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> 2303b047f4SSong Liu #include <linux/types.h> 24f6bed0efSShaohua Li #include "md.h" 25f6bed0efSShaohua Li #include "raid5.h" 261e6d690bSSong Liu #include "bitmap.h" 27f6bed0efSShaohua Li 28f6bed0efSShaohua Li /* 29f6bed0efSShaohua Li * metadata/data stored in disk with 4k size unit (a block) regardless 30f6bed0efSShaohua Li * underneath hardware sector size. only works with PAGE_SIZE == 4096 31f6bed0efSShaohua Li */ 32f6bed0efSShaohua Li #define BLOCK_SECTORS (8) 33effe6ee7SSong Liu #define BLOCK_SECTOR_SHIFT (3) 34f6bed0efSShaohua Li 350576b1c6SShaohua Li /* 36a39f7afdSSong Liu * log->max_free_space is min(1/4 disk size, 10G reclaimable space). 37a39f7afdSSong Liu * 38a39f7afdSSong Liu * In write through mode, the reclaim runs every log->max_free_space. 39a39f7afdSSong Liu * This can prevent the recovery scans for too long 400576b1c6SShaohua Li */ 410576b1c6SShaohua Li #define RECLAIM_MAX_FREE_SPACE (10 * 1024 * 1024 * 2) /* sector */ 420576b1c6SShaohua Li #define RECLAIM_MAX_FREE_SPACE_SHIFT (2) 430576b1c6SShaohua Li 44a39f7afdSSong Liu /* wake up reclaim thread periodically */ 45a39f7afdSSong Liu #define R5C_RECLAIM_WAKEUP_INTERVAL (30 * HZ) 46a39f7afdSSong Liu /* start flush with these full stripes */ 4784890c03SShaohua Li #define R5C_FULL_STRIPE_FLUSH_BATCH(conf) (conf->max_nr_stripes / 4) 48a39f7afdSSong Liu /* reclaim stripes in groups */ 49a39f7afdSSong Liu #define R5C_RECLAIM_STRIPE_GROUP (NR_STRIPE_HASH_LOCKS * 2) 50a39f7afdSSong Liu 51c38d29b3SChristoph Hellwig /* 52c38d29b3SChristoph Hellwig * We only need 2 bios per I/O unit to make progress, but ensure we 53c38d29b3SChristoph Hellwig * have a few more available to not get too tight. 54c38d29b3SChristoph Hellwig */ 55c38d29b3SChristoph Hellwig #define R5L_POOL_SIZE 4 56c38d29b3SChristoph Hellwig 572ded3703SSong Liu /* 582ded3703SSong Liu * r5c journal modes of the array: write-back or write-through. 592ded3703SSong Liu * write-through mode has identical behavior as existing log only 602ded3703SSong Liu * implementation. 612ded3703SSong Liu */ 622ded3703SSong Liu enum r5c_journal_mode { 632ded3703SSong Liu R5C_JOURNAL_MODE_WRITE_THROUGH = 0, 642ded3703SSong Liu R5C_JOURNAL_MODE_WRITE_BACK = 1, 652ded3703SSong Liu }; 662ded3703SSong Liu 672c7da14bSSong Liu static char *r5c_journal_mode_str[] = {"write-through", 682c7da14bSSong Liu "write-back"}; 692ded3703SSong Liu /* 702ded3703SSong Liu * raid5 cache state machine 712ded3703SSong Liu * 729b69173eSJackieLiu * With the RAID cache, each stripe works in two phases: 732ded3703SSong Liu * - caching phase 742ded3703SSong Liu * - writing-out phase 752ded3703SSong Liu * 762ded3703SSong Liu * These two phases are controlled by bit STRIPE_R5C_CACHING: 772ded3703SSong Liu * if STRIPE_R5C_CACHING == 0, the stripe is in writing-out phase 782ded3703SSong Liu * if STRIPE_R5C_CACHING == 1, the stripe is in caching phase 792ded3703SSong Liu * 802ded3703SSong Liu * When there is no journal, or the journal is in write-through mode, 812ded3703SSong Liu * the stripe is always in writing-out phase. 822ded3703SSong Liu * 832ded3703SSong Liu * For write-back journal, the stripe is sent to caching phase on write 842ded3703SSong Liu * (r5c_try_caching_write). r5c_make_stripe_write_out() kicks off 852ded3703SSong Liu * the write-out phase by clearing STRIPE_R5C_CACHING. 862ded3703SSong Liu * 872ded3703SSong Liu * Stripes in caching phase do not write the raid disks. Instead, all 882ded3703SSong Liu * writes are committed from the log device. Therefore, a stripe in 892ded3703SSong Liu * caching phase handles writes as: 902ded3703SSong Liu * - write to log device 912ded3703SSong Liu * - return IO 922ded3703SSong Liu * 932ded3703SSong Liu * Stripes in writing-out phase handle writes as: 942ded3703SSong Liu * - calculate parity 952ded3703SSong Liu * - write pending data and parity to journal 962ded3703SSong Liu * - write data and parity to raid disks 972ded3703SSong Liu * - return IO for pending writes 982ded3703SSong Liu */ 992ded3703SSong Liu 100f6bed0efSShaohua Li struct r5l_log { 101f6bed0efSShaohua Li struct md_rdev *rdev; 102f6bed0efSShaohua Li 103f6bed0efSShaohua Li u32 uuid_checksum; 104f6bed0efSShaohua Li 105f6bed0efSShaohua Li sector_t device_size; /* log device size, round to 106f6bed0efSShaohua Li * BLOCK_SECTORS */ 1070576b1c6SShaohua Li sector_t max_free_space; /* reclaim run if free space is at 1080576b1c6SShaohua Li * this size */ 109f6bed0efSShaohua Li 110f6bed0efSShaohua Li sector_t last_checkpoint; /* log tail. where recovery scan 111f6bed0efSShaohua Li * starts from */ 112f6bed0efSShaohua Li u64 last_cp_seq; /* log tail sequence */ 113f6bed0efSShaohua Li 114f6bed0efSShaohua Li sector_t log_start; /* log head. where new data appends */ 115f6bed0efSShaohua Li u64 seq; /* log head sequence */ 116f6bed0efSShaohua Li 11717036461SChristoph Hellwig sector_t next_checkpoint; 11817036461SChristoph Hellwig 119f6bed0efSShaohua Li struct mutex io_mutex; 120f6bed0efSShaohua Li struct r5l_io_unit *current_io; /* current io_unit accepting new data */ 121f6bed0efSShaohua Li 122f6bed0efSShaohua Li spinlock_t io_list_lock; 123f6bed0efSShaohua Li struct list_head running_ios; /* io_units which are still running, 124f6bed0efSShaohua Li * and have not yet been completely 125f6bed0efSShaohua Li * written to the log */ 126f6bed0efSShaohua Li struct list_head io_end_ios; /* io_units which have been completely 127f6bed0efSShaohua Li * written to the log but not yet written 128f6bed0efSShaohua Li * to the RAID */ 129a8c34f91SShaohua Li struct list_head flushing_ios; /* io_units which are waiting for log 130a8c34f91SShaohua Li * cache flush */ 13104732f74SChristoph Hellwig struct list_head finished_ios; /* io_units which settle down in log disk */ 132a8c34f91SShaohua Li struct bio flush_bio; 133f6bed0efSShaohua Li 1345036c390SChristoph Hellwig struct list_head no_mem_stripes; /* pending stripes, -ENOMEM */ 1355036c390SChristoph Hellwig 136f6bed0efSShaohua Li struct kmem_cache *io_kc; 1375036c390SChristoph Hellwig mempool_t *io_pool; 138c38d29b3SChristoph Hellwig struct bio_set *bs; 139e8deb638SChristoph Hellwig mempool_t *meta_pool; 140f6bed0efSShaohua Li 1410576b1c6SShaohua Li struct md_thread *reclaim_thread; 1420576b1c6SShaohua Li unsigned long reclaim_target; /* number of space that need to be 1430576b1c6SShaohua Li * reclaimed. if it's 0, reclaim spaces 1440576b1c6SShaohua Li * used by io_units which are in 1450576b1c6SShaohua Li * IO_UNIT_STRIPE_END state (eg, reclaim 1460576b1c6SShaohua Li * dones't wait for specific io_unit 1470576b1c6SShaohua Li * switching to IO_UNIT_STRIPE_END 1480576b1c6SShaohua Li * state) */ 1490fd22b45SShaohua Li wait_queue_head_t iounit_wait; 1500576b1c6SShaohua Li 151f6bed0efSShaohua Li struct list_head no_space_stripes; /* pending stripes, log has no space */ 152f6bed0efSShaohua Li spinlock_t no_space_stripes_lock; 15356fef7c6SChristoph Hellwig 15456fef7c6SChristoph Hellwig bool need_cache_flush; 1552ded3703SSong Liu 1562ded3703SSong Liu /* for r5c_cache */ 1572ded3703SSong Liu enum r5c_journal_mode r5c_journal_mode; 158a39f7afdSSong Liu 159a39f7afdSSong Liu /* all stripes in r5cache, in the order of seq at sh->log_start */ 160a39f7afdSSong Liu struct list_head stripe_in_journal_list; 161a39f7afdSSong Liu 162a39f7afdSSong Liu spinlock_t stripe_in_journal_lock; 163a39f7afdSSong Liu atomic_t stripe_in_journal_count; 1643bddb7f8SSong Liu 1653bddb7f8SSong Liu /* to submit async io_units, to fulfill ordering of flush */ 1663bddb7f8SSong Liu struct work_struct deferred_io_work; 1672e38a37fSSong Liu /* to disable write back during in degraded mode */ 1682e38a37fSSong Liu struct work_struct disable_writeback_work; 16903b047f4SSong Liu 17003b047f4SSong Liu /* to for chunk_aligned_read in writeback mode, details below */ 17103b047f4SSong Liu spinlock_t tree_lock; 17203b047f4SSong Liu struct radix_tree_root big_stripe_tree; 173f6bed0efSShaohua Li }; 174f6bed0efSShaohua Li 175f6bed0efSShaohua Li /* 17603b047f4SSong Liu * Enable chunk_aligned_read() with write back cache. 17703b047f4SSong Liu * 17803b047f4SSong Liu * Each chunk may contain more than one stripe (for example, a 256kB 17903b047f4SSong Liu * chunk contains 64 4kB-page, so this chunk contain 64 stripes). For 18003b047f4SSong Liu * chunk_aligned_read, these stripes are grouped into one "big_stripe". 18103b047f4SSong Liu * For each big_stripe, we count how many stripes of this big_stripe 18203b047f4SSong Liu * are in the write back cache. These data are tracked in a radix tree 18303b047f4SSong Liu * (big_stripe_tree). We use radix_tree item pointer as the counter. 18403b047f4SSong Liu * r5c_tree_index() is used to calculate keys for the radix tree. 18503b047f4SSong Liu * 18603b047f4SSong Liu * chunk_aligned_read() calls r5c_big_stripe_cached() to look up 18703b047f4SSong Liu * big_stripe of each chunk in the tree. If this big_stripe is in the 18803b047f4SSong Liu * tree, chunk_aligned_read() aborts. This look up is protected by 18903b047f4SSong Liu * rcu_read_lock(). 19003b047f4SSong Liu * 19103b047f4SSong Liu * It is necessary to remember whether a stripe is counted in 19203b047f4SSong Liu * big_stripe_tree. Instead of adding new flag, we reuses existing flags: 19303b047f4SSong Liu * STRIPE_R5C_PARTIAL_STRIPE and STRIPE_R5C_FULL_STRIPE. If either of these 19403b047f4SSong Liu * two flags are set, the stripe is counted in big_stripe_tree. This 19503b047f4SSong Liu * requires moving set_bit(STRIPE_R5C_PARTIAL_STRIPE) to 19603b047f4SSong Liu * r5c_try_caching_write(); and moving clear_bit of 19703b047f4SSong Liu * STRIPE_R5C_PARTIAL_STRIPE and STRIPE_R5C_FULL_STRIPE to 19803b047f4SSong Liu * r5c_finish_stripe_write_out(). 19903b047f4SSong Liu */ 20003b047f4SSong Liu 20103b047f4SSong Liu /* 20203b047f4SSong Liu * radix tree requests lowest 2 bits of data pointer to be 2b'00. 20303b047f4SSong Liu * So it is necessary to left shift the counter by 2 bits before using it 20403b047f4SSong Liu * as data pointer of the tree. 20503b047f4SSong Liu */ 20603b047f4SSong Liu #define R5C_RADIX_COUNT_SHIFT 2 20703b047f4SSong Liu 20803b047f4SSong Liu /* 20903b047f4SSong Liu * calculate key for big_stripe_tree 21003b047f4SSong Liu * 21103b047f4SSong Liu * sect: align_bi->bi_iter.bi_sector or sh->sector 21203b047f4SSong Liu */ 21303b047f4SSong Liu static inline sector_t r5c_tree_index(struct r5conf *conf, 21403b047f4SSong Liu sector_t sect) 21503b047f4SSong Liu { 21603b047f4SSong Liu sector_t offset; 21703b047f4SSong Liu 21803b047f4SSong Liu offset = sector_div(sect, conf->chunk_sectors); 21903b047f4SSong Liu return sect; 22003b047f4SSong Liu } 22103b047f4SSong Liu 22203b047f4SSong Liu /* 223f6bed0efSShaohua Li * an IO range starts from a meta data block and end at the next meta data 224f6bed0efSShaohua Li * block. The io unit's the meta data block tracks data/parity followed it. io 225f6bed0efSShaohua Li * unit is written to log disk with normal write, as we always flush log disk 226f6bed0efSShaohua Li * first and then start move data to raid disks, there is no requirement to 227f6bed0efSShaohua Li * write io unit with FLUSH/FUA 228f6bed0efSShaohua Li */ 229f6bed0efSShaohua Li struct r5l_io_unit { 230f6bed0efSShaohua Li struct r5l_log *log; 231f6bed0efSShaohua Li 232f6bed0efSShaohua Li struct page *meta_page; /* store meta block */ 233f6bed0efSShaohua Li int meta_offset; /* current offset in meta_page */ 234f6bed0efSShaohua Li 235f6bed0efSShaohua Li struct bio *current_bio;/* current_bio accepting new data */ 236f6bed0efSShaohua Li 237f6bed0efSShaohua Li atomic_t pending_stripe;/* how many stripes not flushed to raid */ 238f6bed0efSShaohua Li u64 seq; /* seq number of the metablock */ 239f6bed0efSShaohua Li sector_t log_start; /* where the io_unit starts */ 240f6bed0efSShaohua Li sector_t log_end; /* where the io_unit ends */ 241f6bed0efSShaohua Li struct list_head log_sibling; /* log->running_ios */ 242f6bed0efSShaohua Li struct list_head stripe_list; /* stripes added to the io_unit */ 243f6bed0efSShaohua Li 244f6bed0efSShaohua Li int state; 2456143e2ceSChristoph Hellwig bool need_split_bio; 2463bddb7f8SSong Liu struct bio *split_bio; 2473bddb7f8SSong Liu 2483bddb7f8SSong Liu unsigned int has_flush:1; /* include flush request */ 2493bddb7f8SSong Liu unsigned int has_fua:1; /* include fua request */ 2503bddb7f8SSong Liu unsigned int has_null_flush:1; /* include empty flush request */ 2513bddb7f8SSong Liu /* 2523bddb7f8SSong Liu * io isn't sent yet, flush/fua request can only be submitted till it's 2533bddb7f8SSong Liu * the first IO in running_ios list 2543bddb7f8SSong Liu */ 2553bddb7f8SSong Liu unsigned int io_deferred:1; 2563bddb7f8SSong Liu 2573bddb7f8SSong Liu struct bio_list flush_barriers; /* size == 0 flush bios */ 258f6bed0efSShaohua Li }; 259f6bed0efSShaohua Li 260f6bed0efSShaohua Li /* r5l_io_unit state */ 261f6bed0efSShaohua Li enum r5l_io_unit_state { 262f6bed0efSShaohua Li IO_UNIT_RUNNING = 0, /* accepting new IO */ 263f6bed0efSShaohua Li IO_UNIT_IO_START = 1, /* io_unit bio start writing to log, 264f6bed0efSShaohua Li * don't accepting new bio */ 265f6bed0efSShaohua Li IO_UNIT_IO_END = 2, /* io_unit bio finish writing to log */ 266a8c34f91SShaohua Li IO_UNIT_STRIPE_END = 3, /* stripes data finished writing to raid */ 267f6bed0efSShaohua Li }; 268f6bed0efSShaohua Li 2692ded3703SSong Liu bool r5c_is_writeback(struct r5l_log *log) 2702ded3703SSong Liu { 2712ded3703SSong Liu return (log != NULL && 2722ded3703SSong Liu log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK); 2732ded3703SSong Liu } 2742ded3703SSong Liu 275f6bed0efSShaohua Li static sector_t r5l_ring_add(struct r5l_log *log, sector_t start, sector_t inc) 276f6bed0efSShaohua Li { 277f6bed0efSShaohua Li start += inc; 278f6bed0efSShaohua Li if (start >= log->device_size) 279f6bed0efSShaohua Li start = start - log->device_size; 280f6bed0efSShaohua Li return start; 281f6bed0efSShaohua Li } 282f6bed0efSShaohua Li 283f6bed0efSShaohua Li static sector_t r5l_ring_distance(struct r5l_log *log, sector_t start, 284f6bed0efSShaohua Li sector_t end) 285f6bed0efSShaohua Li { 286f6bed0efSShaohua Li if (end >= start) 287f6bed0efSShaohua Li return end - start; 288f6bed0efSShaohua Li else 289f6bed0efSShaohua Li return end + log->device_size - start; 290f6bed0efSShaohua Li } 291f6bed0efSShaohua Li 292f6bed0efSShaohua Li static bool r5l_has_free_space(struct r5l_log *log, sector_t size) 293f6bed0efSShaohua Li { 294f6bed0efSShaohua Li sector_t used_size; 295f6bed0efSShaohua Li 296f6bed0efSShaohua Li used_size = r5l_ring_distance(log, log->last_checkpoint, 297f6bed0efSShaohua Li log->log_start); 298f6bed0efSShaohua Li 299f6bed0efSShaohua Li return log->device_size > used_size + size; 300f6bed0efSShaohua Li } 301f6bed0efSShaohua Li 302f6bed0efSShaohua Li static void __r5l_set_io_unit_state(struct r5l_io_unit *io, 303f6bed0efSShaohua Li enum r5l_io_unit_state state) 304f6bed0efSShaohua Li { 305f6bed0efSShaohua Li if (WARN_ON(io->state >= state)) 306f6bed0efSShaohua Li return; 307f6bed0efSShaohua Li io->state = state; 308f6bed0efSShaohua Li } 309f6bed0efSShaohua Li 3101e6d690bSSong Liu static void 3111e6d690bSSong Liu r5c_return_dev_pending_writes(struct r5conf *conf, struct r5dev *dev, 3121e6d690bSSong Liu struct bio_list *return_bi) 3131e6d690bSSong Liu { 3141e6d690bSSong Liu struct bio *wbi, *wbi2; 3151e6d690bSSong Liu 3161e6d690bSSong Liu wbi = dev->written; 3171e6d690bSSong Liu dev->written = NULL; 3181e6d690bSSong Liu while (wbi && wbi->bi_iter.bi_sector < 3191e6d690bSSong Liu dev->sector + STRIPE_SECTORS) { 3201e6d690bSSong Liu wbi2 = r5_next_bio(wbi, dev->sector); 3211e6d690bSSong Liu if (!raid5_dec_bi_active_stripes(wbi)) { 3221e6d690bSSong Liu md_write_end(conf->mddev); 3231e6d690bSSong Liu bio_list_add(return_bi, wbi); 3241e6d690bSSong Liu } 3251e6d690bSSong Liu wbi = wbi2; 3261e6d690bSSong Liu } 3271e6d690bSSong Liu } 3281e6d690bSSong Liu 3291e6d690bSSong Liu void r5c_handle_cached_data_endio(struct r5conf *conf, 3301e6d690bSSong Liu struct stripe_head *sh, int disks, struct bio_list *return_bi) 3311e6d690bSSong Liu { 3321e6d690bSSong Liu int i; 3331e6d690bSSong Liu 3341e6d690bSSong Liu for (i = sh->disks; i--; ) { 3351e6d690bSSong Liu if (sh->dev[i].written) { 3361e6d690bSSong Liu set_bit(R5_UPTODATE, &sh->dev[i].flags); 3371e6d690bSSong Liu r5c_return_dev_pending_writes(conf, &sh->dev[i], 3381e6d690bSSong Liu return_bi); 3391e6d690bSSong Liu bitmap_endwrite(conf->mddev->bitmap, sh->sector, 3401e6d690bSSong Liu STRIPE_SECTORS, 3411e6d690bSSong Liu !test_bit(STRIPE_DEGRADED, &sh->state), 3421e6d690bSSong Liu 0); 3431e6d690bSSong Liu } 3441e6d690bSSong Liu } 3451e6d690bSSong Liu } 3461e6d690bSSong Liu 347ff875738SArtur Paszkiewicz void r5l_wake_reclaim(struct r5l_log *log, sector_t space); 348ff875738SArtur Paszkiewicz 349a39f7afdSSong Liu /* Check whether we should flush some stripes to free up stripe cache */ 350a39f7afdSSong Liu void r5c_check_stripe_cache_usage(struct r5conf *conf) 351a39f7afdSSong Liu { 352a39f7afdSSong Liu int total_cached; 353a39f7afdSSong Liu 354a39f7afdSSong Liu if (!r5c_is_writeback(conf->log)) 355a39f7afdSSong Liu return; 356a39f7afdSSong Liu 357a39f7afdSSong Liu total_cached = atomic_read(&conf->r5c_cached_partial_stripes) + 358a39f7afdSSong Liu atomic_read(&conf->r5c_cached_full_stripes); 359a39f7afdSSong Liu 360a39f7afdSSong Liu /* 361a39f7afdSSong Liu * The following condition is true for either of the following: 362a39f7afdSSong Liu * - stripe cache pressure high: 363a39f7afdSSong Liu * total_cached > 3/4 min_nr_stripes || 364a39f7afdSSong Liu * empty_inactive_list_nr > 0 365a39f7afdSSong Liu * - stripe cache pressure moderate: 366a39f7afdSSong Liu * total_cached > 1/2 min_nr_stripes 367a39f7afdSSong Liu */ 368a39f7afdSSong Liu if (total_cached > conf->min_nr_stripes * 1 / 2 || 369a39f7afdSSong Liu atomic_read(&conf->empty_inactive_list_nr) > 0) 370a39f7afdSSong Liu r5l_wake_reclaim(conf->log, 0); 371a39f7afdSSong Liu } 372a39f7afdSSong Liu 373a39f7afdSSong Liu /* 374a39f7afdSSong Liu * flush cache when there are R5C_FULL_STRIPE_FLUSH_BATCH or more full 375a39f7afdSSong Liu * stripes in the cache 376a39f7afdSSong Liu */ 377a39f7afdSSong Liu void r5c_check_cached_full_stripe(struct r5conf *conf) 378a39f7afdSSong Liu { 379a39f7afdSSong Liu if (!r5c_is_writeback(conf->log)) 380a39f7afdSSong Liu return; 381a39f7afdSSong Liu 382a39f7afdSSong Liu /* 383a39f7afdSSong Liu * wake up reclaim for R5C_FULL_STRIPE_FLUSH_BATCH cached stripes 384a39f7afdSSong Liu * or a full stripe (chunk size / 4k stripes). 385a39f7afdSSong Liu */ 386a39f7afdSSong Liu if (atomic_read(&conf->r5c_cached_full_stripes) >= 38784890c03SShaohua Li min(R5C_FULL_STRIPE_FLUSH_BATCH(conf), 388a39f7afdSSong Liu conf->chunk_sectors >> STRIPE_SHIFT)) 389a39f7afdSSong Liu r5l_wake_reclaim(conf->log, 0); 390a39f7afdSSong Liu } 391a39f7afdSSong Liu 392a39f7afdSSong Liu /* 393a39f7afdSSong Liu * Total log space (in sectors) needed to flush all data in cache 394a39f7afdSSong Liu * 39539b99586SSong Liu * To avoid deadlock due to log space, it is necessary to reserve log 39639b99586SSong Liu * space to flush critical stripes (stripes that occupying log space near 39739b99586SSong Liu * last_checkpoint). This function helps check how much log space is 39839b99586SSong Liu * required to flush all cached stripes. 399a39f7afdSSong Liu * 40039b99586SSong Liu * To reduce log space requirements, two mechanisms are used to give cache 40139b99586SSong Liu * flush higher priorities: 40239b99586SSong Liu * 1. In handle_stripe_dirtying() and schedule_reconstruction(), 40339b99586SSong Liu * stripes ALREADY in journal can be flushed w/o pending writes; 40439b99586SSong Liu * 2. In r5l_write_stripe() and r5c_cache_data(), stripes NOT in journal 40539b99586SSong Liu * can be delayed (r5l_add_no_space_stripe). 406a39f7afdSSong Liu * 40739b99586SSong Liu * In cache flush, the stripe goes through 1 and then 2. For a stripe that 40839b99586SSong Liu * already passed 1, flushing it requires at most (conf->max_degraded + 1) 40939b99586SSong Liu * pages of journal space. For stripes that has not passed 1, flushing it 41039b99586SSong Liu * requires (conf->raid_disks + 1) pages of journal space. There are at 41139b99586SSong Liu * most (conf->group_cnt + 1) stripe that passed 1. So total journal space 41239b99586SSong Liu * required to flush all cached stripes (in pages) is: 41339b99586SSong Liu * 41439b99586SSong Liu * (stripe_in_journal_count - group_cnt - 1) * (max_degraded + 1) + 41539b99586SSong Liu * (group_cnt + 1) * (raid_disks + 1) 41639b99586SSong Liu * or 41739b99586SSong Liu * (stripe_in_journal_count) * (max_degraded + 1) + 41839b99586SSong Liu * (group_cnt + 1) * (raid_disks - max_degraded) 419a39f7afdSSong Liu */ 420a39f7afdSSong Liu static sector_t r5c_log_required_to_flush_cache(struct r5conf *conf) 421a39f7afdSSong Liu { 422a39f7afdSSong Liu struct r5l_log *log = conf->log; 423a39f7afdSSong Liu 424a39f7afdSSong Liu if (!r5c_is_writeback(log)) 425a39f7afdSSong Liu return 0; 426a39f7afdSSong Liu 42739b99586SSong Liu return BLOCK_SECTORS * 42839b99586SSong Liu ((conf->max_degraded + 1) * atomic_read(&log->stripe_in_journal_count) + 42939b99586SSong Liu (conf->raid_disks - conf->max_degraded) * (conf->group_cnt + 1)); 430a39f7afdSSong Liu } 431a39f7afdSSong Liu 432a39f7afdSSong Liu /* 433a39f7afdSSong Liu * evaluate log space usage and update R5C_LOG_TIGHT and R5C_LOG_CRITICAL 434a39f7afdSSong Liu * 435a39f7afdSSong Liu * R5C_LOG_TIGHT is set when free space on the log device is less than 3x of 436a39f7afdSSong Liu * reclaim_required_space. R5C_LOG_CRITICAL is set when free space on the log 437a39f7afdSSong Liu * device is less than 2x of reclaim_required_space. 438a39f7afdSSong Liu */ 439a39f7afdSSong Liu static inline void r5c_update_log_state(struct r5l_log *log) 440a39f7afdSSong Liu { 441a39f7afdSSong Liu struct r5conf *conf = log->rdev->mddev->private; 442a39f7afdSSong Liu sector_t free_space; 443a39f7afdSSong Liu sector_t reclaim_space; 444f687a33eSSong Liu bool wake_reclaim = false; 445a39f7afdSSong Liu 446a39f7afdSSong Liu if (!r5c_is_writeback(log)) 447a39f7afdSSong Liu return; 448a39f7afdSSong Liu 449a39f7afdSSong Liu free_space = r5l_ring_distance(log, log->log_start, 450a39f7afdSSong Liu log->last_checkpoint); 451a39f7afdSSong Liu reclaim_space = r5c_log_required_to_flush_cache(conf); 452a39f7afdSSong Liu if (free_space < 2 * reclaim_space) 453a39f7afdSSong Liu set_bit(R5C_LOG_CRITICAL, &conf->cache_state); 454f687a33eSSong Liu else { 455f687a33eSSong Liu if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state)) 456f687a33eSSong Liu wake_reclaim = true; 457a39f7afdSSong Liu clear_bit(R5C_LOG_CRITICAL, &conf->cache_state); 458f687a33eSSong Liu } 459a39f7afdSSong Liu if (free_space < 3 * reclaim_space) 460a39f7afdSSong Liu set_bit(R5C_LOG_TIGHT, &conf->cache_state); 461a39f7afdSSong Liu else 462a39f7afdSSong Liu clear_bit(R5C_LOG_TIGHT, &conf->cache_state); 463f687a33eSSong Liu 464f687a33eSSong Liu if (wake_reclaim) 465f687a33eSSong Liu r5l_wake_reclaim(log, 0); 466a39f7afdSSong Liu } 467a39f7afdSSong Liu 4682ded3703SSong Liu /* 4692ded3703SSong Liu * Put the stripe into writing-out phase by clearing STRIPE_R5C_CACHING. 4702ded3703SSong Liu * This function should only be called in write-back mode. 4712ded3703SSong Liu */ 472a39f7afdSSong Liu void r5c_make_stripe_write_out(struct stripe_head *sh) 4732ded3703SSong Liu { 4742ded3703SSong Liu struct r5conf *conf = sh->raid_conf; 4752ded3703SSong Liu struct r5l_log *log = conf->log; 4762ded3703SSong Liu 4772ded3703SSong Liu BUG_ON(!r5c_is_writeback(log)); 4782ded3703SSong Liu 4792ded3703SSong Liu WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state)); 4802ded3703SSong Liu clear_bit(STRIPE_R5C_CACHING, &sh->state); 4811e6d690bSSong Liu 4821e6d690bSSong Liu if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state)) 4831e6d690bSSong Liu atomic_inc(&conf->preread_active_stripes); 4841e6d690bSSong Liu } 4851e6d690bSSong Liu 4861e6d690bSSong Liu static void r5c_handle_data_cached(struct stripe_head *sh) 4871e6d690bSSong Liu { 4881e6d690bSSong Liu int i; 4891e6d690bSSong Liu 4901e6d690bSSong Liu for (i = sh->disks; i--; ) 4911e6d690bSSong Liu if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) { 4921e6d690bSSong Liu set_bit(R5_InJournal, &sh->dev[i].flags); 4931e6d690bSSong Liu clear_bit(R5_LOCKED, &sh->dev[i].flags); 4941e6d690bSSong Liu } 4951e6d690bSSong Liu clear_bit(STRIPE_LOG_TRAPPED, &sh->state); 4961e6d690bSSong Liu } 4971e6d690bSSong Liu 4981e6d690bSSong Liu /* 4991e6d690bSSong Liu * this journal write must contain full parity, 5001e6d690bSSong Liu * it may also contain some data pages 5011e6d690bSSong Liu */ 5021e6d690bSSong Liu static void r5c_handle_parity_cached(struct stripe_head *sh) 5031e6d690bSSong Liu { 5041e6d690bSSong Liu int i; 5051e6d690bSSong Liu 5061e6d690bSSong Liu for (i = sh->disks; i--; ) 5071e6d690bSSong Liu if (test_bit(R5_InJournal, &sh->dev[i].flags)) 5081e6d690bSSong Liu set_bit(R5_Wantwrite, &sh->dev[i].flags); 5092ded3703SSong Liu } 5102ded3703SSong Liu 5112ded3703SSong Liu /* 5122ded3703SSong Liu * Setting proper flags after writing (or flushing) data and/or parity to the 5132ded3703SSong Liu * log device. This is called from r5l_log_endio() or r5l_log_flush_endio(). 5142ded3703SSong Liu */ 5152ded3703SSong Liu static void r5c_finish_cache_stripe(struct stripe_head *sh) 5162ded3703SSong Liu { 5172ded3703SSong Liu struct r5l_log *log = sh->raid_conf->log; 5182ded3703SSong Liu 5192ded3703SSong Liu if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) { 5202ded3703SSong Liu BUG_ON(test_bit(STRIPE_R5C_CACHING, &sh->state)); 5212ded3703SSong Liu /* 5222ded3703SSong Liu * Set R5_InJournal for parity dev[pd_idx]. This means 5232ded3703SSong Liu * all data AND parity in the journal. For RAID 6, it is 5242ded3703SSong Liu * NOT necessary to set the flag for dev[qd_idx], as the 5252ded3703SSong Liu * two parities are written out together. 5262ded3703SSong Liu */ 5272ded3703SSong Liu set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags); 5281e6d690bSSong Liu } else if (test_bit(STRIPE_R5C_CACHING, &sh->state)) { 5291e6d690bSSong Liu r5c_handle_data_cached(sh); 5301e6d690bSSong Liu } else { 5311e6d690bSSong Liu r5c_handle_parity_cached(sh); 5321e6d690bSSong Liu set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags); 5331e6d690bSSong Liu } 5342ded3703SSong Liu } 5352ded3703SSong Liu 536d8858f43SChristoph Hellwig static void r5l_io_run_stripes(struct r5l_io_unit *io) 537d8858f43SChristoph Hellwig { 538d8858f43SChristoph Hellwig struct stripe_head *sh, *next; 539d8858f43SChristoph Hellwig 540d8858f43SChristoph Hellwig list_for_each_entry_safe(sh, next, &io->stripe_list, log_list) { 541d8858f43SChristoph Hellwig list_del_init(&sh->log_list); 5422ded3703SSong Liu 5432ded3703SSong Liu r5c_finish_cache_stripe(sh); 5442ded3703SSong Liu 545d8858f43SChristoph Hellwig set_bit(STRIPE_HANDLE, &sh->state); 546d8858f43SChristoph Hellwig raid5_release_stripe(sh); 547d8858f43SChristoph Hellwig } 548d8858f43SChristoph Hellwig } 549d8858f43SChristoph Hellwig 55056fef7c6SChristoph Hellwig static void r5l_log_run_stripes(struct r5l_log *log) 55156fef7c6SChristoph Hellwig { 55256fef7c6SChristoph Hellwig struct r5l_io_unit *io, *next; 55356fef7c6SChristoph Hellwig 55456fef7c6SChristoph Hellwig assert_spin_locked(&log->io_list_lock); 55556fef7c6SChristoph Hellwig 55656fef7c6SChristoph Hellwig list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) { 55756fef7c6SChristoph Hellwig /* don't change list order */ 55856fef7c6SChristoph Hellwig if (io->state < IO_UNIT_IO_END) 55956fef7c6SChristoph Hellwig break; 56056fef7c6SChristoph Hellwig 56156fef7c6SChristoph Hellwig list_move_tail(&io->log_sibling, &log->finished_ios); 56256fef7c6SChristoph Hellwig r5l_io_run_stripes(io); 56356fef7c6SChristoph Hellwig } 56456fef7c6SChristoph Hellwig } 56556fef7c6SChristoph Hellwig 5663848c0bcSChristoph Hellwig static void r5l_move_to_end_ios(struct r5l_log *log) 5673848c0bcSChristoph Hellwig { 5683848c0bcSChristoph Hellwig struct r5l_io_unit *io, *next; 5693848c0bcSChristoph Hellwig 5703848c0bcSChristoph Hellwig assert_spin_locked(&log->io_list_lock); 5713848c0bcSChristoph Hellwig 5723848c0bcSChristoph Hellwig list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) { 5733848c0bcSChristoph Hellwig /* don't change list order */ 5743848c0bcSChristoph Hellwig if (io->state < IO_UNIT_IO_END) 5753848c0bcSChristoph Hellwig break; 5763848c0bcSChristoph Hellwig list_move_tail(&io->log_sibling, &log->io_end_ios); 5773848c0bcSChristoph Hellwig } 5783848c0bcSChristoph Hellwig } 5793848c0bcSChristoph Hellwig 5803bddb7f8SSong Liu static void __r5l_stripe_write_finished(struct r5l_io_unit *io); 581f6bed0efSShaohua Li static void r5l_log_endio(struct bio *bio) 582f6bed0efSShaohua Li { 583f6bed0efSShaohua Li struct r5l_io_unit *io = bio->bi_private; 5843bddb7f8SSong Liu struct r5l_io_unit *io_deferred; 585f6bed0efSShaohua Li struct r5l_log *log = io->log; 586509ffec7SChristoph Hellwig unsigned long flags; 587f6bed0efSShaohua Li 5886e74a9cfSShaohua Li if (bio->bi_error) 5896e74a9cfSShaohua Li md_error(log->rdev->mddev, log->rdev); 5906e74a9cfSShaohua Li 591f6bed0efSShaohua Li bio_put(bio); 592e8deb638SChristoph Hellwig mempool_free(io->meta_page, log->meta_pool); 593f6bed0efSShaohua Li 594509ffec7SChristoph Hellwig spin_lock_irqsave(&log->io_list_lock, flags); 595509ffec7SChristoph Hellwig __r5l_set_io_unit_state(io, IO_UNIT_IO_END); 596*ea17481fSSong Liu if (log->need_cache_flush && !list_empty(&io->stripe_list)) 5973848c0bcSChristoph Hellwig r5l_move_to_end_ios(log); 59856fef7c6SChristoph Hellwig else 59956fef7c6SChristoph Hellwig r5l_log_run_stripes(log); 6003bddb7f8SSong Liu if (!list_empty(&log->running_ios)) { 6013bddb7f8SSong Liu /* 6023bddb7f8SSong Liu * FLUSH/FUA io_unit is deferred because of ordering, now we 6033bddb7f8SSong Liu * can dispatch it 6043bddb7f8SSong Liu */ 6053bddb7f8SSong Liu io_deferred = list_first_entry(&log->running_ios, 6063bddb7f8SSong Liu struct r5l_io_unit, log_sibling); 6073bddb7f8SSong Liu if (io_deferred->io_deferred) 6083bddb7f8SSong Liu schedule_work(&log->deferred_io_work); 6093bddb7f8SSong Liu } 6103bddb7f8SSong Liu 611509ffec7SChristoph Hellwig spin_unlock_irqrestore(&log->io_list_lock, flags); 612509ffec7SChristoph Hellwig 61356fef7c6SChristoph Hellwig if (log->need_cache_flush) 614f6bed0efSShaohua Li md_wakeup_thread(log->rdev->mddev->thread); 6153bddb7f8SSong Liu 6163bddb7f8SSong Liu if (io->has_null_flush) { 6173bddb7f8SSong Liu struct bio *bi; 6183bddb7f8SSong Liu 6193bddb7f8SSong Liu WARN_ON(bio_list_empty(&io->flush_barriers)); 6203bddb7f8SSong Liu while ((bi = bio_list_pop(&io->flush_barriers)) != NULL) { 6213bddb7f8SSong Liu bio_endio(bi); 6223bddb7f8SSong Liu atomic_dec(&io->pending_stripe); 6233bddb7f8SSong Liu } 624*ea17481fSSong Liu } 625*ea17481fSSong Liu 626*ea17481fSSong Liu /* finish flush only io_unit and PAYLOAD_FLUSH only io_unit */ 6273bddb7f8SSong Liu if (atomic_read(&io->pending_stripe) == 0) 6283bddb7f8SSong Liu __r5l_stripe_write_finished(io); 6293bddb7f8SSong Liu } 6303bddb7f8SSong Liu 6313bddb7f8SSong Liu static void r5l_do_submit_io(struct r5l_log *log, struct r5l_io_unit *io) 6323bddb7f8SSong Liu { 6333bddb7f8SSong Liu unsigned long flags; 6343bddb7f8SSong Liu 6353bddb7f8SSong Liu spin_lock_irqsave(&log->io_list_lock, flags); 6363bddb7f8SSong Liu __r5l_set_io_unit_state(io, IO_UNIT_IO_START); 6373bddb7f8SSong Liu spin_unlock_irqrestore(&log->io_list_lock, flags); 6383bddb7f8SSong Liu 6393bddb7f8SSong Liu if (io->has_flush) 64020737738SShaohua Li io->current_bio->bi_opf |= REQ_PREFLUSH; 6413bddb7f8SSong Liu if (io->has_fua) 64220737738SShaohua Li io->current_bio->bi_opf |= REQ_FUA; 6433bddb7f8SSong Liu submit_bio(io->current_bio); 6443bddb7f8SSong Liu 6453bddb7f8SSong Liu if (!io->split_bio) 6463bddb7f8SSong Liu return; 6473bddb7f8SSong Liu 6483bddb7f8SSong Liu if (io->has_flush) 64920737738SShaohua Li io->split_bio->bi_opf |= REQ_PREFLUSH; 6503bddb7f8SSong Liu if (io->has_fua) 65120737738SShaohua Li io->split_bio->bi_opf |= REQ_FUA; 6523bddb7f8SSong Liu submit_bio(io->split_bio); 6533bddb7f8SSong Liu } 6543bddb7f8SSong Liu 6553bddb7f8SSong Liu /* deferred io_unit will be dispatched here */ 6563bddb7f8SSong Liu static void r5l_submit_io_async(struct work_struct *work) 6573bddb7f8SSong Liu { 6583bddb7f8SSong Liu struct r5l_log *log = container_of(work, struct r5l_log, 6593bddb7f8SSong Liu deferred_io_work); 6603bddb7f8SSong Liu struct r5l_io_unit *io = NULL; 6613bddb7f8SSong Liu unsigned long flags; 6623bddb7f8SSong Liu 6633bddb7f8SSong Liu spin_lock_irqsave(&log->io_list_lock, flags); 6643bddb7f8SSong Liu if (!list_empty(&log->running_ios)) { 6653bddb7f8SSong Liu io = list_first_entry(&log->running_ios, struct r5l_io_unit, 6663bddb7f8SSong Liu log_sibling); 6673bddb7f8SSong Liu if (!io->io_deferred) 6683bddb7f8SSong Liu io = NULL; 6693bddb7f8SSong Liu else 6703bddb7f8SSong Liu io->io_deferred = 0; 6713bddb7f8SSong Liu } 6723bddb7f8SSong Liu spin_unlock_irqrestore(&log->io_list_lock, flags); 6733bddb7f8SSong Liu if (io) 6743bddb7f8SSong Liu r5l_do_submit_io(log, io); 675f6bed0efSShaohua Li } 676f6bed0efSShaohua Li 6772e38a37fSSong Liu static void r5c_disable_writeback_async(struct work_struct *work) 6782e38a37fSSong Liu { 6792e38a37fSSong Liu struct r5l_log *log = container_of(work, struct r5l_log, 6802e38a37fSSong Liu disable_writeback_work); 6812e38a37fSSong Liu struct mddev *mddev = log->rdev->mddev; 6822e38a37fSSong Liu 6832e38a37fSSong Liu if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) 6842e38a37fSSong Liu return; 6852e38a37fSSong Liu pr_info("md/raid:%s: Disabling writeback cache for degraded array.\n", 6862e38a37fSSong Liu mdname(mddev)); 6872e38a37fSSong Liu mddev_suspend(mddev); 6882e38a37fSSong Liu log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH; 6892e38a37fSSong Liu mddev_resume(mddev); 6902e38a37fSSong Liu } 6912e38a37fSSong Liu 692f6bed0efSShaohua Li static void r5l_submit_current_io(struct r5l_log *log) 693f6bed0efSShaohua Li { 694f6bed0efSShaohua Li struct r5l_io_unit *io = log->current_io; 6953bddb7f8SSong Liu struct bio *bio; 696f6bed0efSShaohua Li struct r5l_meta_block *block; 697509ffec7SChristoph Hellwig unsigned long flags; 698f6bed0efSShaohua Li u32 crc; 6993bddb7f8SSong Liu bool do_submit = true; 700f6bed0efSShaohua Li 701f6bed0efSShaohua Li if (!io) 702f6bed0efSShaohua Li return; 703f6bed0efSShaohua Li 704f6bed0efSShaohua Li block = page_address(io->meta_page); 705f6bed0efSShaohua Li block->meta_size = cpu_to_le32(io->meta_offset); 7065cb2fbd6SShaohua Li crc = crc32c_le(log->uuid_checksum, block, PAGE_SIZE); 707f6bed0efSShaohua Li block->checksum = cpu_to_le32(crc); 7083bddb7f8SSong Liu bio = io->current_bio; 709f6bed0efSShaohua Li 710f6bed0efSShaohua Li log->current_io = NULL; 711509ffec7SChristoph Hellwig spin_lock_irqsave(&log->io_list_lock, flags); 7123bddb7f8SSong Liu if (io->has_flush || io->has_fua) { 7133bddb7f8SSong Liu if (io != list_first_entry(&log->running_ios, 7143bddb7f8SSong Liu struct r5l_io_unit, log_sibling)) { 7153bddb7f8SSong Liu io->io_deferred = 1; 7163bddb7f8SSong Liu do_submit = false; 7173bddb7f8SSong Liu } 7183bddb7f8SSong Liu } 719509ffec7SChristoph Hellwig spin_unlock_irqrestore(&log->io_list_lock, flags); 7203bddb7f8SSong Liu if (do_submit) 7213bddb7f8SSong Liu r5l_do_submit_io(log, io); 722f6bed0efSShaohua Li } 723f6bed0efSShaohua Li 7246143e2ceSChristoph Hellwig static struct bio *r5l_bio_alloc(struct r5l_log *log) 725b349feb3SChristoph Hellwig { 726c38d29b3SChristoph Hellwig struct bio *bio = bio_alloc_bioset(GFP_NOIO, BIO_MAX_PAGES, log->bs); 727b349feb3SChristoph Hellwig 728796a5cf0SMike Christie bio_set_op_attrs(bio, REQ_OP_WRITE, 0); 729b349feb3SChristoph Hellwig bio->bi_bdev = log->rdev->bdev; 7301e932a37SChristoph Hellwig bio->bi_iter.bi_sector = log->rdev->data_offset + log->log_start; 731b349feb3SChristoph Hellwig 732b349feb3SChristoph Hellwig return bio; 733b349feb3SChristoph Hellwig } 734b349feb3SChristoph Hellwig 735c1b99198SChristoph Hellwig static void r5_reserve_log_entry(struct r5l_log *log, struct r5l_io_unit *io) 736c1b99198SChristoph Hellwig { 737c1b99198SChristoph Hellwig log->log_start = r5l_ring_add(log, log->log_start, BLOCK_SECTORS); 738c1b99198SChristoph Hellwig 739a39f7afdSSong Liu r5c_update_log_state(log); 740c1b99198SChristoph Hellwig /* 741c1b99198SChristoph Hellwig * If we filled up the log device start from the beginning again, 742c1b99198SChristoph Hellwig * which will require a new bio. 743c1b99198SChristoph Hellwig * 744c1b99198SChristoph Hellwig * Note: for this to work properly the log size needs to me a multiple 745c1b99198SChristoph Hellwig * of BLOCK_SECTORS. 746c1b99198SChristoph Hellwig */ 747c1b99198SChristoph Hellwig if (log->log_start == 0) 7486143e2ceSChristoph Hellwig io->need_split_bio = true; 749c1b99198SChristoph Hellwig 750c1b99198SChristoph Hellwig io->log_end = log->log_start; 751c1b99198SChristoph Hellwig } 752c1b99198SChristoph Hellwig 753f6bed0efSShaohua Li static struct r5l_io_unit *r5l_new_meta(struct r5l_log *log) 754f6bed0efSShaohua Li { 755f6bed0efSShaohua Li struct r5l_io_unit *io; 756f6bed0efSShaohua Li struct r5l_meta_block *block; 757f6bed0efSShaohua Li 7585036c390SChristoph Hellwig io = mempool_alloc(log->io_pool, GFP_ATOMIC); 7595036c390SChristoph Hellwig if (!io) 7605036c390SChristoph Hellwig return NULL; 7615036c390SChristoph Hellwig memset(io, 0, sizeof(*io)); 7625036c390SChristoph Hellwig 76351039cd0SChristoph Hellwig io->log = log; 76451039cd0SChristoph Hellwig INIT_LIST_HEAD(&io->log_sibling); 76551039cd0SChristoph Hellwig INIT_LIST_HEAD(&io->stripe_list); 7663bddb7f8SSong Liu bio_list_init(&io->flush_barriers); 76751039cd0SChristoph Hellwig io->state = IO_UNIT_RUNNING; 768f6bed0efSShaohua Li 769e8deb638SChristoph Hellwig io->meta_page = mempool_alloc(log->meta_pool, GFP_NOIO); 770f6bed0efSShaohua Li block = page_address(io->meta_page); 771e8deb638SChristoph Hellwig clear_page(block); 772f6bed0efSShaohua Li block->magic = cpu_to_le32(R5LOG_MAGIC); 773f6bed0efSShaohua Li block->version = R5LOG_VERSION; 774f6bed0efSShaohua Li block->seq = cpu_to_le64(log->seq); 775f6bed0efSShaohua Li block->position = cpu_to_le64(log->log_start); 776f6bed0efSShaohua Li 777f6bed0efSShaohua Li io->log_start = log->log_start; 778f6bed0efSShaohua Li io->meta_offset = sizeof(struct r5l_meta_block); 7792b8ef16eSChristoph Hellwig io->seq = log->seq++; 780f6bed0efSShaohua Li 7816143e2ceSChristoph Hellwig io->current_bio = r5l_bio_alloc(log); 7826143e2ceSChristoph Hellwig io->current_bio->bi_end_io = r5l_log_endio; 7836143e2ceSChristoph Hellwig io->current_bio->bi_private = io; 784b349feb3SChristoph Hellwig bio_add_page(io->current_bio, io->meta_page, PAGE_SIZE, 0); 785f6bed0efSShaohua Li 786c1b99198SChristoph Hellwig r5_reserve_log_entry(log, io); 787f6bed0efSShaohua Li 788f6bed0efSShaohua Li spin_lock_irq(&log->io_list_lock); 789f6bed0efSShaohua Li list_add_tail(&io->log_sibling, &log->running_ios); 790f6bed0efSShaohua Li spin_unlock_irq(&log->io_list_lock); 791f6bed0efSShaohua Li 792f6bed0efSShaohua Li return io; 793f6bed0efSShaohua Li } 794f6bed0efSShaohua Li 795f6bed0efSShaohua Li static int r5l_get_meta(struct r5l_log *log, unsigned int payload_size) 796f6bed0efSShaohua Li { 79722581f58SChristoph Hellwig if (log->current_io && 79822581f58SChristoph Hellwig log->current_io->meta_offset + payload_size > PAGE_SIZE) 799f6bed0efSShaohua Li r5l_submit_current_io(log); 800f6bed0efSShaohua Li 8015036c390SChristoph Hellwig if (!log->current_io) { 802f6bed0efSShaohua Li log->current_io = r5l_new_meta(log); 8035036c390SChristoph Hellwig if (!log->current_io) 8045036c390SChristoph Hellwig return -ENOMEM; 8055036c390SChristoph Hellwig } 8065036c390SChristoph Hellwig 807f6bed0efSShaohua Li return 0; 808f6bed0efSShaohua Li } 809f6bed0efSShaohua Li 810f6bed0efSShaohua Li static void r5l_append_payload_meta(struct r5l_log *log, u16 type, 811f6bed0efSShaohua Li sector_t location, 812f6bed0efSShaohua Li u32 checksum1, u32 checksum2, 813f6bed0efSShaohua Li bool checksum2_valid) 814f6bed0efSShaohua Li { 815f6bed0efSShaohua Li struct r5l_io_unit *io = log->current_io; 816f6bed0efSShaohua Li struct r5l_payload_data_parity *payload; 817f6bed0efSShaohua Li 818f6bed0efSShaohua Li payload = page_address(io->meta_page) + io->meta_offset; 819f6bed0efSShaohua Li payload->header.type = cpu_to_le16(type); 820f6bed0efSShaohua Li payload->header.flags = cpu_to_le16(0); 821f6bed0efSShaohua Li payload->size = cpu_to_le32((1 + !!checksum2_valid) << 822f6bed0efSShaohua Li (PAGE_SHIFT - 9)); 823f6bed0efSShaohua Li payload->location = cpu_to_le64(location); 824f6bed0efSShaohua Li payload->checksum[0] = cpu_to_le32(checksum1); 825f6bed0efSShaohua Li if (checksum2_valid) 826f6bed0efSShaohua Li payload->checksum[1] = cpu_to_le32(checksum2); 827f6bed0efSShaohua Li 828f6bed0efSShaohua Li io->meta_offset += sizeof(struct r5l_payload_data_parity) + 829f6bed0efSShaohua Li sizeof(__le32) * (1 + !!checksum2_valid); 830f6bed0efSShaohua Li } 831f6bed0efSShaohua Li 832f6bed0efSShaohua Li static void r5l_append_payload_page(struct r5l_log *log, struct page *page) 833f6bed0efSShaohua Li { 834f6bed0efSShaohua Li struct r5l_io_unit *io = log->current_io; 835f6bed0efSShaohua Li 8366143e2ceSChristoph Hellwig if (io->need_split_bio) { 8373bddb7f8SSong Liu BUG_ON(io->split_bio); 8383bddb7f8SSong Liu io->split_bio = io->current_bio; 8396143e2ceSChristoph Hellwig io->current_bio = r5l_bio_alloc(log); 8403bddb7f8SSong Liu bio_chain(io->current_bio, io->split_bio); 8413bddb7f8SSong Liu io->need_split_bio = false; 842f6bed0efSShaohua Li } 843f6bed0efSShaohua Li 8446143e2ceSChristoph Hellwig if (!bio_add_page(io->current_bio, page, PAGE_SIZE, 0)) 8456143e2ceSChristoph Hellwig BUG(); 8466143e2ceSChristoph Hellwig 847c1b99198SChristoph Hellwig r5_reserve_log_entry(log, io); 848f6bed0efSShaohua Li } 849f6bed0efSShaohua Li 850*ea17481fSSong Liu static void r5l_append_flush_payload(struct r5l_log *log, sector_t sect) 851*ea17481fSSong Liu { 852*ea17481fSSong Liu struct mddev *mddev = log->rdev->mddev; 853*ea17481fSSong Liu struct r5conf *conf = mddev->private; 854*ea17481fSSong Liu struct r5l_io_unit *io; 855*ea17481fSSong Liu struct r5l_payload_flush *payload; 856*ea17481fSSong Liu int meta_size; 857*ea17481fSSong Liu 858*ea17481fSSong Liu /* 859*ea17481fSSong Liu * payload_flush requires extra writes to the journal. 860*ea17481fSSong Liu * To avoid handling the extra IO in quiesce, just skip 861*ea17481fSSong Liu * flush_payload 862*ea17481fSSong Liu */ 863*ea17481fSSong Liu if (conf->quiesce) 864*ea17481fSSong Liu return; 865*ea17481fSSong Liu 866*ea17481fSSong Liu mutex_lock(&log->io_mutex); 867*ea17481fSSong Liu meta_size = sizeof(struct r5l_payload_flush) + sizeof(__le64); 868*ea17481fSSong Liu 869*ea17481fSSong Liu if (r5l_get_meta(log, meta_size)) { 870*ea17481fSSong Liu mutex_unlock(&log->io_mutex); 871*ea17481fSSong Liu return; 872*ea17481fSSong Liu } 873*ea17481fSSong Liu 874*ea17481fSSong Liu /* current implementation is one stripe per flush payload */ 875*ea17481fSSong Liu io = log->current_io; 876*ea17481fSSong Liu payload = page_address(io->meta_page) + io->meta_offset; 877*ea17481fSSong Liu payload->header.type = cpu_to_le16(R5LOG_PAYLOAD_FLUSH); 878*ea17481fSSong Liu payload->header.flags = cpu_to_le16(0); 879*ea17481fSSong Liu payload->size = cpu_to_le32(sizeof(__le64)); 880*ea17481fSSong Liu payload->flush_stripes[0] = cpu_to_le64(sect); 881*ea17481fSSong Liu io->meta_offset += meta_size; 882*ea17481fSSong Liu mutex_unlock(&log->io_mutex); 883*ea17481fSSong Liu } 884*ea17481fSSong Liu 8855036c390SChristoph Hellwig static int r5l_log_stripe(struct r5l_log *log, struct stripe_head *sh, 886f6bed0efSShaohua Li int data_pages, int parity_pages) 887f6bed0efSShaohua Li { 888f6bed0efSShaohua Li int i; 889f6bed0efSShaohua Li int meta_size; 8905036c390SChristoph Hellwig int ret; 891f6bed0efSShaohua Li struct r5l_io_unit *io; 892f6bed0efSShaohua Li 893f6bed0efSShaohua Li meta_size = 894f6bed0efSShaohua Li ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32)) 895f6bed0efSShaohua Li * data_pages) + 896f6bed0efSShaohua Li sizeof(struct r5l_payload_data_parity) + 897f6bed0efSShaohua Li sizeof(__le32) * parity_pages; 898f6bed0efSShaohua Li 8995036c390SChristoph Hellwig ret = r5l_get_meta(log, meta_size); 9005036c390SChristoph Hellwig if (ret) 9015036c390SChristoph Hellwig return ret; 9025036c390SChristoph Hellwig 903f6bed0efSShaohua Li io = log->current_io; 904f6bed0efSShaohua Li 9053bddb7f8SSong Liu if (test_and_clear_bit(STRIPE_R5C_PREFLUSH, &sh->state)) 9063bddb7f8SSong Liu io->has_flush = 1; 9073bddb7f8SSong Liu 908f6bed0efSShaohua Li for (i = 0; i < sh->disks; i++) { 9091e6d690bSSong Liu if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) || 9101e6d690bSSong Liu test_bit(R5_InJournal, &sh->dev[i].flags)) 911f6bed0efSShaohua Li continue; 912f6bed0efSShaohua Li if (i == sh->pd_idx || i == sh->qd_idx) 913f6bed0efSShaohua Li continue; 9143bddb7f8SSong Liu if (test_bit(R5_WantFUA, &sh->dev[i].flags) && 9153bddb7f8SSong Liu log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK) { 9163bddb7f8SSong Liu io->has_fua = 1; 9173bddb7f8SSong Liu /* 9183bddb7f8SSong Liu * we need to flush journal to make sure recovery can 9193bddb7f8SSong Liu * reach the data with fua flag 9203bddb7f8SSong Liu */ 9213bddb7f8SSong Liu io->has_flush = 1; 9223bddb7f8SSong Liu } 923f6bed0efSShaohua Li r5l_append_payload_meta(log, R5LOG_PAYLOAD_DATA, 924f6bed0efSShaohua Li raid5_compute_blocknr(sh, i, 0), 925f6bed0efSShaohua Li sh->dev[i].log_checksum, 0, false); 926f6bed0efSShaohua Li r5l_append_payload_page(log, sh->dev[i].page); 927f6bed0efSShaohua Li } 928f6bed0efSShaohua Li 9292ded3703SSong Liu if (parity_pages == 2) { 930f6bed0efSShaohua Li r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY, 931f6bed0efSShaohua Li sh->sector, sh->dev[sh->pd_idx].log_checksum, 932f6bed0efSShaohua Li sh->dev[sh->qd_idx].log_checksum, true); 933f6bed0efSShaohua Li r5l_append_payload_page(log, sh->dev[sh->pd_idx].page); 934f6bed0efSShaohua Li r5l_append_payload_page(log, sh->dev[sh->qd_idx].page); 9352ded3703SSong Liu } else if (parity_pages == 1) { 936f6bed0efSShaohua Li r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY, 937f6bed0efSShaohua Li sh->sector, sh->dev[sh->pd_idx].log_checksum, 938f6bed0efSShaohua Li 0, false); 939f6bed0efSShaohua Li r5l_append_payload_page(log, sh->dev[sh->pd_idx].page); 9402ded3703SSong Liu } else /* Just writing data, not parity, in caching phase */ 9412ded3703SSong Liu BUG_ON(parity_pages != 0); 942f6bed0efSShaohua Li 943f6bed0efSShaohua Li list_add_tail(&sh->log_list, &io->stripe_list); 944f6bed0efSShaohua Li atomic_inc(&io->pending_stripe); 945f6bed0efSShaohua Li sh->log_io = io; 9465036c390SChristoph Hellwig 947a39f7afdSSong Liu if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) 948a39f7afdSSong Liu return 0; 949a39f7afdSSong Liu 950a39f7afdSSong Liu if (sh->log_start == MaxSector) { 951a39f7afdSSong Liu BUG_ON(!list_empty(&sh->r5c)); 952a39f7afdSSong Liu sh->log_start = io->log_start; 953a39f7afdSSong Liu spin_lock_irq(&log->stripe_in_journal_lock); 954a39f7afdSSong Liu list_add_tail(&sh->r5c, 955a39f7afdSSong Liu &log->stripe_in_journal_list); 956a39f7afdSSong Liu spin_unlock_irq(&log->stripe_in_journal_lock); 957a39f7afdSSong Liu atomic_inc(&log->stripe_in_journal_count); 958a39f7afdSSong Liu } 9595036c390SChristoph Hellwig return 0; 960f6bed0efSShaohua Li } 961f6bed0efSShaohua Li 962a39f7afdSSong Liu /* add stripe to no_space_stripes, and then wake up reclaim */ 963a39f7afdSSong Liu static inline void r5l_add_no_space_stripe(struct r5l_log *log, 964a39f7afdSSong Liu struct stripe_head *sh) 965a39f7afdSSong Liu { 966a39f7afdSSong Liu spin_lock(&log->no_space_stripes_lock); 967a39f7afdSSong Liu list_add_tail(&sh->log_list, &log->no_space_stripes); 968a39f7afdSSong Liu spin_unlock(&log->no_space_stripes_lock); 969a39f7afdSSong Liu } 970a39f7afdSSong Liu 971f6bed0efSShaohua Li /* 972f6bed0efSShaohua Li * running in raid5d, where reclaim could wait for raid5d too (when it flushes 973f6bed0efSShaohua Li * data from log to raid disks), so we shouldn't wait for reclaim here 974f6bed0efSShaohua Li */ 975f6bed0efSShaohua Li int r5l_write_stripe(struct r5l_log *log, struct stripe_head *sh) 976f6bed0efSShaohua Li { 977a39f7afdSSong Liu struct r5conf *conf = sh->raid_conf; 978f6bed0efSShaohua Li int write_disks = 0; 979f6bed0efSShaohua Li int data_pages, parity_pages; 980f6bed0efSShaohua Li int reserve; 981f6bed0efSShaohua Li int i; 9825036c390SChristoph Hellwig int ret = 0; 983a39f7afdSSong Liu bool wake_reclaim = false; 984f6bed0efSShaohua Li 985f6bed0efSShaohua Li if (!log) 986f6bed0efSShaohua Li return -EAGAIN; 987f6bed0efSShaohua Li /* Don't support stripe batch */ 988f6bed0efSShaohua Li if (sh->log_io || !test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags) || 989f6bed0efSShaohua Li test_bit(STRIPE_SYNCING, &sh->state)) { 990f6bed0efSShaohua Li /* the stripe is written to log, we start writing it to raid */ 991f6bed0efSShaohua Li clear_bit(STRIPE_LOG_TRAPPED, &sh->state); 992f6bed0efSShaohua Li return -EAGAIN; 993f6bed0efSShaohua Li } 994f6bed0efSShaohua Li 9952ded3703SSong Liu WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state)); 9962ded3703SSong Liu 997f6bed0efSShaohua Li for (i = 0; i < sh->disks; i++) { 998f6bed0efSShaohua Li void *addr; 999f6bed0efSShaohua Li 10001e6d690bSSong Liu if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) || 10011e6d690bSSong Liu test_bit(R5_InJournal, &sh->dev[i].flags)) 1002f6bed0efSShaohua Li continue; 10031e6d690bSSong Liu 1004f6bed0efSShaohua Li write_disks++; 1005f6bed0efSShaohua Li /* checksum is already calculated in last run */ 1006f6bed0efSShaohua Li if (test_bit(STRIPE_LOG_TRAPPED, &sh->state)) 1007f6bed0efSShaohua Li continue; 1008f6bed0efSShaohua Li addr = kmap_atomic(sh->dev[i].page); 10095cb2fbd6SShaohua Li sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum, 1010f6bed0efSShaohua Li addr, PAGE_SIZE); 1011f6bed0efSShaohua Li kunmap_atomic(addr); 1012f6bed0efSShaohua Li } 1013f6bed0efSShaohua Li parity_pages = 1 + !!(sh->qd_idx >= 0); 1014f6bed0efSShaohua Li data_pages = write_disks - parity_pages; 1015f6bed0efSShaohua Li 1016f6bed0efSShaohua Li set_bit(STRIPE_LOG_TRAPPED, &sh->state); 1017253f9fd4SShaohua Li /* 1018253f9fd4SShaohua Li * The stripe must enter state machine again to finish the write, so 1019253f9fd4SShaohua Li * don't delay. 1020253f9fd4SShaohua Li */ 1021253f9fd4SShaohua Li clear_bit(STRIPE_DELAYED, &sh->state); 1022f6bed0efSShaohua Li atomic_inc(&sh->count); 1023f6bed0efSShaohua Li 1024f6bed0efSShaohua Li mutex_lock(&log->io_mutex); 1025f6bed0efSShaohua Li /* meta + data */ 1026f6bed0efSShaohua Li reserve = (1 + write_disks) << (PAGE_SHIFT - 9); 1027f6bed0efSShaohua Li 1028a39f7afdSSong Liu if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) { 1029a39f7afdSSong Liu if (!r5l_has_free_space(log, reserve)) { 1030a39f7afdSSong Liu r5l_add_no_space_stripe(log, sh); 1031a39f7afdSSong Liu wake_reclaim = true; 10325036c390SChristoph Hellwig } else { 10335036c390SChristoph Hellwig ret = r5l_log_stripe(log, sh, data_pages, parity_pages); 10345036c390SChristoph Hellwig if (ret) { 10355036c390SChristoph Hellwig spin_lock_irq(&log->io_list_lock); 1036a39f7afdSSong Liu list_add_tail(&sh->log_list, 1037a39f7afdSSong Liu &log->no_mem_stripes); 10385036c390SChristoph Hellwig spin_unlock_irq(&log->io_list_lock); 1039f6bed0efSShaohua Li } 10405036c390SChristoph Hellwig } 1041a39f7afdSSong Liu } else { /* R5C_JOURNAL_MODE_WRITE_BACK */ 1042a39f7afdSSong Liu /* 1043a39f7afdSSong Liu * log space critical, do not process stripes that are 1044a39f7afdSSong Liu * not in cache yet (sh->log_start == MaxSector). 1045a39f7afdSSong Liu */ 1046a39f7afdSSong Liu if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) && 1047a39f7afdSSong Liu sh->log_start == MaxSector) { 1048a39f7afdSSong Liu r5l_add_no_space_stripe(log, sh); 1049a39f7afdSSong Liu wake_reclaim = true; 1050a39f7afdSSong Liu reserve = 0; 1051a39f7afdSSong Liu } else if (!r5l_has_free_space(log, reserve)) { 1052a39f7afdSSong Liu if (sh->log_start == log->last_checkpoint) 1053a39f7afdSSong Liu BUG(); 1054a39f7afdSSong Liu else 1055a39f7afdSSong Liu r5l_add_no_space_stripe(log, sh); 1056a39f7afdSSong Liu } else { 1057a39f7afdSSong Liu ret = r5l_log_stripe(log, sh, data_pages, parity_pages); 1058a39f7afdSSong Liu if (ret) { 1059a39f7afdSSong Liu spin_lock_irq(&log->io_list_lock); 1060a39f7afdSSong Liu list_add_tail(&sh->log_list, 1061a39f7afdSSong Liu &log->no_mem_stripes); 1062a39f7afdSSong Liu spin_unlock_irq(&log->io_list_lock); 1063a39f7afdSSong Liu } 1064a39f7afdSSong Liu } 1065a39f7afdSSong Liu } 1066f6bed0efSShaohua Li 10675036c390SChristoph Hellwig mutex_unlock(&log->io_mutex); 1068a39f7afdSSong Liu if (wake_reclaim) 1069a39f7afdSSong Liu r5l_wake_reclaim(log, reserve); 1070f6bed0efSShaohua Li return 0; 1071f6bed0efSShaohua Li } 1072f6bed0efSShaohua Li 1073f6bed0efSShaohua Li void r5l_write_stripe_run(struct r5l_log *log) 1074f6bed0efSShaohua Li { 1075f6bed0efSShaohua Li if (!log) 1076f6bed0efSShaohua Li return; 1077f6bed0efSShaohua Li mutex_lock(&log->io_mutex); 1078f6bed0efSShaohua Li r5l_submit_current_io(log); 1079f6bed0efSShaohua Li mutex_unlock(&log->io_mutex); 1080f6bed0efSShaohua Li } 1081f6bed0efSShaohua Li 1082828cbe98SShaohua Li int r5l_handle_flush_request(struct r5l_log *log, struct bio *bio) 1083828cbe98SShaohua Li { 1084828cbe98SShaohua Li if (!log) 1085828cbe98SShaohua Li return -ENODEV; 10863bddb7f8SSong Liu 10873bddb7f8SSong Liu if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) { 1088828cbe98SShaohua Li /* 10893bddb7f8SSong Liu * in write through (journal only) 10903bddb7f8SSong Liu * we flush log disk cache first, then write stripe data to 10913bddb7f8SSong Liu * raid disks. So if bio is finished, the log disk cache is 10923bddb7f8SSong Liu * flushed already. The recovery guarantees we can recovery 10933bddb7f8SSong Liu * the bio from log disk, so we don't need to flush again 1094828cbe98SShaohua Li */ 1095828cbe98SShaohua Li if (bio->bi_iter.bi_size == 0) { 1096828cbe98SShaohua Li bio_endio(bio); 1097828cbe98SShaohua Li return 0; 1098828cbe98SShaohua Li } 10991eff9d32SJens Axboe bio->bi_opf &= ~REQ_PREFLUSH; 11003bddb7f8SSong Liu } else { 11013bddb7f8SSong Liu /* write back (with cache) */ 11023bddb7f8SSong Liu if (bio->bi_iter.bi_size == 0) { 11033bddb7f8SSong Liu mutex_lock(&log->io_mutex); 11043bddb7f8SSong Liu r5l_get_meta(log, 0); 11053bddb7f8SSong Liu bio_list_add(&log->current_io->flush_barriers, bio); 11063bddb7f8SSong Liu log->current_io->has_flush = 1; 11073bddb7f8SSong Liu log->current_io->has_null_flush = 1; 11083bddb7f8SSong Liu atomic_inc(&log->current_io->pending_stripe); 11093bddb7f8SSong Liu r5l_submit_current_io(log); 11103bddb7f8SSong Liu mutex_unlock(&log->io_mutex); 11113bddb7f8SSong Liu return 0; 11123bddb7f8SSong Liu } 11133bddb7f8SSong Liu } 1114828cbe98SShaohua Li return -EAGAIN; 1115828cbe98SShaohua Li } 1116828cbe98SShaohua Li 1117f6bed0efSShaohua Li /* This will run after log space is reclaimed */ 1118f6bed0efSShaohua Li static void r5l_run_no_space_stripes(struct r5l_log *log) 1119f6bed0efSShaohua Li { 1120f6bed0efSShaohua Li struct stripe_head *sh; 1121f6bed0efSShaohua Li 1122f6bed0efSShaohua Li spin_lock(&log->no_space_stripes_lock); 1123f6bed0efSShaohua Li while (!list_empty(&log->no_space_stripes)) { 1124f6bed0efSShaohua Li sh = list_first_entry(&log->no_space_stripes, 1125f6bed0efSShaohua Li struct stripe_head, log_list); 1126f6bed0efSShaohua Li list_del_init(&sh->log_list); 1127f6bed0efSShaohua Li set_bit(STRIPE_HANDLE, &sh->state); 1128f6bed0efSShaohua Li raid5_release_stripe(sh); 1129f6bed0efSShaohua Li } 1130f6bed0efSShaohua Li spin_unlock(&log->no_space_stripes_lock); 1131f6bed0efSShaohua Li } 1132f6bed0efSShaohua Li 1133a39f7afdSSong Liu /* 1134a39f7afdSSong Liu * calculate new last_checkpoint 1135a39f7afdSSong Liu * for write through mode, returns log->next_checkpoint 1136a39f7afdSSong Liu * for write back, returns log_start of first sh in stripe_in_journal_list 1137a39f7afdSSong Liu */ 1138a39f7afdSSong Liu static sector_t r5c_calculate_new_cp(struct r5conf *conf) 1139a39f7afdSSong Liu { 1140a39f7afdSSong Liu struct stripe_head *sh; 1141a39f7afdSSong Liu struct r5l_log *log = conf->log; 1142a39f7afdSSong Liu sector_t new_cp; 1143a39f7afdSSong Liu unsigned long flags; 1144a39f7afdSSong Liu 1145a39f7afdSSong Liu if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) 1146a39f7afdSSong Liu return log->next_checkpoint; 1147a39f7afdSSong Liu 1148a39f7afdSSong Liu spin_lock_irqsave(&log->stripe_in_journal_lock, flags); 1149a39f7afdSSong Liu if (list_empty(&conf->log->stripe_in_journal_list)) { 1150a39f7afdSSong Liu /* all stripes flushed */ 1151d3014e21SDan Carpenter spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags); 1152a39f7afdSSong Liu return log->next_checkpoint; 1153a39f7afdSSong Liu } 1154a39f7afdSSong Liu sh = list_first_entry(&conf->log->stripe_in_journal_list, 1155a39f7afdSSong Liu struct stripe_head, r5c); 1156a39f7afdSSong Liu new_cp = sh->log_start; 1157a39f7afdSSong Liu spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags); 1158a39f7afdSSong Liu return new_cp; 1159a39f7afdSSong Liu } 1160a39f7afdSSong Liu 116117036461SChristoph Hellwig static sector_t r5l_reclaimable_space(struct r5l_log *log) 116217036461SChristoph Hellwig { 1163a39f7afdSSong Liu struct r5conf *conf = log->rdev->mddev->private; 1164a39f7afdSSong Liu 116517036461SChristoph Hellwig return r5l_ring_distance(log, log->last_checkpoint, 1166a39f7afdSSong Liu r5c_calculate_new_cp(conf)); 116717036461SChristoph Hellwig } 116817036461SChristoph Hellwig 11695036c390SChristoph Hellwig static void r5l_run_no_mem_stripe(struct r5l_log *log) 11705036c390SChristoph Hellwig { 11715036c390SChristoph Hellwig struct stripe_head *sh; 11725036c390SChristoph Hellwig 11735036c390SChristoph Hellwig assert_spin_locked(&log->io_list_lock); 11745036c390SChristoph Hellwig 11755036c390SChristoph Hellwig if (!list_empty(&log->no_mem_stripes)) { 11765036c390SChristoph Hellwig sh = list_first_entry(&log->no_mem_stripes, 11775036c390SChristoph Hellwig struct stripe_head, log_list); 11785036c390SChristoph Hellwig list_del_init(&sh->log_list); 11795036c390SChristoph Hellwig set_bit(STRIPE_HANDLE, &sh->state); 11805036c390SChristoph Hellwig raid5_release_stripe(sh); 11815036c390SChristoph Hellwig } 11825036c390SChristoph Hellwig } 11835036c390SChristoph Hellwig 118404732f74SChristoph Hellwig static bool r5l_complete_finished_ios(struct r5l_log *log) 118517036461SChristoph Hellwig { 118617036461SChristoph Hellwig struct r5l_io_unit *io, *next; 118717036461SChristoph Hellwig bool found = false; 118817036461SChristoph Hellwig 118917036461SChristoph Hellwig assert_spin_locked(&log->io_list_lock); 119017036461SChristoph Hellwig 119104732f74SChristoph Hellwig list_for_each_entry_safe(io, next, &log->finished_ios, log_sibling) { 119217036461SChristoph Hellwig /* don't change list order */ 119317036461SChristoph Hellwig if (io->state < IO_UNIT_STRIPE_END) 119417036461SChristoph Hellwig break; 119517036461SChristoph Hellwig 119617036461SChristoph Hellwig log->next_checkpoint = io->log_start; 119717036461SChristoph Hellwig 119817036461SChristoph Hellwig list_del(&io->log_sibling); 11995036c390SChristoph Hellwig mempool_free(io, log->io_pool); 12005036c390SChristoph Hellwig r5l_run_no_mem_stripe(log); 120117036461SChristoph Hellwig 120217036461SChristoph Hellwig found = true; 120317036461SChristoph Hellwig } 120417036461SChristoph Hellwig 120517036461SChristoph Hellwig return found; 120617036461SChristoph Hellwig } 120717036461SChristoph Hellwig 1208509ffec7SChristoph Hellwig static void __r5l_stripe_write_finished(struct r5l_io_unit *io) 1209509ffec7SChristoph Hellwig { 1210509ffec7SChristoph Hellwig struct r5l_log *log = io->log; 1211a39f7afdSSong Liu struct r5conf *conf = log->rdev->mddev->private; 1212509ffec7SChristoph Hellwig unsigned long flags; 1213509ffec7SChristoph Hellwig 1214509ffec7SChristoph Hellwig spin_lock_irqsave(&log->io_list_lock, flags); 1215509ffec7SChristoph Hellwig __r5l_set_io_unit_state(io, IO_UNIT_STRIPE_END); 121617036461SChristoph Hellwig 121704732f74SChristoph Hellwig if (!r5l_complete_finished_ios(log)) { 121885f2f9a4SShaohua Li spin_unlock_irqrestore(&log->io_list_lock, flags); 121985f2f9a4SShaohua Li return; 122085f2f9a4SShaohua Li } 1221509ffec7SChristoph Hellwig 1222a39f7afdSSong Liu if (r5l_reclaimable_space(log) > log->max_free_space || 1223a39f7afdSSong Liu test_bit(R5C_LOG_TIGHT, &conf->cache_state)) 1224509ffec7SChristoph Hellwig r5l_wake_reclaim(log, 0); 1225509ffec7SChristoph Hellwig 1226509ffec7SChristoph Hellwig spin_unlock_irqrestore(&log->io_list_lock, flags); 1227509ffec7SChristoph Hellwig wake_up(&log->iounit_wait); 1228509ffec7SChristoph Hellwig } 1229509ffec7SChristoph Hellwig 12300576b1c6SShaohua Li void r5l_stripe_write_finished(struct stripe_head *sh) 12310576b1c6SShaohua Li { 12320576b1c6SShaohua Li struct r5l_io_unit *io; 12330576b1c6SShaohua Li 12340576b1c6SShaohua Li io = sh->log_io; 12350576b1c6SShaohua Li sh->log_io = NULL; 12360576b1c6SShaohua Li 1237509ffec7SChristoph Hellwig if (io && atomic_dec_and_test(&io->pending_stripe)) 1238509ffec7SChristoph Hellwig __r5l_stripe_write_finished(io); 12390576b1c6SShaohua Li } 12400576b1c6SShaohua Li 1241a8c34f91SShaohua Li static void r5l_log_flush_endio(struct bio *bio) 1242a8c34f91SShaohua Li { 1243a8c34f91SShaohua Li struct r5l_log *log = container_of(bio, struct r5l_log, 1244a8c34f91SShaohua Li flush_bio); 1245a8c34f91SShaohua Li unsigned long flags; 1246a8c34f91SShaohua Li struct r5l_io_unit *io; 1247a8c34f91SShaohua Li 12486e74a9cfSShaohua Li if (bio->bi_error) 12496e74a9cfSShaohua Li md_error(log->rdev->mddev, log->rdev); 12506e74a9cfSShaohua Li 1251a8c34f91SShaohua Li spin_lock_irqsave(&log->io_list_lock, flags); 1252d8858f43SChristoph Hellwig list_for_each_entry(io, &log->flushing_ios, log_sibling) 1253d8858f43SChristoph Hellwig r5l_io_run_stripes(io); 125404732f74SChristoph Hellwig list_splice_tail_init(&log->flushing_ios, &log->finished_ios); 1255a8c34f91SShaohua Li spin_unlock_irqrestore(&log->io_list_lock, flags); 1256a8c34f91SShaohua Li } 1257a8c34f91SShaohua Li 12580576b1c6SShaohua Li /* 12590576b1c6SShaohua Li * Starting dispatch IO to raid. 12600576b1c6SShaohua Li * io_unit(meta) consists of a log. There is one situation we want to avoid. A 12610576b1c6SShaohua Li * broken meta in the middle of a log causes recovery can't find meta at the 12620576b1c6SShaohua Li * head of log. If operations require meta at the head persistent in log, we 12630576b1c6SShaohua Li * must make sure meta before it persistent in log too. A case is: 12640576b1c6SShaohua Li * 12650576b1c6SShaohua Li * stripe data/parity is in log, we start write stripe to raid disks. stripe 12660576b1c6SShaohua Li * data/parity must be persistent in log before we do the write to raid disks. 12670576b1c6SShaohua Li * 12680576b1c6SShaohua Li * The solution is we restrictly maintain io_unit list order. In this case, we 12690576b1c6SShaohua Li * only write stripes of an io_unit to raid disks till the io_unit is the first 12700576b1c6SShaohua Li * one whose data/parity is in log. 12710576b1c6SShaohua Li */ 12720576b1c6SShaohua Li void r5l_flush_stripe_to_raid(struct r5l_log *log) 12730576b1c6SShaohua Li { 1274a8c34f91SShaohua Li bool do_flush; 127556fef7c6SChristoph Hellwig 127656fef7c6SChristoph Hellwig if (!log || !log->need_cache_flush) 12770576b1c6SShaohua Li return; 12780576b1c6SShaohua Li 1279a8c34f91SShaohua Li spin_lock_irq(&log->io_list_lock); 1280a8c34f91SShaohua Li /* flush bio is running */ 1281a8c34f91SShaohua Li if (!list_empty(&log->flushing_ios)) { 1282a8c34f91SShaohua Li spin_unlock_irq(&log->io_list_lock); 12830576b1c6SShaohua Li return; 12840576b1c6SShaohua Li } 1285a8c34f91SShaohua Li list_splice_tail_init(&log->io_end_ios, &log->flushing_ios); 1286a8c34f91SShaohua Li do_flush = !list_empty(&log->flushing_ios); 12870576b1c6SShaohua Li spin_unlock_irq(&log->io_list_lock); 1288a8c34f91SShaohua Li 1289a8c34f91SShaohua Li if (!do_flush) 1290a8c34f91SShaohua Li return; 1291a8c34f91SShaohua Li bio_reset(&log->flush_bio); 1292a8c34f91SShaohua Li log->flush_bio.bi_bdev = log->rdev->bdev; 1293a8c34f91SShaohua Li log->flush_bio.bi_end_io = r5l_log_flush_endio; 129470fd7614SChristoph Hellwig log->flush_bio.bi_opf = REQ_OP_WRITE | REQ_PREFLUSH; 12954e49ea4aSMike Christie submit_bio(&log->flush_bio); 12960576b1c6SShaohua Li } 12970576b1c6SShaohua Li 12980576b1c6SShaohua Li static void r5l_write_super(struct r5l_log *log, sector_t cp); 12994b482044SShaohua Li static void r5l_write_super_and_discard_space(struct r5l_log *log, 13004b482044SShaohua Li sector_t end) 13014b482044SShaohua Li { 13024b482044SShaohua Li struct block_device *bdev = log->rdev->bdev; 13034b482044SShaohua Li struct mddev *mddev; 13044b482044SShaohua Li 13054b482044SShaohua Li r5l_write_super(log, end); 13064b482044SShaohua Li 13074b482044SShaohua Li if (!blk_queue_discard(bdev_get_queue(bdev))) 13084b482044SShaohua Li return; 13094b482044SShaohua Li 13104b482044SShaohua Li mddev = log->rdev->mddev; 13114b482044SShaohua Li /* 13128e018c21SShaohua Li * Discard could zero data, so before discard we must make sure 13138e018c21SShaohua Li * superblock is updated to new log tail. Updating superblock (either 13148e018c21SShaohua Li * directly call md_update_sb() or depend on md thread) must hold 13158e018c21SShaohua Li * reconfig mutex. On the other hand, raid5_quiesce is called with 13168e018c21SShaohua Li * reconfig_mutex hold. The first step of raid5_quiesce() is waitting 13178e018c21SShaohua Li * for all IO finish, hence waitting for reclaim thread, while reclaim 13188e018c21SShaohua Li * thread is calling this function and waitting for reconfig mutex. So 13198e018c21SShaohua Li * there is a deadlock. We workaround this issue with a trylock. 13208e018c21SShaohua Li * FIXME: we could miss discard if we can't take reconfig mutex 13214b482044SShaohua Li */ 13222953079cSShaohua Li set_mask_bits(&mddev->sb_flags, 0, 13232953079cSShaohua Li BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING)); 13248e018c21SShaohua Li if (!mddev_trylock(mddev)) 13258e018c21SShaohua Li return; 13264b482044SShaohua Li md_update_sb(mddev, 1); 13278e018c21SShaohua Li mddev_unlock(mddev); 13284b482044SShaohua Li 13296e74a9cfSShaohua Li /* discard IO error really doesn't matter, ignore it */ 13304b482044SShaohua Li if (log->last_checkpoint < end) { 13314b482044SShaohua Li blkdev_issue_discard(bdev, 13324b482044SShaohua Li log->last_checkpoint + log->rdev->data_offset, 13334b482044SShaohua Li end - log->last_checkpoint, GFP_NOIO, 0); 13344b482044SShaohua Li } else { 13354b482044SShaohua Li blkdev_issue_discard(bdev, 13364b482044SShaohua Li log->last_checkpoint + log->rdev->data_offset, 13374b482044SShaohua Li log->device_size - log->last_checkpoint, 13384b482044SShaohua Li GFP_NOIO, 0); 13394b482044SShaohua Li blkdev_issue_discard(bdev, log->rdev->data_offset, end, 13404b482044SShaohua Li GFP_NOIO, 0); 13414b482044SShaohua Li } 13424b482044SShaohua Li } 13434b482044SShaohua Li 1344a39f7afdSSong Liu /* 1345a39f7afdSSong Liu * r5c_flush_stripe moves stripe from cached list to handle_list. When called, 1346a39f7afdSSong Liu * the stripe must be on r5c_cached_full_stripes or r5c_cached_partial_stripes. 1347a39f7afdSSong Liu * 1348a39f7afdSSong Liu * must hold conf->device_lock 1349a39f7afdSSong Liu */ 1350a39f7afdSSong Liu static void r5c_flush_stripe(struct r5conf *conf, struct stripe_head *sh) 1351a39f7afdSSong Liu { 1352a39f7afdSSong Liu BUG_ON(list_empty(&sh->lru)); 1353a39f7afdSSong Liu BUG_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state)); 1354a39f7afdSSong Liu BUG_ON(test_bit(STRIPE_HANDLE, &sh->state)); 1355a39f7afdSSong Liu 1356a39f7afdSSong Liu /* 1357a39f7afdSSong Liu * The stripe is not ON_RELEASE_LIST, so it is safe to call 1358a39f7afdSSong Liu * raid5_release_stripe() while holding conf->device_lock 1359a39f7afdSSong Liu */ 1360a39f7afdSSong Liu BUG_ON(test_bit(STRIPE_ON_RELEASE_LIST, &sh->state)); 1361a39f7afdSSong Liu assert_spin_locked(&conf->device_lock); 1362a39f7afdSSong Liu 1363a39f7afdSSong Liu list_del_init(&sh->lru); 1364a39f7afdSSong Liu atomic_inc(&sh->count); 1365a39f7afdSSong Liu 1366a39f7afdSSong Liu set_bit(STRIPE_HANDLE, &sh->state); 1367a39f7afdSSong Liu atomic_inc(&conf->active_stripes); 1368a39f7afdSSong Liu r5c_make_stripe_write_out(sh); 1369a39f7afdSSong Liu 1370e33fbb9cSShaohua Li if (test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) 1371e33fbb9cSShaohua Li atomic_inc(&conf->r5c_flushing_partial_stripes); 1372e33fbb9cSShaohua Li else 1373e33fbb9cSShaohua Li atomic_inc(&conf->r5c_flushing_full_stripes); 1374a39f7afdSSong Liu raid5_release_stripe(sh); 1375a39f7afdSSong Liu } 1376a39f7afdSSong Liu 1377a39f7afdSSong Liu /* 1378a39f7afdSSong Liu * if num == 0, flush all full stripes 1379a39f7afdSSong Liu * if num > 0, flush all full stripes. If less than num full stripes are 1380a39f7afdSSong Liu * flushed, flush some partial stripes until totally num stripes are 1381a39f7afdSSong Liu * flushed or there is no more cached stripes. 1382a39f7afdSSong Liu */ 1383a39f7afdSSong Liu void r5c_flush_cache(struct r5conf *conf, int num) 1384a39f7afdSSong Liu { 1385a39f7afdSSong Liu int count; 1386a39f7afdSSong Liu struct stripe_head *sh, *next; 1387a39f7afdSSong Liu 1388a39f7afdSSong Liu assert_spin_locked(&conf->device_lock); 1389a39f7afdSSong Liu if (!conf->log) 1390a39f7afdSSong Liu return; 1391a39f7afdSSong Liu 1392a39f7afdSSong Liu count = 0; 1393a39f7afdSSong Liu list_for_each_entry_safe(sh, next, &conf->r5c_full_stripe_list, lru) { 1394a39f7afdSSong Liu r5c_flush_stripe(conf, sh); 1395a39f7afdSSong Liu count++; 1396a39f7afdSSong Liu } 1397a39f7afdSSong Liu 1398a39f7afdSSong Liu if (count >= num) 1399a39f7afdSSong Liu return; 1400a39f7afdSSong Liu list_for_each_entry_safe(sh, next, 1401a39f7afdSSong Liu &conf->r5c_partial_stripe_list, lru) { 1402a39f7afdSSong Liu r5c_flush_stripe(conf, sh); 1403a39f7afdSSong Liu if (++count >= num) 1404a39f7afdSSong Liu break; 1405a39f7afdSSong Liu } 1406a39f7afdSSong Liu } 1407a39f7afdSSong Liu 1408a39f7afdSSong Liu static void r5c_do_reclaim(struct r5conf *conf) 1409a39f7afdSSong Liu { 1410a39f7afdSSong Liu struct r5l_log *log = conf->log; 1411a39f7afdSSong Liu struct stripe_head *sh; 1412a39f7afdSSong Liu int count = 0; 1413a39f7afdSSong Liu unsigned long flags; 1414a39f7afdSSong Liu int total_cached; 1415a39f7afdSSong Liu int stripes_to_flush; 1416e33fbb9cSShaohua Li int flushing_partial, flushing_full; 1417a39f7afdSSong Liu 1418a39f7afdSSong Liu if (!r5c_is_writeback(log)) 1419a39f7afdSSong Liu return; 1420a39f7afdSSong Liu 1421e33fbb9cSShaohua Li flushing_partial = atomic_read(&conf->r5c_flushing_partial_stripes); 1422e33fbb9cSShaohua Li flushing_full = atomic_read(&conf->r5c_flushing_full_stripes); 1423a39f7afdSSong Liu total_cached = atomic_read(&conf->r5c_cached_partial_stripes) + 1424e33fbb9cSShaohua Li atomic_read(&conf->r5c_cached_full_stripes) - 1425e33fbb9cSShaohua Li flushing_full - flushing_partial; 1426a39f7afdSSong Liu 1427a39f7afdSSong Liu if (total_cached > conf->min_nr_stripes * 3 / 4 || 1428a39f7afdSSong Liu atomic_read(&conf->empty_inactive_list_nr) > 0) 1429a39f7afdSSong Liu /* 1430a39f7afdSSong Liu * if stripe cache pressure high, flush all full stripes and 1431a39f7afdSSong Liu * some partial stripes 1432a39f7afdSSong Liu */ 1433a39f7afdSSong Liu stripes_to_flush = R5C_RECLAIM_STRIPE_GROUP; 1434a39f7afdSSong Liu else if (total_cached > conf->min_nr_stripes * 1 / 2 || 1435e33fbb9cSShaohua Li atomic_read(&conf->r5c_cached_full_stripes) - flushing_full > 143684890c03SShaohua Li R5C_FULL_STRIPE_FLUSH_BATCH(conf)) 1437a39f7afdSSong Liu /* 1438a39f7afdSSong Liu * if stripe cache pressure moderate, or if there is many full 1439a39f7afdSSong Liu * stripes,flush all full stripes 1440a39f7afdSSong Liu */ 1441a39f7afdSSong Liu stripes_to_flush = 0; 1442a39f7afdSSong Liu else 1443a39f7afdSSong Liu /* no need to flush */ 1444a39f7afdSSong Liu stripes_to_flush = -1; 1445a39f7afdSSong Liu 1446a39f7afdSSong Liu if (stripes_to_flush >= 0) { 1447a39f7afdSSong Liu spin_lock_irqsave(&conf->device_lock, flags); 1448a39f7afdSSong Liu r5c_flush_cache(conf, stripes_to_flush); 1449a39f7afdSSong Liu spin_unlock_irqrestore(&conf->device_lock, flags); 1450a39f7afdSSong Liu } 1451a39f7afdSSong Liu 1452a39f7afdSSong Liu /* if log space is tight, flush stripes on stripe_in_journal_list */ 1453a39f7afdSSong Liu if (test_bit(R5C_LOG_TIGHT, &conf->cache_state)) { 1454a39f7afdSSong Liu spin_lock_irqsave(&log->stripe_in_journal_lock, flags); 1455a39f7afdSSong Liu spin_lock(&conf->device_lock); 1456a39f7afdSSong Liu list_for_each_entry(sh, &log->stripe_in_journal_list, r5c) { 1457a39f7afdSSong Liu /* 1458a39f7afdSSong Liu * stripes on stripe_in_journal_list could be in any 1459a39f7afdSSong Liu * state of the stripe_cache state machine. In this 1460a39f7afdSSong Liu * case, we only want to flush stripe on 1461a39f7afdSSong Liu * r5c_cached_full/partial_stripes. The following 1462a39f7afdSSong Liu * condition makes sure the stripe is on one of the 1463a39f7afdSSong Liu * two lists. 1464a39f7afdSSong Liu */ 1465a39f7afdSSong Liu if (!list_empty(&sh->lru) && 1466a39f7afdSSong Liu !test_bit(STRIPE_HANDLE, &sh->state) && 1467a39f7afdSSong Liu atomic_read(&sh->count) == 0) { 1468a39f7afdSSong Liu r5c_flush_stripe(conf, sh); 1469a39f7afdSSong Liu if (count++ >= R5C_RECLAIM_STRIPE_GROUP) 1470a39f7afdSSong Liu break; 1471a39f7afdSSong Liu } 1472e8fd52eeSShaohua Li } 1473a39f7afdSSong Liu spin_unlock(&conf->device_lock); 1474a39f7afdSSong Liu spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags); 1475a39f7afdSSong Liu } 1476f687a33eSSong Liu 1477f687a33eSSong Liu if (!test_bit(R5C_LOG_CRITICAL, &conf->cache_state)) 1478f687a33eSSong Liu r5l_run_no_space_stripes(log); 1479f687a33eSSong Liu 1480a39f7afdSSong Liu md_wakeup_thread(conf->mddev->thread); 1481a39f7afdSSong Liu } 14824b482044SShaohua Li 14830576b1c6SShaohua Li static void r5l_do_reclaim(struct r5l_log *log) 14840576b1c6SShaohua Li { 1485a39f7afdSSong Liu struct r5conf *conf = log->rdev->mddev->private; 14860576b1c6SShaohua Li sector_t reclaim_target = xchg(&log->reclaim_target, 0); 148717036461SChristoph Hellwig sector_t reclaimable; 148817036461SChristoph Hellwig sector_t next_checkpoint; 1489a39f7afdSSong Liu bool write_super; 14900576b1c6SShaohua Li 14910576b1c6SShaohua Li spin_lock_irq(&log->io_list_lock); 1492a39f7afdSSong Liu write_super = r5l_reclaimable_space(log) > log->max_free_space || 1493a39f7afdSSong Liu reclaim_target != 0 || !list_empty(&log->no_space_stripes); 14940576b1c6SShaohua Li /* 14950576b1c6SShaohua Li * move proper io_unit to reclaim list. We should not change the order. 14960576b1c6SShaohua Li * reclaimable/unreclaimable io_unit can be mixed in the list, we 14970576b1c6SShaohua Li * shouldn't reuse space of an unreclaimable io_unit 14980576b1c6SShaohua Li */ 14990576b1c6SShaohua Li while (1) { 150017036461SChristoph Hellwig reclaimable = r5l_reclaimable_space(log); 150117036461SChristoph Hellwig if (reclaimable >= reclaim_target || 15020576b1c6SShaohua Li (list_empty(&log->running_ios) && 15030576b1c6SShaohua Li list_empty(&log->io_end_ios) && 1504a8c34f91SShaohua Li list_empty(&log->flushing_ios) && 150504732f74SChristoph Hellwig list_empty(&log->finished_ios))) 15060576b1c6SShaohua Li break; 15070576b1c6SShaohua Li 150817036461SChristoph Hellwig md_wakeup_thread(log->rdev->mddev->thread); 150917036461SChristoph Hellwig wait_event_lock_irq(log->iounit_wait, 151017036461SChristoph Hellwig r5l_reclaimable_space(log) > reclaimable, 151117036461SChristoph Hellwig log->io_list_lock); 15120576b1c6SShaohua Li } 151317036461SChristoph Hellwig 1514a39f7afdSSong Liu next_checkpoint = r5c_calculate_new_cp(conf); 15150576b1c6SShaohua Li spin_unlock_irq(&log->io_list_lock); 15160576b1c6SShaohua Li 1517a39f7afdSSong Liu if (reclaimable == 0 || !write_super) 15180576b1c6SShaohua Li return; 15190576b1c6SShaohua Li 15200576b1c6SShaohua Li /* 15210576b1c6SShaohua Li * write_super will flush cache of each raid disk. We must write super 15220576b1c6SShaohua Li * here, because the log area might be reused soon and we don't want to 15230576b1c6SShaohua Li * confuse recovery 15240576b1c6SShaohua Li */ 15254b482044SShaohua Li r5l_write_super_and_discard_space(log, next_checkpoint); 15260576b1c6SShaohua Li 15270576b1c6SShaohua Li mutex_lock(&log->io_mutex); 152817036461SChristoph Hellwig log->last_checkpoint = next_checkpoint; 1529a39f7afdSSong Liu r5c_update_log_state(log); 15300576b1c6SShaohua Li mutex_unlock(&log->io_mutex); 15310576b1c6SShaohua Li 153217036461SChristoph Hellwig r5l_run_no_space_stripes(log); 15330576b1c6SShaohua Li } 15340576b1c6SShaohua Li 15350576b1c6SShaohua Li static void r5l_reclaim_thread(struct md_thread *thread) 15360576b1c6SShaohua Li { 15370576b1c6SShaohua Li struct mddev *mddev = thread->mddev; 15380576b1c6SShaohua Li struct r5conf *conf = mddev->private; 15390576b1c6SShaohua Li struct r5l_log *log = conf->log; 15400576b1c6SShaohua Li 15410576b1c6SShaohua Li if (!log) 15420576b1c6SShaohua Li return; 1543a39f7afdSSong Liu r5c_do_reclaim(conf); 15440576b1c6SShaohua Li r5l_do_reclaim(log); 15450576b1c6SShaohua Li } 15460576b1c6SShaohua Li 1547a39f7afdSSong Liu void r5l_wake_reclaim(struct r5l_log *log, sector_t space) 1548f6bed0efSShaohua Li { 15490576b1c6SShaohua Li unsigned long target; 15500576b1c6SShaohua Li unsigned long new = (unsigned long)space; /* overflow in theory */ 15510576b1c6SShaohua Li 1552a39f7afdSSong Liu if (!log) 1553a39f7afdSSong Liu return; 15540576b1c6SShaohua Li do { 15550576b1c6SShaohua Li target = log->reclaim_target; 15560576b1c6SShaohua Li if (new < target) 15570576b1c6SShaohua Li return; 15580576b1c6SShaohua Li } while (cmpxchg(&log->reclaim_target, target, new) != target); 15590576b1c6SShaohua Li md_wakeup_thread(log->reclaim_thread); 1560f6bed0efSShaohua Li } 1561f6bed0efSShaohua Li 1562e6c033f7SShaohua Li void r5l_quiesce(struct r5l_log *log, int state) 1563e6c033f7SShaohua Li { 15644b482044SShaohua Li struct mddev *mddev; 1565e6c033f7SShaohua Li if (!log || state == 2) 1566e6c033f7SShaohua Li return; 1567ce1ccd07SShaohua Li if (state == 0) 1568ce1ccd07SShaohua Li kthread_unpark(log->reclaim_thread->tsk); 1569ce1ccd07SShaohua Li else if (state == 1) { 15704b482044SShaohua Li /* make sure r5l_write_super_and_discard_space exits */ 15714b482044SShaohua Li mddev = log->rdev->mddev; 15724b482044SShaohua Li wake_up(&mddev->sb_wait); 1573ce1ccd07SShaohua Li kthread_park(log->reclaim_thread->tsk); 1574a39f7afdSSong Liu r5l_wake_reclaim(log, MaxSector); 1575e6c033f7SShaohua Li r5l_do_reclaim(log); 1576e6c033f7SShaohua Li } 1577e6c033f7SShaohua Li } 1578e6c033f7SShaohua Li 15796e74a9cfSShaohua Li bool r5l_log_disk_error(struct r5conf *conf) 15806e74a9cfSShaohua Li { 1581f6b6ec5cSShaohua Li struct r5l_log *log; 1582f6b6ec5cSShaohua Li bool ret; 15837dde2ad3SShaohua Li /* don't allow write if journal disk is missing */ 1584f6b6ec5cSShaohua Li rcu_read_lock(); 1585f6b6ec5cSShaohua Li log = rcu_dereference(conf->log); 1586f6b6ec5cSShaohua Li 1587f6b6ec5cSShaohua Li if (!log) 1588f6b6ec5cSShaohua Li ret = test_bit(MD_HAS_JOURNAL, &conf->mddev->flags); 1589f6b6ec5cSShaohua Li else 1590f6b6ec5cSShaohua Li ret = test_bit(Faulty, &log->rdev->flags); 1591f6b6ec5cSShaohua Li rcu_read_unlock(); 1592f6b6ec5cSShaohua Li return ret; 15936e74a9cfSShaohua Li } 15946e74a9cfSShaohua Li 1595effe6ee7SSong Liu #define R5L_RECOVERY_PAGE_POOL_SIZE 256 1596effe6ee7SSong Liu 1597355810d1SShaohua Li struct r5l_recovery_ctx { 1598355810d1SShaohua Li struct page *meta_page; /* current meta */ 1599355810d1SShaohua Li sector_t meta_total_blocks; /* total size of current meta and data */ 1600355810d1SShaohua Li sector_t pos; /* recovery position */ 1601355810d1SShaohua Li u64 seq; /* recovery position seq */ 1602b4c625c6SSong Liu int data_parity_stripes; /* number of data_parity stripes */ 1603b4c625c6SSong Liu int data_only_stripes; /* number of data_only stripes */ 1604b4c625c6SSong Liu struct list_head cached_list; 1605effe6ee7SSong Liu 1606effe6ee7SSong Liu /* 1607effe6ee7SSong Liu * read ahead page pool (ra_pool) 1608effe6ee7SSong Liu * in recovery, log is read sequentially. It is not efficient to 1609effe6ee7SSong Liu * read every page with sync_page_io(). The read ahead page pool 1610effe6ee7SSong Liu * reads multiple pages with one IO, so further log read can 1611effe6ee7SSong Liu * just copy data from the pool. 1612effe6ee7SSong Liu */ 1613effe6ee7SSong Liu struct page *ra_pool[R5L_RECOVERY_PAGE_POOL_SIZE]; 1614effe6ee7SSong Liu sector_t pool_offset; /* offset of first page in the pool */ 1615effe6ee7SSong Liu int total_pages; /* total allocated pages */ 1616effe6ee7SSong Liu int valid_pages; /* pages with valid data */ 1617effe6ee7SSong Liu struct bio *ra_bio; /* bio to do the read ahead */ 1618355810d1SShaohua Li }; 1619355810d1SShaohua Li 1620effe6ee7SSong Liu static int r5l_recovery_allocate_ra_pool(struct r5l_log *log, 1621effe6ee7SSong Liu struct r5l_recovery_ctx *ctx) 1622effe6ee7SSong Liu { 1623effe6ee7SSong Liu struct page *page; 1624effe6ee7SSong Liu 1625effe6ee7SSong Liu ctx->ra_bio = bio_alloc_bioset(GFP_KERNEL, BIO_MAX_PAGES, log->bs); 1626effe6ee7SSong Liu if (!ctx->ra_bio) 1627effe6ee7SSong Liu return -ENOMEM; 1628effe6ee7SSong Liu 1629effe6ee7SSong Liu ctx->valid_pages = 0; 1630effe6ee7SSong Liu ctx->total_pages = 0; 1631effe6ee7SSong Liu while (ctx->total_pages < R5L_RECOVERY_PAGE_POOL_SIZE) { 1632effe6ee7SSong Liu page = alloc_page(GFP_KERNEL); 1633effe6ee7SSong Liu 1634effe6ee7SSong Liu if (!page) 1635effe6ee7SSong Liu break; 1636effe6ee7SSong Liu ctx->ra_pool[ctx->total_pages] = page; 1637effe6ee7SSong Liu ctx->total_pages += 1; 1638effe6ee7SSong Liu } 1639effe6ee7SSong Liu 1640effe6ee7SSong Liu if (ctx->total_pages == 0) { 1641effe6ee7SSong Liu bio_put(ctx->ra_bio); 1642effe6ee7SSong Liu return -ENOMEM; 1643effe6ee7SSong Liu } 1644effe6ee7SSong Liu 1645effe6ee7SSong Liu ctx->pool_offset = 0; 1646effe6ee7SSong Liu return 0; 1647effe6ee7SSong Liu } 1648effe6ee7SSong Liu 1649effe6ee7SSong Liu static void r5l_recovery_free_ra_pool(struct r5l_log *log, 1650effe6ee7SSong Liu struct r5l_recovery_ctx *ctx) 1651effe6ee7SSong Liu { 1652effe6ee7SSong Liu int i; 1653effe6ee7SSong Liu 1654effe6ee7SSong Liu for (i = 0; i < ctx->total_pages; ++i) 1655effe6ee7SSong Liu put_page(ctx->ra_pool[i]); 1656effe6ee7SSong Liu bio_put(ctx->ra_bio); 1657effe6ee7SSong Liu } 1658effe6ee7SSong Liu 1659effe6ee7SSong Liu /* 1660effe6ee7SSong Liu * fetch ctx->valid_pages pages from offset 1661effe6ee7SSong Liu * In normal cases, ctx->valid_pages == ctx->total_pages after the call. 1662effe6ee7SSong Liu * However, if the offset is close to the end of the journal device, 1663effe6ee7SSong Liu * ctx->valid_pages could be smaller than ctx->total_pages 1664effe6ee7SSong Liu */ 1665effe6ee7SSong Liu static int r5l_recovery_fetch_ra_pool(struct r5l_log *log, 1666effe6ee7SSong Liu struct r5l_recovery_ctx *ctx, 1667effe6ee7SSong Liu sector_t offset) 1668effe6ee7SSong Liu { 1669effe6ee7SSong Liu bio_reset(ctx->ra_bio); 1670effe6ee7SSong Liu ctx->ra_bio->bi_bdev = log->rdev->bdev; 1671effe6ee7SSong Liu bio_set_op_attrs(ctx->ra_bio, REQ_OP_READ, 0); 1672effe6ee7SSong Liu ctx->ra_bio->bi_iter.bi_sector = log->rdev->data_offset + offset; 1673effe6ee7SSong Liu 1674effe6ee7SSong Liu ctx->valid_pages = 0; 1675effe6ee7SSong Liu ctx->pool_offset = offset; 1676effe6ee7SSong Liu 1677effe6ee7SSong Liu while (ctx->valid_pages < ctx->total_pages) { 1678effe6ee7SSong Liu bio_add_page(ctx->ra_bio, 1679effe6ee7SSong Liu ctx->ra_pool[ctx->valid_pages], PAGE_SIZE, 0); 1680effe6ee7SSong Liu ctx->valid_pages += 1; 1681effe6ee7SSong Liu 1682effe6ee7SSong Liu offset = r5l_ring_add(log, offset, BLOCK_SECTORS); 1683effe6ee7SSong Liu 1684effe6ee7SSong Liu if (offset == 0) /* reached end of the device */ 1685effe6ee7SSong Liu break; 1686effe6ee7SSong Liu } 1687effe6ee7SSong Liu 1688effe6ee7SSong Liu return submit_bio_wait(ctx->ra_bio); 1689effe6ee7SSong Liu } 1690effe6ee7SSong Liu 1691effe6ee7SSong Liu /* 1692effe6ee7SSong Liu * try read a page from the read ahead page pool, if the page is not in the 1693effe6ee7SSong Liu * pool, call r5l_recovery_fetch_ra_pool 1694effe6ee7SSong Liu */ 1695effe6ee7SSong Liu static int r5l_recovery_read_page(struct r5l_log *log, 1696effe6ee7SSong Liu struct r5l_recovery_ctx *ctx, 1697effe6ee7SSong Liu struct page *page, 1698effe6ee7SSong Liu sector_t offset) 1699effe6ee7SSong Liu { 1700effe6ee7SSong Liu int ret; 1701effe6ee7SSong Liu 1702effe6ee7SSong Liu if (offset < ctx->pool_offset || 1703effe6ee7SSong Liu offset >= ctx->pool_offset + ctx->valid_pages * BLOCK_SECTORS) { 1704effe6ee7SSong Liu ret = r5l_recovery_fetch_ra_pool(log, ctx, offset); 1705effe6ee7SSong Liu if (ret) 1706effe6ee7SSong Liu return ret; 1707effe6ee7SSong Liu } 1708effe6ee7SSong Liu 1709effe6ee7SSong Liu BUG_ON(offset < ctx->pool_offset || 1710effe6ee7SSong Liu offset >= ctx->pool_offset + ctx->valid_pages * BLOCK_SECTORS); 1711effe6ee7SSong Liu 1712effe6ee7SSong Liu memcpy(page_address(page), 1713effe6ee7SSong Liu page_address(ctx->ra_pool[(offset - ctx->pool_offset) >> 1714effe6ee7SSong Liu BLOCK_SECTOR_SHIFT]), 1715effe6ee7SSong Liu PAGE_SIZE); 1716effe6ee7SSong Liu return 0; 1717effe6ee7SSong Liu } 1718effe6ee7SSong Liu 17199ed988f5SSong Liu static int r5l_recovery_read_meta_block(struct r5l_log *log, 1720355810d1SShaohua Li struct r5l_recovery_ctx *ctx) 1721355810d1SShaohua Li { 1722355810d1SShaohua Li struct page *page = ctx->meta_page; 1723355810d1SShaohua Li struct r5l_meta_block *mb; 1724355810d1SShaohua Li u32 crc, stored_crc; 1725effe6ee7SSong Liu int ret; 1726355810d1SShaohua Li 1727effe6ee7SSong Liu ret = r5l_recovery_read_page(log, ctx, page, ctx->pos); 1728effe6ee7SSong Liu if (ret != 0) 1729effe6ee7SSong Liu return ret; 1730355810d1SShaohua Li 1731355810d1SShaohua Li mb = page_address(page); 1732355810d1SShaohua Li stored_crc = le32_to_cpu(mb->checksum); 1733355810d1SShaohua Li mb->checksum = 0; 1734355810d1SShaohua Li 1735355810d1SShaohua Li if (le32_to_cpu(mb->magic) != R5LOG_MAGIC || 1736355810d1SShaohua Li le64_to_cpu(mb->seq) != ctx->seq || 1737355810d1SShaohua Li mb->version != R5LOG_VERSION || 1738355810d1SShaohua Li le64_to_cpu(mb->position) != ctx->pos) 1739355810d1SShaohua Li return -EINVAL; 1740355810d1SShaohua Li 17415cb2fbd6SShaohua Li crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE); 1742355810d1SShaohua Li if (stored_crc != crc) 1743355810d1SShaohua Li return -EINVAL; 1744355810d1SShaohua Li 1745355810d1SShaohua Li if (le32_to_cpu(mb->meta_size) > PAGE_SIZE) 1746355810d1SShaohua Li return -EINVAL; 1747355810d1SShaohua Li 1748355810d1SShaohua Li ctx->meta_total_blocks = BLOCK_SECTORS; 1749355810d1SShaohua Li 1750355810d1SShaohua Li return 0; 1751355810d1SShaohua Li } 1752355810d1SShaohua Li 17539ed988f5SSong Liu static void 17549ed988f5SSong Liu r5l_recovery_create_empty_meta_block(struct r5l_log *log, 17559ed988f5SSong Liu struct page *page, 17569ed988f5SSong Liu sector_t pos, u64 seq) 1757355810d1SShaohua Li { 1758355810d1SShaohua Li struct r5l_meta_block *mb; 1759355810d1SShaohua Li 1760355810d1SShaohua Li mb = page_address(page); 17619ed988f5SSong Liu clear_page(mb); 1762355810d1SShaohua Li mb->magic = cpu_to_le32(R5LOG_MAGIC); 1763355810d1SShaohua Li mb->version = R5LOG_VERSION; 1764355810d1SShaohua Li mb->meta_size = cpu_to_le32(sizeof(struct r5l_meta_block)); 1765355810d1SShaohua Li mb->seq = cpu_to_le64(seq); 1766355810d1SShaohua Li mb->position = cpu_to_le64(pos); 1767355810d1SShaohua Li } 1768355810d1SShaohua Li 1769355810d1SShaohua Li static int r5l_log_write_empty_meta_block(struct r5l_log *log, sector_t pos, 1770355810d1SShaohua Li u64 seq) 1771355810d1SShaohua Li { 1772355810d1SShaohua Li struct page *page; 1773355810d1SShaohua Li struct r5l_meta_block *mb; 1774355810d1SShaohua Li 17759ed988f5SSong Liu page = alloc_page(GFP_KERNEL); 1776355810d1SShaohua Li if (!page) 1777355810d1SShaohua Li return -ENOMEM; 17789ed988f5SSong Liu r5l_recovery_create_empty_meta_block(log, page, pos, seq); 1779355810d1SShaohua Li mb = page_address(page); 17805c88f403SSong Liu mb->checksum = cpu_to_le32(crc32c_le(log->uuid_checksum, 17815c88f403SSong Liu mb, PAGE_SIZE)); 1782796a5cf0SMike Christie if (!sync_page_io(log->rdev, pos, PAGE_SIZE, page, REQ_OP_WRITE, 178370fd7614SChristoph Hellwig REQ_FUA, false)) { 1784355810d1SShaohua Li __free_page(page); 1785355810d1SShaohua Li return -EIO; 1786355810d1SShaohua Li } 1787355810d1SShaohua Li __free_page(page); 1788355810d1SShaohua Li return 0; 1789355810d1SShaohua Li } 1790355810d1SShaohua Li 1791b4c625c6SSong Liu /* 1792b4c625c6SSong Liu * r5l_recovery_load_data and r5l_recovery_load_parity uses flag R5_Wantwrite 1793b4c625c6SSong Liu * to mark valid (potentially not flushed) data in the journal. 1794b4c625c6SSong Liu * 1795b4c625c6SSong Liu * We already verified checksum in r5l_recovery_verify_data_checksum_for_mb, 1796b4c625c6SSong Liu * so there should not be any mismatch here. 1797b4c625c6SSong Liu */ 1798b4c625c6SSong Liu static void r5l_recovery_load_data(struct r5l_log *log, 1799b4c625c6SSong Liu struct stripe_head *sh, 1800b4c625c6SSong Liu struct r5l_recovery_ctx *ctx, 1801b4c625c6SSong Liu struct r5l_payload_data_parity *payload, 1802b4c625c6SSong Liu sector_t log_offset) 1803f6bed0efSShaohua Li { 1804b4c625c6SSong Liu struct mddev *mddev = log->rdev->mddev; 1805b4c625c6SSong Liu struct r5conf *conf = mddev->private; 1806b4c625c6SSong Liu int dd_idx; 1807355810d1SShaohua Li 1808b4c625c6SSong Liu raid5_compute_sector(conf, 1809b4c625c6SSong Liu le64_to_cpu(payload->location), 0, 1810b4c625c6SSong Liu &dd_idx, sh); 1811effe6ee7SSong Liu r5l_recovery_read_page(log, ctx, sh->dev[dd_idx].page, log_offset); 1812b4c625c6SSong Liu sh->dev[dd_idx].log_checksum = 1813b4c625c6SSong Liu le32_to_cpu(payload->checksum[0]); 1814b4c625c6SSong Liu ctx->meta_total_blocks += BLOCK_SECTORS; 1815b4c625c6SSong Liu 1816b4c625c6SSong Liu set_bit(R5_Wantwrite, &sh->dev[dd_idx].flags); 1817b4c625c6SSong Liu set_bit(STRIPE_R5C_CACHING, &sh->state); 1818b4c625c6SSong Liu } 1819b4c625c6SSong Liu 1820b4c625c6SSong Liu static void r5l_recovery_load_parity(struct r5l_log *log, 1821b4c625c6SSong Liu struct stripe_head *sh, 1822b4c625c6SSong Liu struct r5l_recovery_ctx *ctx, 1823b4c625c6SSong Liu struct r5l_payload_data_parity *payload, 1824b4c625c6SSong Liu sector_t log_offset) 1825b4c625c6SSong Liu { 1826b4c625c6SSong Liu struct mddev *mddev = log->rdev->mddev; 1827b4c625c6SSong Liu struct r5conf *conf = mddev->private; 1828b4c625c6SSong Liu 1829b4c625c6SSong Liu ctx->meta_total_blocks += BLOCK_SECTORS * conf->max_degraded; 1830effe6ee7SSong Liu r5l_recovery_read_page(log, ctx, sh->dev[sh->pd_idx].page, log_offset); 1831b4c625c6SSong Liu sh->dev[sh->pd_idx].log_checksum = 1832b4c625c6SSong Liu le32_to_cpu(payload->checksum[0]); 1833b4c625c6SSong Liu set_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags); 1834b4c625c6SSong Liu 1835b4c625c6SSong Liu if (sh->qd_idx >= 0) { 1836effe6ee7SSong Liu r5l_recovery_read_page( 1837effe6ee7SSong Liu log, ctx, sh->dev[sh->qd_idx].page, 1838effe6ee7SSong Liu r5l_ring_add(log, log_offset, BLOCK_SECTORS)); 1839b4c625c6SSong Liu sh->dev[sh->qd_idx].log_checksum = 1840b4c625c6SSong Liu le32_to_cpu(payload->checksum[1]); 1841b4c625c6SSong Liu set_bit(R5_Wantwrite, &sh->dev[sh->qd_idx].flags); 1842b4c625c6SSong Liu } 1843b4c625c6SSong Liu clear_bit(STRIPE_R5C_CACHING, &sh->state); 1844b4c625c6SSong Liu } 1845b4c625c6SSong Liu 1846b4c625c6SSong Liu static void r5l_recovery_reset_stripe(struct stripe_head *sh) 1847b4c625c6SSong Liu { 1848b4c625c6SSong Liu int i; 1849b4c625c6SSong Liu 1850b4c625c6SSong Liu sh->state = 0; 1851b4c625c6SSong Liu sh->log_start = MaxSector; 1852b4c625c6SSong Liu for (i = sh->disks; i--; ) 1853b4c625c6SSong Liu sh->dev[i].flags = 0; 1854b4c625c6SSong Liu } 1855b4c625c6SSong Liu 1856b4c625c6SSong Liu static void 1857b4c625c6SSong Liu r5l_recovery_replay_one_stripe(struct r5conf *conf, 1858b4c625c6SSong Liu struct stripe_head *sh, 1859b4c625c6SSong Liu struct r5l_recovery_ctx *ctx) 1860b4c625c6SSong Liu { 1861b4c625c6SSong Liu struct md_rdev *rdev, *rrdev; 1862b4c625c6SSong Liu int disk_index; 1863b4c625c6SSong Liu int data_count = 0; 1864b4c625c6SSong Liu 1865b4c625c6SSong Liu for (disk_index = 0; disk_index < sh->disks; disk_index++) { 1866b4c625c6SSong Liu if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags)) 1867b4c625c6SSong Liu continue; 1868b4c625c6SSong Liu if (disk_index == sh->qd_idx || disk_index == sh->pd_idx) 1869b4c625c6SSong Liu continue; 1870b4c625c6SSong Liu data_count++; 1871b4c625c6SSong Liu } 1872b4c625c6SSong Liu 1873b4c625c6SSong Liu /* 1874b4c625c6SSong Liu * stripes that only have parity must have been flushed 1875b4c625c6SSong Liu * before the crash that we are now recovering from, so 1876b4c625c6SSong Liu * there is nothing more to recovery. 1877b4c625c6SSong Liu */ 1878b4c625c6SSong Liu if (data_count == 0) 1879b4c625c6SSong Liu goto out; 1880b4c625c6SSong Liu 1881b4c625c6SSong Liu for (disk_index = 0; disk_index < sh->disks; disk_index++) { 1882b4c625c6SSong Liu if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags)) 1883b4c625c6SSong Liu continue; 1884b4c625c6SSong Liu 1885b4c625c6SSong Liu /* in case device is broken */ 1886b4c625c6SSong Liu rcu_read_lock(); 1887b4c625c6SSong Liu rdev = rcu_dereference(conf->disks[disk_index].rdev); 1888b4c625c6SSong Liu if (rdev) { 1889b4c625c6SSong Liu atomic_inc(&rdev->nr_pending); 1890b4c625c6SSong Liu rcu_read_unlock(); 1891b4c625c6SSong Liu sync_page_io(rdev, sh->sector, PAGE_SIZE, 1892b4c625c6SSong Liu sh->dev[disk_index].page, REQ_OP_WRITE, 0, 1893b4c625c6SSong Liu false); 1894b4c625c6SSong Liu rdev_dec_pending(rdev, rdev->mddev); 1895b4c625c6SSong Liu rcu_read_lock(); 1896b4c625c6SSong Liu } 1897b4c625c6SSong Liu rrdev = rcu_dereference(conf->disks[disk_index].replacement); 1898b4c625c6SSong Liu if (rrdev) { 1899b4c625c6SSong Liu atomic_inc(&rrdev->nr_pending); 1900b4c625c6SSong Liu rcu_read_unlock(); 1901b4c625c6SSong Liu sync_page_io(rrdev, sh->sector, PAGE_SIZE, 1902b4c625c6SSong Liu sh->dev[disk_index].page, REQ_OP_WRITE, 0, 1903b4c625c6SSong Liu false); 1904b4c625c6SSong Liu rdev_dec_pending(rrdev, rrdev->mddev); 1905b4c625c6SSong Liu rcu_read_lock(); 1906b4c625c6SSong Liu } 1907b4c625c6SSong Liu rcu_read_unlock(); 1908b4c625c6SSong Liu } 1909b4c625c6SSong Liu ctx->data_parity_stripes++; 1910b4c625c6SSong Liu out: 1911b4c625c6SSong Liu r5l_recovery_reset_stripe(sh); 1912b4c625c6SSong Liu } 1913b4c625c6SSong Liu 1914b4c625c6SSong Liu static struct stripe_head * 1915b4c625c6SSong Liu r5c_recovery_alloc_stripe(struct r5conf *conf, 19163c66abbaSSong Liu sector_t stripe_sect) 1917b4c625c6SSong Liu { 1918b4c625c6SSong Liu struct stripe_head *sh; 1919b4c625c6SSong Liu 1920b4c625c6SSong Liu sh = raid5_get_active_stripe(conf, stripe_sect, 0, 1, 0); 1921b4c625c6SSong Liu if (!sh) 1922b4c625c6SSong Liu return NULL; /* no more stripe available */ 1923b4c625c6SSong Liu 1924b4c625c6SSong Liu r5l_recovery_reset_stripe(sh); 1925b4c625c6SSong Liu 1926b4c625c6SSong Liu return sh; 1927b4c625c6SSong Liu } 1928b4c625c6SSong Liu 1929b4c625c6SSong Liu static struct stripe_head * 1930b4c625c6SSong Liu r5c_recovery_lookup_stripe(struct list_head *list, sector_t sect) 1931b4c625c6SSong Liu { 1932b4c625c6SSong Liu struct stripe_head *sh; 1933b4c625c6SSong Liu 1934b4c625c6SSong Liu list_for_each_entry(sh, list, lru) 1935b4c625c6SSong Liu if (sh->sector == sect) 1936b4c625c6SSong Liu return sh; 1937b4c625c6SSong Liu return NULL; 1938b4c625c6SSong Liu } 1939b4c625c6SSong Liu 1940b4c625c6SSong Liu static void 1941b4c625c6SSong Liu r5c_recovery_drop_stripes(struct list_head *cached_stripe_list, 1942b4c625c6SSong Liu struct r5l_recovery_ctx *ctx) 1943b4c625c6SSong Liu { 1944b4c625c6SSong Liu struct stripe_head *sh, *next; 1945b4c625c6SSong Liu 1946b4c625c6SSong Liu list_for_each_entry_safe(sh, next, cached_stripe_list, lru) { 1947b4c625c6SSong Liu r5l_recovery_reset_stripe(sh); 1948b4c625c6SSong Liu list_del_init(&sh->lru); 1949b4c625c6SSong Liu raid5_release_stripe(sh); 1950b4c625c6SSong Liu } 1951b4c625c6SSong Liu } 1952b4c625c6SSong Liu 1953b4c625c6SSong Liu static void 1954b4c625c6SSong Liu r5c_recovery_replay_stripes(struct list_head *cached_stripe_list, 1955b4c625c6SSong Liu struct r5l_recovery_ctx *ctx) 1956b4c625c6SSong Liu { 1957b4c625c6SSong Liu struct stripe_head *sh, *next; 1958b4c625c6SSong Liu 1959b4c625c6SSong Liu list_for_each_entry_safe(sh, next, cached_stripe_list, lru) 1960b4c625c6SSong Liu if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) { 1961b4c625c6SSong Liu r5l_recovery_replay_one_stripe(sh->raid_conf, sh, ctx); 1962b4c625c6SSong Liu list_del_init(&sh->lru); 1963b4c625c6SSong Liu raid5_release_stripe(sh); 1964b4c625c6SSong Liu } 1965b4c625c6SSong Liu } 1966b4c625c6SSong Liu 1967b4c625c6SSong Liu /* if matches return 0; otherwise return -EINVAL */ 1968b4c625c6SSong Liu static int 1969effe6ee7SSong Liu r5l_recovery_verify_data_checksum(struct r5l_log *log, 1970effe6ee7SSong Liu struct r5l_recovery_ctx *ctx, 1971effe6ee7SSong Liu struct page *page, 1972b4c625c6SSong Liu sector_t log_offset, __le32 log_checksum) 1973b4c625c6SSong Liu { 1974b4c625c6SSong Liu void *addr; 1975b4c625c6SSong Liu u32 checksum; 1976b4c625c6SSong Liu 1977effe6ee7SSong Liu r5l_recovery_read_page(log, ctx, page, log_offset); 1978b4c625c6SSong Liu addr = kmap_atomic(page); 1979b4c625c6SSong Liu checksum = crc32c_le(log->uuid_checksum, addr, PAGE_SIZE); 1980b4c625c6SSong Liu kunmap_atomic(addr); 1981b4c625c6SSong Liu return (le32_to_cpu(log_checksum) == checksum) ? 0 : -EINVAL; 1982b4c625c6SSong Liu } 1983b4c625c6SSong Liu 1984b4c625c6SSong Liu /* 1985b4c625c6SSong Liu * before loading data to stripe cache, we need verify checksum for all data, 1986b4c625c6SSong Liu * if there is mismatch for any data page, we drop all data in the mata block 1987b4c625c6SSong Liu */ 1988b4c625c6SSong Liu static int 1989b4c625c6SSong Liu r5l_recovery_verify_data_checksum_for_mb(struct r5l_log *log, 1990b4c625c6SSong Liu struct r5l_recovery_ctx *ctx) 1991b4c625c6SSong Liu { 1992b4c625c6SSong Liu struct mddev *mddev = log->rdev->mddev; 1993b4c625c6SSong Liu struct r5conf *conf = mddev->private; 1994b4c625c6SSong Liu struct r5l_meta_block *mb = page_address(ctx->meta_page); 1995b4c625c6SSong Liu sector_t mb_offset = sizeof(struct r5l_meta_block); 1996b4c625c6SSong Liu sector_t log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS); 1997b4c625c6SSong Liu struct page *page; 1998b4c625c6SSong Liu struct r5l_payload_data_parity *payload; 19992d4f4687SSong Liu struct r5l_payload_flush *payload_flush; 2000b4c625c6SSong Liu 2001b4c625c6SSong Liu page = alloc_page(GFP_KERNEL); 2002b4c625c6SSong Liu if (!page) 2003355810d1SShaohua Li return -ENOMEM; 2004355810d1SShaohua Li 2005b4c625c6SSong Liu while (mb_offset < le32_to_cpu(mb->meta_size)) { 2006b4c625c6SSong Liu payload = (void *)mb + mb_offset; 20072d4f4687SSong Liu payload_flush = (void *)mb + mb_offset; 2008b4c625c6SSong Liu 2009b4c625c6SSong Liu if (payload->header.type == R5LOG_PAYLOAD_DATA) { 2010b4c625c6SSong Liu if (r5l_recovery_verify_data_checksum( 2011effe6ee7SSong Liu log, ctx, page, log_offset, 2012b4c625c6SSong Liu payload->checksum[0]) < 0) 2013b4c625c6SSong Liu goto mismatch; 2014b4c625c6SSong Liu } else if (payload->header.type == R5LOG_PAYLOAD_PARITY) { 2015b4c625c6SSong Liu if (r5l_recovery_verify_data_checksum( 2016effe6ee7SSong Liu log, ctx, page, log_offset, 2017b4c625c6SSong Liu payload->checksum[0]) < 0) 2018b4c625c6SSong Liu goto mismatch; 2019b4c625c6SSong Liu if (conf->max_degraded == 2 && /* q for RAID 6 */ 2020b4c625c6SSong Liu r5l_recovery_verify_data_checksum( 2021effe6ee7SSong Liu log, ctx, page, 2022b4c625c6SSong Liu r5l_ring_add(log, log_offset, 2023b4c625c6SSong Liu BLOCK_SECTORS), 2024b4c625c6SSong Liu payload->checksum[1]) < 0) 2025b4c625c6SSong Liu goto mismatch; 20262d4f4687SSong Liu } else if (payload->header.type == R5LOG_PAYLOAD_FLUSH) { 20272d4f4687SSong Liu /* nothing to do for R5LOG_PAYLOAD_FLUSH here */ 20282d4f4687SSong Liu } else /* not R5LOG_PAYLOAD_DATA/PARITY/FLUSH */ 2029b4c625c6SSong Liu goto mismatch; 2030b4c625c6SSong Liu 20312d4f4687SSong Liu if (payload->header.type == R5LOG_PAYLOAD_FLUSH) { 20322d4f4687SSong Liu mb_offset += sizeof(struct r5l_payload_flush) + 20332d4f4687SSong Liu le32_to_cpu(payload_flush->size); 20342d4f4687SSong Liu } else { 20352d4f4687SSong Liu /* DATA or PARITY payload */ 2036b4c625c6SSong Liu log_offset = r5l_ring_add(log, log_offset, 2037b4c625c6SSong Liu le32_to_cpu(payload->size)); 2038b4c625c6SSong Liu mb_offset += sizeof(struct r5l_payload_data_parity) + 2039b4c625c6SSong Liu sizeof(__le32) * 2040b4c625c6SSong Liu (le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9)); 2041b4c625c6SSong Liu } 2042b4c625c6SSong Liu 20432d4f4687SSong Liu } 20442d4f4687SSong Liu 2045b4c625c6SSong Liu put_page(page); 2046b4c625c6SSong Liu return 0; 2047b4c625c6SSong Liu 2048b4c625c6SSong Liu mismatch: 2049b4c625c6SSong Liu put_page(page); 2050b4c625c6SSong Liu return -EINVAL; 2051b4c625c6SSong Liu } 2052b4c625c6SSong Liu 2053b4c625c6SSong Liu /* 2054b4c625c6SSong Liu * Analyze all data/parity pages in one meta block 2055b4c625c6SSong Liu * Returns: 2056b4c625c6SSong Liu * 0 for success 2057b4c625c6SSong Liu * -EINVAL for unknown playload type 2058b4c625c6SSong Liu * -EAGAIN for checksum mismatch of data page 2059b4c625c6SSong Liu * -ENOMEM for run out of memory (alloc_page failed or run out of stripes) 2060b4c625c6SSong Liu */ 2061b4c625c6SSong Liu static int 2062b4c625c6SSong Liu r5c_recovery_analyze_meta_block(struct r5l_log *log, 2063b4c625c6SSong Liu struct r5l_recovery_ctx *ctx, 2064b4c625c6SSong Liu struct list_head *cached_stripe_list) 2065b4c625c6SSong Liu { 2066b4c625c6SSong Liu struct mddev *mddev = log->rdev->mddev; 2067b4c625c6SSong Liu struct r5conf *conf = mddev->private; 2068b4c625c6SSong Liu struct r5l_meta_block *mb; 2069b4c625c6SSong Liu struct r5l_payload_data_parity *payload; 20702d4f4687SSong Liu struct r5l_payload_flush *payload_flush; 2071b4c625c6SSong Liu int mb_offset; 2072b4c625c6SSong Liu sector_t log_offset; 2073b4c625c6SSong Liu sector_t stripe_sect; 2074b4c625c6SSong Liu struct stripe_head *sh; 2075b4c625c6SSong Liu int ret; 2076b4c625c6SSong Liu 2077b4c625c6SSong Liu /* 2078b4c625c6SSong Liu * for mismatch in data blocks, we will drop all data in this mb, but 2079b4c625c6SSong Liu * we will still read next mb for other data with FLUSH flag, as 2080b4c625c6SSong Liu * io_unit could finish out of order. 2081b4c625c6SSong Liu */ 2082b4c625c6SSong Liu ret = r5l_recovery_verify_data_checksum_for_mb(log, ctx); 2083b4c625c6SSong Liu if (ret == -EINVAL) 2084b4c625c6SSong Liu return -EAGAIN; 2085b4c625c6SSong Liu else if (ret) 2086b4c625c6SSong Liu return ret; /* -ENOMEM duo to alloc_page() failed */ 2087b4c625c6SSong Liu 2088b4c625c6SSong Liu mb = page_address(ctx->meta_page); 2089b4c625c6SSong Liu mb_offset = sizeof(struct r5l_meta_block); 2090b4c625c6SSong Liu log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS); 2091b4c625c6SSong Liu 2092b4c625c6SSong Liu while (mb_offset < le32_to_cpu(mb->meta_size)) { 2093b4c625c6SSong Liu int dd; 2094b4c625c6SSong Liu 2095b4c625c6SSong Liu payload = (void *)mb + mb_offset; 20962d4f4687SSong Liu payload_flush = (void *)mb + mb_offset; 20972d4f4687SSong Liu 20982d4f4687SSong Liu if (payload->header.type == R5LOG_PAYLOAD_FLUSH) { 20992d4f4687SSong Liu int i, count; 21002d4f4687SSong Liu 21012d4f4687SSong Liu count = le32_to_cpu(payload_flush->size) / sizeof(__le64); 21022d4f4687SSong Liu for (i = 0; i < count; ++i) { 21032d4f4687SSong Liu stripe_sect = le64_to_cpu(payload_flush->flush_stripes[i]); 21042d4f4687SSong Liu sh = r5c_recovery_lookup_stripe(cached_stripe_list, 21052d4f4687SSong Liu stripe_sect); 21062d4f4687SSong Liu if (sh) { 21072d4f4687SSong Liu WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state)); 21082d4f4687SSong Liu r5l_recovery_reset_stripe(sh); 21092d4f4687SSong Liu list_del_init(&sh->lru); 21102d4f4687SSong Liu raid5_release_stripe(sh); 21112d4f4687SSong Liu } 21122d4f4687SSong Liu } 21132d4f4687SSong Liu 21142d4f4687SSong Liu mb_offset += sizeof(struct r5l_payload_flush) + 21152d4f4687SSong Liu le32_to_cpu(payload_flush->size); 21162d4f4687SSong Liu continue; 21172d4f4687SSong Liu } 21182d4f4687SSong Liu 21192d4f4687SSong Liu /* DATA or PARITY payload */ 2120b4c625c6SSong Liu stripe_sect = (payload->header.type == R5LOG_PAYLOAD_DATA) ? 2121b4c625c6SSong Liu raid5_compute_sector( 2122b4c625c6SSong Liu conf, le64_to_cpu(payload->location), 0, &dd, 2123b4c625c6SSong Liu NULL) 2124b4c625c6SSong Liu : le64_to_cpu(payload->location); 2125b4c625c6SSong Liu 2126b4c625c6SSong Liu sh = r5c_recovery_lookup_stripe(cached_stripe_list, 2127b4c625c6SSong Liu stripe_sect); 2128b4c625c6SSong Liu 2129b4c625c6SSong Liu if (!sh) { 21303c66abbaSSong Liu sh = r5c_recovery_alloc_stripe(conf, stripe_sect); 2131b4c625c6SSong Liu /* 2132b4c625c6SSong Liu * cannot get stripe from raid5_get_active_stripe 2133b4c625c6SSong Liu * try replay some stripes 2134b4c625c6SSong Liu */ 2135b4c625c6SSong Liu if (!sh) { 2136b4c625c6SSong Liu r5c_recovery_replay_stripes( 2137b4c625c6SSong Liu cached_stripe_list, ctx); 2138b4c625c6SSong Liu sh = r5c_recovery_alloc_stripe( 21393c66abbaSSong Liu conf, stripe_sect); 2140b4c625c6SSong Liu } 2141b4c625c6SSong Liu if (!sh) { 2142b4c625c6SSong Liu pr_debug("md/raid:%s: Increasing stripe cache size to %d to recovery data on journal.\n", 2143b4c625c6SSong Liu mdname(mddev), 2144b4c625c6SSong Liu conf->min_nr_stripes * 2); 2145b4c625c6SSong Liu raid5_set_cache_size(mddev, 2146b4c625c6SSong Liu conf->min_nr_stripes * 2); 21473c66abbaSSong Liu sh = r5c_recovery_alloc_stripe(conf, 21483c66abbaSSong Liu stripe_sect); 2149b4c625c6SSong Liu } 2150b4c625c6SSong Liu if (!sh) { 2151b4c625c6SSong Liu pr_err("md/raid:%s: Cannot get enough stripes due to memory pressure. Recovery failed.\n", 2152b4c625c6SSong Liu mdname(mddev)); 2153b4c625c6SSong Liu return -ENOMEM; 2154b4c625c6SSong Liu } 2155b4c625c6SSong Liu list_add_tail(&sh->lru, cached_stripe_list); 2156b4c625c6SSong Liu } 2157b4c625c6SSong Liu 2158b4c625c6SSong Liu if (payload->header.type == R5LOG_PAYLOAD_DATA) { 2159f7b7bee7SZhengyuan Liu if (!test_bit(STRIPE_R5C_CACHING, &sh->state) && 2160f7b7bee7SZhengyuan Liu test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags)) { 2161b4c625c6SSong Liu r5l_recovery_replay_one_stripe(conf, sh, ctx); 2162b4c625c6SSong Liu list_move_tail(&sh->lru, cached_stripe_list); 2163b4c625c6SSong Liu } 2164b4c625c6SSong Liu r5l_recovery_load_data(log, sh, ctx, payload, 2165b4c625c6SSong Liu log_offset); 2166b4c625c6SSong Liu } else if (payload->header.type == R5LOG_PAYLOAD_PARITY) 2167b4c625c6SSong Liu r5l_recovery_load_parity(log, sh, ctx, payload, 2168b4c625c6SSong Liu log_offset); 2169b4c625c6SSong Liu else 2170b4c625c6SSong Liu return -EINVAL; 2171b4c625c6SSong Liu 2172b4c625c6SSong Liu log_offset = r5l_ring_add(log, log_offset, 2173b4c625c6SSong Liu le32_to_cpu(payload->size)); 2174b4c625c6SSong Liu 2175b4c625c6SSong Liu mb_offset += sizeof(struct r5l_payload_data_parity) + 2176b4c625c6SSong Liu sizeof(__le32) * 2177b4c625c6SSong Liu (le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9)); 2178b4c625c6SSong Liu } 2179b4c625c6SSong Liu 2180b4c625c6SSong Liu return 0; 2181b4c625c6SSong Liu } 2182b4c625c6SSong Liu 2183b4c625c6SSong Liu /* 2184b4c625c6SSong Liu * Load the stripe into cache. The stripe will be written out later by 2185b4c625c6SSong Liu * the stripe cache state machine. 2186b4c625c6SSong Liu */ 2187b4c625c6SSong Liu static void r5c_recovery_load_one_stripe(struct r5l_log *log, 2188b4c625c6SSong Liu struct stripe_head *sh) 2189b4c625c6SSong Liu { 2190b4c625c6SSong Liu struct r5dev *dev; 2191b4c625c6SSong Liu int i; 2192b4c625c6SSong Liu 2193b4c625c6SSong Liu for (i = sh->disks; i--; ) { 2194b4c625c6SSong Liu dev = sh->dev + i; 2195b4c625c6SSong Liu if (test_and_clear_bit(R5_Wantwrite, &dev->flags)) { 2196b4c625c6SSong Liu set_bit(R5_InJournal, &dev->flags); 2197b4c625c6SSong Liu set_bit(R5_UPTODATE, &dev->flags); 2198b4c625c6SSong Liu } 2199b4c625c6SSong Liu } 2200b4c625c6SSong Liu } 2201b4c625c6SSong Liu 2202b4c625c6SSong Liu /* 2203b4c625c6SSong Liu * Scan through the log for all to-be-flushed data 2204b4c625c6SSong Liu * 2205b4c625c6SSong Liu * For stripes with data and parity, namely Data-Parity stripe 2206b4c625c6SSong Liu * (STRIPE_R5C_CACHING == 0), we simply replay all the writes. 2207b4c625c6SSong Liu * 2208b4c625c6SSong Liu * For stripes with only data, namely Data-Only stripe 2209b4c625c6SSong Liu * (STRIPE_R5C_CACHING == 1), we load them to stripe cache state machine. 2210b4c625c6SSong Liu * 2211b4c625c6SSong Liu * For a stripe, if we see data after parity, we should discard all previous 2212b4c625c6SSong Liu * data and parity for this stripe, as these data are already flushed to 2213b4c625c6SSong Liu * the array. 2214b4c625c6SSong Liu * 2215b4c625c6SSong Liu * At the end of the scan, we return the new journal_tail, which points to 2216b4c625c6SSong Liu * first data-only stripe on the journal device, or next invalid meta block. 2217b4c625c6SSong Liu */ 2218b4c625c6SSong Liu static int r5c_recovery_flush_log(struct r5l_log *log, 2219b4c625c6SSong Liu struct r5l_recovery_ctx *ctx) 2220b4c625c6SSong Liu { 2221bc8f167fSJackieLiu struct stripe_head *sh; 2222b4c625c6SSong Liu int ret = 0; 2223b4c625c6SSong Liu 2224b4c625c6SSong Liu /* scan through the log */ 2225b4c625c6SSong Liu while (1) { 2226b4c625c6SSong Liu if (r5l_recovery_read_meta_block(log, ctx)) 2227b4c625c6SSong Liu break; 2228b4c625c6SSong Liu 2229b4c625c6SSong Liu ret = r5c_recovery_analyze_meta_block(log, ctx, 2230b4c625c6SSong Liu &ctx->cached_list); 2231b4c625c6SSong Liu /* 2232b4c625c6SSong Liu * -EAGAIN means mismatch in data block, in this case, we still 2233b4c625c6SSong Liu * try scan the next metablock 2234b4c625c6SSong Liu */ 2235b4c625c6SSong Liu if (ret && ret != -EAGAIN) 2236b4c625c6SSong Liu break; /* ret == -EINVAL or -ENOMEM */ 2237b4c625c6SSong Liu ctx->seq++; 2238b4c625c6SSong Liu ctx->pos = r5l_ring_add(log, ctx->pos, ctx->meta_total_blocks); 2239b4c625c6SSong Liu } 2240b4c625c6SSong Liu 2241b4c625c6SSong Liu if (ret == -ENOMEM) { 2242b4c625c6SSong Liu r5c_recovery_drop_stripes(&ctx->cached_list, ctx); 2243b4c625c6SSong Liu return ret; 2244b4c625c6SSong Liu } 2245b4c625c6SSong Liu 2246b4c625c6SSong Liu /* replay data-parity stripes */ 2247b4c625c6SSong Liu r5c_recovery_replay_stripes(&ctx->cached_list, ctx); 2248b4c625c6SSong Liu 2249b4c625c6SSong Liu /* load data-only stripes to stripe cache */ 2250bc8f167fSJackieLiu list_for_each_entry(sh, &ctx->cached_list, lru) { 2251b4c625c6SSong Liu WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state)); 2252b4c625c6SSong Liu r5c_recovery_load_one_stripe(log, sh); 2253b4c625c6SSong Liu ctx->data_only_stripes++; 2254b4c625c6SSong Liu } 2255b4c625c6SSong Liu 2256b4c625c6SSong Liu return 0; 2257b4c625c6SSong Liu } 2258355810d1SShaohua Li 2259355810d1SShaohua Li /* 2260355810d1SShaohua Li * we did a recovery. Now ctx.pos points to an invalid meta block. New 2261355810d1SShaohua Li * log will start here. but we can't let superblock point to last valid 2262355810d1SShaohua Li * meta block. The log might looks like: 2263355810d1SShaohua Li * | meta 1| meta 2| meta 3| 2264355810d1SShaohua Li * meta 1 is valid, meta 2 is invalid. meta 3 could be valid. If 2265355810d1SShaohua Li * superblock points to meta 1, we write a new valid meta 2n. if crash 2266355810d1SShaohua Li * happens again, new recovery will start from meta 1. Since meta 2n is 2267355810d1SShaohua Li * valid now, recovery will think meta 3 is valid, which is wrong. 2268355810d1SShaohua Li * The solution is we create a new meta in meta2 with its seq == meta 22693c6edc66SSong Liu * 1's seq + 10000 and let superblock points to meta2. The same recovery 22703c6edc66SSong Liu * will not think meta 3 is a valid meta, because its seq doesn't match 2271355810d1SShaohua Li */ 2272355810d1SShaohua Li 2273b4c625c6SSong Liu /* 2274b4c625c6SSong Liu * Before recovery, the log looks like the following 2275b4c625c6SSong Liu * 2276b4c625c6SSong Liu * --------------------------------------------- 2277b4c625c6SSong Liu * | valid log | invalid log | 2278b4c625c6SSong Liu * --------------------------------------------- 2279b4c625c6SSong Liu * ^ 2280b4c625c6SSong Liu * |- log->last_checkpoint 2281b4c625c6SSong Liu * |- log->last_cp_seq 2282b4c625c6SSong Liu * 2283b4c625c6SSong Liu * Now we scan through the log until we see invalid entry 2284b4c625c6SSong Liu * 2285b4c625c6SSong Liu * --------------------------------------------- 2286b4c625c6SSong Liu * | valid log | invalid log | 2287b4c625c6SSong Liu * --------------------------------------------- 2288b4c625c6SSong Liu * ^ ^ 2289b4c625c6SSong Liu * |- log->last_checkpoint |- ctx->pos 2290b4c625c6SSong Liu * |- log->last_cp_seq |- ctx->seq 2291b4c625c6SSong Liu * 2292b4c625c6SSong Liu * From this point, we need to increase seq number by 10 to avoid 2293b4c625c6SSong Liu * confusing next recovery. 2294b4c625c6SSong Liu * 2295b4c625c6SSong Liu * --------------------------------------------- 2296b4c625c6SSong Liu * | valid log | invalid log | 2297b4c625c6SSong Liu * --------------------------------------------- 2298b4c625c6SSong Liu * ^ ^ 2299b4c625c6SSong Liu * |- log->last_checkpoint |- ctx->pos+1 23003c6edc66SSong Liu * |- log->last_cp_seq |- ctx->seq+10001 2301b4c625c6SSong Liu * 2302b4c625c6SSong Liu * However, it is not safe to start the state machine yet, because data only 2303b4c625c6SSong Liu * parities are not yet secured in RAID. To save these data only parities, we 2304b4c625c6SSong Liu * rewrite them from seq+11. 2305b4c625c6SSong Liu * 2306b4c625c6SSong Liu * ----------------------------------------------------------------- 2307b4c625c6SSong Liu * | valid log | data only stripes | invalid log | 2308b4c625c6SSong Liu * ----------------------------------------------------------------- 2309b4c625c6SSong Liu * ^ ^ 2310b4c625c6SSong Liu * |- log->last_checkpoint |- ctx->pos+n 23113c6edc66SSong Liu * |- log->last_cp_seq |- ctx->seq+10000+n 2312b4c625c6SSong Liu * 2313b4c625c6SSong Liu * If failure happens again during this process, the recovery can safe start 2314b4c625c6SSong Liu * again from log->last_checkpoint. 2315b4c625c6SSong Liu * 2316b4c625c6SSong Liu * Once data only stripes are rewritten to journal, we move log_tail 2317b4c625c6SSong Liu * 2318b4c625c6SSong Liu * ----------------------------------------------------------------- 2319b4c625c6SSong Liu * | old log | data only stripes | invalid log | 2320b4c625c6SSong Liu * ----------------------------------------------------------------- 2321b4c625c6SSong Liu * ^ ^ 2322b4c625c6SSong Liu * |- log->last_checkpoint |- ctx->pos+n 23233c6edc66SSong Liu * |- log->last_cp_seq |- ctx->seq+10000+n 2324b4c625c6SSong Liu * 2325b4c625c6SSong Liu * Then we can safely start the state machine. If failure happens from this 2326b4c625c6SSong Liu * point on, the recovery will start from new log->last_checkpoint. 2327b4c625c6SSong Liu */ 2328b4c625c6SSong Liu static int 2329b4c625c6SSong Liu r5c_recovery_rewrite_data_only_stripes(struct r5l_log *log, 2330b4c625c6SSong Liu struct r5l_recovery_ctx *ctx) 2331b4c625c6SSong Liu { 2332a85dd7b8SSong Liu struct stripe_head *sh; 2333b4c625c6SSong Liu struct mddev *mddev = log->rdev->mddev; 2334b4c625c6SSong Liu struct page *page; 23353c66abbaSSong Liu sector_t next_checkpoint = MaxSector; 2336b4c625c6SSong Liu 2337b4c625c6SSong Liu page = alloc_page(GFP_KERNEL); 2338b4c625c6SSong Liu if (!page) { 2339b4c625c6SSong Liu pr_err("md/raid:%s: cannot allocate memory to rewrite data only stripes\n", 2340b4c625c6SSong Liu mdname(mddev)); 2341b4c625c6SSong Liu return -ENOMEM; 2342b4c625c6SSong Liu } 2343b4c625c6SSong Liu 23443c66abbaSSong Liu WARN_ON(list_empty(&ctx->cached_list)); 23453c66abbaSSong Liu 2346a85dd7b8SSong Liu list_for_each_entry(sh, &ctx->cached_list, lru) { 2347b4c625c6SSong Liu struct r5l_meta_block *mb; 2348b4c625c6SSong Liu int i; 2349b4c625c6SSong Liu int offset; 2350b4c625c6SSong Liu sector_t write_pos; 2351b4c625c6SSong Liu 2352b4c625c6SSong Liu WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state)); 2353b4c625c6SSong Liu r5l_recovery_create_empty_meta_block(log, page, 2354b4c625c6SSong Liu ctx->pos, ctx->seq); 2355b4c625c6SSong Liu mb = page_address(page); 2356b4c625c6SSong Liu offset = le32_to_cpu(mb->meta_size); 2357fc833c2aSJackieLiu write_pos = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS); 2358b4c625c6SSong Liu 2359b4c625c6SSong Liu for (i = sh->disks; i--; ) { 2360b4c625c6SSong Liu struct r5dev *dev = &sh->dev[i]; 2361b4c625c6SSong Liu struct r5l_payload_data_parity *payload; 2362b4c625c6SSong Liu void *addr; 2363b4c625c6SSong Liu 2364b4c625c6SSong Liu if (test_bit(R5_InJournal, &dev->flags)) { 2365b4c625c6SSong Liu payload = (void *)mb + offset; 2366b4c625c6SSong Liu payload->header.type = cpu_to_le16( 2367b4c625c6SSong Liu R5LOG_PAYLOAD_DATA); 2368b4c625c6SSong Liu payload->size = BLOCK_SECTORS; 2369b4c625c6SSong Liu payload->location = cpu_to_le64( 2370b4c625c6SSong Liu raid5_compute_blocknr(sh, i, 0)); 2371b4c625c6SSong Liu addr = kmap_atomic(dev->page); 2372b4c625c6SSong Liu payload->checksum[0] = cpu_to_le32( 2373b4c625c6SSong Liu crc32c_le(log->uuid_checksum, addr, 2374b4c625c6SSong Liu PAGE_SIZE)); 2375b4c625c6SSong Liu kunmap_atomic(addr); 2376b4c625c6SSong Liu sync_page_io(log->rdev, write_pos, PAGE_SIZE, 2377b4c625c6SSong Liu dev->page, REQ_OP_WRITE, 0, false); 2378b4c625c6SSong Liu write_pos = r5l_ring_add(log, write_pos, 2379b4c625c6SSong Liu BLOCK_SECTORS); 2380b4c625c6SSong Liu offset += sizeof(__le32) + 2381b4c625c6SSong Liu sizeof(struct r5l_payload_data_parity); 2382b4c625c6SSong Liu 2383b4c625c6SSong Liu } 2384b4c625c6SSong Liu } 2385b4c625c6SSong Liu mb->meta_size = cpu_to_le32(offset); 23865c88f403SSong Liu mb->checksum = cpu_to_le32(crc32c_le(log->uuid_checksum, 23875c88f403SSong Liu mb, PAGE_SIZE)); 2388b4c625c6SSong Liu sync_page_io(log->rdev, ctx->pos, PAGE_SIZE, page, 238920737738SShaohua Li REQ_OP_WRITE, REQ_FUA, false); 2390b4c625c6SSong Liu sh->log_start = ctx->pos; 23913c66abbaSSong Liu list_add_tail(&sh->r5c, &log->stripe_in_journal_list); 23923c66abbaSSong Liu atomic_inc(&log->stripe_in_journal_count); 2393b4c625c6SSong Liu ctx->pos = write_pos; 2394b4c625c6SSong Liu ctx->seq += 1; 23953c66abbaSSong Liu next_checkpoint = sh->log_start; 2396b4c625c6SSong Liu } 23973c66abbaSSong Liu log->next_checkpoint = next_checkpoint; 2398b4c625c6SSong Liu __free_page(page); 2399b4c625c6SSong Liu return 0; 2400b4c625c6SSong Liu } 2401b4c625c6SSong Liu 2402a85dd7b8SSong Liu static void r5c_recovery_flush_data_only_stripes(struct r5l_log *log, 2403a85dd7b8SSong Liu struct r5l_recovery_ctx *ctx) 2404a85dd7b8SSong Liu { 2405a85dd7b8SSong Liu struct mddev *mddev = log->rdev->mddev; 2406a85dd7b8SSong Liu struct r5conf *conf = mddev->private; 2407a85dd7b8SSong Liu struct stripe_head *sh, *next; 2408a85dd7b8SSong Liu 2409a85dd7b8SSong Liu if (ctx->data_only_stripes == 0) 2410a85dd7b8SSong Liu return; 2411a85dd7b8SSong Liu 2412a85dd7b8SSong Liu log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_BACK; 2413a85dd7b8SSong Liu 2414a85dd7b8SSong Liu list_for_each_entry_safe(sh, next, &ctx->cached_list, lru) { 2415a85dd7b8SSong Liu r5c_make_stripe_write_out(sh); 2416a85dd7b8SSong Liu set_bit(STRIPE_HANDLE, &sh->state); 2417a85dd7b8SSong Liu list_del_init(&sh->lru); 2418a85dd7b8SSong Liu raid5_release_stripe(sh); 2419a85dd7b8SSong Liu } 2420a85dd7b8SSong Liu 2421a85dd7b8SSong Liu md_wakeup_thread(conf->mddev->thread); 2422a85dd7b8SSong Liu /* reuse conf->wait_for_quiescent in recovery */ 2423a85dd7b8SSong Liu wait_event(conf->wait_for_quiescent, 2424a85dd7b8SSong Liu atomic_read(&conf->active_stripes) == 0); 2425a85dd7b8SSong Liu 2426a85dd7b8SSong Liu log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH; 2427a85dd7b8SSong Liu } 2428a85dd7b8SSong Liu 2429f6bed0efSShaohua Li static int r5l_recovery_log(struct r5l_log *log) 2430f6bed0efSShaohua Li { 24315aabf7c4SSong Liu struct mddev *mddev = log->rdev->mddev; 2432effe6ee7SSong Liu struct r5l_recovery_ctx *ctx; 24335aabf7c4SSong Liu int ret; 243443b96748SJackieLiu sector_t pos; 2435355810d1SShaohua Li 2436effe6ee7SSong Liu ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); 2437effe6ee7SSong Liu if (!ctx) 2438355810d1SShaohua Li return -ENOMEM; 2439355810d1SShaohua Li 2440effe6ee7SSong Liu ctx->pos = log->last_checkpoint; 2441effe6ee7SSong Liu ctx->seq = log->last_cp_seq; 2442effe6ee7SSong Liu INIT_LIST_HEAD(&ctx->cached_list); 2443effe6ee7SSong Liu ctx->meta_page = alloc_page(GFP_KERNEL); 2444effe6ee7SSong Liu 2445effe6ee7SSong Liu if (!ctx->meta_page) { 2446effe6ee7SSong Liu ret = -ENOMEM; 2447effe6ee7SSong Liu goto meta_page; 2448effe6ee7SSong Liu } 2449effe6ee7SSong Liu 2450effe6ee7SSong Liu if (r5l_recovery_allocate_ra_pool(log, ctx) != 0) { 2451effe6ee7SSong Liu ret = -ENOMEM; 2452effe6ee7SSong Liu goto ra_pool; 2453effe6ee7SSong Liu } 2454effe6ee7SSong Liu 2455effe6ee7SSong Liu ret = r5c_recovery_flush_log(log, ctx); 2456355810d1SShaohua Li 2457355810d1SShaohua Li if (ret) 2458effe6ee7SSong Liu goto error; 24595aabf7c4SSong Liu 2460effe6ee7SSong Liu pos = ctx->pos; 2461effe6ee7SSong Liu ctx->seq += 10000; 246243b96748SJackieLiu 2463effe6ee7SSong Liu if ((ctx->data_only_stripes == 0) && (ctx->data_parity_stripes == 0)) 24645aabf7c4SSong Liu pr_debug("md/raid:%s: starting from clean shutdown\n", 24655aabf7c4SSong Liu mdname(mddev)); 2466a85dd7b8SSong Liu else 246799f17890SColin Ian King pr_debug("md/raid:%s: recovering %d data-only stripes and %d data-parity stripes\n", 2468effe6ee7SSong Liu mdname(mddev), ctx->data_only_stripes, 2469effe6ee7SSong Liu ctx->data_parity_stripes); 24705aabf7c4SSong Liu 2471effe6ee7SSong Liu if (ctx->data_only_stripes == 0) { 2472effe6ee7SSong Liu log->next_checkpoint = ctx->pos; 2473effe6ee7SSong Liu r5l_log_write_empty_meta_block(log, ctx->pos, ctx->seq++); 2474effe6ee7SSong Liu ctx->pos = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS); 2475effe6ee7SSong Liu } else if (r5c_recovery_rewrite_data_only_stripes(log, ctx)) { 24765aabf7c4SSong Liu pr_err("md/raid:%s: failed to rewrite stripes to journal\n", 24775aabf7c4SSong Liu mdname(mddev)); 2478effe6ee7SSong Liu ret = -EIO; 2479effe6ee7SSong Liu goto error; 24805aabf7c4SSong Liu } 24815aabf7c4SSong Liu 2482effe6ee7SSong Liu log->log_start = ctx->pos; 2483effe6ee7SSong Liu log->seq = ctx->seq; 248443b96748SJackieLiu log->last_checkpoint = pos; 248543b96748SJackieLiu r5l_write_super(log, pos); 2486a85dd7b8SSong Liu 2487effe6ee7SSong Liu r5c_recovery_flush_data_only_stripes(log, ctx); 2488effe6ee7SSong Liu ret = 0; 2489effe6ee7SSong Liu error: 2490effe6ee7SSong Liu r5l_recovery_free_ra_pool(log, ctx); 2491effe6ee7SSong Liu ra_pool: 2492effe6ee7SSong Liu __free_page(ctx->meta_page); 2493effe6ee7SSong Liu meta_page: 2494effe6ee7SSong Liu kfree(ctx); 2495effe6ee7SSong Liu return ret; 2496f6bed0efSShaohua Li } 2497f6bed0efSShaohua Li 2498f6bed0efSShaohua Li static void r5l_write_super(struct r5l_log *log, sector_t cp) 2499f6bed0efSShaohua Li { 2500f6bed0efSShaohua Li struct mddev *mddev = log->rdev->mddev; 2501f6bed0efSShaohua Li 2502f6bed0efSShaohua Li log->rdev->journal_tail = cp; 25032953079cSShaohua Li set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags); 2504f6bed0efSShaohua Li } 2505f6bed0efSShaohua Li 25062c7da14bSSong Liu static ssize_t r5c_journal_mode_show(struct mddev *mddev, char *page) 25072c7da14bSSong Liu { 25082c7da14bSSong Liu struct r5conf *conf = mddev->private; 25092c7da14bSSong Liu int ret; 25102c7da14bSSong Liu 25112c7da14bSSong Liu if (!conf->log) 25122c7da14bSSong Liu return 0; 25132c7da14bSSong Liu 25142c7da14bSSong Liu switch (conf->log->r5c_journal_mode) { 25152c7da14bSSong Liu case R5C_JOURNAL_MODE_WRITE_THROUGH: 25162c7da14bSSong Liu ret = snprintf( 25172c7da14bSSong Liu page, PAGE_SIZE, "[%s] %s\n", 25182c7da14bSSong Liu r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH], 25192c7da14bSSong Liu r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]); 25202c7da14bSSong Liu break; 25212c7da14bSSong Liu case R5C_JOURNAL_MODE_WRITE_BACK: 25222c7da14bSSong Liu ret = snprintf( 25232c7da14bSSong Liu page, PAGE_SIZE, "%s [%s]\n", 25242c7da14bSSong Liu r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH], 25252c7da14bSSong Liu r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]); 25262c7da14bSSong Liu break; 25272c7da14bSSong Liu default: 25282c7da14bSSong Liu ret = 0; 25292c7da14bSSong Liu } 25302c7da14bSSong Liu return ret; 25312c7da14bSSong Liu } 25322c7da14bSSong Liu 25332c7da14bSSong Liu static ssize_t r5c_journal_mode_store(struct mddev *mddev, 25342c7da14bSSong Liu const char *page, size_t length) 25352c7da14bSSong Liu { 25362c7da14bSSong Liu struct r5conf *conf = mddev->private; 25372c7da14bSSong Liu struct r5l_log *log = conf->log; 25382c7da14bSSong Liu int val = -1, i; 25392c7da14bSSong Liu int len = length; 25402c7da14bSSong Liu 25412c7da14bSSong Liu if (!log) 25422c7da14bSSong Liu return -ENODEV; 25432c7da14bSSong Liu 25442c7da14bSSong Liu if (len && page[len - 1] == '\n') 25452c7da14bSSong Liu len -= 1; 25462c7da14bSSong Liu for (i = 0; i < ARRAY_SIZE(r5c_journal_mode_str); i++) 25472c7da14bSSong Liu if (strlen(r5c_journal_mode_str[i]) == len && 25482c7da14bSSong Liu strncmp(page, r5c_journal_mode_str[i], len) == 0) { 25492c7da14bSSong Liu val = i; 25502c7da14bSSong Liu break; 25512c7da14bSSong Liu } 25522c7da14bSSong Liu if (val < R5C_JOURNAL_MODE_WRITE_THROUGH || 25532c7da14bSSong Liu val > R5C_JOURNAL_MODE_WRITE_BACK) 25542c7da14bSSong Liu return -EINVAL; 25552c7da14bSSong Liu 25562e38a37fSSong Liu if (raid5_calc_degraded(conf) > 0 && 25572e38a37fSSong Liu val == R5C_JOURNAL_MODE_WRITE_BACK) 25582e38a37fSSong Liu return -EINVAL; 25592e38a37fSSong Liu 25602c7da14bSSong Liu mddev_suspend(mddev); 25612c7da14bSSong Liu conf->log->r5c_journal_mode = val; 25622c7da14bSSong Liu mddev_resume(mddev); 25632c7da14bSSong Liu 25642c7da14bSSong Liu pr_debug("md/raid:%s: setting r5c cache mode to %d: %s\n", 25652c7da14bSSong Liu mdname(mddev), val, r5c_journal_mode_str[val]); 25662c7da14bSSong Liu return length; 25672c7da14bSSong Liu } 25682c7da14bSSong Liu 25692c7da14bSSong Liu struct md_sysfs_entry 25702c7da14bSSong Liu r5c_journal_mode = __ATTR(journal_mode, 0644, 25712c7da14bSSong Liu r5c_journal_mode_show, r5c_journal_mode_store); 25722c7da14bSSong Liu 25732ded3703SSong Liu /* 25742ded3703SSong Liu * Try handle write operation in caching phase. This function should only 25752ded3703SSong Liu * be called in write-back mode. 25762ded3703SSong Liu * 25772ded3703SSong Liu * If all outstanding writes can be handled in caching phase, returns 0 25782ded3703SSong Liu * If writes requires write-out phase, call r5c_make_stripe_write_out() 25792ded3703SSong Liu * and returns -EAGAIN 25802ded3703SSong Liu */ 25812ded3703SSong Liu int r5c_try_caching_write(struct r5conf *conf, 25822ded3703SSong Liu struct stripe_head *sh, 25832ded3703SSong Liu struct stripe_head_state *s, 25842ded3703SSong Liu int disks) 25852ded3703SSong Liu { 25862ded3703SSong Liu struct r5l_log *log = conf->log; 25871e6d690bSSong Liu int i; 25881e6d690bSSong Liu struct r5dev *dev; 25891e6d690bSSong Liu int to_cache = 0; 259003b047f4SSong Liu void **pslot; 259103b047f4SSong Liu sector_t tree_index; 259203b047f4SSong Liu int ret; 259303b047f4SSong Liu uintptr_t refcount; 25942ded3703SSong Liu 25952ded3703SSong Liu BUG_ON(!r5c_is_writeback(log)); 25962ded3703SSong Liu 25971e6d690bSSong Liu if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) { 25981e6d690bSSong Liu /* 25991e6d690bSSong Liu * There are two different scenarios here: 26001e6d690bSSong Liu * 1. The stripe has some data cached, and it is sent to 26011e6d690bSSong Liu * write-out phase for reclaim 26021e6d690bSSong Liu * 2. The stripe is clean, and this is the first write 26031e6d690bSSong Liu * 26041e6d690bSSong Liu * For 1, return -EAGAIN, so we continue with 26051e6d690bSSong Liu * handle_stripe_dirtying(). 26061e6d690bSSong Liu * 26071e6d690bSSong Liu * For 2, set STRIPE_R5C_CACHING and continue with caching 26081e6d690bSSong Liu * write. 26091e6d690bSSong Liu */ 26101e6d690bSSong Liu 26111e6d690bSSong Liu /* case 1: anything injournal or anything in written */ 26121e6d690bSSong Liu if (s->injournal > 0 || s->written > 0) 26131e6d690bSSong Liu return -EAGAIN; 26141e6d690bSSong Liu /* case 2 */ 26151e6d690bSSong Liu set_bit(STRIPE_R5C_CACHING, &sh->state); 26161e6d690bSSong Liu } 26171e6d690bSSong Liu 26182e38a37fSSong Liu /* 26192e38a37fSSong Liu * When run in degraded mode, array is set to write-through mode. 26202e38a37fSSong Liu * This check helps drain pending write safely in the transition to 26212e38a37fSSong Liu * write-through mode. 26222e38a37fSSong Liu */ 26232e38a37fSSong Liu if (s->failed) { 26242e38a37fSSong Liu r5c_make_stripe_write_out(sh); 26252e38a37fSSong Liu return -EAGAIN; 26262e38a37fSSong Liu } 26272e38a37fSSong Liu 26281e6d690bSSong Liu for (i = disks; i--; ) { 26291e6d690bSSong Liu dev = &sh->dev[i]; 26301e6d690bSSong Liu /* if non-overwrite, use writing-out phase */ 26311e6d690bSSong Liu if (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags) && 26321e6d690bSSong Liu !test_bit(R5_InJournal, &dev->flags)) { 26332ded3703SSong Liu r5c_make_stripe_write_out(sh); 26342ded3703SSong Liu return -EAGAIN; 26352ded3703SSong Liu } 26361e6d690bSSong Liu } 26371e6d690bSSong Liu 263803b047f4SSong Liu /* if the stripe is not counted in big_stripe_tree, add it now */ 263903b047f4SSong Liu if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) && 264003b047f4SSong Liu !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) { 264103b047f4SSong Liu tree_index = r5c_tree_index(conf, sh->sector); 264203b047f4SSong Liu spin_lock(&log->tree_lock); 264303b047f4SSong Liu pslot = radix_tree_lookup_slot(&log->big_stripe_tree, 264403b047f4SSong Liu tree_index); 264503b047f4SSong Liu if (pslot) { 264603b047f4SSong Liu refcount = (uintptr_t)radix_tree_deref_slot_protected( 264703b047f4SSong Liu pslot, &log->tree_lock) >> 264803b047f4SSong Liu R5C_RADIX_COUNT_SHIFT; 264903b047f4SSong Liu radix_tree_replace_slot( 265003b047f4SSong Liu &log->big_stripe_tree, pslot, 265103b047f4SSong Liu (void *)((refcount + 1) << R5C_RADIX_COUNT_SHIFT)); 265203b047f4SSong Liu } else { 265303b047f4SSong Liu /* 265403b047f4SSong Liu * this radix_tree_insert can fail safely, so no 265503b047f4SSong Liu * need to call radix_tree_preload() 265603b047f4SSong Liu */ 265703b047f4SSong Liu ret = radix_tree_insert( 265803b047f4SSong Liu &log->big_stripe_tree, tree_index, 265903b047f4SSong Liu (void *)(1 << R5C_RADIX_COUNT_SHIFT)); 266003b047f4SSong Liu if (ret) { 266103b047f4SSong Liu spin_unlock(&log->tree_lock); 266203b047f4SSong Liu r5c_make_stripe_write_out(sh); 266303b047f4SSong Liu return -EAGAIN; 266403b047f4SSong Liu } 266503b047f4SSong Liu } 266603b047f4SSong Liu spin_unlock(&log->tree_lock); 266703b047f4SSong Liu 266803b047f4SSong Liu /* 266903b047f4SSong Liu * set STRIPE_R5C_PARTIAL_STRIPE, this shows the stripe is 267003b047f4SSong Liu * counted in the radix tree 267103b047f4SSong Liu */ 267203b047f4SSong Liu set_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state); 267303b047f4SSong Liu atomic_inc(&conf->r5c_cached_partial_stripes); 267403b047f4SSong Liu } 267503b047f4SSong Liu 26761e6d690bSSong Liu for (i = disks; i--; ) { 26771e6d690bSSong Liu dev = &sh->dev[i]; 26781e6d690bSSong Liu if (dev->towrite) { 26791e6d690bSSong Liu set_bit(R5_Wantwrite, &dev->flags); 26801e6d690bSSong Liu set_bit(R5_Wantdrain, &dev->flags); 26811e6d690bSSong Liu set_bit(R5_LOCKED, &dev->flags); 26821e6d690bSSong Liu to_cache++; 26831e6d690bSSong Liu } 26841e6d690bSSong Liu } 26851e6d690bSSong Liu 26861e6d690bSSong Liu if (to_cache) { 26871e6d690bSSong Liu set_bit(STRIPE_OP_BIODRAIN, &s->ops_request); 26881e6d690bSSong Liu /* 26891e6d690bSSong Liu * set STRIPE_LOG_TRAPPED, which triggers r5c_cache_data() 26901e6d690bSSong Liu * in ops_run_io(). STRIPE_LOG_TRAPPED will be cleared in 26911e6d690bSSong Liu * r5c_handle_data_cached() 26921e6d690bSSong Liu */ 26931e6d690bSSong Liu set_bit(STRIPE_LOG_TRAPPED, &sh->state); 26941e6d690bSSong Liu } 26951e6d690bSSong Liu 26961e6d690bSSong Liu return 0; 26971e6d690bSSong Liu } 26981e6d690bSSong Liu 26991e6d690bSSong Liu /* 27001e6d690bSSong Liu * free extra pages (orig_page) we allocated for prexor 27011e6d690bSSong Liu */ 27021e6d690bSSong Liu void r5c_release_extra_page(struct stripe_head *sh) 27031e6d690bSSong Liu { 2704d7bd398eSSong Liu struct r5conf *conf = sh->raid_conf; 27051e6d690bSSong Liu int i; 2706d7bd398eSSong Liu bool using_disk_info_extra_page; 2707d7bd398eSSong Liu 2708d7bd398eSSong Liu using_disk_info_extra_page = 2709d7bd398eSSong Liu sh->dev[0].orig_page == conf->disks[0].extra_page; 27101e6d690bSSong Liu 27111e6d690bSSong Liu for (i = sh->disks; i--; ) 27121e6d690bSSong Liu if (sh->dev[i].page != sh->dev[i].orig_page) { 27131e6d690bSSong Liu struct page *p = sh->dev[i].orig_page; 27141e6d690bSSong Liu 27151e6d690bSSong Liu sh->dev[i].orig_page = sh->dev[i].page; 271686aa1397SSong Liu clear_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags); 271786aa1397SSong Liu 2718d7bd398eSSong Liu if (!using_disk_info_extra_page) 27191e6d690bSSong Liu put_page(p); 27201e6d690bSSong Liu } 2721d7bd398eSSong Liu 2722d7bd398eSSong Liu if (using_disk_info_extra_page) { 2723d7bd398eSSong Liu clear_bit(R5C_EXTRA_PAGE_IN_USE, &conf->cache_state); 2724d7bd398eSSong Liu md_wakeup_thread(conf->mddev->thread); 2725d7bd398eSSong Liu } 2726d7bd398eSSong Liu } 2727d7bd398eSSong Liu 2728d7bd398eSSong Liu void r5c_use_extra_page(struct stripe_head *sh) 2729d7bd398eSSong Liu { 2730d7bd398eSSong Liu struct r5conf *conf = sh->raid_conf; 2731d7bd398eSSong Liu int i; 2732d7bd398eSSong Liu struct r5dev *dev; 2733d7bd398eSSong Liu 2734d7bd398eSSong Liu for (i = sh->disks; i--; ) { 2735d7bd398eSSong Liu dev = &sh->dev[i]; 2736d7bd398eSSong Liu if (dev->orig_page != dev->page) 2737d7bd398eSSong Liu put_page(dev->orig_page); 2738d7bd398eSSong Liu dev->orig_page = conf->disks[i].extra_page; 2739d7bd398eSSong Liu } 27401e6d690bSSong Liu } 27412ded3703SSong Liu 27422ded3703SSong Liu /* 27432ded3703SSong Liu * clean up the stripe (clear R5_InJournal for dev[pd_idx] etc.) after the 27442ded3703SSong Liu * stripe is committed to RAID disks. 27452ded3703SSong Liu */ 27462ded3703SSong Liu void r5c_finish_stripe_write_out(struct r5conf *conf, 27472ded3703SSong Liu struct stripe_head *sh, 27482ded3703SSong Liu struct stripe_head_state *s) 27492ded3703SSong Liu { 275003b047f4SSong Liu struct r5l_log *log = conf->log; 27511e6d690bSSong Liu int i; 27521e6d690bSSong Liu int do_wakeup = 0; 275303b047f4SSong Liu sector_t tree_index; 275403b047f4SSong Liu void **pslot; 275503b047f4SSong Liu uintptr_t refcount; 27561e6d690bSSong Liu 275703b047f4SSong Liu if (!log || !test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags)) 27582ded3703SSong Liu return; 27592ded3703SSong Liu 27602ded3703SSong Liu WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state)); 27612ded3703SSong Liu clear_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags); 27622ded3703SSong Liu 276303b047f4SSong Liu if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) 27642ded3703SSong Liu return; 27651e6d690bSSong Liu 27661e6d690bSSong Liu for (i = sh->disks; i--; ) { 27671e6d690bSSong Liu clear_bit(R5_InJournal, &sh->dev[i].flags); 27681e6d690bSSong Liu if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags)) 27691e6d690bSSong Liu do_wakeup = 1; 27701e6d690bSSong Liu } 27711e6d690bSSong Liu 27721e6d690bSSong Liu /* 27731e6d690bSSong Liu * analyse_stripe() runs before r5c_finish_stripe_write_out(), 27741e6d690bSSong Liu * We updated R5_InJournal, so we also update s->injournal. 27751e6d690bSSong Liu */ 27761e6d690bSSong Liu s->injournal = 0; 27771e6d690bSSong Liu 27781e6d690bSSong Liu if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state)) 27791e6d690bSSong Liu if (atomic_dec_and_test(&conf->pending_full_writes)) 27801e6d690bSSong Liu md_wakeup_thread(conf->mddev->thread); 27811e6d690bSSong Liu 27821e6d690bSSong Liu if (do_wakeup) 27831e6d690bSSong Liu wake_up(&conf->wait_for_overlap); 2784a39f7afdSSong Liu 278503b047f4SSong Liu spin_lock_irq(&log->stripe_in_journal_lock); 2786a39f7afdSSong Liu list_del_init(&sh->r5c); 278703b047f4SSong Liu spin_unlock_irq(&log->stripe_in_journal_lock); 2788a39f7afdSSong Liu sh->log_start = MaxSector; 278903b047f4SSong Liu 279003b047f4SSong Liu atomic_dec(&log->stripe_in_journal_count); 279103b047f4SSong Liu r5c_update_log_state(log); 279203b047f4SSong Liu 279303b047f4SSong Liu /* stop counting this stripe in big_stripe_tree */ 279403b047f4SSong Liu if (test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) || 279503b047f4SSong Liu test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) { 279603b047f4SSong Liu tree_index = r5c_tree_index(conf, sh->sector); 279703b047f4SSong Liu spin_lock(&log->tree_lock); 279803b047f4SSong Liu pslot = radix_tree_lookup_slot(&log->big_stripe_tree, 279903b047f4SSong Liu tree_index); 280003b047f4SSong Liu BUG_ON(pslot == NULL); 280103b047f4SSong Liu refcount = (uintptr_t)radix_tree_deref_slot_protected( 280203b047f4SSong Liu pslot, &log->tree_lock) >> 280303b047f4SSong Liu R5C_RADIX_COUNT_SHIFT; 280403b047f4SSong Liu if (refcount == 1) 280503b047f4SSong Liu radix_tree_delete(&log->big_stripe_tree, tree_index); 280603b047f4SSong Liu else 280703b047f4SSong Liu radix_tree_replace_slot( 280803b047f4SSong Liu &log->big_stripe_tree, pslot, 280903b047f4SSong Liu (void *)((refcount - 1) << R5C_RADIX_COUNT_SHIFT)); 281003b047f4SSong Liu spin_unlock(&log->tree_lock); 281103b047f4SSong Liu } 281203b047f4SSong Liu 281303b047f4SSong Liu if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) { 281403b047f4SSong Liu BUG_ON(atomic_read(&conf->r5c_cached_partial_stripes) == 0); 2815e33fbb9cSShaohua Li atomic_dec(&conf->r5c_flushing_partial_stripes); 281603b047f4SSong Liu atomic_dec(&conf->r5c_cached_partial_stripes); 281703b047f4SSong Liu } 281803b047f4SSong Liu 281903b047f4SSong Liu if (test_and_clear_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) { 282003b047f4SSong Liu BUG_ON(atomic_read(&conf->r5c_cached_full_stripes) == 0); 2821e33fbb9cSShaohua Li atomic_dec(&conf->r5c_flushing_full_stripes); 282203b047f4SSong Liu atomic_dec(&conf->r5c_cached_full_stripes); 282303b047f4SSong Liu } 2824*ea17481fSSong Liu 2825*ea17481fSSong Liu r5l_append_flush_payload(log, sh->sector); 28261e6d690bSSong Liu } 28271e6d690bSSong Liu 2828ff875738SArtur Paszkiewicz int r5c_cache_data(struct r5l_log *log, struct stripe_head *sh) 28291e6d690bSSong Liu { 2830a39f7afdSSong Liu struct r5conf *conf = sh->raid_conf; 28311e6d690bSSong Liu int pages = 0; 28321e6d690bSSong Liu int reserve; 28331e6d690bSSong Liu int i; 28341e6d690bSSong Liu int ret = 0; 28351e6d690bSSong Liu 28361e6d690bSSong Liu BUG_ON(!log); 28371e6d690bSSong Liu 28381e6d690bSSong Liu for (i = 0; i < sh->disks; i++) { 28391e6d690bSSong Liu void *addr; 28401e6d690bSSong Liu 28411e6d690bSSong Liu if (!test_bit(R5_Wantwrite, &sh->dev[i].flags)) 28421e6d690bSSong Liu continue; 28431e6d690bSSong Liu addr = kmap_atomic(sh->dev[i].page); 28441e6d690bSSong Liu sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum, 28451e6d690bSSong Liu addr, PAGE_SIZE); 28461e6d690bSSong Liu kunmap_atomic(addr); 28471e6d690bSSong Liu pages++; 28481e6d690bSSong Liu } 28491e6d690bSSong Liu WARN_ON(pages == 0); 28501e6d690bSSong Liu 28511e6d690bSSong Liu /* 28521e6d690bSSong Liu * The stripe must enter state machine again to call endio, so 28531e6d690bSSong Liu * don't delay. 28541e6d690bSSong Liu */ 28551e6d690bSSong Liu clear_bit(STRIPE_DELAYED, &sh->state); 28561e6d690bSSong Liu atomic_inc(&sh->count); 28571e6d690bSSong Liu 28581e6d690bSSong Liu mutex_lock(&log->io_mutex); 28591e6d690bSSong Liu /* meta + data */ 28601e6d690bSSong Liu reserve = (1 + pages) << (PAGE_SHIFT - 9); 28611e6d690bSSong Liu 2862a39f7afdSSong Liu if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) && 2863a39f7afdSSong Liu sh->log_start == MaxSector) 2864a39f7afdSSong Liu r5l_add_no_space_stripe(log, sh); 2865a39f7afdSSong Liu else if (!r5l_has_free_space(log, reserve)) { 2866a39f7afdSSong Liu if (sh->log_start == log->last_checkpoint) 2867a39f7afdSSong Liu BUG(); 2868a39f7afdSSong Liu else 2869a39f7afdSSong Liu r5l_add_no_space_stripe(log, sh); 28701e6d690bSSong Liu } else { 28711e6d690bSSong Liu ret = r5l_log_stripe(log, sh, pages, 0); 28721e6d690bSSong Liu if (ret) { 28731e6d690bSSong Liu spin_lock_irq(&log->io_list_lock); 28741e6d690bSSong Liu list_add_tail(&sh->log_list, &log->no_mem_stripes); 28751e6d690bSSong Liu spin_unlock_irq(&log->io_list_lock); 28761e6d690bSSong Liu } 28771e6d690bSSong Liu } 28781e6d690bSSong Liu 28791e6d690bSSong Liu mutex_unlock(&log->io_mutex); 28801e6d690bSSong Liu return 0; 2881f6bed0efSShaohua Li } 2882f6bed0efSShaohua Li 288303b047f4SSong Liu /* check whether this big stripe is in write back cache. */ 288403b047f4SSong Liu bool r5c_big_stripe_cached(struct r5conf *conf, sector_t sect) 288503b047f4SSong Liu { 288603b047f4SSong Liu struct r5l_log *log = conf->log; 288703b047f4SSong Liu sector_t tree_index; 288803b047f4SSong Liu void *slot; 288903b047f4SSong Liu 289003b047f4SSong Liu if (!log) 289103b047f4SSong Liu return false; 289203b047f4SSong Liu 289303b047f4SSong Liu WARN_ON_ONCE(!rcu_read_lock_held()); 289403b047f4SSong Liu tree_index = r5c_tree_index(conf, sect); 289503b047f4SSong Liu slot = radix_tree_lookup(&log->big_stripe_tree, tree_index); 289603b047f4SSong Liu return slot != NULL; 289703b047f4SSong Liu } 289803b047f4SSong Liu 2899f6bed0efSShaohua Li static int r5l_load_log(struct r5l_log *log) 2900f6bed0efSShaohua Li { 2901f6bed0efSShaohua Li struct md_rdev *rdev = log->rdev; 2902f6bed0efSShaohua Li struct page *page; 2903f6bed0efSShaohua Li struct r5l_meta_block *mb; 2904f6bed0efSShaohua Li sector_t cp = log->rdev->journal_tail; 2905f6bed0efSShaohua Li u32 stored_crc, expected_crc; 2906f6bed0efSShaohua Li bool create_super = false; 2907d30dfeb9SJackieLiu int ret = 0; 2908f6bed0efSShaohua Li 2909f6bed0efSShaohua Li /* Make sure it's valid */ 2910f6bed0efSShaohua Li if (cp >= rdev->sectors || round_down(cp, BLOCK_SECTORS) != cp) 2911f6bed0efSShaohua Li cp = 0; 2912f6bed0efSShaohua Li page = alloc_page(GFP_KERNEL); 2913f6bed0efSShaohua Li if (!page) 2914f6bed0efSShaohua Li return -ENOMEM; 2915f6bed0efSShaohua Li 2916796a5cf0SMike Christie if (!sync_page_io(rdev, cp, PAGE_SIZE, page, REQ_OP_READ, 0, false)) { 2917f6bed0efSShaohua Li ret = -EIO; 2918f6bed0efSShaohua Li goto ioerr; 2919f6bed0efSShaohua Li } 2920f6bed0efSShaohua Li mb = page_address(page); 2921f6bed0efSShaohua Li 2922f6bed0efSShaohua Li if (le32_to_cpu(mb->magic) != R5LOG_MAGIC || 2923f6bed0efSShaohua Li mb->version != R5LOG_VERSION) { 2924f6bed0efSShaohua Li create_super = true; 2925f6bed0efSShaohua Li goto create; 2926f6bed0efSShaohua Li } 2927f6bed0efSShaohua Li stored_crc = le32_to_cpu(mb->checksum); 2928f6bed0efSShaohua Li mb->checksum = 0; 29295cb2fbd6SShaohua Li expected_crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE); 2930f6bed0efSShaohua Li if (stored_crc != expected_crc) { 2931f6bed0efSShaohua Li create_super = true; 2932f6bed0efSShaohua Li goto create; 2933f6bed0efSShaohua Li } 2934f6bed0efSShaohua Li if (le64_to_cpu(mb->position) != cp) { 2935f6bed0efSShaohua Li create_super = true; 2936f6bed0efSShaohua Li goto create; 2937f6bed0efSShaohua Li } 2938f6bed0efSShaohua Li create: 2939f6bed0efSShaohua Li if (create_super) { 2940f6bed0efSShaohua Li log->last_cp_seq = prandom_u32(); 2941f6bed0efSShaohua Li cp = 0; 294256056c2eSZhengyuan Liu r5l_log_write_empty_meta_block(log, cp, log->last_cp_seq); 2943f6bed0efSShaohua Li /* 2944f6bed0efSShaohua Li * Make sure super points to correct address. Log might have 2945f6bed0efSShaohua Li * data very soon. If super hasn't correct log tail address, 2946f6bed0efSShaohua Li * recovery can't find the log 2947f6bed0efSShaohua Li */ 2948f6bed0efSShaohua Li r5l_write_super(log, cp); 2949f6bed0efSShaohua Li } else 2950f6bed0efSShaohua Li log->last_cp_seq = le64_to_cpu(mb->seq); 2951f6bed0efSShaohua Li 2952f6bed0efSShaohua Li log->device_size = round_down(rdev->sectors, BLOCK_SECTORS); 29530576b1c6SShaohua Li log->max_free_space = log->device_size >> RECLAIM_MAX_FREE_SPACE_SHIFT; 29540576b1c6SShaohua Li if (log->max_free_space > RECLAIM_MAX_FREE_SPACE) 29550576b1c6SShaohua Li log->max_free_space = RECLAIM_MAX_FREE_SPACE; 2956f6bed0efSShaohua Li log->last_checkpoint = cp; 2957f6bed0efSShaohua Li 2958f6bed0efSShaohua Li __free_page(page); 2959f6bed0efSShaohua Li 2960d30dfeb9SJackieLiu if (create_super) { 2961d30dfeb9SJackieLiu log->log_start = r5l_ring_add(log, cp, BLOCK_SECTORS); 2962d30dfeb9SJackieLiu log->seq = log->last_cp_seq + 1; 2963d30dfeb9SJackieLiu log->next_checkpoint = cp; 2964d30dfeb9SJackieLiu } else 29653d7e7e1dSZhengyuan Liu ret = r5l_recovery_log(log); 2966d30dfeb9SJackieLiu 29673d7e7e1dSZhengyuan Liu r5c_update_log_state(log); 29683d7e7e1dSZhengyuan Liu return ret; 2969f6bed0efSShaohua Li ioerr: 2970f6bed0efSShaohua Li __free_page(page); 2971f6bed0efSShaohua Li return ret; 2972f6bed0efSShaohua Li } 2973f6bed0efSShaohua Li 29742e38a37fSSong Liu void r5c_update_on_rdev_error(struct mddev *mddev) 29752e38a37fSSong Liu { 29762e38a37fSSong Liu struct r5conf *conf = mddev->private; 29772e38a37fSSong Liu struct r5l_log *log = conf->log; 29782e38a37fSSong Liu 29792e38a37fSSong Liu if (!log) 29802e38a37fSSong Liu return; 29812e38a37fSSong Liu 29822e38a37fSSong Liu if (raid5_calc_degraded(conf) > 0 && 29832e38a37fSSong Liu conf->log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK) 29842e38a37fSSong Liu schedule_work(&log->disable_writeback_work); 29852e38a37fSSong Liu } 29862e38a37fSSong Liu 2987f6bed0efSShaohua Li int r5l_init_log(struct r5conf *conf, struct md_rdev *rdev) 2988f6bed0efSShaohua Li { 2989c888a8f9SJens Axboe struct request_queue *q = bdev_get_queue(rdev->bdev); 2990f6bed0efSShaohua Li struct r5l_log *log; 2991ff875738SArtur Paszkiewicz char b[BDEVNAME_SIZE]; 2992ff875738SArtur Paszkiewicz 2993ff875738SArtur Paszkiewicz pr_debug("md/raid:%s: using device %s as journal\n", 2994ff875738SArtur Paszkiewicz mdname(conf->mddev), bdevname(rdev->bdev, b)); 2995f6bed0efSShaohua Li 2996f6bed0efSShaohua Li if (PAGE_SIZE != 4096) 2997f6bed0efSShaohua Li return -EINVAL; 2998c757ec95SSong Liu 2999c757ec95SSong Liu /* 3000c757ec95SSong Liu * The PAGE_SIZE must be big enough to hold 1 r5l_meta_block and 3001c757ec95SSong Liu * raid_disks r5l_payload_data_parity. 3002c757ec95SSong Liu * 3003c757ec95SSong Liu * Write journal and cache does not work for very big array 3004c757ec95SSong Liu * (raid_disks > 203) 3005c757ec95SSong Liu */ 3006c757ec95SSong Liu if (sizeof(struct r5l_meta_block) + 3007c757ec95SSong Liu ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32)) * 3008c757ec95SSong Liu conf->raid_disks) > PAGE_SIZE) { 3009c757ec95SSong Liu pr_err("md/raid:%s: write journal/cache doesn't work for array with %d disks\n", 3010c757ec95SSong Liu mdname(conf->mddev), conf->raid_disks); 3011c757ec95SSong Liu return -EINVAL; 3012c757ec95SSong Liu } 3013c757ec95SSong Liu 3014f6bed0efSShaohua Li log = kzalloc(sizeof(*log), GFP_KERNEL); 3015f6bed0efSShaohua Li if (!log) 3016f6bed0efSShaohua Li return -ENOMEM; 3017f6bed0efSShaohua Li log->rdev = rdev; 3018f6bed0efSShaohua Li 3019c888a8f9SJens Axboe log->need_cache_flush = test_bit(QUEUE_FLAG_WC, &q->queue_flags) != 0; 302056fef7c6SChristoph Hellwig 30215cb2fbd6SShaohua Li log->uuid_checksum = crc32c_le(~0, rdev->mddev->uuid, 3022f6bed0efSShaohua Li sizeof(rdev->mddev->uuid)); 3023f6bed0efSShaohua Li 3024f6bed0efSShaohua Li mutex_init(&log->io_mutex); 3025f6bed0efSShaohua Li 3026f6bed0efSShaohua Li spin_lock_init(&log->io_list_lock); 3027f6bed0efSShaohua Li INIT_LIST_HEAD(&log->running_ios); 30280576b1c6SShaohua Li INIT_LIST_HEAD(&log->io_end_ios); 3029a8c34f91SShaohua Li INIT_LIST_HEAD(&log->flushing_ios); 303004732f74SChristoph Hellwig INIT_LIST_HEAD(&log->finished_ios); 30313a83f467SMing Lei bio_init(&log->flush_bio, NULL, 0); 3032f6bed0efSShaohua Li 3033f6bed0efSShaohua Li log->io_kc = KMEM_CACHE(r5l_io_unit, 0); 3034f6bed0efSShaohua Li if (!log->io_kc) 3035f6bed0efSShaohua Li goto io_kc; 3036f6bed0efSShaohua Li 30375036c390SChristoph Hellwig log->io_pool = mempool_create_slab_pool(R5L_POOL_SIZE, log->io_kc); 30385036c390SChristoph Hellwig if (!log->io_pool) 30395036c390SChristoph Hellwig goto io_pool; 30405036c390SChristoph Hellwig 3041c38d29b3SChristoph Hellwig log->bs = bioset_create(R5L_POOL_SIZE, 0); 3042c38d29b3SChristoph Hellwig if (!log->bs) 3043c38d29b3SChristoph Hellwig goto io_bs; 3044c38d29b3SChristoph Hellwig 3045e8deb638SChristoph Hellwig log->meta_pool = mempool_create_page_pool(R5L_POOL_SIZE, 0); 3046e8deb638SChristoph Hellwig if (!log->meta_pool) 3047e8deb638SChristoph Hellwig goto out_mempool; 3048e8deb638SChristoph Hellwig 304903b047f4SSong Liu spin_lock_init(&log->tree_lock); 305003b047f4SSong Liu INIT_RADIX_TREE(&log->big_stripe_tree, GFP_NOWAIT | __GFP_NOWARN); 305103b047f4SSong Liu 30520576b1c6SShaohua Li log->reclaim_thread = md_register_thread(r5l_reclaim_thread, 30530576b1c6SShaohua Li log->rdev->mddev, "reclaim"); 30540576b1c6SShaohua Li if (!log->reclaim_thread) 30550576b1c6SShaohua Li goto reclaim_thread; 3056a39f7afdSSong Liu log->reclaim_thread->timeout = R5C_RECLAIM_WAKEUP_INTERVAL; 3057a39f7afdSSong Liu 30580fd22b45SShaohua Li init_waitqueue_head(&log->iounit_wait); 30590576b1c6SShaohua Li 30605036c390SChristoph Hellwig INIT_LIST_HEAD(&log->no_mem_stripes); 30615036c390SChristoph Hellwig 3062f6bed0efSShaohua Li INIT_LIST_HEAD(&log->no_space_stripes); 3063f6bed0efSShaohua Li spin_lock_init(&log->no_space_stripes_lock); 3064f6bed0efSShaohua Li 30653bddb7f8SSong Liu INIT_WORK(&log->deferred_io_work, r5l_submit_io_async); 30662e38a37fSSong Liu INIT_WORK(&log->disable_writeback_work, r5c_disable_writeback_async); 30673bddb7f8SSong Liu 30682ded3703SSong Liu log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH; 3069a39f7afdSSong Liu INIT_LIST_HEAD(&log->stripe_in_journal_list); 3070a39f7afdSSong Liu spin_lock_init(&log->stripe_in_journal_lock); 3071a39f7afdSSong Liu atomic_set(&log->stripe_in_journal_count, 0); 30722ded3703SSong Liu 3073d2250f10SSong Liu rcu_assign_pointer(conf->log, log); 3074d2250f10SSong Liu 3075f6bed0efSShaohua Li if (r5l_load_log(log)) 3076f6bed0efSShaohua Li goto error; 3077f6bed0efSShaohua Li 3078a62ab49eSShaohua Li set_bit(MD_HAS_JOURNAL, &conf->mddev->flags); 3079f6bed0efSShaohua Li return 0; 3080e8deb638SChristoph Hellwig 3081f6bed0efSShaohua Li error: 3082d2250f10SSong Liu rcu_assign_pointer(conf->log, NULL); 30830576b1c6SShaohua Li md_unregister_thread(&log->reclaim_thread); 30840576b1c6SShaohua Li reclaim_thread: 3085e8deb638SChristoph Hellwig mempool_destroy(log->meta_pool); 3086e8deb638SChristoph Hellwig out_mempool: 3087c38d29b3SChristoph Hellwig bioset_free(log->bs); 3088c38d29b3SChristoph Hellwig io_bs: 30895036c390SChristoph Hellwig mempool_destroy(log->io_pool); 30905036c390SChristoph Hellwig io_pool: 3091f6bed0efSShaohua Li kmem_cache_destroy(log->io_kc); 3092f6bed0efSShaohua Li io_kc: 3093f6bed0efSShaohua Li kfree(log); 3094f6bed0efSShaohua Li return -EINVAL; 3095f6bed0efSShaohua Li } 3096f6bed0efSShaohua Li 3097ff875738SArtur Paszkiewicz void r5l_exit_log(struct r5conf *conf) 3098f6bed0efSShaohua Li { 3099ff875738SArtur Paszkiewicz struct r5l_log *log = conf->log; 3100ff875738SArtur Paszkiewicz 3101ff875738SArtur Paszkiewicz conf->log = NULL; 3102ff875738SArtur Paszkiewicz synchronize_rcu(); 3103ff875738SArtur Paszkiewicz 31042e38a37fSSong Liu flush_work(&log->disable_writeback_work); 31050576b1c6SShaohua Li md_unregister_thread(&log->reclaim_thread); 3106e8deb638SChristoph Hellwig mempool_destroy(log->meta_pool); 3107c38d29b3SChristoph Hellwig bioset_free(log->bs); 31085036c390SChristoph Hellwig mempool_destroy(log->io_pool); 3109f6bed0efSShaohua Li kmem_cache_destroy(log->io_kc); 3110f6bed0efSShaohua Li kfree(log); 3111f6bed0efSShaohua Li } 3112