3703SSong Liu 4762ded3703SSong Liu r5c_finish_cache_stripe(sh); 4772ded3703SSong Liu 478d8858f43SChristoph Hellwig set_bit(STRIPE_HANDLE, &sh->state); 479d8858f43SChristoph Hellwig raid5_release_stripe(sh); 480d8858f43SChristoph Hellwig } 481d8858f43SChristoph Hellwig } 482d8858f43SChristoph Hellwig 48356fef7c6SChristoph Hellwig static void r5l_log_run_stripes(struct r5l_log *log) 48456fef7c6SChristoph Hellwig { 48556fef7c6SChristoph Hellwig struct r5l_io_unit *io, *next; 48656fef7c6SChristoph Hellwig 48756fef7c6SChristoph Hellwig assert_spin_locked(&log->io_list_lock); 48856fef7c6SChristoph Hellwig 48956fef7c6SChristoph Hellwig list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) { 49056fef7c6SChristoph Hellwig /* don't change list order */ 49156fef7c6SChristoph Hellwig if (io->state < IO_UNIT_IO_END) 49256fef7c6SChristoph Hellwig break; 49356fef7c6SChristoph Hellwig 49456fef7c6SChristoph Hellwig list_move_tail(&io->log_sibling, &log->finished_ios); 49556fef7c6SChristoph Hellwig r5l_io_run_stripes(io); 49656fef7c6SChristoph Hellwig } 49756fef7c6SChristoph Hellwig } 49856fef7c6SChristoph Hellwig 4993848c0bcSChristoph Hellwig static void r5l_move_to_end_ios(struct r5l_log *log) 5003848c0bcSChristoph Hellwig { 5013848c0bcSChristoph Hellwig struct r5l_io_unit *io, *next; 5023848c0bcSChristoph Hellwig 5033848c0bcSChristoph Hellwig assert_spin_locked(&log->io_list_lock); 5043848c0bcSChristoph Hellwig 5053848c0bcSChristoph Hellwig list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) { 5063848c0bcSChristoph Hellwig /* don't change list order */ 5073848c0bcSChristoph Hellwig if (io->state < IO_UNIT_IO_END) 5083848c0bcSChristoph Hellwig break; 5093848c0bcSChristoph Hellwig list_move_tail(&io->log_sibling, &log->io_end_ios); 5103848c0bcSChristoph Hellwig } 5113848c0bcSChristoph Hellwig } 5123848c0bcSChristoph Hellwig 5133bddb7f8SSong Liu static void __r5l_stripe_write_finished(struct r5l_io_unit *io); 514f6bed0efSShaohua Li static void r5l_log_endio(struct bio *bio) 515f6bed0efSShaohua Li { 516f6bed0efSShaohua Li struct r5l_io_unit *io = bio->bi_private; 5173bddb7f8SSong Liu struct r5l_io_unit *io_deferred; 518f6bed0efSShaohua Li struct r5l_log *log = io->log; 519509ffec7SChristoph Hellwig unsigned long flags; 520f6bed0efSShaohua Li 5216e74a9cfSShaohua Li if (bio->bi_error) 5226e74a9cfSShaohua Li md_error(log->rdev->mddev, log->rdev); 5236e74a9cfSShaohua Li 524f6bed0efSShaohua Li bio_put(bio); 525e8deb638SChristoph Hellwig mempool_free(io->meta_page, log->meta_pool); 526f6bed0efSShaohua Li 527509ffec7SChristoph Hellwig spin_lock_irqsave(&log->io_list_lock, flags); 528509ffec7SChristoph Hellwig __r5l_set_io_unit_state(io, IO_UNIT_IO_END); 52956fef7c6SChristoph Hellwig if (log->need_cache_flush) 5303848c0bcSChristoph Hellwig r5l_move_to_end_ios(log); 53156fef7c6SChristoph Hellwig else 53256fef7c6SChristoph Hellwig r5l_log_run_stripes(log); 5333bddb7f8SSong Liu if (!list_empty(&log->running_ios)) { 5343bddb7f8SSong Liu /* 5353bddb7f8SSong Liu * FLUSH/FUA io_unit is deferred because of ordering, now we 5363bddb7f8SSong Liu * can dispatch it 5373bddb7f8SSong Liu */ 5383bddb7f8SSong Liu io_deferred = list_first_entry(&log->running_ios, 5393bddb7f8SSong Liu struct r5l_io_unit, log_sibling); 5403bddb7f8SSong Liu if (io_deferred->io_deferred) 5413bddb7f8SSong Liu schedule_work(&log->deferred_io_work); 5423bddb7f8SSong Liu } 5433bddb7f8SSong Liu 544509ffec7SChristoph Hellwig spin_unlock_irqrestore(&log->io_list_lock, flags); 545509ffec7SChristoph Hellwig 54656fef7c6SChristoph Hellwig if (log->need_cache_flush) 547f6bed0efSShaohua Li md_wakeup_thread(log->rdev->mddev->thread); 5483bddb7f8SSong Liu 5493bddb7f8SSong Liu if (io->has_null_flush) { 5503bddb7f8SSong Liu struct bio *bi; 5513bddb7f8SSong Liu 5523bddb7f8SSong Liu WARN_ON(bio_list_empty(&io->flush_barriers)); 5533bddb7f8SSong Liu while ((bi = bio_list_pop(&io->flush_barriers)) != NULL) { 5543bddb7f8SSong Liu bio_endio(bi); 5553bddb7f8SSong Liu atomic_dec(&io->pending_stripe); 5563bddb7f8SSong Liu } 5573bddb7f8SSong Liu if (atomic_read(&io->pending_stripe) == 0) 5583bddb7f8SSong Liu __r5l_stripe_write_finished(io); 5593bddb7f8SSong Liu } 5603bddb7f8SSong Liu } 5613bddb7f8SSong Liu 5623bddb7f8SSong Liu static void r5l_do_submit_io(struct r5l_log *log, struct r5l_io_unit *io) 5633bddb7f8SSong Liu { 5643bddb7f8SSong Liu unsigned long flags; 5653bddb7f8SSong Liu 5663bddb7f8SSong Liu spin_lock_irqsave(&log->io_list_lock, flags); 5673bddb7f8SSong Liu __r5l_set_io_unit_state(io, IO_UNIT_IO_START); 5683bddb7f8SSong Liu spin_unlock_irqrestore(&log->io_list_lock, flags); 5693bddb7f8SSong Liu 5703bddb7f8SSong Liu if (io->has_flush) 5713bddb7f8SSong Liu bio_set_op_attrs(io->current_bio, REQ_OP_WRITE, WRITE_FLUSH); 5723bddb7f8SSong Liu if (io->has_fua) 5733bddb7f8SSong Liu bio_set_op_attrs(io->current_bio, REQ_OP_WRITE, WRITE_FUA); 5743bddb7f8SSong Liu submit_bio(io->current_bio); 5753bddb7f8SSong Liu 5763bddb7f8SSong Liu if (!io->split_bio) 5773bddb7f8SSong Liu return; 5783bddb7f8SSong Liu 5793bddb7f8SSong Liu if (io->has_flush) 5803bddb7f8SSong Liu bio_set_op_attrs(io->split_bio, REQ_OP_WRITE, WRITE_FLUSH); 5813bddb7f8SSong Liu if (io->has_fua) 5823bddb7f8SSong Liu bio_set_op_attrs(io->split_bio, REQ_OP_WRITE, WRITE_FUA); 5833bddb7f8SSong Liu submit_bio(io->split_bio); 5843bddb7f8SSong Liu } 5853bddb7f8SSong Liu 5863bddb7f8SSong Liu /* deferred io_unit will be dispatched here */ 5873bddb7f8SSong Liu static void r5l_submit_io_async(struct work_struct *work) 5883bddb7f8SSong Liu { 5893bddb7f8SSong Liu struct r5l_log *log = container_of(work, struct r5l_log, 5903bddb7f8SSong Liu deferred_io_work); 5913bddb7f8SSong Liu struct r5l_io_unit *io = NULL; 5923bddb7f8SSong Liu unsigned long flags; 5933bddb7f8SSong Liu 5943bddb7f8SSong Liu spin_lock_irqsave(&log->io_list_lock, flags); 5953bddb7f8SSong Liu if (!list_empty(&log->running_ios)) { 5963bddb7f8SSong Liu io = list_first_entry(&log->running_ios, struct r5l_io_unit, 5973bddb7f8SSong Liu log_sibling); 5983bddb7f8SSong Liu if (!io->io_deferred) 5993bddb7f8SSong Liu io = NULL; 6003bddb7f8SSong Liu else 6013bddb7f8SSong Liu io->io_deferred = 0; 6023bddb7f8SSong Liu } 6033bddb7f8SSong Liu spin_unlock_irqrestore(&log->io_list_lock, flags); 6043bddb7f8SSong Liu if (io) 6053bddb7f8SSong Liu r5l_do_submit_io(log, io); 606f6bed0efSShaohua Li } 607f6bed0efSShaohua Li 608f6bed0efSShaohua Li static void r5l_submit_current_io(struct r5l_log *log) 609f6bed0efSShaohua Li { 610f6bed0efSShaohua Li struct r5l_io_unit *io = log->current_io; 6113bddb7f8SSong Liu struct bio *bio; 612f6bed0efSShaohua Li struct r5l_meta_block *block; 613509ffec7SChristoph Hellwig unsigned long flags; 614f6bed0efSShaohua Li u32 crc; 6153bddb7f8SSong Liu bool do_submit = true; 616f6bed0efSShaohua Li 617f6bed0efSShaohua Li if (!io) 618f6bed0efSShaohua Li return; 619f6bed0efSShaohua Li 620f6bed0efSShaohua Li block = page_address(io->meta_page); 621f6bed0efSShaohua Li block->meta_size = cpu_to_le32(io->meta_offset); 6225cb2fbd6SShaohua Li crc = crc32c_le(log->uuid_checksum, block, PAGE_SIZE); 623f6bed0efSShaohua Li block->checksum = cpu_to_le32(crc); 6243bddb7f8SSong Liu bio = io->current_bio; 625f6bed0efSShaohua Li 626f6bed0efSShaohua Li log->current_io = NULL; 627509ffec7SChristoph Hellwig spin_lock_irqsave(&log->io_list_lock, flags); 6283bddb7f8SSong Liu if (io->has_flush || io->has_fua) { 6293bddb7f8SSong Liu if (io != list_first_entry(&log->running_ios, 6303bddb7f8SSong Liu struct r5l_io_unit, log_sibling)) { 6313bddb7f8SSong Liu io->io_deferred = 1; 6323bddb7f8SSong Liu do_submit = false; 6333bddb7f8SSong Liu } 6343bddb7f8SSong Liu } 635509ffec7SChristoph Hellwig spin_unlock_irqrestore(&log->io_list_lock, flags); 6363bddb7f8SSong Liu if (do_submit) 6373bddb7f8SSong Liu r5l_do_submit_io(log, io); 638f6bed0efSShaohua Li } 639f6bed0efSShaohua Li 6406143e2ceSChristoph Hellwig static struct bio *r5l_bio_alloc(struct r5l_log *log) 641b349feb3SChristoph Hellwig { 642c38d29b3SChristoph Hellwig struct bio *bio = bio_alloc_bioset(GFP_NOIO, BIO_MAX_PAGES, log->bs); 643b349feb3SChristoph Hellwig 644796a5cf0SMike Christie bio_set_op_attrs(bio, REQ_OP_WRITE, 0); 645b349feb3SChristoph Hellwig bio->bi_bdev = log->rdev->bdev; 6461e932a37SChristoph Hellwig bio->bi_iter.bi_sector = log->rdev->data_offset + log->log_start; 647b349feb3SChristoph Hellwig 648b349feb3SChristoph Hellwig return bio; 649b349feb3SChristoph Hellwig } 650b349feb3SChristoph Hellwig 651c1b99198SChristoph Hellwig static void r5_reserve_log_entry(struct r5l_log *log, struct r5l_io_unit *io) 652c1b99198SChristoph Hellwig { 653c1b99198SChristoph Hellwig log->log_start = r5l_ring_add(log, log->log_start, BLOCK_SECTORS); 654c1b99198SChristoph Hellwig 655a39f7afdSSong Liu r5c_update_log_state(log); 656c1b99198SChristoph Hellwig /* 657c1b99198SChristoph Hellwig * If we filled up the log device start from the beginning again, 658c1b99198SChristoph Hellwig * which will require a new bio. 659c1b99198SChristoph Hellwig * 660c1b99198SChristoph Hellwig * Note: for this to work properly the log size needs to me a multiple 661c1b99198SChristoph Hellwig * of BLOCK_SECTORS. 662c1b99198SChristoph Hellwig */ 663c1b99198SChristoph Hellwig if (log->log_start == 0) 6646143e2ceSChristoph Hellwig io->need_split_bio = true; 665c1b99198SChristoph Hellwig 666c1b99198SChristoph Hellwig io->log_end = log->log_start; 667c1b99198SChristoph Hellwig } 668c1b99198SChristoph Hellwig 669f6bed0efSShaohua Li static struct r5l_io_unit *r5l_new_meta(struct r5l_log *log) 670f6bed0efSShaohua Li { 671f6bed0efSShaohua Li struct r5l_io_unit *io; 672f6bed0efSShaohua Li struct r5l_meta_block *block; 673f6bed0efSShaohua Li 6745036c390SChristoph Hellwig io = mempool_alloc(log->io_pool, GFP_ATOMIC); 6755036c390SChristoph Hellwig if (!io) 6765036c390SChristoph Hellwig return NULL; 6775036c390SChristoph Hellwig memset(io, 0, sizeof(*io)); 6785036c390SChristoph Hellwig 67951039cd0SChristoph Hellwig io->log = log; 68051039cd0SChristoph Hellwig INIT_LIST_HEAD(&io->log_sibling); 68151039cd0SChristoph Hellwig INIT_LIST_HEAD(&io->stripe_list); 6823bddb7f8SSong Liu bio_list_init(&io->flush_barriers); 68351039cd0SChristoph Hellwig io->state = IO_UNIT_RUNNING; 684f6bed0efSShaohua Li 685e8deb638SChristoph Hellwig io->meta_page = mempool_alloc(log->meta_pool, GFP_NOIO); 686f6bed0efSShaohua Li block = page_address(io->meta_page); 687e8deb638SChristoph Hellwig clear_page(block); 688f6bed0efSShaohua Li block->magic = cpu_to_le32(R5LOG_MAGIC); 689f6bed0efSShaohua Li block->version = R5LOG_VERSION; 690f6bed0efSShaohua Li block->seq = cpu_to_le64(log->seq); 691f6bed0efSShaohua Li block->position = cpu_to_le64(log->log_start); 692f6bed0efSShaohua Li 693f6bed0efSShaohua Li io->log_start = log->log_start; 694f6bed0efSShaohua Li io->meta_offset = sizeof(struct r5l_meta_block); 6952b8ef16eSChristoph Hellwig io->seq = log->seq++; 696f6bed0efSShaohua Li 6976143e2ceSChristoph Hellwig io->current_bio = r5l_bio_alloc(log); 6986143e2ceSChristoph Hellwig io->current_bio->bi_end_io = r5l_log_endio; 6996143e2ceSChristoph Hellwig io->current_bio->bi_private = io; 700b349feb3SChristoph Hellwig bio_add_page(io->current_bio, io->meta_page, PAGE_SIZE, 0); 701f6bed0efSShaohua Li 702c1b99198SChristoph Hellwig r5_reserve_log_entry(log, io); 703f6bed0efSShaohua Li 704f6bed0efSShaohua Li spin_lock_irq(&log->io_list_lock); 705f6bed0efSShaohua Li list_add_tail(&io->log_sibling, &log->running_ios); 706f6bed0efSShaohua Li spin_unlock_irq(&log->io_list_lock); 707f6bed0efSShaohua Li 708f6bed0efSShaohua Li return io; 709f6bed0efSShaohua Li } 710f6bed0efSShaohua Li 711f6bed0efSShaohua Li static int r5l_get_meta(struct r5l_log *log, unsigned int payload_size) 712f6bed0efSShaohua Li { 71322581f58SChristoph Hellwig if (log->current_io && 71422581f58SChristoph Hellwig log->current_io->meta_offset + payload_size > PAGE_SIZE) 715f6bed0efSShaohua Li r5l_submit_current_io(log); 716f6bed0efSShaohua Li 7175036c390SChristoph Hellwig if (!log->current_io) { 718f6bed0efSShaohua Li log->current_io = r5l_new_meta(log); 7195036c390SChristoph Hellwig if (!log->current_io) 7205036c390SChristoph Hellwig return -ENOMEM; 7215036c390SChristoph Hellwig } 7225036c390SChristoph Hellwig 723f6bed0efSShaohua Li return 0; 724f6bed0efSShaohua Li } 725f6bed0efSShaohua Li 726f6bed0efSShaohua Li static void r5l_append_payload_meta(struct r5l_log *log, u16 type, 727f6bed0efSShaohua Li sector_t location, 728f6bed0efSShaohua Li u32 checksum1, u32 checksum2, 729f6bed0efSShaohua Li bool checksum2_valid) 730f6bed0efSShaohua Li { 731f6bed0efSShaohua Li struct r5l_io_unit *io = log->current_io; 732f6bed0efSShaohua Li struct r5l_payload_data_parity *payload; 733f6bed0efSShaohua Li 734f6bed0efSShaohua Li payload = page_address(io->meta_page) + io->meta_offset; 735f6bed0efSShaohua Li payload->header.type = cpu_to_le16(type); 736f6bed0efSShaohua Li payload->header.flags = cpu_to_le16(0); 737f6bed0efSShaohua Li payload->size = cpu_to_le32((1 + !!checksum2_valid) << 738f6bed0efSShaohua Li (PAGE_SHIFT - 9)); 739f6bed0efSShaohua Li payload->location = cpu_to_le64(location); 740f6bed0efSShaohua Li payload->checksum[0] = cpu_to_le32(checksum1); 741f6bed0efSShaohua Li if (checksum2_valid) 742f6bed0efSShaohua Li payload->checksum[1] = cpu_to_le32(checksum2); 743f6bed0efSShaohua Li 744f6bed0efSShaohua Li io->meta_offset += sizeof(struct r5l_payload_data_parity) + 745f6bed0efSShaohua Li sizeof(__le32) * (1 + !!checksum2_valid); 746f6bed0efSShaohua Li } 747f6bed0efSShaohua Li 748f6bed0efSShaohua Li static void r5l_append_payload_page(struct r5l_log *log, struct page *page) 749f6bed0efSShaohua Li { 750f6bed0efSShaohua Li struct r5l_io_unit *io = log->current_io; 751f6bed0efSShaohua Li 7526143e2ceSChristoph Hellwig if (io->need_split_bio) { 7533bddb7f8SSong Liu BUG_ON(io->split_bio); 7543bddb7f8SSong Liu io->split_bio = io->current_bio; 7556143e2ceSChristoph Hellwig io->current_bio = r5l_bio_alloc(log); 7563bddb7f8SSong Liu bio_chain(io->current_bio, io->split_bio); 7573bddb7f8SSong Liu io->need_split_bio = false; 758f6bed0efSShaohua Li } 759f6bed0efSShaohua Li 7606143e2ceSChristoph Hellwig if (!bio_add_page(io->current_bio, page, PAGE_SIZE, 0)) 7616143e2ceSChristoph Hellwig BUG(); 7626143e2ceSChristoph Hellwig 763c1b99198SChristoph Hellwig r5_reserve_log_entry(log, io); 764f6bed0efSShaohua Li } 765f6bed0efSShaohua Li 7665036c390SChristoph Hellwig static int r5l_log_stripe(struct r5l_log *log, struct stripe_head *sh, 767f6bed0efSShaohua Li int data_pages, int parity_pages) 768f6bed0efSShaohua Li { 769f6bed0efSShaohua Li int i; 770f6bed0efSShaohua Li int meta_size; 7715036c390SChristoph Hellwig int ret; 772f6bed0efSShaohua Li struct r5l_io_unit *io; 773f6bed0efSShaohua Li 774f6bed0efSShaohua Li meta_size = 775f6bed0efSShaohua Li ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32)) 776f6bed0efSShaohua Li * data_pages) + 777f6bed0efSShaohua Li sizeof(struct r5l_payload_data_parity) + 778f6bed0efSShaohua Li sizeof(__le32) * parity_pages; 779f6bed0efSShaohua Li 7805036c390SChristoph Hellwig ret = r5l_get_meta(log, meta_size); 7815036c390SChristoph Hellwig if (ret) 7825036c390SChristoph Hellwig return ret; 7835036c390SChristoph Hellwig 784f6bed0efSShaohua Li io = log->current_io; 785f6bed0efSShaohua Li 7863bddb7f8SSong Liu if (test_and_clear_bit(STRIPE_R5C_PREFLUSH, &sh->state)) 7873bddb7f8SSong Liu io->has_flush = 1; 7883bddb7f8SSong Liu 789f6bed0efSShaohua Li for (i = 0; i < sh->disks; i++) { 7901e6d690bSSong Liu if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) || 7911e6d690bSSong Liu test_bit(R5_InJournal, &sh->dev[i].flags)) 792f6bed0efSShaohua Li continue; 793f6bed0efSShaohua Li if (i == sh->pd_idx || i == sh->qd_idx) 794f6bed0efSShaohua Li continue; 7953bddb7f8SSong Liu if (test_bit(R5_WantFUA, &sh->dev[i].flags) && 7963bddb7f8SSong Liu log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK) { 7973bddb7f8SSong Liu io->has_fua = 1; 7983bddb7f8SSong Liu /* 7993bddb7f8SSong Liu * we need to flush journal to make sure recovery can 8003bddb7f8SSong Liu * reach the data with fua flag 8013bddb7f8SSong Liu */ 8023bddb7f8SSong Liu io->has_flush = 1; 8033bddb7f8SSong Liu } 804f6bed0efSShaohua Li r5l_append_payload_meta(log, R5LOG_PAYLOAD_DATA, 805f6bed0efSShaohua Li raid5_compute_blocknr(sh, i, 0), 806f6bed0efSShaohua Li sh->dev[i].log_checksum, 0, false); 807f6bed0efSShaohua Li r5l_append_payload_page(log, sh->dev[i].page); 808f6bed0efSShaohua Li } 809f6bed0efSShaohua Li 8102ded3703SSong Liu if (parity_pages == 2) { 811f6bed0efSShaohua Li r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY, 812f6bed0efSShaohua Li sh->sector, sh->dev[sh->pd_idx].log_checksum, 813f6bed0efSShaohua Li sh->dev[sh->qd_idx].log_checksum, true); 814f6bed0efSShaohua Li r5l_append_payload_page(log, sh->dev[sh->pd_idx].page); 815f6bed0efSShaohua Li r5l_append_payload_page(log, sh->dev[sh->qd_idx].page); 8162ded3703SSong Liu } else if (parity_pages == 1) { 817f6bed0efSShaohua Li r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY, 818f6bed0efSShaohua Li sh->sector, sh->dev[sh->pd_idx].log_checksum, 819f6bed0efSShaohua Li 0, false); 820f6bed0efSShaohua Li r5l_append_payload_page(log, sh->dev[sh->pd_idx].page); 8212ded3703SSong Liu } else /* Just writing data, not parity, in caching phase */ 8222ded3703SSong Liu BUG_ON(parity_pages != 0); 823f6bed0efSShaohua Li 824f6bed0efSShaohua Li list_add_tail(&sh->log_list, &io->stripe_list); 825f6bed0efSShaohua Li atomic_inc(&io->pending_stripe); 826f6bed0efSShaohua Li sh->log_io = io; 8275036c390SChristoph Hellwig 828a39f7afdSSong Liu if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) 829a39f7afdSSong Liu return 0; 830a39f7afdSSong Liu 831a39f7afdSSong Liu if (sh->log_start == MaxSector) { 832a39f7afdSSong Liu BUG_ON(!list_empty(&sh->r5c)); 833a39f7afdSSong Liu sh->log_start = io->log_start; 834a39f7afdSSong Liu spin_lock_irq(&log->stripe_in_journal_lock); 835a39f7afdSSong Liu list_add_tail(&sh->r5c, 836a39f7afdSSong Liu &log->stripe_in_journal_list); 837a39f7afdSSong Liu spin_unlock_irq(&log->stripe_in_journal_lock); 838a39f7afdSSong Liu atomic_inc(&log->stripe_in_journal_count); 839a39f7afdSSong Liu } 8405036c390SChristoph Hellwig return 0; 841f6bed0efSShaohua Li } 842f6bed0efSShaohua Li 843a39f7afdSSong Liu /* add stripe to no_space_stripes, and then wake up reclaim */ 844a39f7afdSSong Liu static inline void r5l_add_no_space_stripe(struct r5l_log *log, 845a39f7afdSSong Liu struct stripe_head *sh) 846a39f7afdSSong Liu { 847a39f7afdSSong Liu spin_lock(&log->no_space_stripes_lock); 848a39f7afdSSong Liu list_add_tail(&sh->log_list, &log->no_space_stripes); 849a39f7afdSSong Liu spin_unlock(&log->no_space_stripes_lock); 850a39f7afdSSong Liu } 851a39f7afdSSong Liu 852f6bed0efSShaohua Li /* 853f6bed0efSShaohua Li * running in raid5d, where reclaim could wait for raid5d too (when it flushes 854f6bed0efSShaohua Li * data from log to raid disks), so we shouldn't wait for reclaim here 855f6bed0efSShaohua Li */ 856f6bed0efSShaohua Li int r5l_write_stripe(struct r5l_log *log, struct stripe_head *sh) 857f6bed0efSShaohua Li { 858a39f7afdSSong Liu struct r5conf *conf = sh->raid_conf; 859f6bed0efSShaohua Li int write_disks = 0; 860f6bed0efSShaohua Li int data_pages, parity_pages; 861f6bed0efSShaohua Li int reserve; 862f6bed0efSShaohua Li int i; 8635036c390SChristoph Hellwig int ret = 0; 864a39f7afdSSong Liu bool wake_reclaim = false; 865f6bed0efSShaohua Li 866f6bed0efSShaohua Li if (!log) 867f6bed0efSShaohua Li return -EAGAIN; 868f6bed0efSShaohua Li /* Don't support stripe batch */ 869f6bed0efSShaohua Li if (sh->log_io || !test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags) || 870f6bed0efSShaohua Li test_bit(STRIPE_SYNCING, &sh->state)) { 871f6bed0efSShaohua Li /* the stripe is written to log, we start writing it to raid */ 872f6bed0efSShaohua Li clear_bit(STRIPE_LOG_TRAPPED, &sh->state); 873f6bed0efSShaohua Li return -EAGAIN; 874f6bed0efSShaohua Li } 875f6bed0efSShaohua Li 8762ded3703SSong Liu WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state)); 8772ded3703SSong Liu 878f6bed0efSShaohua Li for (i = 0; i < sh->disks; i++) { 879f6bed0efSShaohua Li void *addr; 880f6bed0efSShaohua Li 8811e6d690bSSong Liu if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) || 8821e6d690bSSong Liu test_bit(R5_InJournal, &sh->dev[i].flags)) 883f6bed0efSShaohua Li continue; 8841e6d690bSSong Liu 885f6bed0efSShaohua Li write_disks++; 886f6bed0efSShaohua Li /* checksum is already calculated in last run */ 887f6bed0efSShaohua Li if (test_bit(STRIPE_LOG_TRAPPED, &sh->state)) 888f6bed0efSShaohua Li continue; 889f6bed0efSShaohua Li addr = kmap_atomic(sh->dev[i].page); 8905cb2fbd6SShaohua Li sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum, 891f6bed0efSShaohua Li addr, PAGE_SIZE); 892f6bed0efSShaohua Li kunmap_atomic(addr); 893f6bed0efSShaohua Li } 894f6bed0efSShaohua Li parity_pages = 1 + !!(sh->qd_idx >= 0); 895f6bed0efSShaohua Li data_pages = write_disks - parity_pages; 896f6bed0efSShaohua Li 897f6bed0efSShaohua Li set_bit(STRIPE_LOG_TRAPPED, &sh->state); 898253f9fd4SShaohua Li /* 899253f9fd4SShaohua Li * The stripe must enter state machine again to finish the write, so 900253f9fd4SShaohua Li * don't delay. 901253f9fd4SShaohua Li */ 902253f9fd4SShaohua Li clear_bit(STRIPE_DELAYED, &sh->state); 903f6bed0efSShaohua Li atomic_inc(&sh->count); 904f6bed0efSShaohua Li 905f6bed0efSShaohua Li mutex_lock(&log->io_mutex); 906f6bed0efSShaohua Li /* meta + data */ 907f6bed0efSShaohua Li reserve = (1 + write_disks) << (PAGE_SHIFT - 9); 908f6bed0efSShaohua Li 909a39f7afdSSong Liu if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) { 910a39f7afdSSong Liu if (!r5l_has_free_space(log, reserve)) { 911a39f7afdSSong Liu r5l_add_no_space_stripe(log, sh); 912a39f7afdSSong Liu wake_reclaim = true; 9135036c390SChristoph Hellwig } else { 9145036c390SChristoph Hellwig ret = r5l_log_stripe(log, sh, data_pages, parity_pages); 9155036c390SChristoph Hellwig if (ret) { 9165036c390SChristoph Hellwig spin_lock_irq(&log->io_list_lock); 917a39f7afdSSong Liu list_add_tail(&sh->log_list, 918a39f7afdSSong Liu &log->no_mem_stripes); 9195036c390SChristoph Hellwig spin_unlock_irq(&log->io_list_lock); 920f6bed0efSShaohua Li } 9215036c390SChristoph Hellwig } 922a39f7afdSSong Liu } else { /* R5C_JOURNAL_MODE_WRITE_BACK */ 923a39f7afdSSong Liu /* 924a39f7afdSSong Liu * log space critical, do not process stripes that are 925a39f7afdSSong Liu * not in cache yet (sh->log_start == MaxSector). 926a39f7afdSSong Liu */ 927a39f7afdSSong Liu if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) && 928a39f7afdSSong Liu sh->log_start == MaxSector) { 929a39f7afdSSong Liu r5l_add_no_space_stripe(log, sh); 930a39f7afdSSong Liu wake_reclaim = true; 931a39f7afdSSong Liu reserve = 0; 932a39f7afdSSong Liu } else if (!r5l_has_free_space(log, reserve)) { 933a39f7afdSSong Liu if (sh->log_start == log->last_checkpoint) 934a39f7afdSSong Liu BUG(); 935a39f7afdSSong Liu else 936a39f7afdSSong Liu r5l_add_no_space_stripe(log, sh); 937a39f7afdSSong Liu } else { 938a39f7afdSSong Liu ret = r5l_log_stripe(log, sh, data_pages, parity_pages); 939a39f7afdSSong Liu if (ret) { 940a39f7afdSSong Liu spin_lock_irq(&log->io_list_lock); 941a39f7afdSSong Liu list_add_tail(&sh->log_list, 942a39f7afdSSong Liu &log->no_mem_stripes); 943a39f7afdSSong Liu spin_unlock_irq(&log->io_list_lock); 944a39f7afdSSong Liu } 945a39f7afdSSong Liu } 946a39f7afdSSong Liu } 947f6bed0efSShaohua Li 9485036c390SChristoph Hellwig mutex_unlock(&log->io_mutex); 949a39f7afdSSong Liu if (wake_reclaim) 950a39f7afdSSong Liu r5l_wake_reclaim(log, reserve); 951f6bed0efSShaohua Li return 0; 952f6bed0efSShaohua Li } 953f6bed0efSShaohua Li 954f6bed0efSShaohua Li void r5l_write_stripe_run(struct r5l_log *log) 955f6bed0efSShaohua Li { 956f6bed0efSShaohua Li if (!log) 957f6bed0efSShaohua Li return; 958f6bed0efSShaohua Li mutex_lock(&log->io_mutex); 959f6bed0efSShaohua Li r5l_submit_current_io(log); 960f6bed0efSShaohua Li mutex_unlock(&log->io_mutex); 961f6bed0efSShaohua Li } 962f6bed0efSShaohua Li 963828cbe98SShaohua Li int r5l_handle_flush_request(struct r5l_log *log, struct bio *bio) 964828cbe98SShaohua Li { 965828cbe98SShaohua Li if (!log) 966828cbe98SShaohua Li return -ENODEV; 9673bddb7f8SSong Liu 9683bddb7f8SSong Liu if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) { 969828cbe98SShaohua Li /* 9703bddb7f8SSong Liu * in write through (journal only) 9713bddb7f8SSong Liu * we flush log disk cache first, then write stripe data to 9723bddb7f8SSong Liu * raid disks. So if bio is finished, the log disk cache is 9733bddb7f8SSong Liu * flushed already. The recovery guarantees we can recovery 9743bddb7f8SSong Liu * the bio from log disk, so we don't need to flush again 975828cbe98SShaohua Li */ 976828cbe98SShaohua Li if (bio->bi_iter.bi_size == 0) { 977828cbe98SShaohua Li bio_endio(bio); 978828cbe98SShaohua Li return 0; 979828cbe98SShaohua Li } 9801eff9d32SJens Axboe bio->bi_opf &= ~REQ_PREFLUSH; 9813bddb7f8SSong Liu } else { 9823bddb7f8SSong Liu /* write back (with cache) */ 9833bddb7f8SSong Liu if (bio->bi_iter.bi_size == 0) { 9843bddb7f8SSong Liu mutex_lock(&log->io_mutex); 9853bddb7f8SSong Liu r5l_get_meta(log, 0); 9863bddb7f8SSong Liu bio_list_add(&log->current_io->flush_barriers, bio); 9873bddb7f8SSong Liu log->current_io->has_flush = 1; 9883bddb7f8SSong Liu log->current_io->has_null_flush = 1; 9893bddb7f8SSong Liu atomic_inc(&log->current_io->pending_stripe); 9903bddb7f8SSong Liu r5l_submit_current_io(log); 9913bddb7f8SSong Liu mutex_unlock(&log->io_mutex); 9923bddb7f8SSong Liu return 0; 9933bddb7f8SSong Liu } 9943bddb7f8SSong Liu } 995828cbe98SShaohua Li return -EAGAIN; 996828cbe98SShaohua Li } 997828cbe98SShaohua Li 998f6bed0efSShaohua Li /* This will run after log space is reclaimed */ 999f6bed0efSShaohua Li static void r5l_run_no_space_stripes(struct r5l_log *log) 1000f6bed0efSShaohua Li { 1001f6bed0efSShaohua Li struct stripe_head *sh; 1002f6bed0efSShaohua Li 1003f6bed0efSShaohua Li spin_lock(&log->no_space_stripes_lock); 1004f6bed0efSShaohua Li while (!list_empty(&log->no_space_stripes)) { 1005f6bed0efSShaohua Li sh = list_first_entry(&log->no_space_stripes, 1006f6bed0efSShaohua Li struct stripe_head, log_list); 1007f6bed0efSShaohua Li list_del_init(&sh->log_list); 1008f6bed0efSShaohua Li set_bit(STRIPE_HANDLE, &sh->state); 1009f6bed0efSShaohua Li raid5_release_stripe(sh); 1010f6bed0efSShaohua Li } 1011f6bed0efSShaohua Li spin_unlock(&log->no_space_stripes_lock); 1012f6bed0efSShaohua Li } 1013f6bed0efSShaohua Li 1014a39f7afdSSong Liu /* 1015a39f7afdSSong Liu * calculate new last_checkpoint 1016a39f7afdSSong Liu * for write through mode, returns log->next_checkpoint 1017a39f7afdSSong Liu * for write back, returns log_start of first sh in stripe_in_journal_list 1018a39f7afdSSong Liu */ 1019a39f7afdSSong Liu static sector_t r5c_calculate_new_cp(struct r5conf *conf) 1020a39f7afdSSong Liu { 1021a39f7afdSSong Liu struct stripe_head *sh; 1022a39f7afdSSong Liu struct r5l_log *log = conf->log; 1023a39f7afdSSong Liu sector_t new_cp; 1024a39f7afdSSong Liu unsigned long flags; 1025a39f7afdSSong Liu 1026a39f7afdSSong Liu if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) 1027a39f7afdSSong Liu return log->next_checkpoint; 1028a39f7afdSSong Liu 1029a39f7afdSSong Liu spin_lock_irqsave(&log->stripe_in_journal_lock, flags); 1030a39f7afdSSong Liu if (list_empty(&conf->log->stripe_in_journal_list)) { 1031a39f7afdSSong Liu /* all stripes flushed */ 1032a39f7afdSSong Liu spin_unlock(&log->stripe_in_journal_lock); 1033a39f7afdSSong Liu return log->next_checkpoint; 1034a39f7afdSSong Liu } 1035a39f7afdSSong Liu sh = list_first_entry(&conf->log->stripe_in_journal_list, 1036a39f7afdSSong Liu struct stripe_head, r5c); 1037a39f7afdSSong Liu new_cp = sh->log_start; 1038a39f7afdSSong Liu spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags); 1039a39f7afdSSong Liu return new_cp; 1040a39f7afdSSong Liu } 1041a39f7afdSSong Liu 104217036461SChristoph Hellwig static sector_t r5l_reclaimable_space(struct r5l_log *log) 104317036461SChristoph Hellwig { 1044a39f7afdSSong Liu struct r5conf *conf = log->rdev->mddev->private; 1045a39f7afdSSong Liu 104617036461SChristoph Hellwig return r5l_ring_distance(log, log->last_checkpoint, 1047a39f7afdSSong Liu r5c_calculate_new_cp(conf)); 104817036461SChristoph Hellwig } 104917036461SChristoph Hellwig 10505036c390SChristoph Hellwig static void r5l_run_no_mem_stripe(struct r5l_log *log) 10515036c390SChristoph Hellwig { 10525036c390SChristoph Hellwig struct stripe_head *sh; 10535036c390SChristoph Hellwig 10545036c390SChristoph Hellwig assert_spin_locked(&log->io_list_lock); 10555036c390SChristoph Hellwig 10565036c390SChristoph Hellwig if (!list_empty(&log->no_mem_stripes)) { 10575036c390SChristoph Hellwig sh = list_first_entry(&log->no_mem_stripes, 10585036c390SChristoph Hellwig struct stripe_head, log_list); 10595036c390SChristoph Hellwig list_del_init(&sh->log_list); 10605036c390SChristoph Hellwig set_bit(STRIPE_HANDLE, &sh->state); 10615036c390SChristoph Hellwig raid5_release_stripe(sh); 10625036c390SChristoph Hellwig } 10635036c390SChristoph Hellwig } 10645036c390SChristoph Hellwig 106504732f74SChristoph Hellwig static bool r5l_complete_finished_ios(struct r5l_log *log) 106617036461SChristoph Hellwig { 106717036461SChristoph Hellwig struct r5l_io_unit *io, *next; 106817036461SChristoph Hellwig bool found = false; 106917036461SChristoph Hellwig 107017036461SChristoph Hellwig assert_spin_locked(&log->io_list_lock); 107117036461SChristoph Hellwig 107204732f74SChristoph Hellwig list_for_each_entry_safe(io, next, &log->finished_ios, log_sibling) { 107317036461SChristoph Hellwig /* don't change list order */ 107417036461SChristoph Hellwig if (io->state < IO_UNIT_STRIPE_END) 107517036461SChristoph Hellwig break; 107617036461SChristoph Hellwig 107717036461SChristoph Hellwig log->next_checkpoint = io->log_start; 107817036461SChristoph Hellwig log->next_cp_seq = io->seq; 107917036461SChristoph Hellwig 108017036461SChristoph Hellwig list_del(&io->log_sibling); 10815036c390SChristoph Hellwig mempool_free(io, log->io_pool); 10825036c390SChristoph Hellwig r5l_run_no_mem_stripe(log); 108317036461SChristoph Hellwig 108417036461SChristoph Hellwig found = true; 108517036461SChristoph Hellwig } 108617036461SChristoph Hellwig 108717036461SChristoph Hellwig return found; 108817036461SChristoph Hellwig } 108917036461SChristoph Hellwig 1090509ffec7SChristoph Hellwig static void __r5l_stripe_write_finished(struct r5l_io_unit *io) 1091509ffec7SChristoph Hellwig { 1092509ffec7SChristoph Hellwig struct r5l_log *log = io->log; 1093a39f7afdSSong Liu struct r5conf *conf = log->rdev->mddev->private; 1094509ffec7SChristoph Hellwig unsigned long flags; 1095509ffec7SChristoph Hellwig 1096509ffec7SChristoph Hellwig spin_lock_irqsave(&log->io_list_lock, flags); 1097509ffec7SChristoph Hellwig __r5l_set_io_unit_state(io, IO_UNIT_STRIPE_END); 109817036461SChristoph Hellwig 109904732f74SChristoph Hellwig if (!r5l_complete_finished_ios(log)) { 110085f2f9a4SShaohua Li spin_unlock_irqrestore(&log->io_list_lock, flags); 110185f2f9a4SShaohua Li return; 110285f2f9a4SShaohua Li } 1103509ffec7SChristoph Hellwig 1104a39f7afdSSong Liu if (r5l_reclaimable_space(log) > log->max_free_space || 1105a39f7afdSSong Liu test_bit(R5C_LOG_TIGHT, &conf->cache_state)) 1106509ffec7SChristoph Hellwig r5l_wake_reclaim(log, 0); 1107509ffec7SChristoph Hellwig 1108509ffec7SChristoph Hellwig spin_unlock_irqrestore(&log->io_list_lock, flags); 1109509ffec7SChristoph Hellwig wake_up(&log->iounit_wait); 1110509ffec7SChristoph Hellwig } 1111509ffec7SChristoph Hellwig 11120576b1c6SShaohua Li void r5l_stripe_write_finished(struct stripe_head *sh) 11130576b1c6SShaohua Li { 11140576b1c6SShaohua Li struct r5l_io_unit *io; 11150576b1c6SShaohua Li 11160576b1c6SShaohua Li io = sh->log_io; 11170576b1c6SShaohua Li sh->log_io = NULL; 11180576b1c6SShaohua Li 1119509ffec7SChristoph Hellwig if (io && atomic_dec_and_test(&io->pending_stripe)) 1120509ffec7SChristoph Hellwig __r5l_stripe_write_finished(io); 11210576b1c6SShaohua Li } 11220576b1c6SShaohua Li 1123a8c34f91SShaohua Li static void r5l_log_flush_endio(struct bio *bio) 1124a8c34f91SShaohua Li { 1125a8c34f91SShaohua Li struct r5l_log *log = container_of(bio, struct r5l_log, 1126a8c34f91SShaohua Li flush_bio); 1127a8c34f91SShaohua Li unsigned long flags; 1128a8c34f91SShaohua Li struct r5l_io_unit *io; 1129a8c34f91SShaohua Li 11306e74a9cfSShaohua Li if (bio->bi_error) 11316e74a9cfSShaohua Li md_error(log->rdev->mddev, log->rdev); 11326e74a9cfSShaohua Li 1133a8c34f91SShaohua Li spin_lock_irqsave(&log->io_list_lock, flags); 1134d8858f43SChristoph Hellwig list_for_each_entry(io, &log->flushing_ios, log_sibling) 1135d8858f43SChristoph Hellwig r5l_io_run_stripes(io); 113604732f74SChristoph Hellwig list_splice_tail_init(&log->flushing_ios, &log->finished_ios); 1137a8c34f91SShaohua Li spin_unlock_irqrestore(&log->io_list_lock, flags); 1138a8c34f91SShaohua Li } 1139a8c34f91SShaohua Li 11400576b1c6SShaohua Li /* 11410576b1c6SShaohua Li * Starting dispatch IO to raid. 11420576b1c6SShaohua Li * io_unit(meta) consists of a log. There is one situation we want to avoid. A 11430576b1c6SShaohua Li * broken meta in the middle of a log causes recovery can't find meta at the 11440576b1c6SShaohua Li * head of log. If operations require meta at the head persistent in log, we 11450576b1c6SShaohua Li * must make sure meta before it persistent in log too. A case is: 11460576b1c6SShaohua Li * 11470576b1c6SShaohua Li * stripe data/parity is in log, we start write stripe to raid disks. stripe 11480576b1c6SShaohua Li * data/parity must be persistent in log before we do the write to raid disks. 11490576b1c6SShaohua Li * 11500576b1c6SShaohua Li * The solution is we restrictly maintain io_unit list order. In this case, we 11510576b1c6SShaohua Li * only write stripes of an io_unit to raid disks till the io_unit is the first 11520576b1c6SShaohua Li * one whose data/parity is in log. 11530576b1c6SShaohua Li */ 11540576b1c6SShaohua Li void r5l_flush_stripe_to_raid(struct r5l_log *log) 11550576b1c6SShaohua Li { 1156a8c34f91SShaohua Li bool do_flush; 115756fef7c6SChristoph Hellwig 115856fef7c6SChristoph Hellwig if (!log || !log->need_cache_flush) 11590576b1c6SShaohua Li return; 11600576b1c6SShaohua Li 1161a8c34f91SShaohua Li spin_lock_irq(&log->io_list_lock); 1162a8c34f91SShaohua Li /* flush bio is running */ 1163a8c34f91SShaohua Li if (!list_empty(&log->flushing_ios)) { 1164a8c34f91SShaohua Li spin_unlock_irq(&log->io_list_lock); 11650576b1c6SShaohua Li return; 11660576b1c6SShaohua Li } 1167a8c34f91SShaohua Li list_splice_tail_init(&log->io_end_ios, &log->flushing_ios); 1168a8c34f91SShaohua Li do_flush = !list_empty(&log->flushing_ios); 11690576b1c6SShaohua Li spin_unlock_irq(&log->io_list_lock); 1170a8c34f91SShaohua Li 1171a8c34f91SShaohua Li if (!do_flush) 1172a8c34f91SShaohua Li return; 1173a8c34f91SShaohua Li bio_reset(&log->flush_bio); 1174a8c34f91SShaohua Li log->flush_bio.bi_bdev = log->rdev->bdev; 1175a8c34f91SShaohua Li log->flush_bio.bi_end_io = r5l_log_flush_endio; 1176796a5cf0SMike Christie bio_set_op_attrs(&log->flush_bio, REQ_OP_WRITE, WRITE_FLUSH); 11774e49ea4aSMike Christie submit_bio(&log->flush_bio); 11780576b1c6SShaohua Li } 11790576b1c6SShaohua Li 11800576b1c6SShaohua Li static void r5l_write_super(struct r5l_log *log, sector_t cp); 11814b482044SShaohua Li static void r5l_write_super_and_discard_space(struct r5l_log *log, 11824b482044SShaohua Li sector_t end) 11834b482044SShaohua Li { 11844b482044SShaohua Li struct block_device *bdev = log->rdev->bdev; 11854b482044SShaohua Li struct mddev *mddev; 11864b482044SShaohua Li 11874b482044SShaohua Li r5l_write_super(log, end); 11884b482044SShaohua Li 11894b482044SShaohua Li if (!blk_queue_discard(bdev_get_queue(bdev))) 11904b482044SShaohua Li return; 11914b482044SShaohua Li 11924b482044SShaohua Li mddev = log->rdev->mddev; 11934b482044SShaohua Li /* 11948e018c21SShaohua Li * Discard could zero data, so before discard we must make sure 11958e018c21SShaohua Li * superblock is updated to new log tail. Updating superblock (either 11968e018c21SShaohua Li * directly call md_update_sb() or depend on md thread) must hold 11978e018c21SShaohua Li * reconfig mutex. On the other hand, raid5_quiesce is called with 11988e018c21SShaohua Li * reconfig_mutex hold. The first step of raid5_quiesce() is waitting 11998e018c21SShaohua Li * for all IO finish, hence waitting for reclaim thread, while reclaim 12008e018c21SShaohua Li * thread is calling this function and waitting for reconfig mutex. So 12018e018c21SShaohua Li * there is a deadlock. We workaround this issue with a trylock. 12028e018c21SShaohua Li * FIXME: we could miss discard if we can't take reconfig mutex 12034b482044SShaohua Li */ 120485ad1d13SGuoqing Jiang set_mask_bits(&mddev->flags, 0, 120585ad1d13SGuoqing Jiang BIT(MD_CHANGE_DEVS) | BIT(MD_CHANGE_PENDING)); 12068e018c21SShaohua Li if (!mddev_trylock(mddev)) 12078e018c21SShaohua Li return; 12084b482044SShaohua Li md_update_sb(mddev, 1); 12098e018c21SShaohua Li mddev_unlock(mddev); 12104b482044SShaohua Li 12116e74a9cfSShaohua Li /* discard IO error really doesn't matter, ignore it */ 12124b482044SShaohua Li if (log->last_checkpoint < end) { 12134b482044SShaohua Li blkdev_issue_discard(bdev, 12144b482044SShaohua Li log->last_checkpoint + log->rdev->data_offset, 12154b482044SShaohua Li end - log->last_checkpoint, GFP_NOIO, 0); 12164b482044SShaohua Li } else { 12174b482044SShaohua Li blkdev_issue_discard(bdev, 12184b482044SShaohua Li log->last_checkpoint + log->rdev->data_offset, 12194b482044SShaohua Li log->device_size - log->last_checkpoint, 12204b482044SShaohua Li GFP_NOIO, 0); 12214b482044SShaohua Li blkdev_issue_discard(bdev, log->rdev->data_offset, end, 12224b482044SShaohua Li GFP_NOIO, 0); 12234b482044SShaohua Li } 12244b482044SShaohua Li } 12254b482044SShaohua Li 1226a39f7afdSSong Liu /* 1227a39f7afdSSong Liu * r5c_flush_stripe moves stripe from cached list to handle_list. When called, 1228a39f7afdSSong Liu * the stripe must be on r5c_cached_full_stripes or r5c_cached_partial_stripes. 1229a39f7afdSSong Liu * 1230a39f7afdSSong Liu * must hold conf->device_lock 1231a39f7afdSSong Liu */ 1232a39f7afdSSong Liu static void r5c_flush_stripe(struct r5conf *conf, struct stripe_head *sh) 1233a39f7afdSSong Liu { 1234a39f7afdSSong Liu BUG_ON(list_empty(&sh->lru)); 1235a39f7afdSSong Liu BUG_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state)); 1236a39f7afdSSong Liu BUG_ON(test_bit(STRIPE_HANDLE, &sh->state)); 1237a39f7afdSSong Liu 1238a39f7afdSSong Liu /* 1239a39f7afdSSong Liu * The stripe is not ON_RELEASE_LIST, so it is safe to call 1240a39f7afdSSong Liu * raid5_release_stripe() while holding conf->device_lock 1241a39f7afdSSong Liu */ 1242a39f7afdSSong Liu BUG_ON(test_bit(STRIPE_ON_RELEASE_LIST, &sh->state)); 1243a39f7afdSSong Liu assert_spin_locked(&conf->device_lock); 1244a39f7afdSSong Liu 1245a39f7afdSSong Liu list_del_init(&sh->lru); 1246a39f7afdSSong Liu atomic_inc(&sh->count); 1247a39f7afdSSong Liu 1248a39f7afdSSong Liu set_bit(STRIPE_HANDLE, &sh->state); 1249a39f7afdSSong Liu atomic_inc(&conf->active_stripes); 1250a39f7afdSSong Liu r5c_make_stripe_write_out(sh); 1251a39f7afdSSong Liu 1252a39f7afdSSong Liu if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) 1253a39f7afdSSong Liu atomic_inc(&conf->preread_active_stripes); 1254a39f7afdSSong Liu raid5_release_stripe(sh); 1255a39f7afdSSong Liu } 1256a39f7afdSSong Liu 1257a39f7afdSSong Liu /* 1258a39f7afdSSong Liu * if num == 0, flush all full stripes 1259a39f7afdSSong Liu * if num > 0, flush all full stripes. If less than num full stripes are 1260a39f7afdSSong Liu * flushed, flush some partial stripes until totally num stripes are 1261a39f7afdSSong Liu * flushed or there is no more cached stripes. 1262a39f7afdSSong Liu */ 1263a39f7afdSSong Liu void r5c_flush_cache(struct r5conf *conf, int num) 1264a39f7afdSSong Liu { 1265a39f7afdSSong Liu int count; 1266a39f7afdSSong Liu struct stripe_head *sh, *next; 1267a39f7afdSSong Liu 1268a39f7afdSSong Liu assert_spin_locked(&conf->device_lock); 1269a39f7afdSSong Liu if (!conf->log) 1270a39f7afdSSong Liu return; 1271a39f7afdSSong Liu 1272a39f7afdSSong Liu count = 0; 1273a39f7afdSSong Liu list_for_each_entry_safe(sh, next, &conf->r5c_full_stripe_list, lru) { 1274a39f7afdSSong Liu r5c_flush_stripe(conf, sh); 1275a39f7afdSSong Liu count++; 1276a39f7afdSSong Liu } 1277a39f7afdSSong Liu 1278a39f7afdSSong Liu if (count >= num) 1279a39f7afdSSong Liu return; 1280a39f7afdSSong Liu list_for_each_entry_safe(sh, next, 1281a39f7afdSSong Liu &conf->r5c_partial_stripe_list, lru) { 1282a39f7afdSSong Liu r5c_flush_stripe(conf, sh); 1283a39f7afdSSong Liu if (++count >= num) 1284a39f7afdSSong Liu break; 1285a39f7afdSSong Liu } 1286a39f7afdSSong Liu } 1287a39f7afdSSong Liu 1288a39f7afdSSong Liu static void r5c_do_reclaim(struct r5conf *conf) 1289a39f7afdSSong Liu { 1290a39f7afdSSong Liu struct r5l_log *log = conf->log; 1291a39f7afdSSong Liu struct stripe_head *sh; 1292a39f7afdSSong Liu int count = 0; 1293a39f7afdSSong Liu unsigned long flags; 1294a39f7afdSSong Liu int total_cached; 1295a39f7afdSSong Liu int stripes_to_flush; 1296a39f7afdSSong Liu 1297a39f7afdSSong Liu if (!r5c_is_writeback(log)) 1298a39f7afdSSong Liu return; 1299a39f7afdSSong Liu 1300a39f7afdSSong Liu total_cached = atomic_read(&conf->r5c_cached_partial_stripes) + 1301a39f7afdSSong Liu atomic_read(&conf->r5c_cached_full_stripes); 1302a39f7afdSSong Liu 1303a39f7afdSSong Liu if (total_cached > conf->min_nr_stripes * 3 / 4 || 1304a39f7afdSSong Liu atomic_read(&conf->empty_inactive_list_nr) > 0) 1305a39f7afdSSong Liu /* 1306a39f7afdSSong Liu * if stripe cache pressure high, flush all full stripes and 1307a39f7afdSSong Liu * some partial stripes 1308a39f7afdSSong Liu */ 1309a39f7afdSSong Liu stripes_to_flush = R5C_RECLAIM_STRIPE_GROUP; 1310a39f7afdSSong Liu else if (total_cached > conf->min_nr_stripes * 1 / 2 || 1311a39f7afdSSong Liu atomic_read(&conf->r5c_cached_full_stripes) > 1312a39f7afdSSong Liu R5C_FULL_STRIPE_FLUSH_BATCH) 1313a39f7afdSSong Liu /* 1314a39f7afdSSong Liu * if stripe cache pressure moderate, or if there is many full 1315a39f7afdSSong Liu * stripes,flush all full stripes 1316a39f7afdSSong Liu */ 1317a39f7afdSSong Liu stripes_to_flush = 0; 1318a39f7afdSSong Liu else 1319a39f7afdSSong Liu /* no need to flush */ 1320a39f7afdSSong Liu stripes_to_flush = -1; 1321a39f7afdSSong Liu 1322a39f7afdSSong Liu if (stripes_to_flush >= 0) { 1323a39f7afdSSong Liu spin_lock_irqsave(&conf->device_lock, flags); 1324a39f7afdSSong Liu r5c_flush_cache(conf, stripes_to_flush); 1325a39f7afdSSong Liu spin_unlock_irqrestore(&conf->device_lock, flags); 1326a39f7afdSSong Liu } 1327a39f7afdSSong Liu 1328a39f7afdSSong Liu /* if log space is tight, flush stripes on stripe_in_journal_list */ 1329a39f7afdSSong Liu if (test_bit(R5C_LOG_TIGHT, &conf->cache_state)) { 1330a39f7afdSSong Liu spin_lock_irqsave(&log->stripe_in_journal_lock, flags); 1331a39f7afdSSong Liu spin_lock(&conf->device_lock); 1332a39f7afdSSong Liu list_for_each_entry(sh, &log->stripe_in_journal_list, r5c) { 1333a39f7afdSSong Liu /* 1334a39f7afdSSong Liu * stripes on stripe_in_journal_list could be in any 1335a39f7afdSSong Liu * state of the stripe_cache state machine. In this 1336a39f7afdSSong Liu * case, we only want to flush stripe on 1337a39f7afdSSong Liu * r5c_cached_full/partial_stripes. The following 1338a39f7afdSSong Liu * condition makes sure the stripe is on one of the 1339a39f7afdSSong Liu * two lists. 1340a39f7afdSSong Liu */ 1341a39f7afdSSong Liu if (!list_empty(&sh->lru) && 1342a39f7afdSSong Liu !test_bit(STRIPE_HANDLE, &sh->state) && 1343a39f7afdSSong Liu atomic_read(&sh->count) == 0) { 1344a39f7afdSSong Liu r5c_flush_stripe(conf, sh); 1345a39f7afdSSong Liu } 1346a39f7afdSSong Liu if (count++ >= R5C_RECLAIM_STRIPE_GROUP) 1347a39f7afdSSong Liu break; 1348a39f7afdSSong Liu } 1349a39f7afdSSong Liu spin_unlock(&conf->device_lock); 1350a39f7afdSSong Liu spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags); 1351a39f7afdSSong Liu } 1352a39f7afdSSong Liu md_wakeup_thread(conf->mddev->thread); 1353a39f7afdSSong Liu } 1354a39f7afdSSong Liu 13550576b1c6SShaohua Li static void r5l_do_reclaim(struct r5l_log *log) 13560576b1c6SShaohua Li { 1357a39f7afdSSong Liu struct r5conf *conf = log->rdev->mddev->private; 13580576b1c6SShaohua Li sector_t reclaim_target = xchg(&log->reclaim_target, 0); 135917036461SChristoph Hellwig sector_t reclaimable; 136017036461SChristoph Hellwig sector_t next_checkpoint; 1361a39f7afdSSong Liu bool write_super; 13620576b1c6SShaohua Li 13630576b1c6SShaohua Li spin_lock_irq(&log->io_list_lock); 1364a39f7afdSSong Liu write_super = r5l_reclaimable_space(log) > log->max_free_space || 1365a39f7afdSSong Liu reclaim_target != 0 || !list_empty(&log->no_space_stripes); 13660576b1c6SShaohua Li /* 13670576b1c6SShaohua Li * move proper io_unit to reclaim list. We should not change the order. 13680576b1c6SShaohua Li * reclaimable/unreclaimable io_unit can be mixed in the list, we 13690576b1c6SShaohua Li * shouldn't reuse space of an unreclaimable io_unit 13700576b1c6SShaohua Li */ 13710576b1c6SShaohua Li while (1) { 137217036461SChristoph Hellwig reclaimable = r5l_reclaimable_space(log); 137317036461SChristoph Hellwig if (reclaimable >= reclaim_target || 13740576b1c6SShaohua Li (list_empty(&log->running_ios) && 13750576b1c6SShaohua Li list_empty(&log->io_end_ios) && 1376a8c34f91SShaohua Li list_empty(&log->flushing_ios) && 137704732f74SChristoph Hellwig list_empty(&log->finished_ios))) 13780576b1c6SShaohua Li break; 13790576b1c6SShaohua Li 138017036461SChristoph Hellwig md_wakeup_thread(log->rdev->mddev->thread); 138117036461SChristoph Hellwig wait_event_lock_irq(log->iounit_wait, 138217036461SChristoph Hellwig r5l_reclaimable_space(log) > reclaimable, 138317036461SChristoph Hellwig log->io_list_lock); 13840576b1c6SShaohua Li } 138517036461SChristoph Hellwig 1386a39f7afdSSong Liu next_checkpoint = r5c_calculate_new_cp(conf); 13870576b1c6SShaohua Li spin_unlock_irq(&log->io_list_lock); 13880576b1c6SShaohua Li 138917036461SChristoph Hellwig BUG_ON(reclaimable < 0); 1390a39f7afdSSong Liu 1391a39f7afdSSong Liu if (reclaimable == 0 || !write_super) 13920576b1c6SShaohua Li return; 13930576b1c6SShaohua Li 13940576b1c6SShaohua Li /* 13950576b1c6SShaohua Li * write_super will flush cache of each raid disk. We must write super 13960576b1c6SShaohua Li * here, because the log area might be reused soon and we don't want to 13970576b1c6SShaohua Li * confuse recovery 13980576b1c6SShaohua Li */ 13994b482044SShaohua Li r5l_write_super_and_discard_space(log, next_checkpoint); 14000576b1c6SShaohua Li 14010576b1c6SShaohua Li mutex_lock(&log->io_mutex); 140217036461SChristoph Hellwig log->last_checkpoint = next_checkpoint; 1403a39f7afdSSong Liu r5c_update_log_state(log); 14040576b1c6SShaohua Li mutex_unlock(&log->io_mutex); 14050576b1c6SShaohua Li 140617036461SChristoph Hellwig r5l_run_no_space_stripes(log); 14070576b1c6SShaohua Li } 14080576b1c6SShaohua Li 14090576b1c6SShaohua Li static void r5l_reclaim_thread(struct md_thread *thread) 14100576b1c6SShaohua Li { 14110576b1c6SShaohua Li struct mddev *mddev = thread->mddev; 14120576b1c6SShaohua Li struct r5conf *conf = mddev->private; 14130576b1c6SShaohua Li struct r5l_log *log = conf->log; 14140576b1c6SShaohua Li 14150576b1c6SShaohua Li if (!log) 14160576b1c6SShaohua Li return; 1417a39f7afdSSong Liu r5c_do_reclaim(conf); 14180576b1c6SShaohua Li r5l_do_reclaim(log); 14190576b1c6SShaohua Li } 14200576b1c6SShaohua Li 1421a39f7afdSSong Liu void r5l_wake_reclaim(struct r5l_log *log, sector_t space) 1422f6bed0efSShaohua Li { 14230576b1c6SShaohua Li unsigned long target; 14240576b1c6SShaohua Li unsigned long new = (unsigned long)space; /* overflow in theory */ 14250576b1c6SShaohua Li 1426a39f7afdSSong Liu if (!log) 1427a39f7afdSSong Liu return; 14280576b1c6SShaohua Li do { 14290576b1c6SShaohua Li target = log->reclaim_target; 14300576b1c6SShaohua Li if (new < target) 14310576b1c6SShaohua Li return; 14320576b1c6SShaohua Li } while (cmpxchg(&log->reclaim_target, target, new) != target); 14330576b1c6SShaohua Li md_wakeup_thread(log->reclaim_thread); 1434f6bed0efSShaohua Li } 1435f6bed0efSShaohua Li 1436e6c033f7SShaohua Li void r5l_quiesce(struct r5l_log *log, int state) 1437e6c033f7SShaohua Li { 14384b482044SShaohua Li struct mddev *mddev; 1439e6c033f7SShaohua Li if (!log || state == 2) 1440e6c033f7SShaohua Li return; 1441ce1ccd07SShaohua Li if (state == 0) 1442ce1ccd07SShaohua Li kthread_unpark(log->reclaim_thread->tsk); 1443ce1ccd07SShaohua Li else if (state == 1) { 14444b482044SShaohua Li /* make sure r5l_write_super_and_discard_space exits */ 14454b482044SShaohua Li mddev = log->rdev->mddev; 14464b482044SShaohua Li wake_up(&mddev->sb_wait); 1447ce1ccd07SShaohua Li kthread_park(log->reclaim_thread->tsk); 1448a39f7afdSSong Liu r5l_wake_reclaim(log, MaxSector); 1449e6c033f7SShaohua Li r5l_do_reclaim(log); 1450e6c033f7SShaohua Li } 1451e6c033f7SShaohua Li } 1452e6c033f7SShaohua Li 14536e74a9cfSShaohua Li bool r5l_log_disk_error(struct r5conf *conf) 14546e74a9cfSShaohua Li { 1455f6b6ec5cSShaohua Li struct r5l_log *log; 1456f6b6ec5cSShaohua Li bool ret; 14577dde2ad3SShaohua Li /* don't allow write if journal disk is missing */ 1458f6b6ec5cSShaohua Li rcu_read_lock(); 1459f6b6ec5cSShaohua Li log = rcu_dereference(conf->log); 1460f6b6ec5cSShaohua Li 1461f6b6ec5cSShaohua Li if (!log) 1462f6b6ec5cSShaohua Li ret = test_bit(MD_HAS_JOURNAL, &conf->mddev->flags); 1463f6b6ec5cSShaohua Li else 1464f6b6ec5cSShaohua Li ret = test_bit(Faulty, &log->rdev->flags); 1465f6b6ec5cSShaohua Li rcu_read_unlock(); 1466f6b6ec5cSShaohua Li return ret; 14676e74a9cfSShaohua Li } 14686e74a9cfSShaohua Li 1469355810d1SShaohua Li struct r5l_recovery_ctx { 1470355810d1SShaohua Li struct page *meta_page; /* current meta */ 1471355810d1SShaohua Li sector_t meta_total_blocks; /* total size of current meta and data */ 1472355810d1SShaohua Li sector_t pos; /* recovery position */ 1473355810d1SShaohua Li u64 seq; /* recovery position seq */ 1474b4c625c6SSong Liu int data_parity_stripes; /* number of data_parity stripes */ 1475b4c625c6SSong Liu int data_only_stripes; /* number of data_only stripes */ 1476b4c625c6SSong Liu struct list_head cached_list; 1477355810d1SShaohua Li }; 1478355810d1SShaohua Li 14799ed988f5SSong Liu static int r5l_recovery_read_meta_block(struct r5l_log *log, 1480355810d1SShaohua Li struct r5l_recovery_ctx *ctx) 1481355810d1SShaohua Li { 1482355810d1SShaohua Li struct page *page = ctx->meta_page; 1483355810d1SShaohua Li struct r5l_meta_block *mb; 1484355810d1SShaohua Li u32 crc, stored_crc; 1485355810d1SShaohua Li 1486796a5cf0SMike Christie if (!sync_page_io(log->rdev, ctx->pos, PAGE_SIZE, page, REQ_OP_READ, 0, 1487796a5cf0SMike Christie false)) 1488355810d1SShaohua Li return -EIO; 1489355810d1SShaohua Li 1490355810d1SShaohua Li mb = page_address(page); 1491355810d1SShaohua Li stored_crc = le32_to_cpu(mb->checksum); 1492355810d1SShaohua Li mb->checksum = 0; 1493355810d1SShaohua Li 1494355810d1SShaohua Li if (le32_to_cpu(mb->magic) != R5LOG_MAGIC || 1495355810d1SShaohua Li le64_to_cpu(mb->seq) != ctx->seq || 1496355810d1SShaohua Li mb->version != R5LOG_VERSION || 1497355810d1SShaohua Li le64_to_cpu(mb->position) != ctx->pos) 1498355810d1SShaohua Li return -EINVAL; 1499355810d1SShaohua Li 15005cb2fbd6SShaohua Li crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE); 1501355810d1SShaohua Li if (stored_crc != crc) 1502355810d1SShaohua Li return -EINVAL; 1503355810d1SShaohua Li 1504355810d1SShaohua Li if (le32_to_cpu(mb->meta_size) > PAGE_SIZE) 1505355810d1SShaohua Li return -EINVAL; 1506355810d1SShaohua Li 1507355810d1SShaohua Li ctx->meta_total_blocks = BLOCK_SECTORS; 1508355810d1SShaohua Li 1509355810d1SShaohua Li return 0; 1510355810d1SShaohua Li } 1511355810d1SShaohua Li 15129ed988f5SSong Liu static void 15139ed988f5SSong Liu r5l_recovery_create_empty_meta_block(struct r5l_log *log, 15149ed988f5SSong Liu struct page *page, 15159ed988f5SSong Liu sector_t pos, u64 seq) 1516355810d1SShaohua Li { 1517355810d1SShaohua Li struct r5l_meta_block *mb; 1518355810d1SShaohua Li u32 crc; 1519355810d1SShaohua Li 1520355810d1SShaohua Li mb = page_address(page); 15219ed988f5SSong Liu clear_page(mb); 1522355810d1SShaohua Li mb->magic = cpu_to_le32(R5LOG_MAGIC); 1523355810d1SShaohua Li mb->version = R5LOG_VERSION; 1524355810d1SShaohua Li mb->meta_size = cpu_to_le32(sizeof(struct r5l_meta_block)); 1525355810d1SShaohua Li mb->seq = cpu_to_le64(seq); 1526355810d1SShaohua Li mb->position = cpu_to_le64(pos); 15275cb2fbd6SShaohua Li crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE); 1528355810d1SShaohua Li mb->checksum = cpu_to_le32(crc); 15299ed988f5SSong Liu } 1530355810d1SShaohua Li 15319ed988f5SSong Liu static int r5l_log_write_empty_meta_block(struct r5l_log *log, sector_t pos, 15329ed988f5SSong Liu u64 seq) 15339ed988f5SSong Liu { 15349ed988f5SSong Liu struct page *page; 15359ed988f5SSong Liu 15369ed988f5SSong Liu page = alloc_page(GFP_KERNEL); 15379ed988f5SSong Liu if (!page) 15389ed988f5SSong Liu return -ENOMEM; 15399ed988f5SSong Liu r5l_recovery_create_empty_meta_block(log, page, pos, seq); 1540796a5cf0SMike Christie if (!sync_page_io(log->rdev, pos, PAGE_SIZE, page, REQ_OP_WRITE, 1541796a5cf0SMike Christie WRITE_FUA, false)) { 1542355810d1SShaohua Li __free_page(page); 1543355810d1SShaohua Li return -EIO; 1544355810d1SShaohua Li } 1545355810d1SShaohua Li __free_page(page); 1546355810d1SShaohua Li return 0; 1547355810d1SShaohua Li } 1548355810d1SShaohua Li 1549b4c625c6SSong Liu /* 1550b4c625c6SSong Liu * r5l_recovery_load_data and r5l_recovery_load_parity uses flag R5_Wantwrite 1551b4c625c6SSong Liu * to mark valid (potentially not flushed) data in the journal. 1552b4c625c6SSong Liu * 1553b4c625c6SSong Liu * We already verified checksum in r5l_recovery_verify_data_checksum_for_mb, 1554b4c625c6SSong Liu * so there should not be any mismatch here. 1555b4c625c6SSong Liu */ 1556b4c625c6SSong Liu static void r5l_recovery_load_data(struct r5l_log *log, 1557b4c625c6SSong Liu struct stripe_head *sh, 1558b4c625c6SSong Liu struct r5l_recovery_ctx *ctx, 1559b4c625c6SSong Liu struct r5l_payload_data_parity *payload, 1560b4c625c6SSong Liu sector_t log_offset) 1561b4c625c6SSong Liu { 1562b4c625c6SSong Liu struct mddev *mddev = log->rdev->mddev; 1563b4c625c6SSong Liu struct r5conf *conf = mddev->private; 1564b4c625c6SSong Liu int dd_idx; 1565b4c625c6SSong Liu 1566b4c625c6SSong Liu raid5_compute_sector(conf, 1567b4c625c6SSong Liu le64_to_cpu(payload->location), 0, 1568b4c625c6SSong Liu &dd_idx, sh); 1569b4c625c6SSong Liu sync_page_io(log->rdev, log_offset, PAGE_SIZE, 1570b4c625c6SSong Liu sh->dev[dd_idx].page, REQ_OP_READ, 0, false); 1571b4c625c6SSong Liu sh->dev[dd_idx].log_checksum = 1572b4c625c6SSong Liu le32_to_cpu(payload->checksum[0]); 1573b4c625c6SSong Liu ctx->meta_total_blocks += BLOCK_SECTORS; 1574b4c625c6SSong Liu 1575b4c625c6SSong Liu set_bit(R5_Wantwrite, &sh->dev[dd_idx].flags); 1576b4c625c6SSong Liu set_bit(STRIPE_R5C_CACHING, &sh->state); 1577b4c625c6SSong Liu } 1578b4c625c6SSong Liu 1579b4c625c6SSong Liu static void r5l_recovery_load_parity(struct r5l_log *log, 1580b4c625c6SSong Liu struct stripe_head *sh, 1581b4c625c6SSong Liu struct r5l_recovery_ctx *ctx, 1582b4c625c6SSong Liu struct r5l_payload_data_parity *payload, 1583b4c625c6SSong Liu sector_t log_offset) 1584b4c625c6SSong Liu { 1585b4c625c6SSong Liu struct mddev *mddev = log->rdev->mddev; 1586b4c625c6SSong Liu struct r5conf *conf = mddev->private; 1587b4c625c6SSong Liu 1588b4c625c6SSong Liu ctx->meta_total_blocks += BLOCK_SECTORS * conf->max_degraded; 1589b4c625c6SSong Liu sync_page_io(log->rdev, log_offset, PAGE_SIZE, 1590b4c625c6SSong Liu sh->dev[sh->pd_idx].page, REQ_OP_READ, 0, false); 1591b4c625c6SSong Liu sh->dev[sh->pd_idx].log_checksum = 1592b4c625c6SSong Liu le32_to_cpu(payload->checksum[0]); 1593b4c625c6SSong Liu set_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags); 1594b4c625c6SSong Liu 1595b4c625c6SSong Liu if (sh->qd_idx >= 0) { 1596b4c625c6SSong Liu sync_page_io(log->rdev, 1597b4c625c6SSong Liu r5l_ring_add(log, log_offset, BLOCK_SECTORS), 1598b4c625c6SSong Liu PAGE_SIZE, sh->dev[sh->qd_idx].page, 1599b4c625c6SSong Liu REQ_OP_READ, 0, false); 1600b4c625c6SSong Liu sh->dev[sh->qd_idx].log_checksum = 1601b4c625c6SSong Liu le32_to_cpu(payload->checksum[1]); 1602b4c625c6SSong Liu set_bit(R5_Wantwrite, &sh->dev[sh->qd_idx].flags); 1603b4c625c6SSong Liu } 1604b4c625c6SSong Liu clear_bit(STRIPE_R5C_CACHING, &sh->state); 1605b4c625c6SSong Liu } 1606b4c625c6SSong Liu 1607b4c625c6SSong Liu static void r5l_recovery_reset_stripe(struct stripe_head *sh) 1608b4c625c6SSong Liu { 1609b4c625c6SSong Liu int i; 1610b4c625c6SSong Liu 1611b4c625c6SSong Liu sh->state = 0; 1612b4c625c6SSong Liu sh->log_start = MaxSector; 1613b4c625c6SSong Liu for (i = sh->disks; i--; ) 1614b4c625c6SSong Liu sh->dev[i].flags = 0; 1615b4c625c6SSong Liu } 1616b4c625c6SSong Liu 1617b4c625c6SSong Liu static void 1618b4c625c6SSong Liu r5l_recovery_replay_one_stripe(struct r5conf *conf, 1619b4c625c6SSong Liu struct stripe_head *sh, 1620b4c625c6SSong Liu struct r5l_recovery_ctx *ctx) 1621b4c625c6SSong Liu { 1622b4c625c6SSong Liu struct md_rdev *rdev, *rrdev; 1623b4c625c6SSong Liu int disk_index; 1624b4c625c6SSong Liu int data_count = 0; 1625b4c625c6SSong Liu 1626b4c625c6SSong Liu for (disk_index = 0; disk_index < sh->disks; disk_index++) { 1627b4c625c6SSong Liu if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags)) 1628b4c625c6SSong Liu continue; 1629b4c625c6SSong Liu if (disk_index == sh->qd_idx || disk_index == sh->pd_idx) 1630b4c625c6SSong Liu continue; 1631b4c625c6SSong Liu data_count++; 1632b4c625c6SSong Liu } 1633b4c625c6SSong Liu 1634b4c625c6SSong Liu /* 1635b4c625c6SSong Liu * stripes that only have parity must have been flushed 1636b4c625c6SSong Liu * before the crash that we are now recovering from, so 1637b4c625c6SSong Liu * there is nothing more to recovery. 1638b4c625c6SSong Liu */ 1639b4c625c6SSong Liu if (data_count == 0) 1640b4c625c6SSong Liu goto out; 1641b4c625c6SSong Liu 1642b4c625c6SSong Liu for (disk_index = 0; disk_index < sh->disks; disk_index++) { 1643b4c625c6SSong Liu if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags)) 1644b4c625c6SSong Liu continue; 1645b4c625c6SSong Liu 1646b4c625c6SSong Liu /* in case device is broken */ 1647b4c625c6SSong Liu rcu_read_lock(); 1648b4c625c6SSong Liu rdev = rcu_dereference(conf->disks[disk_index].rdev); 1649b4c625c6SSong Liu if (rdev) { 1650b4c625c6SSong Liu atomic_inc(&rdev->nr_pending); 1651b4c625c6SSong Liu rcu_read_unlock(); 1652b4c625c6SSong Liu sync_page_io(rdev, sh->sector, PAGE_SIZE, 1653b4c625c6SSong Liu sh->dev[disk_index].page, REQ_OP_WRITE, 0, 1654b4c625c6SSong Liu false); 1655b4c625c6SSong Liu rdev_dec_pending(rdev, rdev->mddev); 1656b4c625c6SSong Liu rcu_read_lock(); 1657b4c625c6SSong Liu } 1658b4c625c6SSong Liu rrdev = rcu_dereference(conf->disks[disk_index].replacement); 1659b4c625c6SSong Liu if (rrdev) { 1660b4c625c6SSong Liu atomic_inc(&rrdev->nr_pending); 1661b4c625c6SSong Liu rcu_read_unlock(); 1662b4c625c6SSong Liu sync_page_io(rrdev, sh->sector, PAGE_SIZE, 1663b4c625c6SSong Liu sh->dev[disk_index].page, REQ_OP_WRITE, 0, 1664b4c625c6SSong Liu false); 1665b4c625c6SSong Liu rdev_dec_pending(rrdev, rrdev->mddev); 1666b4c625c6SSong Liu rcu_read_lock(); 1667b4c625c6SSong Liu } 1668b4c625c6SSong Liu rcu_read_unlock(); 1669b4c625c6SSong Liu } 1670b4c625c6SSong Liu ctx->data_parity_stripes++; 1671b4c625c6SSong Liu out: 1672b4c625c6SSong Liu r5l_recovery_reset_stripe(sh); 1673b4c625c6SSong Liu } 1674b4c625c6SSong Liu 1675b4c625c6SSong Liu static struct stripe_head * 1676b4c625c6SSong Liu r5c_recovery_alloc_stripe(struct r5conf *conf, 1677b4c625c6SSong Liu struct list_head *recovery_list, 1678b4c625c6SSong Liu sector_t stripe_sect, 1679b4c625c6SSong Liu sector_t log_start) 1680b4c625c6SSong Liu { 1681b4c625c6SSong Liu struct stripe_head *sh; 1682b4c625c6SSong Liu 1683b4c625c6SSong Liu sh = raid5_get_active_stripe(conf, stripe_sect, 0, 1, 0); 1684b4c625c6SSong Liu if (!sh) 1685b4c625c6SSong Liu return NULL; /* no more stripe available */ 1686b4c625c6SSong Liu 1687b4c625c6SSong Liu r5l_recovery_reset_stripe(sh); 1688b4c625c6SSong Liu sh->log_start = log_start; 1689b4c625c6SSong Liu 1690b4c625c6SSong Liu return sh; 1691b4c625c6SSong Liu } 1692b4c625c6SSong Liu 1693b4c625c6SSong Liu static struct stripe_head * 1694b4c625c6SSong Liu r5c_recovery_lookup_stripe(struct list_head *list, sector_t sect) 1695b4c625c6SSong Liu { 1696b4c625c6SSong Liu struct stripe_head *sh; 1697b4c625c6SSong Liu 1698b4c625c6SSong Liu list_for_each_entry(sh, list, lru) 1699b4c625c6SSong Liu if (sh->sector == sect) 1700b4c625c6SSong Liu return sh; 1701b4c625c6SSong Liu return NULL; 1702b4c625c6SSong Liu } 1703b4c625c6SSong Liu 1704b4c625c6SSong Liu static void 1705b4c625c6SSong Liu r5c_recovery_drop_stripes(struct list_head *cached_stripe_list, 1706b4c625c6SSong Liu struct r5l_recovery_ctx *ctx) 1707b4c625c6SSong Liu { 1708b4c625c6SSong Liu struct stripe_head *sh, *next; 1709b4c625c6SSong Liu 1710b4c625c6SSong Liu list_for_each_entry_safe(sh, next, cached_stripe_list, lru) { 1711b4c625c6SSong Liu r5l_recovery_reset_stripe(sh); 1712b4c625c6SSong Liu list_del_init(&sh->lru); 1713b4c625c6SSong Liu raid5_release_stripe(sh); 1714b4c625c6SSong Liu } 1715b4c625c6SSong Liu } 1716b4c625c6SSong Liu 1717b4c625c6SSong Liu static void 1718b4c625c6SSong Liu r5c_recovery_replay_stripes(struct list_head *cached_stripe_list, 1719b4c625c6SSong Liu struct r5l_recovery_ctx *ctx) 1720b4c625c6SSong Liu { 1721b4c625c6SSong Liu struct stripe_head *sh, *next; 1722b4c625c6SSong Liu 1723b4c625c6SSong Liu list_for_each_entry_safe(sh, next, cached_stripe_list, lru) 1724b4c625c6SSong Liu if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) { 1725b4c625c6SSong Liu r5l_recovery_replay_one_stripe(sh->raid_conf, sh, ctx); 1726b4c625c6SSong Liu list_del_init(&sh->lru); 1727b4c625c6SSong Liu raid5_release_stripe(sh); 1728b4c625c6SSong Liu } 1729b4c625c6SSong Liu } 1730b4c625c6SSong Liu 1731b4c625c6SSong Liu /* if matches return 0; otherwise return -EINVAL */ 1732b4c625c6SSong Liu static int 1733b4c625c6SSong Liu r5l_recovery_verify_data_checksum(struct r5l_log *log, struct page *page, 1734b4c625c6SSong Liu sector_t log_offset, __le32 log_checksum) 1735b4c625c6SSong Liu { 1736b4c625c6SSong Liu void *addr; 1737b4c625c6SSong Liu u32 checksum; 1738b4c625c6SSong Liu 1739b4c625c6SSong Liu sync_page_io(log->rdev, log_offset, PAGE_SIZE, 1740b4c625c6SSong Liu page, REQ_OP_READ, 0, false); 1741b4c625c6SSong Liu addr = kmap_atomic(page); 1742b4c625c6SSong Liu checksum = crc32c_le(log->uuid_checksum, addr, PAGE_SIZE); 1743b4c625c6SSong Liu kunmap_atomic(addr); 1744b4c625c6SSong Liu return (le32_to_cpu(log_checksum) == checksum) ? 0 : -EINVAL; 1745b4c625c6SSong Liu } 1746b4c625c6SSong Liu 1747b4c625c6SSong Liu /* 1748b4c625c6SSong Liu * before loading data to stripe cache, we need verify checksum for all data, 1749b4c625c6SSong Liu * if there is mismatch for any data page, we drop all data in the mata block 1750b4c625c6SSong Liu */ 1751b4c625c6SSong Liu static int 1752b4c625c6SSong Liu r5l_recovery_verify_data_checksum_for_mb(struct r5l_log *log, 1753b4c625c6SSong Liu struct r5l_recovery_ctx *ctx) 1754b4c625c6SSong Liu { 1755b4c625c6SSong Liu struct mddev *mddev = log->rdev->mddev; 1756b4c625c6SSong Liu struct r5conf *conf = mddev->private; 1757b4c625c6SSong Liu struct r5l_meta_block *mb = page_address(ctx->meta_page); 1758b4c625c6SSong Liu sector_t mb_offset = sizeof(struct r5l_meta_block); 1759b4c625c6SSong Liu sector_t log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS); 1760b4c625c6SSong Liu struct page *page; 1761b4c625c6SSong Liu struct r5l_payload_data_parity *payload; 1762b4c625c6SSong Liu 1763b4c625c6SSong Liu page = alloc_page(GFP_KERNEL); 1764b4c625c6SSong Liu if (!page) 1765b4c625c6SSong Liu return -ENOMEM; 1766b4c625c6SSong Liu 1767b4c625c6SSong Liu while (mb_offset < le32_to_cpu(mb->meta_size)) { 1768b4c625c6SSong Liu payload = (void *)mb + mb_offset; 1769b4c625c6SSong Liu 1770b4c625c6SSong Liu if (payload->header.type == R5LOG_PAYLOAD_DATA) { 1771b4c625c6SSong Liu if (r5l_recovery_verify_data_checksum( 1772b4c625c6SSong Liu log, page, log_offset, 1773b4c625c6SSong Liu payload->checksum[0]) < 0) 1774b4c625c6SSong Liu goto mismatch; 1775b4c625c6SSong Liu } else if (payload->header.type == R5LOG_PAYLOAD_PARITY) { 1776b4c625c6SSong Liu if (r5l_recovery_verify_data_checksum( 1777b4c625c6SSong Liu log, page, log_offset, 1778b4c625c6SSong Liu payload->checksum[0]) < 0) 1779b4c625c6SSong Liu goto mismatch; 1780b4c625c6SSong Liu if (conf->max_degraded == 2 && /* q for RAID 6 */ 1781b4c625c6SSong Liu r5l_recovery_verify_data_checksum( 1782b4c625c6SSong Liu log, page, 1783b4c625c6SSong Liu r5l_ring_add(log, log_offset, 1784b4c625c6SSong Liu BLOCK_SECTORS), 1785b4c625c6SSong Liu payload->checksum[1]) < 0) 1786b4c625c6SSong Liu goto mismatch; 1787b4c625c6SSong Liu } else /* not R5LOG_PAYLOAD_DATA or R5LOG_PAYLOAD_PARITY */ 1788b4c625c6SSong Liu goto mismatch; 1789b4c625c6SSong Liu 1790b4c625c6SSong Liu log_offset = r5l_ring_add(log, log_offset, 1791b4c625c6SSong Liu le32_to_cpu(payload->size)); 1792b4c625c6SSong Liu 1793b4c625c6SSong Liu mb_offset += sizeof(struct r5l_payload_data_parity) + 1794b4c625c6SSong Liu sizeof(__le32) * 1795b4c625c6SSong Liu (le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9)); 1796b4c625c6SSong Liu } 1797b4c625c6SSong Liu 1798b4c625c6SSong Liu put_page(page); 1799b4c625c6SSong Liu return 0; 1800b4c625c6SSong Liu 1801b4c625c6SSong Liu mismatch: 1802b4c625c6SSong Liu put_page(page); 1803b4c625c6SSong Liu return -EINVAL; 1804b4c625c6SSong Liu } 1805b4c625c6SSong Liu 1806b4c625c6SSong Liu /* 1807b4c625c6SSong Liu * Analyze all data/parity pages in one meta block 1808b4c625c6SSong Liu * Returns: 1809b4c625c6SSong Liu * 0 for success 1810b4c625c6SSong Liu * -EINVAL for unknown playload type 1811b4c625c6SSong Liu * -EAGAIN for checksum mismatch of data page 1812b4c625c6SSong Liu * -ENOMEM for run out of memory (alloc_page failed or run out of stripes) 1813b4c625c6SSong Liu */ 1814b4c625c6SSong Liu static int 1815b4c625c6SSong Liu r5c_recovery_analyze_meta_block(struct r5l_log *log, 1816b4c625c6SSong Liu struct r5l_recovery_ctx *ctx, 1817b4c625c6SSong Liu struct list_head *cached_stripe_list) 1818b4c625c6SSong Liu { 1819b4c625c6SSong Liu struct mddev *mddev = log->rdev->mddev; 1820b4c625c6SSong Liu struct r5conf *conf = mddev->private; 1821b4c625c6SSong Liu struct r5l_meta_block *mb; 1822b4c625c6SSong Liu struct r5l_payload_data_parity *payload; 1823b4c625c6SSong Liu int mb_offset; 1824b4c625c6SSong Liu sector_t log_offset; 1825b4c625c6SSong Liu sector_t stripe_sect; 1826b4c625c6SSong Liu struct stripe_head *sh; 1827b4c625c6SSong Liu int ret; 1828b4c625c6SSong Liu 1829b4c625c6SSong Liu /* 1830b4c625c6SSong Liu * for mismatch in data blocks, we will drop all data in this mb, but 1831b4c625c6SSong Liu * we will still read next mb for other data with FLUSH flag, as 1832b4c625c6SSong Liu * io_unit could finish out of order. 1833b4c625c6SSong Liu */ 1834b4c625c6SSong Liu ret = r5l_recovery_verify_data_checksum_for_mb(log, ctx); 1835b4c625c6SSong Liu if (ret == -EINVAL) 1836b4c625c6SSong Liu return -EAGAIN; 1837b4c625c6SSong Liu else if (ret) 1838b4c625c6SSong Liu return ret; /* -ENOMEM duo to alloc_page() failed */ 1839b4c625c6SSong Liu 1840b4c625c6SSong Liu mb = page_address(ctx->meta_page); 1841b4c625c6SSong Liu mb_offset = sizeof(struct r5l_meta_block); 1842b4c625c6SSong Liu log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS); 1843b4c625c6SSong Liu 1844b4c625c6SSong Liu while (mb_offset < le32_to_cpu(mb->meta_size)) { 1845b4c625c6SSong Liu int dd; 1846b4c625c6SSong Liu 1847b4c625c6SSong Liu payload = (void *)mb + mb_offset; 1848b4c625c6SSong Liu stripe_sect = (payload->header.type == R5LOG_PAYLOAD_DATA) ? 1849b4c625c6SSong Liu raid5_compute_sector( 1850b4c625c6SSong Liu conf, le64_to_cpu(payload->location), 0, &dd, 1851b4c625c6SSong Liu NULL) 1852b4c625c6SSong Liu : le64_to_cpu(payload->location); 1853b4c625c6SSong Liu 1854b4c625c6SSong Liu sh = r5c_recovery_lookup_stripe(cached_stripe_list, 1855b4c625c6SSong Liu stripe_sect); 1856b4c625c6SSong Liu 1857b4c625c6SSong Liu if (!sh) { 1858b4c625c6SSong Liu sh = r5c_recovery_alloc_stripe(conf, cached_stripe_list, 1859b4c625c6SSong Liu stripe_sect, ctx->pos); 1860b4c625c6SSong Liu /* 1861b4c625c6SSong Liu * cannot get stripe from raid5_get_active_stripe 1862b4c625c6SSong Liu * try replay some stripes 1863b4c625c6SSong Liu */ 1864b4c625c6SSong Liu if (!sh) { 1865b4c625c6SSong Liu r5c_recovery_replay_stripes( 1866b4c625c6SSong Liu cached_stripe_list, ctx); 1867b4c625c6SSong Liu sh = r5c_recovery_alloc_stripe( 1868b4c625c6SSong Liu conf, cached_stripe_list, 1869b4c625c6SSong Liu stripe_sect, ctx->pos); 1870b4c625c6SSong Liu } 1871b4c625c6SSong Liu if (!sh) { 1872b4c625c6SSong Liu pr_debug("md/raid:%s: Increasing stripe cache size to %d to recovery data on journal.\n", 1873b4c625c6SSong Liu mdname(mddev), 1874b4c625c6SSong Liu conf->min_nr_stripes * 2); 1875b4c625c6SSong Liu raid5_set_cache_size(mddev, 1876b4c625c6SSong Liu conf->min_nr_stripes * 2); 1877b4c625c6SSong Liu sh = r5c_recovery_alloc_stripe( 1878b4c625c6SSong Liu conf, cached_stripe_list, stripe_sect, 1879b4c625c6SSong Liu ctx->pos); 1880b4c625c6SSong Liu } 1881b4c625c6SSong Liu if (!sh) { 1882b4c625c6SSong Liu pr_err("md/raid:%s: Cannot get enough stripes due to memory pressure. Recovery failed.\n", 1883b4c625c6SSong Liu mdname(mddev)); 1884b4c625c6SSong Liu return -ENOMEM; 1885b4c625c6SSong Liu } 1886b4c625c6SSong Liu list_add_tail(&sh->lru, cached_stripe_list); 1887b4c625c6SSong Liu } 1888b4c625c6SSong Liu 1889b4c625c6SSong Liu if (payload->header.type == R5LOG_PAYLOAD_DATA) { 1890b4c625c6SSong Liu if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) { 1891b4c625c6SSong Liu r5l_recovery_replay_one_stripe(conf, sh, ctx); 1892b4c625c6SSong Liu r5l_recovery_reset_stripe(sh); 1893b4c625c6SSong Liu sh->log_start = ctx->pos; 1894b4c625c6SSong Liu list_move_tail(&sh->lru, cached_stripe_list); 1895b4c625c6SSong Liu } 1896b4c625c6SSong Liu r5l_recovery_load_data(log, sh, ctx, payload, 1897b4c625c6SSong Liu log_offset); 1898b4c625c6SSong Liu } else if (payload->header.type == R5LOG_PAYLOAD_PARITY) 1899b4c625c6SSong Liu r5l_recovery_load_parity(log, sh, ctx, payload, 1900b4c625c6SSong Liu log_offset); 1901b4c625c6SSong Liu else 1902b4c625c6SSong Liu return -EINVAL; 1903b4c625c6SSong Liu 1904b4c625c6SSong Liu log_offset = r5l_ring_add(log, log_offset, 1905b4c625c6SSong Liu le32_to_cpu(payload->size)); 1906b4c625c6SSong Liu 1907b4c625c6SSong Liu mb_offset += sizeof(struct r5l_payload_data_parity) + 1908b4c625c6SSong Liu sizeof(__le32) * 1909b4c625c6SSong Liu (le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9)); 1910b4c625c6SSong Liu } 1911b4c625c6SSong Liu 1912b4c625c6SSong Liu return 0; 1913b4c625c6SSong Liu } 1914b4c625c6SSong Liu 1915b4c625c6SSong Liu /* 1916b4c625c6SSong Liu * Load the stripe into cache. The stripe will be written out later by 1917b4c625c6SSong Liu * the stripe cache state machine. 1918b4c625c6SSong Liu */ 1919b4c625c6SSong Liu static void r5c_recovery_load_one_stripe(struct r5l_log *log, 1920b4c625c6SSong Liu struct stripe_head *sh) 1921b4c625c6SSong Liu { 1922b4c625c6SSong Liu struct r5conf *conf = sh->raid_conf; 1923b4c625c6SSong Liu struct r5dev *dev; 1924b4c625c6SSong Liu int i; 1925b4c625c6SSong Liu 1926b4c625c6SSong Liu for (i = sh->disks; i--; ) { 1927b4c625c6SSong Liu dev = sh->dev + i; 1928b4c625c6SSong Liu if (test_and_clear_bit(R5_Wantwrite, &dev->flags)) { 1929b4c625c6SSong Liu set_bit(R5_InJournal, &dev->flags); 1930b4c625c6SSong Liu set_bit(R5_UPTODATE, &dev->flags); 1931b4c625c6SSong Liu } 1932b4c625c6SSong Liu } 1933b4c625c6SSong Liu set_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state); 1934b4c625c6SSong Liu atomic_inc(&conf->r5c_cached_partial_stripes); 1935b4c625c6SSong Liu list_add_tail(&sh->r5c, &log->stripe_in_journal_list); 1936b4c625c6SSong Liu } 1937b4c625c6SSong Liu 1938b4c625c6SSong Liu /* 1939b4c625c6SSong Liu * Scan through the log for all to-be-flushed data 1940b4c625c6SSong Liu * 1941b4c625c6SSong Liu * For stripes with data and parity, namely Data-Parity stripe 1942b4c625c6SSong Liu * (STRIPE_R5C_CACHING == 0), we simply replay all the writes. 1943b4c625c6SSong Liu * 1944b4c625c6SSong Liu * For stripes with only data, namely Data-Only stripe 1945b4c625c6SSong Liu * (STRIPE_R5C_CACHING == 1), we load them to stripe cache state machine. 1946b4c625c6SSong Liu * 1947b4c625c6SSong Liu * For a stripe, if we see data after parity, we should discard all previous 1948b4c625c6SSong Liu * data and parity for this stripe, as these data are already flushed to 1949b4c625c6SSong Liu * the array. 1950b4c625c6SSong Liu * 1951b4c625c6SSong Liu * At the end of the scan, we return the new journal_tail, which points to 1952b4c625c6SSong Liu * first data-only stripe on the journal device, or next invalid meta block. 1953b4c625c6SSong Liu */ 1954b4c625c6SSong Liu static int r5c_recovery_flush_log(struct r5l_log *log, 1955b4c625c6SSong Liu struct r5l_recovery_ctx *ctx) 1956b4c625c6SSong Liu { 1957b4c625c6SSong Liu struct stripe_head *sh, *next; 1958b4c625c6SSong Liu int ret = 0; 1959b4c625c6SSong Liu 1960b4c625c6SSong Liu /* scan through the log */ 1961b4c625c6SSong Liu while (1) { 1962b4c625c6SSong Liu if (r5l_recovery_read_meta_block(log, ctx)) 1963b4c625c6SSong Liu break; 1964b4c625c6SSong Liu 1965b4c625c6SSong Liu ret = r5c_recovery_analyze_meta_block(log, ctx, 1966b4c625c6SSong Liu &ctx->cached_list); 1967b4c625c6SSong Liu /* 1968b4c625c6SSong Liu * -EAGAIN means mismatch in data block, in this case, we still 1969b4c625c6SSong Liu * try scan the next metablock 1970b4c625c6SSong Liu */ 1971b4c625c6SSong Liu if (ret && ret != -EAGAIN) 1972b4c625c6SSong Liu break; /* ret == -EINVAL or -ENOMEM */ 1973b4c625c6SSong Liu ctx->seq++; 1974b4c625c6SSong Liu ctx->pos = r5l_ring_add(log, ctx->pos, ctx->meta_total_blocks); 1975b4c625c6SSong Liu } 1976b4c625c6SSong Liu 1977b4c625c6SSong Liu if (ret == -ENOMEM) { 1978b4c625c6SSong Liu r5c_recovery_drop_stripes(&ctx->cached_list, ctx); 1979b4c625c6SSong Liu return ret; 1980b4c625c6SSong Liu } 1981b4c625c6SSong Liu 1982b4c625c6SSong Liu /* replay data-parity stripes */ 1983b4c625c6SSong Liu r5c_recovery_replay_stripes(&ctx->cached_list, ctx); 1984b4c625c6SSong Liu 1985b4c625c6SSong Liu /* load data-only stripes to stripe cache */ 1986b4c625c6SSong Liu list_for_each_entry_safe(sh, next, &ctx->cached_list, lru) { 1987b4c625c6SSong Liu WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state)); 1988b4c625c6SSong Liu r5c_recovery_load_one_stripe(log, sh); 1989b4c625c6SSong Liu list_del_init(&sh->lru); 1990b4c625c6SSong Liu raid5_release_stripe(sh); 1991b4c625c6SSong Liu ctx->data_only_stripes++; 1992b4c625c6SSong Liu } 1993b4c625c6SSong Liu 1994b4c625c6SSong Liu return 0; 1995b4c625c6SSong Liu } 1996b4c625c6SSong Liu 1997b4c625c6SSong Liu /* 1998b4c625c6SSong Liu * we did a recovery. Now ctx.pos points to an invalid meta block. New 1999b4c625c6SSong Liu * log will start here. but we can't let superblock point to last valid 2000b4c625c6SSong Liu * meta block. The log might looks like: 2001b4c625c6SSong Liu * | meta 1| meta 2| meta 3| 2002b4c625c6SSong Liu * meta 1 is valid, meta 2 is invalid. meta 3 could be valid. If 2003b4c625c6SSong Liu * superblock points to meta 1, we write a new valid meta 2n. if crash 2004b4c625c6SSong Liu * happens again, new recovery will start from meta 1. Since meta 2n is 2005b4c625c6SSong Liu * valid now, recovery will think meta 3 is valid, which is wrong. 2006b4c625c6SSong Liu * The solution is we create a new meta in meta2 with its seq == meta 2007b4c625c6SSong Liu * 1's seq + 10 and let superblock points to meta2. The same recovery will 2008b4c625c6SSong Liu * not think meta 3 is a valid meta, because its seq doesn't match 2009b4c625c6SSong Liu */ 2010b4c625c6SSong Liu 2011b4c625c6SSong Liu /* 2012b4c625c6SSong Liu * Before recovery, the log looks like the following 2013b4c625c6SSong Liu * 2014b4c625c6SSong Liu * --------------------------------------------- 2015b4c625c6SSong Liu * | valid log | invalid log | 2016b4c625c6SSong Liu * --------------------------------------------- 2017b4c625c6SSong Liu * ^ 2018b4c625c6SSong Liu * |- log->last_checkpoint 2019b4c625c6SSong Liu * |- log->last_cp_seq 2020b4c625c6SSong Liu * 2021b4c625c6SSong Liu * Now we scan through the log until we see invalid entry 2022b4c625c6SSong Liu * 2023b4c625c6SSong Liu * --------------------------------------------- 2024b4c625c6SSong Liu * | valid log | invalid log | 2025b4c625c6SSong Liu * --------------------------------------------- 2026b4c625c6SSong Liu * ^ ^ 2027b4c625c6SSong Liu * |- log->last_checkpoint |- ctx->pos 2028b4c625c6SSong Liu * |- log->last_cp_seq |- ctx->seq 2029b4c625c6SSong Liu * 2030b4c625c6SSong Liu * From this point, we need to increase seq number by 10 to avoid 2031b4c625c6SSong Liu * confusing next recovery. 2032b4c625c6SSong Liu * 2033b4c625c6SSong Liu * --------------------------------------------- 2034b4c625c6SSong Liu * | valid log | invalid log | 2035b4c625c6SSong Liu * --------------------------------------------- 2036b4c625c6SSong Liu * ^ ^ 2037b4c625c6SSong Liu * |- log->last_checkpoint |- ctx->pos+1 2038b4c625c6SSong Liu * |- log->last_cp_seq |- ctx->seq+11 2039b4c625c6SSong Liu * 2040b4c625c6SSong Liu * However, it is not safe to start the state machine yet, because data only 2041b4c625c6SSong Liu * parities are not yet secured in RAID. To save these data only parities, we 2042b4c625c6SSong Liu * rewrite them from seq+11. 2043b4c625c6SSong Liu * 2044b4c625c6SSong Liu * ----------------------------------------------------------------- 2045b4c625c6SSong Liu * | valid log | data only stripes | invalid log | 2046b4c625c6SSong Liu * ----------------------------------------------------------------- 2047b4c625c6SSong Liu * ^ ^ 2048b4c625c6SSong Liu * |- log->last_checkpoint |- ctx->pos+n 2049b4c625c6SSong Liu * |- log->last_cp_seq |- ctx->seq+10+n 2050b4c625c6SSong Liu * 2051b4c625c6SSong Liu * If failure happens again during this process, the recovery can safe start 2052b4c625c6SSong Liu * again from log->last_checkpoint. 2053b4c625c6SSong Liu * 2054b4c625c6SSong Liu * Once data only stripes are rewritten to journal, we move log_tail 2055b4c625c6SSong Liu * 2056b4c625c6SSong Liu * ----------------------------------------------------------------- 2057b4c625c6SSong Liu * | old log | data only stripes | invalid log | 2058b4c625c6SSong Liu * ----------------------------------------------------------------- 2059b4c625c6SSong Liu * ^ ^ 2060b4c625c6SSong Liu * |- log->last_checkpoint |- ctx->pos+n 2061b4c625c6SSong Liu * |- log->last_cp_seq |- ctx->seq+10+n 2062b4c625c6SSong Liu * 2063b4c625c6SSong Liu * Then we can safely start the state machine. If failure happens from this 2064b4c625c6SSong Liu * point on, the recovery will start from new log->last_checkpoint. 2065b4c625c6SSong Liu */ 2066b4c625c6SSong Liu static int 2067b4c625c6SSong Liu r5c_recovery_rewrite_data_only_stripes(struct r5l_log *log, 2068b4c625c6SSong Liu struct r5l_recovery_ctx *ctx) 2069b4c625c6SSong Liu { 2070b4c625c6SSong Liu struct stripe_head *sh; 2071b4c625c6SSong Liu struct mddev *mddev = log->rdev->mddev; 2072b4c625c6SSong Liu struct page *page; 2073b4c625c6SSong Liu 2074b4c625c6SSong Liu page = alloc_page(GFP_KERNEL); 2075b4c625c6SSong Liu if (!page) { 2076b4c625c6SSong Liu pr_err("md/raid:%s: cannot allocate memory to rewrite data only stripes\n", 2077b4c625c6SSong Liu mdname(mddev)); 2078b4c625c6SSong Liu return -ENOMEM; 2079b4c625c6SSong Liu } 2080b4c625c6SSong Liu 2081b4c625c6SSong Liu ctx->seq += 10; 2082b4c625c6SSong Liu list_for_each_entry(sh, &ctx->cached_list, lru) { 2083b4c625c6SSong Liu struct r5l_meta_block *mb; 2084b4c625c6SSong Liu int i; 2085b4c625c6SSong Liu int offset; 2086b4c625c6SSong Liu sector_t write_pos; 2087b4c625c6SSong Liu 2088b4c625c6SSong Liu WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state)); 2089b4c625c6SSong Liu r5l_recovery_create_empty_meta_block(log, page, 2090b4c625c6SSong Liu ctx->pos, ctx->seq); 2091b4c625c6SSong Liu mb = page_address(page); 2092b4c625c6SSong Liu offset = le32_to_cpu(mb->meta_size); 2093b4c625c6SSong Liu write_pos = ctx->pos + BLOCK_SECTORS; 2094b4c625c6SSong Liu 2095b4c625c6SSong Liu for (i = sh->disks; i--; ) { 2096b4c625c6SSong Liu struct r5dev *dev = &sh->dev[i]; 2097b4c625c6SSong Liu struct r5l_payload_data_parity *payload; 2098b4c625c6SSong Liu void *addr; 2099b4c625c6SSong Liu 2100b4c625c6SSong Liu if (test_bit(R5_InJournal, &dev->flags)) { 2101b4c625c6SSong Liu payload = (void *)mb + offset; 2102b4c625c6SSong Liu payload->header.type = cpu_to_le16( 2103b4c625c6SSong Liu R5LOG_PAYLOAD_DATA); 2104b4c625c6SSong Liu payload->size = BLOCK_SECTORS; 2105b4c625c6SSong Liu payload->location = cpu_to_le64( 2106b4c625c6SSong Liu raid5_compute_blocknr(sh, i, 0)); 2107b4c625c6SSong Liu addr = kmap_atomic(dev->page); 2108b4c625c6SSong Liu payload->checksum[0] = cpu_to_le32( 2109b4c625c6SSong Liu crc32c_le(log->uuid_checksum, addr, 2110b4c625c6SSong Liu PAGE_SIZE)); 2111b4c625c6SSong Liu kunmap_atomic(addr); 2112b4c625c6SSong Liu sync_page_io(log->rdev, write_pos, PAGE_SIZE, 2113b4c625c6SSong Liu dev->page, REQ_OP_WRITE, 0, false); 2114b4c625c6SSong Liu write_pos = r5l_ring_add(log, write_pos, 2115b4c625c6SSong Liu BLOCK_SECTORS); 2116b4c625c6SSong Liu offset += sizeof(__le32) + 2117b4c625c6SSong Liu sizeof(struct r5l_payload_data_parity); 2118b4c625c6SSong Liu 2119b4c625c6SSong Liu } 2120b4c625c6SSong Liu } 2121b4c625c6SSong Liu mb->meta_size = cpu_to_le32(offset); 2122b4c625c6SSong Liu mb->checksum = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE); 2123b4c625c6SSong Liu sync_page_io(log->rdev, ctx->pos, PAGE_SIZE, page, 2124b4c625c6SSong Liu REQ_OP_WRITE, WRITE_FUA, false); 2125b4c625c6SSong Liu sh->log_start = ctx->pos; 2126b4c625c6SSong Liu ctx->pos = write_pos; 2127b4c625c6SSong Liu ctx->seq += 1; 2128b4c625c6SSong Liu } 2129b4c625c6SSong Liu __free_page(page); 2130b4c625c6SSong Liu return 0; 2131b4c625c6SSong Liu } 2132b4c625c6SSong Liu 2133f6bed0efSShaohua Li static int r5l_recovery_log(struct r5l_log *log) 2134f6bed0efSShaohua Li { 21355aabf7c4SSong Liu struct mddev *mddev = log->rdev->mddev; 2136355810d1SShaohua Li struct r5l_recovery_ctx ctx; 21375aabf7c4SSong Liu int ret; 2138355810d1SShaohua Li 2139355810d1SShaohua Li ctx.pos = log->last_checkpoint; 2140355810d1SShaohua Li ctx.seq = log->last_cp_seq; 2141355810d1SShaohua Li ctx.meta_page = alloc_page(GFP_KERNEL); 2142b4c625c6SSong Liu ctx.data_only_stripes = 0; 2143b4c625c6SSong Liu ctx.data_parity_stripes = 0; 2144b4c625c6SSong Liu INIT_LIST_HEAD(&ctx.cached_list); 2145b4c625c6SSong Liu 2146355810d1SShaohua Li if (!ctx.meta_page) 2147355810d1SShaohua Li return -ENOMEM; 2148355810d1SShaohua Li 21495aabf7c4SSong Liu ret = r5c_recovery_flush_log(log, &ctx); 2150355810d1SShaohua Li __free_page(ctx.meta_page); 2151355810d1SShaohua Li 2152355810d1SShaohua Li if (ret) 2153355810d1SShaohua Li return ret; 21545aabf7c4SSong Liu 21555aabf7c4SSong Liu if ((ctx.data_only_stripes == 0) && (ctx.data_parity_stripes == 0)) 21565aabf7c4SSong Liu pr_debug("md/raid:%s: starting from clean shutdown\n", 21575aabf7c4SSong Liu mdname(mddev)); 21585aabf7c4SSong Liu else { 21595aabf7c4SSong Liu pr_debug("md/raid:%s: recoverying %d data-only stripes and %d data-parity stripes\n", 21605aabf7c4SSong Liu mdname(mddev), ctx.data_only_stripes, 21615aabf7c4SSong Liu ctx.data_parity_stripes); 21625aabf7c4SSong Liu 21635aabf7c4SSong Liu if (ctx.data_only_stripes > 0) 21645aabf7c4SSong Liu if (r5c_recovery_rewrite_data_only_stripes(log, &ctx)) { 21655aabf7c4SSong Liu pr_err("md/raid:%s: failed to rewrite stripes to journal\n", 21665aabf7c4SSong Liu mdname(mddev)); 21675aabf7c4SSong Liu return -EIO; 21685aabf7c4SSong Liu } 21695aabf7c4SSong Liu } 21705aabf7c4SSong Liu 2171355810d1SShaohua Li log->log_start = ctx.pos; 21725aabf7c4SSong Liu log->next_checkpoint = ctx.pos; 2173355810d1SShaohua Li log->seq = ctx.seq; 21745aabf7c4SSong Liu r5l_log_write_empty_meta_block(log, ctx.pos, ctx.seq); 21755aabf7c4SSong Liu r5l_write_super(log, ctx.pos); 2176f6bed0efSShaohua Li return 0; 2177f6bed0efSShaohua Li } 2178f6bed0efSShaohua Li 2179f6bed0efSShaohua Li static void r5l_write_super(struct r5l_log *log, sector_t cp) 2180f6bed0efSShaohua Li { 2181f6bed0efSShaohua Li struct mddev *mddev = log->rdev->mddev; 2182f6bed0efSShaohua Li 2183f6bed0efSShaohua Li log->rdev->journal_tail = cp; 2184f6bed0efSShaohua Li set_bit(MD_CHANGE_DEVS, &mddev->flags); 2185f6bed0efSShaohua Li } 2186f6bed0efSShaohua Li 21872c7da14bSSong Liu static ssize_t r5c_journal_mode_show(struct mddev *mddev, char *page) 21882c7da14bSSong Liu { 21892c7da14bSSong Liu struct r5conf *conf = mddev->private; 21902c7da14bSSong Liu int ret; 21912c7da14bSSong Liu 21922c7da14bSSong Liu if (!conf->log) 21932c7da14bSSong Liu return 0; 21942c7da14bSSong Liu 21952c7da14bSSong Liu switch (conf->log->r5c_journal_mode) { 21962c7da14bSSong Liu case R5C_JOURNAL_MODE_WRITE_THROUGH: 21972c7da14bSSong Liu ret = snprintf( 21982c7da14bSSong Liu page, PAGE_SIZE, "[%s] %s\n", 21992c7da14bSSong Liu r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH], 22002c7da14bSSong Liu r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]); 22012c7da14bSSong Liu break; 22022c7da14bSSong Liu case R5C_JOURNAL_MODE_WRITE_BACK: 22032c7da14bSSong Liu ret = snprintf( 22042c7da14bSSong Liu page, PAGE_SIZE, "%s [%s]\n", 22052c7da14bSSong Liu r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH], 22062c7da14bSSong Liu r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]); 22072c7da14bSSong Liu break; 22082c7da14bSSong Liu default: 22092c7da14bSSong Liu ret = 0; 22102c7da14bSSong Liu } 22112c7da14bSSong Liu return ret; 22122c7da14bSSong Liu } 22132c7da14bSSong Liu 22142c7da14bSSong Liu static ssize_t r5c_journal_mode_store(struct mddev *mddev, 22152c7da14bSSong Liu const char *page, size_t length) 22162c7da14bSSong Liu { 22172c7da14bSSong Liu struct r5conf *conf = mddev->private; 22182c7da14bSSong Liu struct r5l_log *log = conf->log; 22192c7da14bSSong Liu int val = -1, i; 22202c7da14bSSong Liu int len = length; 22212c7da14bSSong Liu 22222c7da14bSSong Liu if (!log) 22232c7da14bSSong Liu return -ENODEV; 22242c7da14bSSong Liu 22252c7da14bSSong Liu if (len && page[len - 1] == '\n') 22262c7da14bSSong Liu len -= 1; 22272c7da14bSSong Liu for (i = 0; i < ARRAY_SIZE(r5c_journal_mode_str); i++) 22282c7da14bSSong Liu if (strlen(r5c_journal_mode_str[i]) == len && 22292c7da14bSSong Liu strncmp(page, r5c_journal_mode_str[i], len) == 0) { 22302c7da14bSSong Liu val = i; 22312c7da14bSSong Liu break; 22322c7da14bSSong Liu } 22332c7da14bSSong Liu if (val < R5C_JOURNAL_MODE_WRITE_THROUGH || 22342c7da14bSSong Liu val > R5C_JOURNAL_MODE_WRITE_BACK) 22352c7da14bSSong Liu return -EINVAL; 22362c7da14bSSong Liu 22372c7da14bSSong Liu mddev_suspend(mddev); 22382c7da14bSSong Liu conf->log->r5c_journal_mode = val; 22392c7da14bSSong Liu mddev_resume(mddev); 22402c7da14bSSong Liu 22412c7da14bSSong Liu pr_debug("md/raid:%s: setting r5c cache mode to %d: %s\n", 22422c7da14bSSong Liu mdname(mddev), val, r5c_journal_mode_str[val]); 22432c7da14bSSong Liu return length; 22442c7da14bSSong Liu } 22452c7da14bSSong Liu 22462c7da14bSSong Liu struct md_sysfs_entry 22472c7da14bSSong Liu r5c_journal_mode = __ATTR(journal_mode, 0644, 22482c7da14bSSong Liu r5c_journal_mode_show, r5c_journal_mode_store); 22492c7da14bSSong Liu 22502ded3703SSong Liu /* 22512ded3703SSong Liu * Try handle write operation in caching phase. This function should only 22522ded3703SSong Liu * be called in write-back mode. 22532ded3703SSong Liu * 22542ded3703SSong Liu * If all outstanding writes can be handled in caching phase, returns 0 22552ded3703SSong Liu * If writes requires write-out phase, call r5c_make_stripe_write_out() 22562ded3703SSong Liu * and returns -EAGAIN 22572ded3703SSong Liu */ 22582ded3703SSong Liu int r5c_try_caching_write(struct r5conf *conf, 22592ded3703SSong Liu struct stripe_head *sh, 22602ded3703SSong Liu struct stripe_head_state *s, 22612ded3703SSong Liu int disks) 22622ded3703SSong Liu { 22632ded3703SSong Liu struct r5l_log *log = conf->log; 22641e6d690bSSong Liu int i; 22651e6d690bSSong Liu struct r5dev *dev; 22661e6d690bSSong Liu int to_cache = 0; 22672ded3703SSong Liu 22682ded3703SSong Liu BUG_ON(!r5c_is_writeback(log)); 22692ded3703SSong Liu 22701e6d690bSSong Liu if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) { 22711e6d690bSSong Liu /* 22721e6d690bSSong Liu * There are two different scenarios here: 22731e6d690bSSong Liu * 1. The stripe has some data cached, and it is sent to 22741e6d690bSSong Liu * write-out phase for reclaim 22751e6d690bSSong Liu * 2. The stripe is clean, and this is the first write 22761e6d690bSSong Liu * 22771e6d690bSSong Liu * For 1, return -EAGAIN, so we continue with 22781e6d690bSSong Liu * handle_stripe_dirtying(). 22791e6d690bSSong Liu * 22801e6d690bSSong Liu * For 2, set STRIPE_R5C_CACHING and continue with caching 22811e6d690bSSong Liu * write. 22821e6d690bSSong Liu */ 22831e6d690bSSong Liu 22841e6d690bSSong Liu /* case 1: anything injournal or anything in written */ 22851e6d690bSSong Liu if (s->injournal > 0 || s->written > 0) 22861e6d690bSSong Liu return -EAGAIN; 22871e6d690bSSong Liu /* case 2 */ 22881e6d690bSSong Liu set_bit(STRIPE_R5C_CACHING, &sh->state); 22891e6d690bSSong Liu } 22901e6d690bSSong Liu 22911e6d690bSSong Liu for (i = disks; i--; ) { 22921e6d690bSSong Liu dev = &sh->dev[i]; 22931e6d690bSSong Liu /* if non-overwrite, use writing-out phase */ 22941e6d690bSSong Liu if (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags) && 22951e6d690bSSong Liu !test_bit(R5_InJournal, &dev->flags)) { 22962ded3703SSong Liu r5c_make_stripe_write_out(sh); 22972ded3703SSong Liu return -EAGAIN; 22982ded3703SSong Liu } 22991e6d690bSSong Liu } 23001e6d690bSSong Liu 23011e6d690bSSong Liu for (i = disks; i--; ) { 23021e6d690bSSong Liu dev = &sh->dev[i]; 23031e6d690bSSong Liu if (dev->towrite) { 23041e6d690bSSong Liu set_bit(R5_Wantwrite, &dev->flags); 23051e6d690bSSong Liu set_bit(R5_Wantdrain, &dev->flags); 23061e6d690bSSong Liu set_bit(R5_LOCKED, &dev->flags); 23071e6d690bSSong Liu to_cache++; 23081e6d690bSSong Liu } 23091e6d690bSSong Liu } 23101e6d690bSSong Liu 23111e6d690bSSong Liu if (to_cache) { 23121e6d690bSSong Liu set_bit(STRIPE_OP_BIODRAIN, &s->ops_request); 23131e6d690bSSong Liu /* 23141e6d690bSSong Liu * set STRIPE_LOG_TRAPPED, which triggers r5c_cache_data() 23151e6d690bSSong Liu * in ops_run_io(). STRIPE_LOG_TRAPPED will be cleared in 23161e6d690bSSong Liu * r5c_handle_data_cached() 23171e6d690bSSong Liu */ 23181e6d690bSSong Liu set_bit(STRIPE_LOG_TRAPPED, &sh->state); 23191e6d690bSSong Liu } 23201e6d690bSSong Liu 23211e6d690bSSong Liu return 0; 23221e6d690bSSong Liu } 23231e6d690bSSong Liu 23241e6d690bSSong Liu /* 23251e6d690bSSong Liu * free extra pages (orig_page) we allocated for prexor 23261e6d690bSSong Liu */ 23271e6d690bSSong Liu void r5c_release_extra_page(struct stripe_head *sh) 23281e6d690bSSong Liu { 2329*d7bd398eSSong Liu struct r5conf *conf = sh->raid_conf; 23301e6d690bSSong Liu int i; 2331*d7bd398eSSong Liu bool using_disk_info_extra_page; 2332*d7bd398eSSong Liu 2333*d7bd398eSSong Liu using_disk_info_extra_page = 2334*d7bd398eSSong Liu sh->dev[0].orig_page == conf->disks[0].extra_page; 23351e6d690bSSong Liu 23361e6d690bSSong Liu for (i = sh->disks; i--; ) 23371e6d690bSSong Liu if (sh->dev[i].page != sh->dev[i].orig_page) { 23381e6d690bSSong Liu struct page *p = sh->dev[i].orig_page; 23391e6d690bSSong Liu 23401e6d690bSSong Liu sh->dev[i].orig_page = sh->dev[i].page; 2341*d7bd398eSSong Liu if (!using_disk_info_extra_page) 23421e6d690bSSong Liu put_page(p); 23431e6d690bSSong Liu } 2344*d7bd398eSSong Liu 2345*d7bd398eSSong Liu if (using_disk_info_extra_page) { 2346*d7bd398eSSong Liu clear_bit(R5C_EXTRA_PAGE_IN_USE, &conf->cache_state); 2347*d7bd398eSSong Liu md_wakeup_thread(conf->mddev->thread); 2348*d7bd398eSSong Liu } 2349*d7bd398eSSong Liu } 2350*d7bd398eSSong Liu 2351*d7bd398eSSong Liu void r5c_use_extra_page(struct stripe_head *sh) 2352*d7bd398eSSong Liu { 2353*d7bd398eSSong Liu struct r5conf *conf = sh->raid_conf; 2354*d7bd398eSSong Liu int i; 2355*d7bd398eSSong Liu struct r5dev *dev; 2356*d7bd398eSSong Liu 2357*d7bd398eSSong Liu for (i = sh->disks; i--; ) { 2358*d7bd398eSSong Liu dev = &sh->dev[i]; 2359*d7bd398eSSong Liu if (dev->orig_page != dev->page) 2360*d7bd398eSSong Liu put_page(dev->orig_page); 2361*d7bd398eSSong Liu dev->orig_page = conf->disks[i].extra_page; 2362*d7bd398eSSong Liu } 23631e6d690bSSong Liu } 23642ded3703SSong Liu 23652ded3703SSong Liu /* 23662ded3703SSong Liu * clean up the stripe (clear R5_InJournal for dev[pd_idx] etc.) after the 23672ded3703SSong Liu * stripe is committed to RAID disks. 23682ded3703SSong Liu */ 23692ded3703SSong Liu void r5c_finish_stripe_write_out(struct r5conf *conf, 23702ded3703SSong Liu struct stripe_head *sh, 23712ded3703SSong Liu struct stripe_head_state *s) 23722ded3703SSong Liu { 23731e6d690bSSong Liu int i; 23741e6d690bSSong Liu int do_wakeup = 0; 23751e6d690bSSong Liu 23762ded3703SSong Liu if (!conf->log || 23772ded3703SSong Liu !test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags)) 23782ded3703SSong Liu return; 23792ded3703SSong Liu 23802ded3703SSong Liu WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state)); 23812ded3703SSong Liu clear_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags); 23822ded3703SSong Liu 23832ded3703SSong Liu if (conf->log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) 23842ded3703SSong Liu return; 23851e6d690bSSong Liu 23861e6d690bSSong Liu for (i = sh->disks; i--; ) { 23871e6d690bSSong Liu clear_bit(R5_InJournal, &sh->dev[i].flags); 23881e6d690bSSong Liu if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags)) 23891e6d690bSSong Liu do_wakeup = 1; 23901e6d690bSSong Liu } 23911e6d690bSSong Liu 23921e6d690bSSong Liu /* 23931e6d690bSSong Liu * analyse_stripe() runs before r5c_finish_stripe_write_out(), 23941e6d690bSSong Liu * We updated R5_InJournal, so we also update s->injournal. 23951e6d690bSSong Liu */ 23961e6d690bSSong Liu s->injournal = 0; 23971e6d690bSSong Liu 23981e6d690bSSong Liu if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state)) 23991e6d690bSSong Liu if (atomic_dec_and_test(&conf->pending_full_writes)) 24001e6d690bSSong Liu md_wakeup_thread(conf->mddev->thread); 24011e6d690bSSong Liu 24021e6d690bSSong Liu if (do_wakeup) 24031e6d690bSSong Liu wake_up(&conf->wait_for_overlap); 2404a39f7afdSSong Liu 2405a39f7afdSSong Liu if (conf->log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) 2406a39f7afdSSong Liu return; 2407a39f7afdSSong Liu 2408a39f7afdSSong Liu spin_lock_irq(&conf->log->stripe_in_journal_lock); 2409a39f7afdSSong Liu list_del_init(&sh->r5c); 2410a39f7afdSSong Liu spin_unlock_irq(&conf->log->stripe_in_journal_lock); 2411a39f7afdSSong Liu sh->log_start = MaxSector; 2412a39f7afdSSong Liu atomic_dec(&conf->log->stripe_in_journal_count); 24131e6d690bSSong Liu } 24141e6d690bSSong Liu 24151e6d690bSSong Liu int 24161e6d690bSSong Liu r5c_cache_data(struct r5l_log *log, struct stripe_head *sh, 24171e6d690bSSong Liu struct stripe_head_state *s) 24181e6d690bSSong Liu { 2419a39f7afdSSong Liu struct r5conf *conf = sh->raid_conf; 24201e6d690bSSong Liu int pages = 0; 24211e6d690bSSong Liu int reserve; 24221e6d690bSSong Liu int i; 24231e6d690bSSong Liu int ret = 0; 24241e6d690bSSong Liu 24251e6d690bSSong Liu BUG_ON(!log); 24261e6d690bSSong Liu 24271e6d690bSSong Liu for (i = 0; i < sh->disks; i++) { 24281e6d690bSSong Liu void *addr; 24291e6d690bSSong Liu 24301e6d690bSSong Liu if (!test_bit(R5_Wantwrite, &sh->dev[i].flags)) 24311e6d690bSSong Liu continue; 24321e6d690bSSong Liu addr = kmap_atomic(sh->dev[i].page); 24331e6d690bSSong Liu sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum, 24341e6d690bSSong Liu addr, PAGE_SIZE); 24351e6d690bSSong Liu kunmap_atomic(addr); 24361e6d690bSSong Liu pages++; 24371e6d690bSSong Liu } 24381e6d690bSSong Liu WARN_ON(pages == 0); 24391e6d690bSSong Liu 24401e6d690bSSong Liu /* 24411e6d690bSSong Liu * The stripe must enter state machine again to call endio, so 24421e6d690bSSong Liu * don't delay. 24431e6d690bSSong Liu */ 24441e6d690bSSong Liu clear_bit(STRIPE_DELAYED, &sh->state); 24451e6d690bSSong Liu atomic_inc(&sh->count); 24461e6d690bSSong Liu 24471e6d690bSSong Liu mutex_lock(&log->io_mutex); 24481e6d690bSSong Liu /* meta + data */ 24491e6d690bSSong Liu reserve = (1 + pages) << (PAGE_SHIFT - 9); 24501e6d690bSSong Liu 2451a39f7afdSSong Liu if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) && 2452a39f7afdSSong Liu sh->log_start == MaxSector) 2453a39f7afdSSong Liu r5l_add_no_space_stripe(log, sh); 2454a39f7afdSSong Liu else if (!r5l_has_free_space(log, reserve)) { 2455a39f7afdSSong Liu if (sh->log_start == log->last_checkpoint) 2456a39f7afdSSong Liu BUG(); 2457a39f7afdSSong Liu else 2458a39f7afdSSong Liu r5l_add_no_space_stripe(log, sh); 24591e6d690bSSong Liu } else { 24601e6d690bSSong Liu ret = r5l_log_stripe(log, sh, pages, 0); 24611e6d690bSSong Liu if (ret) { 24621e6d690bSSong Liu spin_lock_irq(&log->io_list_lock); 24631e6d690bSSong Liu list_add_tail(&sh->log_list, &log->no_mem_stripes); 24641e6d690bSSong Liu spin_unlock_irq(&log->io_list_lock); 24651e6d690bSSong Liu } 24661e6d690bSSong Liu } 24671e6d690bSSong Liu 24681e6d690bSSong Liu mutex_unlock(&log->io_mutex); 24691e6d690bSSong Liu return 0; 24702ded3703SSong Liu } 24712ded3703SSong Liu 2472f6bed0efSShaohua Li static int r5l_load_log(struct r5l_log *log) 2473f6bed0efSShaohua Li { 2474f6bed0efSShaohua Li struct md_rdev *rdev = log->rdev; 2475f6bed0efSShaohua Li struct page *page; 2476f6bed0efSShaohua Li struct r5l_meta_block *mb; 2477f6bed0efSShaohua Li sector_t cp = log->rdev->journal_tail; 2478f6bed0efSShaohua Li u32 stored_crc, expected_crc; 2479f6bed0efSShaohua Li bool create_super = false; 2480f6bed0efSShaohua Li int ret; 2481f6bed0efSShaohua Li 2482f6bed0efSShaohua Li /* Make sure it's valid */ 2483f6bed0efSShaohua Li if (cp >= rdev->sectors || round_down(cp, BLOCK_SECTORS) != cp) 2484f6bed0efSShaohua Li cp = 0; 2485f6bed0efSShaohua Li page = alloc_page(GFP_KERNEL); 2486f6bed0efSShaohua Li if (!page) 2487f6bed0efSShaohua Li return -ENOMEM; 2488f6bed0efSShaohua Li 2489796a5cf0SMike Christie if (!sync_page_io(rdev, cp, PAGE_SIZE, page, REQ_OP_READ, 0, false)) { 2490f6bed0efSShaohua Li ret = -EIO; 2491f6bed0efSShaohua Li goto ioerr; 2492f6bed0efSShaohua Li } 2493f6bed0efSShaohua Li mb = page_address(page); 2494f6bed0efSShaohua Li 2495f6bed0efSShaohua Li if (le32_to_cpu(mb->magic) != R5LOG_MAGIC || 2496f6bed0efSShaohua Li mb->version != R5LOG_VERSION) { 2497f6bed0efSShaohua Li create_super = true; 2498f6bed0efSShaohua Li goto create; 2499f6bed0efSShaohua Li } 2500f6bed0efSShaohua Li stored_crc = le32_to_cpu(mb->checksum); 2501f6bed0efSShaohua Li mb->checksum = 0; 25025cb2fbd6SShaohua Li expected_crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE); 2503f6bed0efSShaohua Li if (stored_crc != expected_crc) { 2504f6bed0efSShaohua Li create_super = true; 2505f6bed0efSShaohua Li goto create; 2506f6bed0efSShaohua Li } 2507f6bed0efSShaohua Li if (le64_to_cpu(mb->position) != cp) { 2508f6bed0efSShaohua Li create_super = true; 2509f6bed0efSShaohua Li goto create; 2510f6bed0efSShaohua Li } 2511f6bed0efSShaohua Li create: 2512f6bed0efSShaohua Li if (create_super) { 2513f6bed0efSShaohua Li log->last_cp_seq = prandom_u32(); 2514f6bed0efSShaohua Li cp = 0; 251556056c2eSZhengyuan Liu r5l_log_write_empty_meta_block(log, cp, log->last_cp_seq); 2516f6bed0efSShaohua Li /* 2517f6bed0efSShaohua Li * Make sure super points to correct address. Log might have 2518f6bed0efSShaohua Li * data very soon. If super hasn't correct log tail address, 2519f6bed0efSShaohua Li * recovery can't find the log 2520f6bed0efSShaohua Li */ 2521f6bed0efSShaohua Li r5l_write_super(log, cp); 2522f6bed0efSShaohua Li } else 2523f6bed0efSShaohua Li log->last_cp_seq = le64_to_cpu(mb->seq); 2524f6bed0efSShaohua Li 2525f6bed0efSShaohua Li log->device_size = round_down(rdev->sectors, BLOCK_SECTORS); 25260576b1c6SShaohua Li log->max_free_space = log->device_size >> RECLAIM_MAX_FREE_SPACE_SHIFT; 25270576b1c6SShaohua Li if (log->max_free_space > RECLAIM_MAX_FREE_SPACE) 25280576b1c6SShaohua Li log->max_free_space = RECLAIM_MAX_FREE_SPACE; 2529f6bed0efSShaohua Li log->last_checkpoint = cp; 253028cd88e2SZhengyuan Liu log->next_checkpoint = cp; 2531a39f7afdSSong Liu mutex_lock(&log->io_mutex); 2532a39f7afdSSong Liu r5c_update_log_state(log); 2533a39f7afdSSong Liu mutex_unlock(&log->io_mutex); 2534f6bed0efSShaohua Li 2535f6bed0efSShaohua Li __free_page(page); 2536f6bed0efSShaohua Li 2537f6bed0efSShaohua Li return r5l_recovery_log(log); 2538f6bed0efSShaohua Li ioerr: 2539f6bed0efSShaohua Li __free_page(page); 2540f6bed0efSShaohua Li return ret; 2541f6bed0efSShaohua Li } 2542f6bed0efSShaohua Li 2543f6bed0efSShaohua Li int r5l_init_log(struct r5conf *conf, struct md_rdev *rdev) 2544f6bed0efSShaohua Li { 2545c888a8f9SJens Axboe struct request_queue *q = bdev_get_queue(rdev->bdev); 2546f6bed0efSShaohua Li struct r5l_log *log; 2547f6bed0efSShaohua Li 2548f6bed0efSShaohua Li if (PAGE_SIZE != 4096) 2549f6bed0efSShaohua Li return -EINVAL; 2550c757ec95SSong Liu 2551c757ec95SSong Liu /* 2552c757ec95SSong Liu * The PAGE_SIZE must be big enough to hold 1 r5l_meta_block and 2553c757ec95SSong Liu * raid_disks r5l_payload_data_parity. 2554c757ec95SSong Liu * 2555c757ec95SSong Liu * Write journal and cache does not work for very big array 2556c757ec95SSong Liu * (raid_disks > 203) 2557c757ec95SSong Liu */ 2558c757ec95SSong Liu if (sizeof(struct r5l_meta_block) + 2559c757ec95SSong Liu ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32)) * 2560c757ec95SSong Liu conf->raid_disks) > PAGE_SIZE) { 2561c757ec95SSong Liu pr_err("md/raid:%s: write journal/cache doesn't work for array with %d disks\n", 2562c757ec95SSong Liu mdname(conf->mddev), conf->raid_disks); 2563c757ec95SSong Liu return -EINVAL; 2564c757ec95SSong Liu } 2565c757ec95SSong Liu 2566f6bed0efSShaohua Li log = kzalloc(sizeof(*log), GFP_KERNEL); 2567f6bed0efSShaohua Li if (!log) 2568f6bed0efSShaohua Li return -ENOMEM; 2569f6bed0efSShaohua Li log->rdev = rdev; 2570f6bed0efSShaohua Li 2571c888a8f9SJens Axboe log->need_cache_flush = test_bit(QUEUE_FLAG_WC, &q->queue_flags) != 0; 257256fef7c6SChristoph Hellwig 25735cb2fbd6SShaohua Li log->uuid_checksum = crc32c_le(~0, rdev->mddev->uuid, 2574f6bed0efSShaohua Li sizeof(rdev->mddev->uuid)); 2575f6bed0efSShaohua Li 2576f6bed0efSShaohua Li mutex_init(&log->io_mutex); 2577f6bed0efSShaohua Li 2578f6bed0efSShaohua Li spin_lock_init(&log->io_list_lock); 2579f6bed0efSShaohua Li INIT_LIST_HEAD(&log->running_ios); 25800576b1c6SShaohua Li INIT_LIST_HEAD(&log->io_end_ios); 2581a8c34f91SShaohua Li INIT_LIST_HEAD(&log->flushing_ios); 258204732f74SChristoph Hellwig INIT_LIST_HEAD(&log->finished_ios); 2583a8c34f91SShaohua Li bio_init(&log->flush_bio); 2584f6bed0efSShaohua Li 2585f6bed0efSShaohua Li log->io_kc = KMEM_CACHE(r5l_io_unit, 0); 2586f6bed0efSShaohua Li if (!log->io_kc) 2587f6bed0efSShaohua Li goto io_kc; 2588f6bed0efSShaohua Li 25895036c390SChristoph Hellwig log->io_pool = mempool_create_slab_pool(R5L_POOL_SIZE, log->io_kc); 25905036c390SChristoph Hellwig if (!log->io_pool) 25915036c390SChristoph Hellwig goto io_pool; 25925036c390SChristoph Hellwig 2593c38d29b3SChristoph Hellwig log->bs = bioset_create(R5L_POOL_SIZE, 0); 2594c38d29b3SChristoph Hellwig if (!log->bs) 2595c38d29b3SChristoph Hellwig goto io_bs; 2596c38d29b3SChristoph Hellwig 2597e8deb638SChristoph Hellwig log->meta_pool = mempool_create_page_pool(R5L_POOL_SIZE, 0); 2598e8deb638SChristoph Hellwig if (!log->meta_pool) 2599e8deb638SChristoph Hellwig goto out_mempool; 2600e8deb638SChristoph Hellwig 26010576b1c6SShaohua Li log->reclaim_thread = md_register_thread(r5l_reclaim_thread, 26020576b1c6SShaohua Li log->rdev->mddev, "reclaim"); 26030576b1c6SShaohua Li if (!log->reclaim_thread) 26040576b1c6SShaohua Li goto reclaim_thread; 2605a39f7afdSSong Liu log->reclaim_thread->timeout = R5C_RECLAIM_WAKEUP_INTERVAL; 2606a39f7afdSSong Liu 26070fd22b45SShaohua Li init_waitqueue_head(&log->iounit_wait); 26080576b1c6SShaohua Li 26095036c390SChristoph Hellwig INIT_LIST_HEAD(&log->no_mem_stripes); 26105036c390SChristoph Hellwig 2611f6bed0efSShaohua Li INIT_LIST_HEAD(&log->no_space_stripes); 2612f6bed0efSShaohua Li spin_lock_init(&log->no_space_stripes_lock); 2613f6bed0efSShaohua Li 26143bddb7f8SSong Liu INIT_WORK(&log->deferred_io_work, r5l_submit_io_async); 26153bddb7f8SSong Liu 26162ded3703SSong Liu log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH; 2617a39f7afdSSong Liu INIT_LIST_HEAD(&log->stripe_in_journal_list); 2618a39f7afdSSong Liu spin_lock_init(&log->stripe_in_journal_lock); 2619a39f7afdSSong Liu atomic_set(&log->stripe_in_journal_count, 0); 26202ded3703SSong Liu 2621f6bed0efSShaohua Li if (r5l_load_log(log)) 2622f6bed0efSShaohua Li goto error; 2623f6bed0efSShaohua Li 2624f6b6ec5cSShaohua Li rcu_assign_pointer(conf->log, log); 2625a62ab49eSShaohua Li set_bit(MD_HAS_JOURNAL, &conf->mddev->flags); 2626f6bed0efSShaohua Li return 0; 2627e8deb638SChristoph Hellwig 2628f6bed0efSShaohua Li error: 26290576b1c6SShaohua Li md_unregister_thread(&log->reclaim_thread); 26300576b1c6SShaohua Li reclaim_thread: 2631e8deb638SChristoph Hellwig mempool_destroy(log->meta_pool); 2632e8deb638SChristoph Hellwig out_mempool: 2633c38d29b3SChristoph Hellwig bioset_free(log->bs); 2634c38d29b3SChristoph Hellwig io_bs: 26355036c390SChristoph Hellwig mempool_destroy(log->io_pool); 26365036c390SChristoph Hellwig io_pool: 2637f6bed0efSShaohua Li kmem_cache_destroy(log->io_kc); 2638f6bed0efSShaohua Li io_kc: 2639f6bed0efSShaohua Li kfree(log); 2640f6bed0efSShaohua Li return -EINVAL; 2641f6bed0efSShaohua Li } 2642f6bed0efSShaohua Li 2643f6bed0efSShaohua Li void r5l_exit_log(struct r5l_log *log) 2644f6bed0efSShaohua Li { 26450576b1c6SShaohua Li md_unregister_thread(&log->reclaim_thread); 2646e8deb638SChristoph Hellwig mempool_destroy(log->meta_pool); 2647c38d29b3SChristoph Hellwig bioset_free(log->bs); 26485036c390SChristoph Hellwig mempool_destroy(log->io_pool); 2649f6bed0efSShaohua Li kmem_cache_destroy(log->io_kc); 2650f6bed0efSShaohua Li kfree(log); 2651f6bed0efSShaohua Li } 2652