1f6bed0efSShaohua Li /* 2f6bed0efSShaohua Li * Copyright (C) 2015 Shaohua Li <shli@fb.com> 3f6bed0efSShaohua Li * 4f6bed0efSShaohua Li * This program is free software; you can redistribute it and/or modify it 5f6bed0efSShaohua Li * under the terms and conditions of the GNU General Public License, 6f6bed0efSShaohua Li * version 2, as published by the Free Software Foundation. 7f6bed0efSShaohua Li * 8f6bed0efSShaohua Li * This program is distributed in the hope it will be useful, but WITHOUT 9f6bed0efSShaohua Li * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 10f6bed0efSShaohua Li * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for 11f6bed0efSShaohua Li * more details. 12f6bed0efSShaohua Li * 13f6bed0efSShaohua Li */ 14f6bed0efSShaohua Li #include <linux/kernel.h> 15f6bed0efSShaohua Li #include <linux/wait.h> 16f6bed0efSShaohua Li #include <linux/blkdev.h> 17f6bed0efSShaohua Li #include <linux/slab.h> 18f6bed0efSShaohua Li #include <linux/raid/md_p.h> 195cb2fbd6SShaohua Li #include <linux/crc32c.h> 20f6bed0efSShaohua Li #include <linux/random.h> 21f6bed0efSShaohua Li #include "md.h" 22f6bed0efSShaohua Li #include "raid5.h" 23f6bed0efSShaohua Li 24f6bed0efSShaohua Li /* 25f6bed0efSShaohua Li * metadata/data stored in disk with 4k size unit (a block) regardless 26f6bed0efSShaohua Li * underneath hardware sector size. only works with PAGE_SIZE == 4096 27f6bed0efSShaohua Li */ 28f6bed0efSShaohua Li #define BLOCK_SECTORS (8) 29f6bed0efSShaohua Li 300576b1c6SShaohua Li /* 310576b1c6SShaohua Li * reclaim runs every 1/4 disk size or 10G reclaimable space. This can prevent 320576b1c6SShaohua Li * recovery scans a very long log 330576b1c6SShaohua Li */ 340576b1c6SShaohua Li #define RECLAIM_MAX_FREE_SPACE (10 * 1024 * 1024 * 2) /* sector */ 350576b1c6SShaohua Li #define RECLAIM_MAX_FREE_SPACE_SHIFT (2) 360576b1c6SShaohua Li 37f6bed0efSShaohua Li struct r5l_log { 38f6bed0efSShaohua Li struct md_rdev *rdev; 39f6bed0efSShaohua Li 40f6bed0efSShaohua Li u32 uuid_checksum; 41f6bed0efSShaohua Li 42f6bed0efSShaohua Li sector_t device_size; /* log device size, round to 43f6bed0efSShaohua Li * BLOCK_SECTORS */ 440576b1c6SShaohua Li sector_t max_free_space; /* reclaim run if free space is at 450576b1c6SShaohua Li * this size */ 46f6bed0efSShaohua Li 47f6bed0efSShaohua Li sector_t last_checkpoint; /* log tail. where recovery scan 48f6bed0efSShaohua Li * starts from */ 49f6bed0efSShaohua Li u64 last_cp_seq; /* log tail sequence */ 50f6bed0efSShaohua Li 51f6bed0efSShaohua Li sector_t log_start; /* log head. where new data appends */ 52f6bed0efSShaohua Li u64 seq; /* log head sequence */ 53f6bed0efSShaohua Li 5417036461SChristoph Hellwig sector_t next_checkpoint; 5517036461SChristoph Hellwig u64 next_cp_seq; 5617036461SChristoph Hellwig 57f6bed0efSShaohua Li struct mutex io_mutex; 58f6bed0efSShaohua Li struct r5l_io_unit *current_io; /* current io_unit accepting new data */ 59f6bed0efSShaohua Li 60f6bed0efSShaohua Li spinlock_t io_list_lock; 61f6bed0efSShaohua Li struct list_head running_ios; /* io_units which are still running, 62f6bed0efSShaohua Li * and have not yet been completely 63f6bed0efSShaohua Li * written to the log */ 64f6bed0efSShaohua Li struct list_head io_end_ios; /* io_units which have been completely 65f6bed0efSShaohua Li * written to the log but not yet written 66f6bed0efSShaohua Li * to the RAID */ 67a8c34f91SShaohua Li struct list_head flushing_ios; /* io_units which are waiting for log 68a8c34f91SShaohua Li * cache flush */ 6904732f74SChristoph Hellwig struct list_head finished_ios; /* io_units which settle down in log disk */ 70a8c34f91SShaohua Li struct bio flush_bio; 71f6bed0efSShaohua Li 72f6bed0efSShaohua Li struct kmem_cache *io_kc; 73f6bed0efSShaohua Li 740576b1c6SShaohua Li struct md_thread *reclaim_thread; 750576b1c6SShaohua Li unsigned long reclaim_target; /* number of space that need to be 760576b1c6SShaohua Li * reclaimed. if it's 0, reclaim spaces 770576b1c6SShaohua Li * used by io_units which are in 780576b1c6SShaohua Li * IO_UNIT_STRIPE_END state (eg, reclaim 790576b1c6SShaohua Li * dones't wait for specific io_unit 800576b1c6SShaohua Li * switching to IO_UNIT_STRIPE_END 810576b1c6SShaohua Li * state) */ 820fd22b45SShaohua Li wait_queue_head_t iounit_wait; 830576b1c6SShaohua Li 84f6bed0efSShaohua Li struct list_head no_space_stripes; /* pending stripes, log has no space */ 85f6bed0efSShaohua Li spinlock_t no_space_stripes_lock; 8656fef7c6SChristoph Hellwig 8756fef7c6SChristoph Hellwig bool need_cache_flush; 88f6bed0efSShaohua Li }; 89f6bed0efSShaohua Li 90f6bed0efSShaohua Li /* 91f6bed0efSShaohua Li * an IO range starts from a meta data block and end at the next meta data 92f6bed0efSShaohua Li * block. The io unit's the meta data block tracks data/parity followed it. io 93f6bed0efSShaohua Li * unit is written to log disk with normal write, as we always flush log disk 94f6bed0efSShaohua Li * first and then start move data to raid disks, there is no requirement to 95f6bed0efSShaohua Li * write io unit with FLUSH/FUA 96f6bed0efSShaohua Li */ 97f6bed0efSShaohua Li struct r5l_io_unit { 98f6bed0efSShaohua Li struct r5l_log *log; 99f6bed0efSShaohua Li 100f6bed0efSShaohua Li struct page *meta_page; /* store meta block */ 101f6bed0efSShaohua Li int meta_offset; /* current offset in meta_page */ 102f6bed0efSShaohua Li 103f6bed0efSShaohua Li struct bio_list bios; 104f6bed0efSShaohua Li atomic_t pending_io; /* pending bios not written to log yet */ 105f6bed0efSShaohua Li struct bio *current_bio;/* current_bio accepting new data */ 106f6bed0efSShaohua Li 107f6bed0efSShaohua Li atomic_t pending_stripe;/* how many stripes not flushed to raid */ 108f6bed0efSShaohua Li u64 seq; /* seq number of the metablock */ 109f6bed0efSShaohua Li sector_t log_start; /* where the io_unit starts */ 110f6bed0efSShaohua Li sector_t log_end; /* where the io_unit ends */ 111f6bed0efSShaohua Li struct list_head log_sibling; /* log->running_ios */ 112f6bed0efSShaohua Li struct list_head stripe_list; /* stripes added to the io_unit */ 113f6bed0efSShaohua Li 114f6bed0efSShaohua Li int state; 115f6bed0efSShaohua Li }; 116f6bed0efSShaohua Li 117f6bed0efSShaohua Li /* r5l_io_unit state */ 118f6bed0efSShaohua Li enum r5l_io_unit_state { 119f6bed0efSShaohua Li IO_UNIT_RUNNING = 0, /* accepting new IO */ 120f6bed0efSShaohua Li IO_UNIT_IO_START = 1, /* io_unit bio start writing to log, 121f6bed0efSShaohua Li * don't accepting new bio */ 122f6bed0efSShaohua Li IO_UNIT_IO_END = 2, /* io_unit bio finish writing to log */ 123a8c34f91SShaohua Li IO_UNIT_STRIPE_END = 3, /* stripes data finished writing to raid */ 124f6bed0efSShaohua Li }; 125f6bed0efSShaohua Li 126f6bed0efSShaohua Li static sector_t r5l_ring_add(struct r5l_log *log, sector_t start, sector_t inc) 127f6bed0efSShaohua Li { 128f6bed0efSShaohua Li start += inc; 129f6bed0efSShaohua Li if (start >= log->device_size) 130f6bed0efSShaohua Li start = start - log->device_size; 131f6bed0efSShaohua Li return start; 132f6bed0efSShaohua Li } 133f6bed0efSShaohua Li 134f6bed0efSShaohua Li static sector_t r5l_ring_distance(struct r5l_log *log, sector_t start, 135f6bed0efSShaohua Li sector_t end) 136f6bed0efSShaohua Li { 137f6bed0efSShaohua Li if (end >= start) 138f6bed0efSShaohua Li return end - start; 139f6bed0efSShaohua Li else 140f6bed0efSShaohua Li return end + log->device_size - start; 141f6bed0efSShaohua Li } 142f6bed0efSShaohua Li 143f6bed0efSShaohua Li static bool r5l_has_free_space(struct r5l_log *log, sector_t size) 144f6bed0efSShaohua Li { 145f6bed0efSShaohua Li sector_t used_size; 146f6bed0efSShaohua Li 147f6bed0efSShaohua Li used_size = r5l_ring_distance(log, log->last_checkpoint, 148f6bed0efSShaohua Li log->log_start); 149f6bed0efSShaohua Li 150f6bed0efSShaohua Li return log->device_size > used_size + size; 151f6bed0efSShaohua Li } 152f6bed0efSShaohua Li 153f6bed0efSShaohua Li static void r5l_free_io_unit(struct r5l_log *log, struct r5l_io_unit *io) 154f6bed0efSShaohua Li { 155f6bed0efSShaohua Li __free_page(io->meta_page); 156f6bed0efSShaohua Li kmem_cache_free(log->io_kc, io); 157f6bed0efSShaohua Li } 158f6bed0efSShaohua Li 159f6bed0efSShaohua Li static void r5l_move_io_unit_list(struct list_head *from, struct list_head *to, 160f6bed0efSShaohua Li enum r5l_io_unit_state state) 161f6bed0efSShaohua Li { 162f6bed0efSShaohua Li struct r5l_io_unit *io; 163f6bed0efSShaohua Li 164f6bed0efSShaohua Li while (!list_empty(from)) { 165f6bed0efSShaohua Li io = list_first_entry(from, struct r5l_io_unit, log_sibling); 166f6bed0efSShaohua Li /* don't change list order */ 167f6bed0efSShaohua Li if (io->state >= state) 168f6bed0efSShaohua Li list_move_tail(&io->log_sibling, to); 169f6bed0efSShaohua Li else 170f6bed0efSShaohua Li break; 171f6bed0efSShaohua Li } 172f6bed0efSShaohua Li } 173f6bed0efSShaohua Li 174f6bed0efSShaohua Li static void __r5l_set_io_unit_state(struct r5l_io_unit *io, 175f6bed0efSShaohua Li enum r5l_io_unit_state state) 176f6bed0efSShaohua Li { 177f6bed0efSShaohua Li if (WARN_ON(io->state >= state)) 178f6bed0efSShaohua Li return; 179f6bed0efSShaohua Li io->state = state; 180f6bed0efSShaohua Li } 181f6bed0efSShaohua Li 182d8858f43SChristoph Hellwig static void r5l_io_run_stripes(struct r5l_io_unit *io) 183d8858f43SChristoph Hellwig { 184d8858f43SChristoph Hellwig struct stripe_head *sh, *next; 185d8858f43SChristoph Hellwig 186d8858f43SChristoph Hellwig list_for_each_entry_safe(sh, next, &io->stripe_list, log_list) { 187d8858f43SChristoph Hellwig list_del_init(&sh->log_list); 188d8858f43SChristoph Hellwig set_bit(STRIPE_HANDLE, &sh->state); 189d8858f43SChristoph Hellwig raid5_release_stripe(sh); 190d8858f43SChristoph Hellwig } 191d8858f43SChristoph Hellwig } 192d8858f43SChristoph Hellwig 193f6bed0efSShaohua Li /* XXX: totally ignores I/O errors */ 19456fef7c6SChristoph Hellwig static void r5l_log_run_stripes(struct r5l_log *log) 19556fef7c6SChristoph Hellwig { 19656fef7c6SChristoph Hellwig struct r5l_io_unit *io, *next; 19756fef7c6SChristoph Hellwig 19856fef7c6SChristoph Hellwig assert_spin_locked(&log->io_list_lock); 19956fef7c6SChristoph Hellwig 20056fef7c6SChristoph Hellwig list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) { 20156fef7c6SChristoph Hellwig /* don't change list order */ 20256fef7c6SChristoph Hellwig if (io->state < IO_UNIT_IO_END) 20356fef7c6SChristoph Hellwig break; 20456fef7c6SChristoph Hellwig 20556fef7c6SChristoph Hellwig list_move_tail(&io->log_sibling, &log->finished_ios); 20656fef7c6SChristoph Hellwig r5l_io_run_stripes(io); 20756fef7c6SChristoph Hellwig } 20856fef7c6SChristoph Hellwig } 20956fef7c6SChristoph Hellwig 210f6bed0efSShaohua Li static void r5l_log_endio(struct bio *bio) 211f6bed0efSShaohua Li { 212f6bed0efSShaohua Li struct r5l_io_unit *io = bio->bi_private; 213f6bed0efSShaohua Li struct r5l_log *log = io->log; 214509ffec7SChristoph Hellwig unsigned long flags; 215f6bed0efSShaohua Li 216f6bed0efSShaohua Li bio_put(bio); 217f6bed0efSShaohua Li 218f6bed0efSShaohua Li if (!atomic_dec_and_test(&io->pending_io)) 219f6bed0efSShaohua Li return; 220f6bed0efSShaohua Li 221509ffec7SChristoph Hellwig spin_lock_irqsave(&log->io_list_lock, flags); 222509ffec7SChristoph Hellwig __r5l_set_io_unit_state(io, IO_UNIT_IO_END); 22356fef7c6SChristoph Hellwig if (log->need_cache_flush) 224509ffec7SChristoph Hellwig r5l_move_io_unit_list(&log->running_ios, &log->io_end_ios, 225509ffec7SChristoph Hellwig IO_UNIT_IO_END); 22656fef7c6SChristoph Hellwig else 22756fef7c6SChristoph Hellwig r5l_log_run_stripes(log); 228509ffec7SChristoph Hellwig spin_unlock_irqrestore(&log->io_list_lock, flags); 229509ffec7SChristoph Hellwig 23056fef7c6SChristoph Hellwig if (log->need_cache_flush) 231f6bed0efSShaohua Li md_wakeup_thread(log->rdev->mddev->thread); 232f6bed0efSShaohua Li } 233f6bed0efSShaohua Li 234f6bed0efSShaohua Li static void r5l_submit_current_io(struct r5l_log *log) 235f6bed0efSShaohua Li { 236f6bed0efSShaohua Li struct r5l_io_unit *io = log->current_io; 237f6bed0efSShaohua Li struct r5l_meta_block *block; 238f6bed0efSShaohua Li struct bio *bio; 239509ffec7SChristoph Hellwig unsigned long flags; 240f6bed0efSShaohua Li u32 crc; 241f6bed0efSShaohua Li 242f6bed0efSShaohua Li if (!io) 243f6bed0efSShaohua Li return; 244f6bed0efSShaohua Li 245f6bed0efSShaohua Li block = page_address(io->meta_page); 246f6bed0efSShaohua Li block->meta_size = cpu_to_le32(io->meta_offset); 2475cb2fbd6SShaohua Li crc = crc32c_le(log->uuid_checksum, block, PAGE_SIZE); 248f6bed0efSShaohua Li block->checksum = cpu_to_le32(crc); 249f6bed0efSShaohua Li 250f6bed0efSShaohua Li log->current_io = NULL; 251509ffec7SChristoph Hellwig spin_lock_irqsave(&log->io_list_lock, flags); 252509ffec7SChristoph Hellwig __r5l_set_io_unit_state(io, IO_UNIT_IO_START); 253509ffec7SChristoph Hellwig spin_unlock_irqrestore(&log->io_list_lock, flags); 254f6bed0efSShaohua Li 2551e932a37SChristoph Hellwig while ((bio = bio_list_pop(&io->bios))) 256f6bed0efSShaohua Li submit_bio(WRITE, bio); 257f6bed0efSShaohua Li } 258f6bed0efSShaohua Li 259b349feb3SChristoph Hellwig static struct bio *r5l_bio_alloc(struct r5l_log *log, struct r5l_io_unit *io) 260b349feb3SChristoph Hellwig { 261b349feb3SChristoph Hellwig struct bio *bio = bio_kmalloc(GFP_NOIO | __GFP_NOFAIL, BIO_MAX_PAGES); 262b349feb3SChristoph Hellwig 263b349feb3SChristoph Hellwig bio->bi_rw = WRITE; 264b349feb3SChristoph Hellwig bio->bi_bdev = log->rdev->bdev; 2651e932a37SChristoph Hellwig bio->bi_iter.bi_sector = log->rdev->data_offset + log->log_start; 266b349feb3SChristoph Hellwig bio->bi_end_io = r5l_log_endio; 267b349feb3SChristoph Hellwig bio->bi_private = io; 268b349feb3SChristoph Hellwig 269b349feb3SChristoph Hellwig bio_list_add(&io->bios, bio); 270b349feb3SChristoph Hellwig atomic_inc(&io->pending_io); 271b349feb3SChristoph Hellwig return bio; 272b349feb3SChristoph Hellwig } 273b349feb3SChristoph Hellwig 274c1b99198SChristoph Hellwig static void r5_reserve_log_entry(struct r5l_log *log, struct r5l_io_unit *io) 275c1b99198SChristoph Hellwig { 276c1b99198SChristoph Hellwig log->log_start = r5l_ring_add(log, log->log_start, BLOCK_SECTORS); 277c1b99198SChristoph Hellwig 278c1b99198SChristoph Hellwig /* 279c1b99198SChristoph Hellwig * If we filled up the log device start from the beginning again, 280c1b99198SChristoph Hellwig * which will require a new bio. 281c1b99198SChristoph Hellwig * 282c1b99198SChristoph Hellwig * Note: for this to work properly the log size needs to me a multiple 283c1b99198SChristoph Hellwig * of BLOCK_SECTORS. 284c1b99198SChristoph Hellwig */ 285c1b99198SChristoph Hellwig if (log->log_start == 0) 286c1b99198SChristoph Hellwig io->current_bio = NULL; 287c1b99198SChristoph Hellwig 288c1b99198SChristoph Hellwig io->log_end = log->log_start; 289c1b99198SChristoph Hellwig } 290c1b99198SChristoph Hellwig 291f6bed0efSShaohua Li static struct r5l_io_unit *r5l_new_meta(struct r5l_log *log) 292f6bed0efSShaohua Li { 293f6bed0efSShaohua Li struct r5l_io_unit *io; 294f6bed0efSShaohua Li struct r5l_meta_block *block; 295f6bed0efSShaohua Li 29651039cd0SChristoph Hellwig /* We can't handle memory allocate failure so far */ 29751039cd0SChristoph Hellwig io = kmem_cache_zalloc(log->io_kc, GFP_NOIO | __GFP_NOFAIL); 29851039cd0SChristoph Hellwig io->log = log; 29951039cd0SChristoph Hellwig bio_list_init(&io->bios); 30051039cd0SChristoph Hellwig INIT_LIST_HEAD(&io->log_sibling); 30151039cd0SChristoph Hellwig INIT_LIST_HEAD(&io->stripe_list); 30251039cd0SChristoph Hellwig io->state = IO_UNIT_RUNNING; 303f6bed0efSShaohua Li 30451039cd0SChristoph Hellwig io->meta_page = alloc_page(GFP_NOIO | __GFP_NOFAIL | __GFP_ZERO); 305f6bed0efSShaohua Li block = page_address(io->meta_page); 306f6bed0efSShaohua Li block->magic = cpu_to_le32(R5LOG_MAGIC); 307f6bed0efSShaohua Li block->version = R5LOG_VERSION; 308f6bed0efSShaohua Li block->seq = cpu_to_le64(log->seq); 309f6bed0efSShaohua Li block->position = cpu_to_le64(log->log_start); 310f6bed0efSShaohua Li 311f6bed0efSShaohua Li io->log_start = log->log_start; 312f6bed0efSShaohua Li io->meta_offset = sizeof(struct r5l_meta_block); 313*2b8ef16eSChristoph Hellwig io->seq = log->seq++; 314f6bed0efSShaohua Li 315b349feb3SChristoph Hellwig io->current_bio = r5l_bio_alloc(log, io); 316b349feb3SChristoph Hellwig bio_add_page(io->current_bio, io->meta_page, PAGE_SIZE, 0); 317f6bed0efSShaohua Li 318c1b99198SChristoph Hellwig r5_reserve_log_entry(log, io); 319f6bed0efSShaohua Li 320f6bed0efSShaohua Li spin_lock_irq(&log->io_list_lock); 321f6bed0efSShaohua Li list_add_tail(&io->log_sibling, &log->running_ios); 322f6bed0efSShaohua Li spin_unlock_irq(&log->io_list_lock); 323f6bed0efSShaohua Li 324f6bed0efSShaohua Li return io; 325f6bed0efSShaohua Li } 326f6bed0efSShaohua Li 327f6bed0efSShaohua Li static int r5l_get_meta(struct r5l_log *log, unsigned int payload_size) 328f6bed0efSShaohua Li { 32922581f58SChristoph Hellwig if (log->current_io && 33022581f58SChristoph Hellwig log->current_io->meta_offset + payload_size > PAGE_SIZE) 331f6bed0efSShaohua Li r5l_submit_current_io(log); 332f6bed0efSShaohua Li 33322581f58SChristoph Hellwig if (!log->current_io) 334f6bed0efSShaohua Li log->current_io = r5l_new_meta(log); 335f6bed0efSShaohua Li return 0; 336f6bed0efSShaohua Li } 337f6bed0efSShaohua Li 338f6bed0efSShaohua Li static void r5l_append_payload_meta(struct r5l_log *log, u16 type, 339f6bed0efSShaohua Li sector_t location, 340f6bed0efSShaohua Li u32 checksum1, u32 checksum2, 341f6bed0efSShaohua Li bool checksum2_valid) 342f6bed0efSShaohua Li { 343f6bed0efSShaohua Li struct r5l_io_unit *io = log->current_io; 344f6bed0efSShaohua Li struct r5l_payload_data_parity *payload; 345f6bed0efSShaohua Li 346f6bed0efSShaohua Li payload = page_address(io->meta_page) + io->meta_offset; 347f6bed0efSShaohua Li payload->header.type = cpu_to_le16(type); 348f6bed0efSShaohua Li payload->header.flags = cpu_to_le16(0); 349f6bed0efSShaohua Li payload->size = cpu_to_le32((1 + !!checksum2_valid) << 350f6bed0efSShaohua Li (PAGE_SHIFT - 9)); 351f6bed0efSShaohua Li payload->location = cpu_to_le64(location); 352f6bed0efSShaohua Li payload->checksum[0] = cpu_to_le32(checksum1); 353f6bed0efSShaohua Li if (checksum2_valid) 354f6bed0efSShaohua Li payload->checksum[1] = cpu_to_le32(checksum2); 355f6bed0efSShaohua Li 356f6bed0efSShaohua Li io->meta_offset += sizeof(struct r5l_payload_data_parity) + 357f6bed0efSShaohua Li sizeof(__le32) * (1 + !!checksum2_valid); 358f6bed0efSShaohua Li } 359f6bed0efSShaohua Li 360f6bed0efSShaohua Li static void r5l_append_payload_page(struct r5l_log *log, struct page *page) 361f6bed0efSShaohua Li { 362f6bed0efSShaohua Li struct r5l_io_unit *io = log->current_io; 363f6bed0efSShaohua Li 364f6bed0efSShaohua Li alloc_bio: 365b349feb3SChristoph Hellwig if (!io->current_bio) 366b349feb3SChristoph Hellwig io->current_bio = r5l_bio_alloc(log, io); 367f6bed0efSShaohua Li 368f6bed0efSShaohua Li if (!bio_add_page(io->current_bio, page, PAGE_SIZE, 0)) { 369f6bed0efSShaohua Li io->current_bio = NULL; 370f6bed0efSShaohua Li goto alloc_bio; 371f6bed0efSShaohua Li } 372f6bed0efSShaohua Li 373c1b99198SChristoph Hellwig r5_reserve_log_entry(log, io); 374f6bed0efSShaohua Li } 375f6bed0efSShaohua Li 376f6bed0efSShaohua Li static void r5l_log_stripe(struct r5l_log *log, struct stripe_head *sh, 377f6bed0efSShaohua Li int data_pages, int parity_pages) 378f6bed0efSShaohua Li { 379f6bed0efSShaohua Li int i; 380f6bed0efSShaohua Li int meta_size; 381f6bed0efSShaohua Li struct r5l_io_unit *io; 382f6bed0efSShaohua Li 383f6bed0efSShaohua Li meta_size = 384f6bed0efSShaohua Li ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32)) 385f6bed0efSShaohua Li * data_pages) + 386f6bed0efSShaohua Li sizeof(struct r5l_payload_data_parity) + 387f6bed0efSShaohua Li sizeof(__le32) * parity_pages; 388f6bed0efSShaohua Li 389f6bed0efSShaohua Li r5l_get_meta(log, meta_size); 390f6bed0efSShaohua Li io = log->current_io; 391f6bed0efSShaohua Li 392f6bed0efSShaohua Li for (i = 0; i < sh->disks; i++) { 393f6bed0efSShaohua Li if (!test_bit(R5_Wantwrite, &sh->dev[i].flags)) 394f6bed0efSShaohua Li continue; 395f6bed0efSShaohua Li if (i == sh->pd_idx || i == sh->qd_idx) 396f6bed0efSShaohua Li continue; 397f6bed0efSShaohua Li r5l_append_payload_meta(log, R5LOG_PAYLOAD_DATA, 398f6bed0efSShaohua Li raid5_compute_blocknr(sh, i, 0), 399f6bed0efSShaohua Li sh->dev[i].log_checksum, 0, false); 400f6bed0efSShaohua Li r5l_append_payload_page(log, sh->dev[i].page); 401f6bed0efSShaohua Li } 402f6bed0efSShaohua Li 403f6bed0efSShaohua Li if (sh->qd_idx >= 0) { 404f6bed0efSShaohua Li r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY, 405f6bed0efSShaohua Li sh->sector, sh->dev[sh->pd_idx].log_checksum, 406f6bed0efSShaohua Li sh->dev[sh->qd_idx].log_checksum, true); 407f6bed0efSShaohua Li r5l_append_payload_page(log, sh->dev[sh->pd_idx].page); 408f6bed0efSShaohua Li r5l_append_payload_page(log, sh->dev[sh->qd_idx].page); 409f6bed0efSShaohua Li } else { 410f6bed0efSShaohua Li r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY, 411f6bed0efSShaohua Li sh->sector, sh->dev[sh->pd_idx].log_checksum, 412f6bed0efSShaohua Li 0, false); 413f6bed0efSShaohua Li r5l_append_payload_page(log, sh->dev[sh->pd_idx].page); 414f6bed0efSShaohua Li } 415f6bed0efSShaohua Li 416f6bed0efSShaohua Li list_add_tail(&sh->log_list, &io->stripe_list); 417f6bed0efSShaohua Li atomic_inc(&io->pending_stripe); 418f6bed0efSShaohua Li sh->log_io = io; 419f6bed0efSShaohua Li } 420f6bed0efSShaohua Li 421509ffec7SChristoph Hellwig static void r5l_wake_reclaim(struct r5l_log *log, sector_t space); 422f6bed0efSShaohua Li /* 423f6bed0efSShaohua Li * running in raid5d, where reclaim could wait for raid5d too (when it flushes 424f6bed0efSShaohua Li * data from log to raid disks), so we shouldn't wait for reclaim here 425f6bed0efSShaohua Li */ 426f6bed0efSShaohua Li int r5l_write_stripe(struct r5l_log *log, struct stripe_head *sh) 427f6bed0efSShaohua Li { 428f6bed0efSShaohua Li int write_disks = 0; 429f6bed0efSShaohua Li int data_pages, parity_pages; 430f6bed0efSShaohua Li int meta_size; 431f6bed0efSShaohua Li int reserve; 432f6bed0efSShaohua Li int i; 433f6bed0efSShaohua Li 434f6bed0efSShaohua Li if (!log) 435f6bed0efSShaohua Li return -EAGAIN; 436f6bed0efSShaohua Li /* Don't support stripe batch */ 437f6bed0efSShaohua Li if (sh->log_io || !test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags) || 438f6bed0efSShaohua Li test_bit(STRIPE_SYNCING, &sh->state)) { 439f6bed0efSShaohua Li /* the stripe is written to log, we start writing it to raid */ 440f6bed0efSShaohua Li clear_bit(STRIPE_LOG_TRAPPED, &sh->state); 441f6bed0efSShaohua Li return -EAGAIN; 442f6bed0efSShaohua Li } 443f6bed0efSShaohua Li 444f6bed0efSShaohua Li for (i = 0; i < sh->disks; i++) { 445f6bed0efSShaohua Li void *addr; 446f6bed0efSShaohua Li 447f6bed0efSShaohua Li if (!test_bit(R5_Wantwrite, &sh->dev[i].flags)) 448f6bed0efSShaohua Li continue; 449f6bed0efSShaohua Li write_disks++; 450f6bed0efSShaohua Li /* checksum is already calculated in last run */ 451f6bed0efSShaohua Li if (test_bit(STRIPE_LOG_TRAPPED, &sh->state)) 452f6bed0efSShaohua Li continue; 453f6bed0efSShaohua Li addr = kmap_atomic(sh->dev[i].page); 4545cb2fbd6SShaohua Li sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum, 455f6bed0efSShaohua Li addr, PAGE_SIZE); 456f6bed0efSShaohua Li kunmap_atomic(addr); 457f6bed0efSShaohua Li } 458f6bed0efSShaohua Li parity_pages = 1 + !!(sh->qd_idx >= 0); 459f6bed0efSShaohua Li data_pages = write_disks - parity_pages; 460f6bed0efSShaohua Li 461f6bed0efSShaohua Li meta_size = 462f6bed0efSShaohua Li ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32)) 463f6bed0efSShaohua Li * data_pages) + 464f6bed0efSShaohua Li sizeof(struct r5l_payload_data_parity) + 465f6bed0efSShaohua Li sizeof(__le32) * parity_pages; 466f6bed0efSShaohua Li /* Doesn't work with very big raid array */ 467f6bed0efSShaohua Li if (meta_size + sizeof(struct r5l_meta_block) > PAGE_SIZE) 468f6bed0efSShaohua Li return -EINVAL; 469f6bed0efSShaohua Li 470f6bed0efSShaohua Li set_bit(STRIPE_LOG_TRAPPED, &sh->state); 471253f9fd4SShaohua Li /* 472253f9fd4SShaohua Li * The stripe must enter state machine again to finish the write, so 473253f9fd4SShaohua Li * don't delay. 474253f9fd4SShaohua Li */ 475253f9fd4SShaohua Li clear_bit(STRIPE_DELAYED, &sh->state); 476f6bed0efSShaohua Li atomic_inc(&sh->count); 477f6bed0efSShaohua Li 478f6bed0efSShaohua Li mutex_lock(&log->io_mutex); 479f6bed0efSShaohua Li /* meta + data */ 480f6bed0efSShaohua Li reserve = (1 + write_disks) << (PAGE_SHIFT - 9); 481f6bed0efSShaohua Li if (r5l_has_free_space(log, reserve)) 482f6bed0efSShaohua Li r5l_log_stripe(log, sh, data_pages, parity_pages); 483f6bed0efSShaohua Li else { 484f6bed0efSShaohua Li spin_lock(&log->no_space_stripes_lock); 485f6bed0efSShaohua Li list_add_tail(&sh->log_list, &log->no_space_stripes); 486f6bed0efSShaohua Li spin_unlock(&log->no_space_stripes_lock); 487f6bed0efSShaohua Li 488f6bed0efSShaohua Li r5l_wake_reclaim(log, reserve); 489f6bed0efSShaohua Li } 490f6bed0efSShaohua Li mutex_unlock(&log->io_mutex); 491f6bed0efSShaohua Li 492f6bed0efSShaohua Li return 0; 493f6bed0efSShaohua Li } 494f6bed0efSShaohua Li 495f6bed0efSShaohua Li void r5l_write_stripe_run(struct r5l_log *log) 496f6bed0efSShaohua Li { 497f6bed0efSShaohua Li if (!log) 498f6bed0efSShaohua Li return; 499f6bed0efSShaohua Li mutex_lock(&log->io_mutex); 500f6bed0efSShaohua Li r5l_submit_current_io(log); 501f6bed0efSShaohua Li mutex_unlock(&log->io_mutex); 502f6bed0efSShaohua Li } 503f6bed0efSShaohua Li 504828cbe98SShaohua Li int r5l_handle_flush_request(struct r5l_log *log, struct bio *bio) 505828cbe98SShaohua Li { 506828cbe98SShaohua Li if (!log) 507828cbe98SShaohua Li return -ENODEV; 508828cbe98SShaohua Li /* 509828cbe98SShaohua Li * we flush log disk cache first, then write stripe data to raid disks. 510828cbe98SShaohua Li * So if bio is finished, the log disk cache is flushed already. The 511828cbe98SShaohua Li * recovery guarantees we can recovery the bio from log disk, so we 512828cbe98SShaohua Li * don't need to flush again 513828cbe98SShaohua Li */ 514828cbe98SShaohua Li if (bio->bi_iter.bi_size == 0) { 515828cbe98SShaohua Li bio_endio(bio); 516828cbe98SShaohua Li return 0; 517828cbe98SShaohua Li } 518828cbe98SShaohua Li bio->bi_rw &= ~REQ_FLUSH; 519828cbe98SShaohua Li return -EAGAIN; 520828cbe98SShaohua Li } 521828cbe98SShaohua Li 522f6bed0efSShaohua Li /* This will run after log space is reclaimed */ 523f6bed0efSShaohua Li static void r5l_run_no_space_stripes(struct r5l_log *log) 524f6bed0efSShaohua Li { 525f6bed0efSShaohua Li struct stripe_head *sh; 526f6bed0efSShaohua Li 527f6bed0efSShaohua Li spin_lock(&log->no_space_stripes_lock); 528f6bed0efSShaohua Li while (!list_empty(&log->no_space_stripes)) { 529f6bed0efSShaohua Li sh = list_first_entry(&log->no_space_stripes, 530f6bed0efSShaohua Li struct stripe_head, log_list); 531f6bed0efSShaohua Li list_del_init(&sh->log_list); 532f6bed0efSShaohua Li set_bit(STRIPE_HANDLE, &sh->state); 533f6bed0efSShaohua Li raid5_release_stripe(sh); 534f6bed0efSShaohua Li } 535f6bed0efSShaohua Li spin_unlock(&log->no_space_stripes_lock); 536f6bed0efSShaohua Li } 537f6bed0efSShaohua Li 53817036461SChristoph Hellwig static sector_t r5l_reclaimable_space(struct r5l_log *log) 53917036461SChristoph Hellwig { 54017036461SChristoph Hellwig return r5l_ring_distance(log, log->last_checkpoint, 54117036461SChristoph Hellwig log->next_checkpoint); 54217036461SChristoph Hellwig } 54317036461SChristoph Hellwig 54404732f74SChristoph Hellwig static bool r5l_complete_finished_ios(struct r5l_log *log) 54517036461SChristoph Hellwig { 54617036461SChristoph Hellwig struct r5l_io_unit *io, *next; 54717036461SChristoph Hellwig bool found = false; 54817036461SChristoph Hellwig 54917036461SChristoph Hellwig assert_spin_locked(&log->io_list_lock); 55017036461SChristoph Hellwig 55104732f74SChristoph Hellwig list_for_each_entry_safe(io, next, &log->finished_ios, log_sibling) { 55217036461SChristoph Hellwig /* don't change list order */ 55317036461SChristoph Hellwig if (io->state < IO_UNIT_STRIPE_END) 55417036461SChristoph Hellwig break; 55517036461SChristoph Hellwig 55617036461SChristoph Hellwig log->next_checkpoint = io->log_start; 55717036461SChristoph Hellwig log->next_cp_seq = io->seq; 55817036461SChristoph Hellwig 55917036461SChristoph Hellwig list_del(&io->log_sibling); 56017036461SChristoph Hellwig r5l_free_io_unit(log, io); 56117036461SChristoph Hellwig 56217036461SChristoph Hellwig found = true; 56317036461SChristoph Hellwig } 56417036461SChristoph Hellwig 56517036461SChristoph Hellwig return found; 56617036461SChristoph Hellwig } 56717036461SChristoph Hellwig 568509ffec7SChristoph Hellwig static void __r5l_stripe_write_finished(struct r5l_io_unit *io) 569509ffec7SChristoph Hellwig { 570509ffec7SChristoph Hellwig struct r5l_log *log = io->log; 571509ffec7SChristoph Hellwig unsigned long flags; 572509ffec7SChristoph Hellwig 573509ffec7SChristoph Hellwig spin_lock_irqsave(&log->io_list_lock, flags); 574509ffec7SChristoph Hellwig __r5l_set_io_unit_state(io, IO_UNIT_STRIPE_END); 57517036461SChristoph Hellwig 57604732f74SChristoph Hellwig if (!r5l_complete_finished_ios(log)) { 57785f2f9a4SShaohua Li spin_unlock_irqrestore(&log->io_list_lock, flags); 57885f2f9a4SShaohua Li return; 57985f2f9a4SShaohua Li } 580509ffec7SChristoph Hellwig 58117036461SChristoph Hellwig if (r5l_reclaimable_space(log) > log->max_free_space) 582509ffec7SChristoph Hellwig r5l_wake_reclaim(log, 0); 583509ffec7SChristoph Hellwig 584509ffec7SChristoph Hellwig spin_unlock_irqrestore(&log->io_list_lock, flags); 585509ffec7SChristoph Hellwig wake_up(&log->iounit_wait); 586509ffec7SChristoph Hellwig } 587509ffec7SChristoph Hellwig 5880576b1c6SShaohua Li void r5l_stripe_write_finished(struct stripe_head *sh) 5890576b1c6SShaohua Li { 5900576b1c6SShaohua Li struct r5l_io_unit *io; 5910576b1c6SShaohua Li 5920576b1c6SShaohua Li io = sh->log_io; 5930576b1c6SShaohua Li sh->log_io = NULL; 5940576b1c6SShaohua Li 595509ffec7SChristoph Hellwig if (io && atomic_dec_and_test(&io->pending_stripe)) 596509ffec7SChristoph Hellwig __r5l_stripe_write_finished(io); 5970576b1c6SShaohua Li } 5980576b1c6SShaohua Li 599a8c34f91SShaohua Li static void r5l_log_flush_endio(struct bio *bio) 600a8c34f91SShaohua Li { 601a8c34f91SShaohua Li struct r5l_log *log = container_of(bio, struct r5l_log, 602a8c34f91SShaohua Li flush_bio); 603a8c34f91SShaohua Li unsigned long flags; 604a8c34f91SShaohua Li struct r5l_io_unit *io; 605a8c34f91SShaohua Li 606a8c34f91SShaohua Li spin_lock_irqsave(&log->io_list_lock, flags); 607d8858f43SChristoph Hellwig list_for_each_entry(io, &log->flushing_ios, log_sibling) 608d8858f43SChristoph Hellwig r5l_io_run_stripes(io); 60904732f74SChristoph Hellwig list_splice_tail_init(&log->flushing_ios, &log->finished_ios); 610a8c34f91SShaohua Li spin_unlock_irqrestore(&log->io_list_lock, flags); 611a8c34f91SShaohua Li } 612a8c34f91SShaohua Li 6130576b1c6SShaohua Li /* 6140576b1c6SShaohua Li * Starting dispatch IO to raid. 6150576b1c6SShaohua Li * io_unit(meta) consists of a log. There is one situation we want to avoid. A 6160576b1c6SShaohua Li * broken meta in the middle of a log causes recovery can't find meta at the 6170576b1c6SShaohua Li * head of log. If operations require meta at the head persistent in log, we 6180576b1c6SShaohua Li * must make sure meta before it persistent in log too. A case is: 6190576b1c6SShaohua Li * 6200576b1c6SShaohua Li * stripe data/parity is in log, we start write stripe to raid disks. stripe 6210576b1c6SShaohua Li * data/parity must be persistent in log before we do the write to raid disks. 6220576b1c6SShaohua Li * 6230576b1c6SShaohua Li * The solution is we restrictly maintain io_unit list order. In this case, we 6240576b1c6SShaohua Li * only write stripes of an io_unit to raid disks till the io_unit is the first 6250576b1c6SShaohua Li * one whose data/parity is in log. 6260576b1c6SShaohua Li */ 6270576b1c6SShaohua Li void r5l_flush_stripe_to_raid(struct r5l_log *log) 6280576b1c6SShaohua Li { 629a8c34f91SShaohua Li bool do_flush; 63056fef7c6SChristoph Hellwig 63156fef7c6SChristoph Hellwig if (!log || !log->need_cache_flush) 6320576b1c6SShaohua Li return; 6330576b1c6SShaohua Li 634a8c34f91SShaohua Li spin_lock_irq(&log->io_list_lock); 635a8c34f91SShaohua Li /* flush bio is running */ 636a8c34f91SShaohua Li if (!list_empty(&log->flushing_ios)) { 637a8c34f91SShaohua Li spin_unlock_irq(&log->io_list_lock); 6380576b1c6SShaohua Li return; 6390576b1c6SShaohua Li } 640a8c34f91SShaohua Li list_splice_tail_init(&log->io_end_ios, &log->flushing_ios); 641a8c34f91SShaohua Li do_flush = !list_empty(&log->flushing_ios); 6420576b1c6SShaohua Li spin_unlock_irq(&log->io_list_lock); 643a8c34f91SShaohua Li 644a8c34f91SShaohua Li if (!do_flush) 645a8c34f91SShaohua Li return; 646a8c34f91SShaohua Li bio_reset(&log->flush_bio); 647a8c34f91SShaohua Li log->flush_bio.bi_bdev = log->rdev->bdev; 648a8c34f91SShaohua Li log->flush_bio.bi_end_io = r5l_log_flush_endio; 649a8c34f91SShaohua Li submit_bio(WRITE_FLUSH, &log->flush_bio); 6500576b1c6SShaohua Li } 6510576b1c6SShaohua Li 6520576b1c6SShaohua Li static void r5l_write_super(struct r5l_log *log, sector_t cp); 6530576b1c6SShaohua Li static void r5l_do_reclaim(struct r5l_log *log) 6540576b1c6SShaohua Li { 6550576b1c6SShaohua Li sector_t reclaim_target = xchg(&log->reclaim_target, 0); 65617036461SChristoph Hellwig sector_t reclaimable; 65717036461SChristoph Hellwig sector_t next_checkpoint; 65817036461SChristoph Hellwig u64 next_cp_seq; 6590576b1c6SShaohua Li 6600576b1c6SShaohua Li spin_lock_irq(&log->io_list_lock); 6610576b1c6SShaohua Li /* 6620576b1c6SShaohua Li * move proper io_unit to reclaim list. We should not change the order. 6630576b1c6SShaohua Li * reclaimable/unreclaimable io_unit can be mixed in the list, we 6640576b1c6SShaohua Li * shouldn't reuse space of an unreclaimable io_unit 6650576b1c6SShaohua Li */ 6660576b1c6SShaohua Li while (1) { 66717036461SChristoph Hellwig reclaimable = r5l_reclaimable_space(log); 66817036461SChristoph Hellwig if (reclaimable >= reclaim_target || 6690576b1c6SShaohua Li (list_empty(&log->running_ios) && 6700576b1c6SShaohua Li list_empty(&log->io_end_ios) && 671a8c34f91SShaohua Li list_empty(&log->flushing_ios) && 67204732f74SChristoph Hellwig list_empty(&log->finished_ios))) 6730576b1c6SShaohua Li break; 6740576b1c6SShaohua Li 67517036461SChristoph Hellwig md_wakeup_thread(log->rdev->mddev->thread); 67617036461SChristoph Hellwig wait_event_lock_irq(log->iounit_wait, 67717036461SChristoph Hellwig r5l_reclaimable_space(log) > reclaimable, 67817036461SChristoph Hellwig log->io_list_lock); 6790576b1c6SShaohua Li } 68017036461SChristoph Hellwig 68117036461SChristoph Hellwig next_checkpoint = log->next_checkpoint; 68217036461SChristoph Hellwig next_cp_seq = log->next_cp_seq; 6830576b1c6SShaohua Li spin_unlock_irq(&log->io_list_lock); 6840576b1c6SShaohua Li 68517036461SChristoph Hellwig BUG_ON(reclaimable < 0); 68617036461SChristoph Hellwig if (reclaimable == 0) 6870576b1c6SShaohua Li return; 6880576b1c6SShaohua Li 6890576b1c6SShaohua Li /* 6900576b1c6SShaohua Li * write_super will flush cache of each raid disk. We must write super 6910576b1c6SShaohua Li * here, because the log area might be reused soon and we don't want to 6920576b1c6SShaohua Li * confuse recovery 6930576b1c6SShaohua Li */ 69417036461SChristoph Hellwig r5l_write_super(log, next_checkpoint); 6950576b1c6SShaohua Li 6960576b1c6SShaohua Li mutex_lock(&log->io_mutex); 69717036461SChristoph Hellwig log->last_checkpoint = next_checkpoint; 69817036461SChristoph Hellwig log->last_cp_seq = next_cp_seq; 6990576b1c6SShaohua Li mutex_unlock(&log->io_mutex); 7000576b1c6SShaohua Li 70117036461SChristoph Hellwig r5l_run_no_space_stripes(log); 7020576b1c6SShaohua Li } 7030576b1c6SShaohua Li 7040576b1c6SShaohua Li static void r5l_reclaim_thread(struct md_thread *thread) 7050576b1c6SShaohua Li { 7060576b1c6SShaohua Li struct mddev *mddev = thread->mddev; 7070576b1c6SShaohua Li struct r5conf *conf = mddev->private; 7080576b1c6SShaohua Li struct r5l_log *log = conf->log; 7090576b1c6SShaohua Li 7100576b1c6SShaohua Li if (!log) 7110576b1c6SShaohua Li return; 7120576b1c6SShaohua Li r5l_do_reclaim(log); 7130576b1c6SShaohua Li } 7140576b1c6SShaohua Li 715f6bed0efSShaohua Li static void r5l_wake_reclaim(struct r5l_log *log, sector_t space) 716f6bed0efSShaohua Li { 7170576b1c6SShaohua Li unsigned long target; 7180576b1c6SShaohua Li unsigned long new = (unsigned long)space; /* overflow in theory */ 7190576b1c6SShaohua Li 7200576b1c6SShaohua Li do { 7210576b1c6SShaohua Li target = log->reclaim_target; 7220576b1c6SShaohua Li if (new < target) 7230576b1c6SShaohua Li return; 7240576b1c6SShaohua Li } while (cmpxchg(&log->reclaim_target, target, new) != target); 7250576b1c6SShaohua Li md_wakeup_thread(log->reclaim_thread); 726f6bed0efSShaohua Li } 727f6bed0efSShaohua Li 728e6c033f7SShaohua Li void r5l_quiesce(struct r5l_log *log, int state) 729e6c033f7SShaohua Li { 730e6c033f7SShaohua Li if (!log || state == 2) 731e6c033f7SShaohua Li return; 732e6c033f7SShaohua Li if (state == 0) { 733e6c033f7SShaohua Li log->reclaim_thread = md_register_thread(r5l_reclaim_thread, 734e6c033f7SShaohua Li log->rdev->mddev, "reclaim"); 735e6c033f7SShaohua Li } else if (state == 1) { 736e6c033f7SShaohua Li /* 737e6c033f7SShaohua Li * at this point all stripes are finished, so io_unit is at 738e6c033f7SShaohua Li * least in STRIPE_END state 739e6c033f7SShaohua Li */ 740e6c033f7SShaohua Li r5l_wake_reclaim(log, -1L); 741e6c033f7SShaohua Li md_unregister_thread(&log->reclaim_thread); 742e6c033f7SShaohua Li r5l_do_reclaim(log); 743e6c033f7SShaohua Li } 744e6c033f7SShaohua Li } 745e6c033f7SShaohua Li 746355810d1SShaohua Li struct r5l_recovery_ctx { 747355810d1SShaohua Li struct page *meta_page; /* current meta */ 748355810d1SShaohua Li sector_t meta_total_blocks; /* total size of current meta and data */ 749355810d1SShaohua Li sector_t pos; /* recovery position */ 750355810d1SShaohua Li u64 seq; /* recovery position seq */ 751355810d1SShaohua Li }; 752355810d1SShaohua Li 753355810d1SShaohua Li static int r5l_read_meta_block(struct r5l_log *log, 754355810d1SShaohua Li struct r5l_recovery_ctx *ctx) 755355810d1SShaohua Li { 756355810d1SShaohua Li struct page *page = ctx->meta_page; 757355810d1SShaohua Li struct r5l_meta_block *mb; 758355810d1SShaohua Li u32 crc, stored_crc; 759355810d1SShaohua Li 760355810d1SShaohua Li if (!sync_page_io(log->rdev, ctx->pos, PAGE_SIZE, page, READ, false)) 761355810d1SShaohua Li return -EIO; 762355810d1SShaohua Li 763355810d1SShaohua Li mb = page_address(page); 764355810d1SShaohua Li stored_crc = le32_to_cpu(mb->checksum); 765355810d1SShaohua Li mb->checksum = 0; 766355810d1SShaohua Li 767355810d1SShaohua Li if (le32_to_cpu(mb->magic) != R5LOG_MAGIC || 768355810d1SShaohua Li le64_to_cpu(mb->seq) != ctx->seq || 769355810d1SShaohua Li mb->version != R5LOG_VERSION || 770355810d1SShaohua Li le64_to_cpu(mb->position) != ctx->pos) 771355810d1SShaohua Li return -EINVAL; 772355810d1SShaohua Li 7735cb2fbd6SShaohua Li crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE); 774355810d1SShaohua Li if (stored_crc != crc) 775355810d1SShaohua Li return -EINVAL; 776355810d1SShaohua Li 777355810d1SShaohua Li if (le32_to_cpu(mb->meta_size) > PAGE_SIZE) 778355810d1SShaohua Li return -EINVAL; 779355810d1SShaohua Li 780355810d1SShaohua Li ctx->meta_total_blocks = BLOCK_SECTORS; 781355810d1SShaohua Li 782355810d1SShaohua Li return 0; 783355810d1SShaohua Li } 784355810d1SShaohua Li 785355810d1SShaohua Li static int r5l_recovery_flush_one_stripe(struct r5l_log *log, 786355810d1SShaohua Li struct r5l_recovery_ctx *ctx, 787355810d1SShaohua Li sector_t stripe_sect, 788355810d1SShaohua Li int *offset, sector_t *log_offset) 789355810d1SShaohua Li { 790355810d1SShaohua Li struct r5conf *conf = log->rdev->mddev->private; 791355810d1SShaohua Li struct stripe_head *sh; 792355810d1SShaohua Li struct r5l_payload_data_parity *payload; 793355810d1SShaohua Li int disk_index; 794355810d1SShaohua Li 795355810d1SShaohua Li sh = raid5_get_active_stripe(conf, stripe_sect, 0, 0, 0); 796355810d1SShaohua Li while (1) { 797355810d1SShaohua Li payload = page_address(ctx->meta_page) + *offset; 798355810d1SShaohua Li 799355810d1SShaohua Li if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) { 800355810d1SShaohua Li raid5_compute_sector(conf, 801355810d1SShaohua Li le64_to_cpu(payload->location), 0, 802355810d1SShaohua Li &disk_index, sh); 803355810d1SShaohua Li 804355810d1SShaohua Li sync_page_io(log->rdev, *log_offset, PAGE_SIZE, 805355810d1SShaohua Li sh->dev[disk_index].page, READ, false); 806355810d1SShaohua Li sh->dev[disk_index].log_checksum = 807355810d1SShaohua Li le32_to_cpu(payload->checksum[0]); 808355810d1SShaohua Li set_bit(R5_Wantwrite, &sh->dev[disk_index].flags); 809355810d1SShaohua Li ctx->meta_total_blocks += BLOCK_SECTORS; 810355810d1SShaohua Li } else { 811355810d1SShaohua Li disk_index = sh->pd_idx; 812355810d1SShaohua Li sync_page_io(log->rdev, *log_offset, PAGE_SIZE, 813355810d1SShaohua Li sh->dev[disk_index].page, READ, false); 814355810d1SShaohua Li sh->dev[disk_index].log_checksum = 815355810d1SShaohua Li le32_to_cpu(payload->checksum[0]); 816355810d1SShaohua Li set_bit(R5_Wantwrite, &sh->dev[disk_index].flags); 817355810d1SShaohua Li 818355810d1SShaohua Li if (sh->qd_idx >= 0) { 819355810d1SShaohua Li disk_index = sh->qd_idx; 820355810d1SShaohua Li sync_page_io(log->rdev, 821355810d1SShaohua Li r5l_ring_add(log, *log_offset, BLOCK_SECTORS), 822355810d1SShaohua Li PAGE_SIZE, sh->dev[disk_index].page, 823355810d1SShaohua Li READ, false); 824355810d1SShaohua Li sh->dev[disk_index].log_checksum = 825355810d1SShaohua Li le32_to_cpu(payload->checksum[1]); 826355810d1SShaohua Li set_bit(R5_Wantwrite, 827355810d1SShaohua Li &sh->dev[disk_index].flags); 828355810d1SShaohua Li } 829355810d1SShaohua Li ctx->meta_total_blocks += BLOCK_SECTORS * conf->max_degraded; 830355810d1SShaohua Li } 831355810d1SShaohua Li 832355810d1SShaohua Li *log_offset = r5l_ring_add(log, *log_offset, 833355810d1SShaohua Li le32_to_cpu(payload->size)); 834355810d1SShaohua Li *offset += sizeof(struct r5l_payload_data_parity) + 835355810d1SShaohua Li sizeof(__le32) * 836355810d1SShaohua Li (le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9)); 837355810d1SShaohua Li if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_PARITY) 838355810d1SShaohua Li break; 839355810d1SShaohua Li } 840355810d1SShaohua Li 841355810d1SShaohua Li for (disk_index = 0; disk_index < sh->disks; disk_index++) { 842355810d1SShaohua Li void *addr; 843355810d1SShaohua Li u32 checksum; 844355810d1SShaohua Li 845355810d1SShaohua Li if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags)) 846355810d1SShaohua Li continue; 847355810d1SShaohua Li addr = kmap_atomic(sh->dev[disk_index].page); 8485cb2fbd6SShaohua Li checksum = crc32c_le(log->uuid_checksum, addr, PAGE_SIZE); 849355810d1SShaohua Li kunmap_atomic(addr); 850355810d1SShaohua Li if (checksum != sh->dev[disk_index].log_checksum) 851355810d1SShaohua Li goto error; 852355810d1SShaohua Li } 853355810d1SShaohua Li 854355810d1SShaohua Li for (disk_index = 0; disk_index < sh->disks; disk_index++) { 855355810d1SShaohua Li struct md_rdev *rdev, *rrdev; 856355810d1SShaohua Li 857355810d1SShaohua Li if (!test_and_clear_bit(R5_Wantwrite, 858355810d1SShaohua Li &sh->dev[disk_index].flags)) 859355810d1SShaohua Li continue; 860355810d1SShaohua Li 861355810d1SShaohua Li /* in case device is broken */ 862355810d1SShaohua Li rdev = rcu_dereference(conf->disks[disk_index].rdev); 863355810d1SShaohua Li if (rdev) 864355810d1SShaohua Li sync_page_io(rdev, stripe_sect, PAGE_SIZE, 865355810d1SShaohua Li sh->dev[disk_index].page, WRITE, false); 866355810d1SShaohua Li rrdev = rcu_dereference(conf->disks[disk_index].replacement); 867355810d1SShaohua Li if (rrdev) 868355810d1SShaohua Li sync_page_io(rrdev, stripe_sect, PAGE_SIZE, 869355810d1SShaohua Li sh->dev[disk_index].page, WRITE, false); 870355810d1SShaohua Li } 871355810d1SShaohua Li raid5_release_stripe(sh); 872355810d1SShaohua Li return 0; 873355810d1SShaohua Li 874355810d1SShaohua Li error: 875355810d1SShaohua Li for (disk_index = 0; disk_index < sh->disks; disk_index++) 876355810d1SShaohua Li sh->dev[disk_index].flags = 0; 877355810d1SShaohua Li raid5_release_stripe(sh); 878355810d1SShaohua Li return -EINVAL; 879355810d1SShaohua Li } 880355810d1SShaohua Li 881355810d1SShaohua Li static int r5l_recovery_flush_one_meta(struct r5l_log *log, 882355810d1SShaohua Li struct r5l_recovery_ctx *ctx) 883355810d1SShaohua Li { 884355810d1SShaohua Li struct r5conf *conf = log->rdev->mddev->private; 885355810d1SShaohua Li struct r5l_payload_data_parity *payload; 886355810d1SShaohua Li struct r5l_meta_block *mb; 887355810d1SShaohua Li int offset; 888355810d1SShaohua Li sector_t log_offset; 889355810d1SShaohua Li sector_t stripe_sector; 890355810d1SShaohua Li 891355810d1SShaohua Li mb = page_address(ctx->meta_page); 892355810d1SShaohua Li offset = sizeof(struct r5l_meta_block); 893355810d1SShaohua Li log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS); 894355810d1SShaohua Li 895355810d1SShaohua Li while (offset < le32_to_cpu(mb->meta_size)) { 896355810d1SShaohua Li int dd; 897355810d1SShaohua Li 898355810d1SShaohua Li payload = (void *)mb + offset; 899355810d1SShaohua Li stripe_sector = raid5_compute_sector(conf, 900355810d1SShaohua Li le64_to_cpu(payload->location), 0, &dd, NULL); 901355810d1SShaohua Li if (r5l_recovery_flush_one_stripe(log, ctx, stripe_sector, 902355810d1SShaohua Li &offset, &log_offset)) 903355810d1SShaohua Li return -EINVAL; 904355810d1SShaohua Li } 905355810d1SShaohua Li return 0; 906355810d1SShaohua Li } 907355810d1SShaohua Li 908355810d1SShaohua Li /* copy data/parity from log to raid disks */ 909355810d1SShaohua Li static void r5l_recovery_flush_log(struct r5l_log *log, 910355810d1SShaohua Li struct r5l_recovery_ctx *ctx) 911355810d1SShaohua Li { 912355810d1SShaohua Li while (1) { 913355810d1SShaohua Li if (r5l_read_meta_block(log, ctx)) 914355810d1SShaohua Li return; 915355810d1SShaohua Li if (r5l_recovery_flush_one_meta(log, ctx)) 916355810d1SShaohua Li return; 917355810d1SShaohua Li ctx->seq++; 918355810d1SShaohua Li ctx->pos = r5l_ring_add(log, ctx->pos, ctx->meta_total_blocks); 919355810d1SShaohua Li } 920355810d1SShaohua Li } 921355810d1SShaohua Li 922355810d1SShaohua Li static int r5l_log_write_empty_meta_block(struct r5l_log *log, sector_t pos, 923355810d1SShaohua Li u64 seq) 924355810d1SShaohua Li { 925355810d1SShaohua Li struct page *page; 926355810d1SShaohua Li struct r5l_meta_block *mb; 927355810d1SShaohua Li u32 crc; 928355810d1SShaohua Li 929355810d1SShaohua Li page = alloc_page(GFP_KERNEL | __GFP_ZERO); 930355810d1SShaohua Li if (!page) 931355810d1SShaohua Li return -ENOMEM; 932355810d1SShaohua Li mb = page_address(page); 933355810d1SShaohua Li mb->magic = cpu_to_le32(R5LOG_MAGIC); 934355810d1SShaohua Li mb->version = R5LOG_VERSION; 935355810d1SShaohua Li mb->meta_size = cpu_to_le32(sizeof(struct r5l_meta_block)); 936355810d1SShaohua Li mb->seq = cpu_to_le64(seq); 937355810d1SShaohua Li mb->position = cpu_to_le64(pos); 9385cb2fbd6SShaohua Li crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE); 939355810d1SShaohua Li mb->checksum = cpu_to_le32(crc); 940355810d1SShaohua Li 941355810d1SShaohua Li if (!sync_page_io(log->rdev, pos, PAGE_SIZE, page, WRITE_FUA, false)) { 942355810d1SShaohua Li __free_page(page); 943355810d1SShaohua Li return -EIO; 944355810d1SShaohua Li } 945355810d1SShaohua Li __free_page(page); 946355810d1SShaohua Li return 0; 947355810d1SShaohua Li } 948355810d1SShaohua Li 949f6bed0efSShaohua Li static int r5l_recovery_log(struct r5l_log *log) 950f6bed0efSShaohua Li { 951355810d1SShaohua Li struct r5l_recovery_ctx ctx; 952355810d1SShaohua Li 953355810d1SShaohua Li ctx.pos = log->last_checkpoint; 954355810d1SShaohua Li ctx.seq = log->last_cp_seq; 955355810d1SShaohua Li ctx.meta_page = alloc_page(GFP_KERNEL); 956355810d1SShaohua Li if (!ctx.meta_page) 957355810d1SShaohua Li return -ENOMEM; 958355810d1SShaohua Li 959355810d1SShaohua Li r5l_recovery_flush_log(log, &ctx); 960355810d1SShaohua Li __free_page(ctx.meta_page); 961355810d1SShaohua Li 962355810d1SShaohua Li /* 963355810d1SShaohua Li * we did a recovery. Now ctx.pos points to an invalid meta block. New 964355810d1SShaohua Li * log will start here. but we can't let superblock point to last valid 965355810d1SShaohua Li * meta block. The log might looks like: 966355810d1SShaohua Li * | meta 1| meta 2| meta 3| 967355810d1SShaohua Li * meta 1 is valid, meta 2 is invalid. meta 3 could be valid. If 968355810d1SShaohua Li * superblock points to meta 1, we write a new valid meta 2n. if crash 969355810d1SShaohua Li * happens again, new recovery will start from meta 1. Since meta 2n is 970355810d1SShaohua Li * valid now, recovery will think meta 3 is valid, which is wrong. 971355810d1SShaohua Li * The solution is we create a new meta in meta2 with its seq == meta 972355810d1SShaohua Li * 1's seq + 10 and let superblock points to meta2. The same recovery will 973355810d1SShaohua Li * not think meta 3 is a valid meta, because its seq doesn't match 974355810d1SShaohua Li */ 975355810d1SShaohua Li if (ctx.seq > log->last_cp_seq + 1) { 976355810d1SShaohua Li int ret; 977355810d1SShaohua Li 978355810d1SShaohua Li ret = r5l_log_write_empty_meta_block(log, ctx.pos, ctx.seq + 10); 979355810d1SShaohua Li if (ret) 980355810d1SShaohua Li return ret; 981355810d1SShaohua Li log->seq = ctx.seq + 11; 982355810d1SShaohua Li log->log_start = r5l_ring_add(log, ctx.pos, BLOCK_SECTORS); 983355810d1SShaohua Li r5l_write_super(log, ctx.pos); 984355810d1SShaohua Li } else { 985355810d1SShaohua Li log->log_start = ctx.pos; 986355810d1SShaohua Li log->seq = ctx.seq; 987355810d1SShaohua Li } 988f6bed0efSShaohua Li return 0; 989f6bed0efSShaohua Li } 990f6bed0efSShaohua Li 991f6bed0efSShaohua Li static void r5l_write_super(struct r5l_log *log, sector_t cp) 992f6bed0efSShaohua Li { 993f6bed0efSShaohua Li struct mddev *mddev = log->rdev->mddev; 994f6bed0efSShaohua Li 995f6bed0efSShaohua Li log->rdev->journal_tail = cp; 996f6bed0efSShaohua Li set_bit(MD_CHANGE_DEVS, &mddev->flags); 997f6bed0efSShaohua Li } 998f6bed0efSShaohua Li 999f6bed0efSShaohua Li static int r5l_load_log(struct r5l_log *log) 1000f6bed0efSShaohua Li { 1001f6bed0efSShaohua Li struct md_rdev *rdev = log->rdev; 1002f6bed0efSShaohua Li struct page *page; 1003f6bed0efSShaohua Li struct r5l_meta_block *mb; 1004f6bed0efSShaohua Li sector_t cp = log->rdev->journal_tail; 1005f6bed0efSShaohua Li u32 stored_crc, expected_crc; 1006f6bed0efSShaohua Li bool create_super = false; 1007f6bed0efSShaohua Li int ret; 1008f6bed0efSShaohua Li 1009f6bed0efSShaohua Li /* Make sure it's valid */ 1010f6bed0efSShaohua Li if (cp >= rdev->sectors || round_down(cp, BLOCK_SECTORS) != cp) 1011f6bed0efSShaohua Li cp = 0; 1012f6bed0efSShaohua Li page = alloc_page(GFP_KERNEL); 1013f6bed0efSShaohua Li if (!page) 1014f6bed0efSShaohua Li return -ENOMEM; 1015f6bed0efSShaohua Li 1016f6bed0efSShaohua Li if (!sync_page_io(rdev, cp, PAGE_SIZE, page, READ, false)) { 1017f6bed0efSShaohua Li ret = -EIO; 1018f6bed0efSShaohua Li goto ioerr; 1019f6bed0efSShaohua Li } 1020f6bed0efSShaohua Li mb = page_address(page); 1021f6bed0efSShaohua Li 1022f6bed0efSShaohua Li if (le32_to_cpu(mb->magic) != R5LOG_MAGIC || 1023f6bed0efSShaohua Li mb->version != R5LOG_VERSION) { 1024f6bed0efSShaohua Li create_super = true; 1025f6bed0efSShaohua Li goto create; 1026f6bed0efSShaohua Li } 1027f6bed0efSShaohua Li stored_crc = le32_to_cpu(mb->checksum); 1028f6bed0efSShaohua Li mb->checksum = 0; 10295cb2fbd6SShaohua Li expected_crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE); 1030f6bed0efSShaohua Li if (stored_crc != expected_crc) { 1031f6bed0efSShaohua Li create_super = true; 1032f6bed0efSShaohua Li goto create; 1033f6bed0efSShaohua Li } 1034f6bed0efSShaohua Li if (le64_to_cpu(mb->position) != cp) { 1035f6bed0efSShaohua Li create_super = true; 1036f6bed0efSShaohua Li goto create; 1037f6bed0efSShaohua Li } 1038f6bed0efSShaohua Li create: 1039f6bed0efSShaohua Li if (create_super) { 1040f6bed0efSShaohua Li log->last_cp_seq = prandom_u32(); 1041f6bed0efSShaohua Li cp = 0; 1042f6bed0efSShaohua Li /* 1043f6bed0efSShaohua Li * Make sure super points to correct address. Log might have 1044f6bed0efSShaohua Li * data very soon. If super hasn't correct log tail address, 1045f6bed0efSShaohua Li * recovery can't find the log 1046f6bed0efSShaohua Li */ 1047f6bed0efSShaohua Li r5l_write_super(log, cp); 1048f6bed0efSShaohua Li } else 1049f6bed0efSShaohua Li log->last_cp_seq = le64_to_cpu(mb->seq); 1050f6bed0efSShaohua Li 1051f6bed0efSShaohua Li log->device_size = round_down(rdev->sectors, BLOCK_SECTORS); 10520576b1c6SShaohua Li log->max_free_space = log->device_size >> RECLAIM_MAX_FREE_SPACE_SHIFT; 10530576b1c6SShaohua Li if (log->max_free_space > RECLAIM_MAX_FREE_SPACE) 10540576b1c6SShaohua Li log->max_free_space = RECLAIM_MAX_FREE_SPACE; 1055f6bed0efSShaohua Li log->last_checkpoint = cp; 1056f6bed0efSShaohua Li 1057f6bed0efSShaohua Li __free_page(page); 1058f6bed0efSShaohua Li 1059f6bed0efSShaohua Li return r5l_recovery_log(log); 1060f6bed0efSShaohua Li ioerr: 1061f6bed0efSShaohua Li __free_page(page); 1062f6bed0efSShaohua Li return ret; 1063f6bed0efSShaohua Li } 1064f6bed0efSShaohua Li 1065f6bed0efSShaohua Li int r5l_init_log(struct r5conf *conf, struct md_rdev *rdev) 1066f6bed0efSShaohua Li { 1067f6bed0efSShaohua Li struct r5l_log *log; 1068f6bed0efSShaohua Li 1069f6bed0efSShaohua Li if (PAGE_SIZE != 4096) 1070f6bed0efSShaohua Li return -EINVAL; 1071f6bed0efSShaohua Li log = kzalloc(sizeof(*log), GFP_KERNEL); 1072f6bed0efSShaohua Li if (!log) 1073f6bed0efSShaohua Li return -ENOMEM; 1074f6bed0efSShaohua Li log->rdev = rdev; 1075f6bed0efSShaohua Li 107656fef7c6SChristoph Hellwig log->need_cache_flush = (rdev->bdev->bd_disk->queue->flush_flags != 0); 107756fef7c6SChristoph Hellwig 10785cb2fbd6SShaohua Li log->uuid_checksum = crc32c_le(~0, rdev->mddev->uuid, 1079f6bed0efSShaohua Li sizeof(rdev->mddev->uuid)); 1080f6bed0efSShaohua Li 1081f6bed0efSShaohua Li mutex_init(&log->io_mutex); 1082f6bed0efSShaohua Li 1083f6bed0efSShaohua Li spin_lock_init(&log->io_list_lock); 1084f6bed0efSShaohua Li INIT_LIST_HEAD(&log->running_ios); 10850576b1c6SShaohua Li INIT_LIST_HEAD(&log->io_end_ios); 1086a8c34f91SShaohua Li INIT_LIST_HEAD(&log->flushing_ios); 108704732f74SChristoph Hellwig INIT_LIST_HEAD(&log->finished_ios); 1088a8c34f91SShaohua Li bio_init(&log->flush_bio); 1089f6bed0efSShaohua Li 1090f6bed0efSShaohua Li log->io_kc = KMEM_CACHE(r5l_io_unit, 0); 1091f6bed0efSShaohua Li if (!log->io_kc) 1092f6bed0efSShaohua Li goto io_kc; 1093f6bed0efSShaohua Li 10940576b1c6SShaohua Li log->reclaim_thread = md_register_thread(r5l_reclaim_thread, 10950576b1c6SShaohua Li log->rdev->mddev, "reclaim"); 10960576b1c6SShaohua Li if (!log->reclaim_thread) 10970576b1c6SShaohua Li goto reclaim_thread; 10980fd22b45SShaohua Li init_waitqueue_head(&log->iounit_wait); 10990576b1c6SShaohua Li 1100f6bed0efSShaohua Li INIT_LIST_HEAD(&log->no_space_stripes); 1101f6bed0efSShaohua Li spin_lock_init(&log->no_space_stripes_lock); 1102f6bed0efSShaohua Li 1103f6bed0efSShaohua Li if (r5l_load_log(log)) 1104f6bed0efSShaohua Li goto error; 1105f6bed0efSShaohua Li 1106f6bed0efSShaohua Li conf->log = log; 1107f6bed0efSShaohua Li return 0; 1108f6bed0efSShaohua Li error: 11090576b1c6SShaohua Li md_unregister_thread(&log->reclaim_thread); 11100576b1c6SShaohua Li reclaim_thread: 1111f6bed0efSShaohua Li kmem_cache_destroy(log->io_kc); 1112f6bed0efSShaohua Li io_kc: 1113f6bed0efSShaohua Li kfree(log); 1114f6bed0efSShaohua Li return -EINVAL; 1115f6bed0efSShaohua Li } 1116f6bed0efSShaohua Li 1117f6bed0efSShaohua Li void r5l_exit_log(struct r5l_log *log) 1118f6bed0efSShaohua Li { 11190576b1c6SShaohua Li md_unregister_thread(&log->reclaim_thread); 1120f6bed0efSShaohua Li kmem_cache_destroy(log->io_kc); 1121f6bed0efSShaohua Li kfree(log); 1122f6bed0efSShaohua Li } 1123