Linux Audio

Check our new training course

Loading...
v6.8
   1// SPDX-License-Identifier: GPL-2.0-only
   2/*
   3 * Copyright (C) 2015 Shaohua Li <shli@fb.com>
   4 * Copyright (C) 2016 Song Liu <songliubraving@fb.com>
 
 
 
 
 
 
 
 
 
 
   5 */
   6#include <linux/kernel.h>
   7#include <linux/wait.h>
   8#include <linux/blkdev.h>
   9#include <linux/slab.h>
  10#include <linux/raid/md_p.h>
  11#include <linux/crc32c.h>
  12#include <linux/random.h>
  13#include <linux/kthread.h>
  14#include <linux/types.h>
  15#include "md.h"
  16#include "raid5.h"
  17#include "md-bitmap.h"
  18#include "raid5-log.h"
  19
  20/*
  21 * metadata/data stored in disk with 4k size unit (a block) regardless
  22 * underneath hardware sector size. only works with PAGE_SIZE == 4096
  23 */
  24#define BLOCK_SECTORS (8)
  25#define BLOCK_SECTOR_SHIFT (3)
  26
  27/*
  28 * log->max_free_space is min(1/4 disk size, 10G reclaimable space).
  29 *
  30 * In write through mode, the reclaim runs every log->max_free_space.
  31 * This can prevent the recovery scans for too long
  32 */
  33#define RECLAIM_MAX_FREE_SPACE (10 * 1024 * 1024 * 2) /* sector */
  34#define RECLAIM_MAX_FREE_SPACE_SHIFT (2)
  35
  36/* wake up reclaim thread periodically */
  37#define R5C_RECLAIM_WAKEUP_INTERVAL (30 * HZ)
  38/* start flush with these full stripes */
  39#define R5C_FULL_STRIPE_FLUSH_BATCH(conf) (conf->max_nr_stripes / 4)
  40/* reclaim stripes in groups */
  41#define R5C_RECLAIM_STRIPE_GROUP (NR_STRIPE_HASH_LOCKS * 2)
  42
  43/*
  44 * We only need 2 bios per I/O unit to make progress, but ensure we
  45 * have a few more available to not get too tight.
  46 */
  47#define R5L_POOL_SIZE	4
  48
  49static char *r5c_journal_mode_str[] = {"write-through",
  50				       "write-back"};
  51/*
  52 * raid5 cache state machine
  53 *
  54 * With the RAID cache, each stripe works in two phases:
  55 *	- caching phase
  56 *	- writing-out phase
  57 *
  58 * These two phases are controlled by bit STRIPE_R5C_CACHING:
  59 *   if STRIPE_R5C_CACHING == 0, the stripe is in writing-out phase
  60 *   if STRIPE_R5C_CACHING == 1, the stripe is in caching phase
  61 *
  62 * When there is no journal, or the journal is in write-through mode,
  63 * the stripe is always in writing-out phase.
  64 *
  65 * For write-back journal, the stripe is sent to caching phase on write
  66 * (r5c_try_caching_write). r5c_make_stripe_write_out() kicks off
  67 * the write-out phase by clearing STRIPE_R5C_CACHING.
  68 *
  69 * Stripes in caching phase do not write the raid disks. Instead, all
  70 * writes are committed from the log device. Therefore, a stripe in
  71 * caching phase handles writes as:
  72 *	- write to log device
  73 *	- return IO
  74 *
  75 * Stripes in writing-out phase handle writes as:
  76 *	- calculate parity
  77 *	- write pending data and parity to journal
  78 *	- write data and parity to raid disks
  79 *	- return IO for pending writes
  80 */
  81
  82struct r5l_log {
  83	struct md_rdev *rdev;
  84
  85	u32 uuid_checksum;
  86
  87	sector_t device_size;		/* log device size, round to
  88					 * BLOCK_SECTORS */
  89	sector_t max_free_space;	/* reclaim run if free space is at
  90					 * this size */
  91
  92	sector_t last_checkpoint;	/* log tail. where recovery scan
  93					 * starts from */
  94	u64 last_cp_seq;		/* log tail sequence */
  95
  96	sector_t log_start;		/* log head. where new data appends */
  97	u64 seq;			/* log head sequence */
  98
  99	sector_t next_checkpoint;
 100
 101	struct mutex io_mutex;
 102	struct r5l_io_unit *current_io;	/* current io_unit accepting new data */
 103
 104	spinlock_t io_list_lock;
 105	struct list_head running_ios;	/* io_units which are still running,
 106					 * and have not yet been completely
 107					 * written to the log */
 108	struct list_head io_end_ios;	/* io_units which have been completely
 109					 * written to the log but not yet written
 110					 * to the RAID */
 111	struct list_head flushing_ios;	/* io_units which are waiting for log
 112					 * cache flush */
 113	struct list_head finished_ios;	/* io_units which settle down in log disk */
 114	struct bio flush_bio;
 115
 116	struct list_head no_mem_stripes;   /* pending stripes, -ENOMEM */
 117
 118	struct kmem_cache *io_kc;
 119	mempool_t io_pool;
 120	struct bio_set bs;
 121	mempool_t meta_pool;
 122
 123	struct md_thread __rcu *reclaim_thread;
 124	unsigned long reclaim_target;	/* number of space that need to be
 125					 * reclaimed.  if it's 0, reclaim spaces
 126					 * used by io_units which are in
 127					 * IO_UNIT_STRIPE_END state (eg, reclaim
 128					 * doesn't wait for specific io_unit
 129					 * switching to IO_UNIT_STRIPE_END
 130					 * state) */
 131	wait_queue_head_t iounit_wait;
 132
 133	struct list_head no_space_stripes; /* pending stripes, log has no space */
 134	spinlock_t no_space_stripes_lock;
 135
 136	bool need_cache_flush;
 137
 138	/* for r5c_cache */
 139	enum r5c_journal_mode r5c_journal_mode;
 140
 141	/* all stripes in r5cache, in the order of seq at sh->log_start */
 142	struct list_head stripe_in_journal_list;
 143
 144	spinlock_t stripe_in_journal_lock;
 145	atomic_t stripe_in_journal_count;
 146
 147	/* to submit async io_units, to fulfill ordering of flush */
 148	struct work_struct deferred_io_work;
 149	/* to disable write back during in degraded mode */
 150	struct work_struct disable_writeback_work;
 151
 152	/* to for chunk_aligned_read in writeback mode, details below */
 153	spinlock_t tree_lock;
 154	struct radix_tree_root big_stripe_tree;
 155};
 156
 157/*
 158 * Enable chunk_aligned_read() with write back cache.
 159 *
 160 * Each chunk may contain more than one stripe (for example, a 256kB
 161 * chunk contains 64 4kB-page, so this chunk contain 64 stripes). For
 162 * chunk_aligned_read, these stripes are grouped into one "big_stripe".
 163 * For each big_stripe, we count how many stripes of this big_stripe
 164 * are in the write back cache. These data are tracked in a radix tree
 165 * (big_stripe_tree). We use radix_tree item pointer as the counter.
 166 * r5c_tree_index() is used to calculate keys for the radix tree.
 167 *
 168 * chunk_aligned_read() calls r5c_big_stripe_cached() to look up
 169 * big_stripe of each chunk in the tree. If this big_stripe is in the
 170 * tree, chunk_aligned_read() aborts. This look up is protected by
 171 * rcu_read_lock().
 172 *
 173 * It is necessary to remember whether a stripe is counted in
 174 * big_stripe_tree. Instead of adding new flag, we reuses existing flags:
 175 * STRIPE_R5C_PARTIAL_STRIPE and STRIPE_R5C_FULL_STRIPE. If either of these
 176 * two flags are set, the stripe is counted in big_stripe_tree. This
 177 * requires moving set_bit(STRIPE_R5C_PARTIAL_STRIPE) to
 178 * r5c_try_caching_write(); and moving clear_bit of
 179 * STRIPE_R5C_PARTIAL_STRIPE and STRIPE_R5C_FULL_STRIPE to
 180 * r5c_finish_stripe_write_out().
 181 */
 182
 183/*
 184 * radix tree requests lowest 2 bits of data pointer to be 2b'00.
 185 * So it is necessary to left shift the counter by 2 bits before using it
 186 * as data pointer of the tree.
 187 */
 188#define R5C_RADIX_COUNT_SHIFT 2
 189
 190/*
 191 * calculate key for big_stripe_tree
 192 *
 193 * sect: align_bi->bi_iter.bi_sector or sh->sector
 194 */
 195static inline sector_t r5c_tree_index(struct r5conf *conf,
 196				      sector_t sect)
 197{
 198	sector_div(sect, conf->chunk_sectors);
 
 
 199	return sect;
 200}
 201
 202/*
 203 * an IO range starts from a meta data block and end at the next meta data
 204 * block. The io unit's the meta data block tracks data/parity followed it. io
 205 * unit is written to log disk with normal write, as we always flush log disk
 206 * first and then start move data to raid disks, there is no requirement to
 207 * write io unit with FLUSH/FUA
 208 */
 209struct r5l_io_unit {
 210	struct r5l_log *log;
 211
 212	struct page *meta_page;	/* store meta block */
 213	int meta_offset;	/* current offset in meta_page */
 214
 215	struct bio *current_bio;/* current_bio accepting new data */
 216
 217	atomic_t pending_stripe;/* how many stripes not flushed to raid */
 218	u64 seq;		/* seq number of the metablock */
 219	sector_t log_start;	/* where the io_unit starts */
 220	sector_t log_end;	/* where the io_unit ends */
 221	struct list_head log_sibling; /* log->running_ios */
 222	struct list_head stripe_list; /* stripes added to the io_unit */
 223
 224	int state;
 225	bool need_split_bio;
 226	struct bio *split_bio;
 227
 228	unsigned int has_flush:1;		/* include flush request */
 229	unsigned int has_fua:1;			/* include fua request */
 230	unsigned int has_null_flush:1;		/* include null flush request */
 231	unsigned int has_flush_payload:1;	/* include flush payload  */
 232	/*
 233	 * io isn't sent yet, flush/fua request can only be submitted till it's
 234	 * the first IO in running_ios list
 235	 */
 236	unsigned int io_deferred:1;
 237
 238	struct bio_list flush_barriers;   /* size == 0 flush bios */
 239};
 240
 241/* r5l_io_unit state */
 242enum r5l_io_unit_state {
 243	IO_UNIT_RUNNING = 0,	/* accepting new IO */
 244	IO_UNIT_IO_START = 1,	/* io_unit bio start writing to log,
 245				 * don't accepting new bio */
 246	IO_UNIT_IO_END = 2,	/* io_unit bio finish writing to log */
 247	IO_UNIT_STRIPE_END = 3,	/* stripes data finished writing to raid */
 248};
 249
 250bool r5c_is_writeback(struct r5l_log *log)
 251{
 252	return (log != NULL &&
 253		log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK);
 254}
 255
 256static sector_t r5l_ring_add(struct r5l_log *log, sector_t start, sector_t inc)
 257{
 258	start += inc;
 259	if (start >= log->device_size)
 260		start = start - log->device_size;
 261	return start;
 262}
 263
 264static sector_t r5l_ring_distance(struct r5l_log *log, sector_t start,
 265				  sector_t end)
 266{
 267	if (end >= start)
 268		return end - start;
 269	else
 270		return end + log->device_size - start;
 271}
 272
 273static bool r5l_has_free_space(struct r5l_log *log, sector_t size)
 274{
 275	sector_t used_size;
 276
 277	used_size = r5l_ring_distance(log, log->last_checkpoint,
 278					log->log_start);
 279
 280	return log->device_size > used_size + size;
 281}
 282
 283static void __r5l_set_io_unit_state(struct r5l_io_unit *io,
 284				    enum r5l_io_unit_state state)
 285{
 286	if (WARN_ON(io->state >= state))
 287		return;
 288	io->state = state;
 289}
 290
 291static void
 292r5c_return_dev_pending_writes(struct r5conf *conf, struct r5dev *dev)
 293{
 294	struct bio *wbi, *wbi2;
 295
 296	wbi = dev->written;
 297	dev->written = NULL;
 298	while (wbi && wbi->bi_iter.bi_sector <
 299	       dev->sector + RAID5_STRIPE_SECTORS(conf)) {
 300		wbi2 = r5_next_bio(conf, wbi, dev->sector);
 301		md_write_end(conf->mddev);
 302		bio_endio(wbi);
 303		wbi = wbi2;
 304	}
 305}
 306
 307void r5c_handle_cached_data_endio(struct r5conf *conf,
 308				  struct stripe_head *sh, int disks)
 309{
 310	int i;
 311
 312	for (i = sh->disks; i--; ) {
 313		if (sh->dev[i].written) {
 314			set_bit(R5_UPTODATE, &sh->dev[i].flags);
 315			r5c_return_dev_pending_writes(conf, &sh->dev[i]);
 316			md_bitmap_endwrite(conf->mddev->bitmap, sh->sector,
 317					   RAID5_STRIPE_SECTORS(conf),
 318					   !test_bit(STRIPE_DEGRADED, &sh->state),
 319					   0);
 320		}
 321	}
 322}
 323
 324void r5l_wake_reclaim(struct r5l_log *log, sector_t space);
 325
 326/* Check whether we should flush some stripes to free up stripe cache */
 327void r5c_check_stripe_cache_usage(struct r5conf *conf)
 328{
 329	int total_cached;
 330	struct r5l_log *log = READ_ONCE(conf->log);
 331
 332	if (!r5c_is_writeback(log))
 333		return;
 334
 335	total_cached = atomic_read(&conf->r5c_cached_partial_stripes) +
 336		atomic_read(&conf->r5c_cached_full_stripes);
 337
 338	/*
 339	 * The following condition is true for either of the following:
 340	 *   - stripe cache pressure high:
 341	 *          total_cached > 3/4 min_nr_stripes ||
 342	 *          empty_inactive_list_nr > 0
 343	 *   - stripe cache pressure moderate:
 344	 *          total_cached > 1/2 min_nr_stripes
 345	 */
 346	if (total_cached > conf->min_nr_stripes * 1 / 2 ||
 347	    atomic_read(&conf->empty_inactive_list_nr) > 0)
 348		r5l_wake_reclaim(log, 0);
 349}
 350
 351/*
 352 * flush cache when there are R5C_FULL_STRIPE_FLUSH_BATCH or more full
 353 * stripes in the cache
 354 */
 355void r5c_check_cached_full_stripe(struct r5conf *conf)
 356{
 357	struct r5l_log *log = READ_ONCE(conf->log);
 358
 359	if (!r5c_is_writeback(log))
 360		return;
 361
 362	/*
 363	 * wake up reclaim for R5C_FULL_STRIPE_FLUSH_BATCH cached stripes
 364	 * or a full stripe (chunk size / 4k stripes).
 365	 */
 366	if (atomic_read(&conf->r5c_cached_full_stripes) >=
 367	    min(R5C_FULL_STRIPE_FLUSH_BATCH(conf),
 368		conf->chunk_sectors >> RAID5_STRIPE_SHIFT(conf)))
 369		r5l_wake_reclaim(log, 0);
 370}
 371
 372/*
 373 * Total log space (in sectors) needed to flush all data in cache
 374 *
 375 * To avoid deadlock due to log space, it is necessary to reserve log
 376 * space to flush critical stripes (stripes that occupying log space near
 377 * last_checkpoint). This function helps check how much log space is
 378 * required to flush all cached stripes.
 379 *
 380 * To reduce log space requirements, two mechanisms are used to give cache
 381 * flush higher priorities:
 382 *    1. In handle_stripe_dirtying() and schedule_reconstruction(),
 383 *       stripes ALREADY in journal can be flushed w/o pending writes;
 384 *    2. In r5l_write_stripe() and r5c_cache_data(), stripes NOT in journal
 385 *       can be delayed (r5l_add_no_space_stripe).
 386 *
 387 * In cache flush, the stripe goes through 1 and then 2. For a stripe that
 388 * already passed 1, flushing it requires at most (conf->max_degraded + 1)
 389 * pages of journal space. For stripes that has not passed 1, flushing it
 390 * requires (conf->raid_disks + 1) pages of journal space. There are at
 391 * most (conf->group_cnt + 1) stripe that passed 1. So total journal space
 392 * required to flush all cached stripes (in pages) is:
 393 *
 394 *     (stripe_in_journal_count - group_cnt - 1) * (max_degraded + 1) +
 395 *     (group_cnt + 1) * (raid_disks + 1)
 396 * or
 397 *     (stripe_in_journal_count) * (max_degraded + 1) +
 398 *     (group_cnt + 1) * (raid_disks - max_degraded)
 399 */
 400static sector_t r5c_log_required_to_flush_cache(struct r5conf *conf)
 401{
 402	struct r5l_log *log = READ_ONCE(conf->log);
 403
 404	if (!r5c_is_writeback(log))
 405		return 0;
 406
 407	return BLOCK_SECTORS *
 408		((conf->max_degraded + 1) * atomic_read(&log->stripe_in_journal_count) +
 409		 (conf->raid_disks - conf->max_degraded) * (conf->group_cnt + 1));
 410}
 411
 412/*
 413 * evaluate log space usage and update R5C_LOG_TIGHT and R5C_LOG_CRITICAL
 414 *
 415 * R5C_LOG_TIGHT is set when free space on the log device is less than 3x of
 416 * reclaim_required_space. R5C_LOG_CRITICAL is set when free space on the log
 417 * device is less than 2x of reclaim_required_space.
 418 */
 419static inline void r5c_update_log_state(struct r5l_log *log)
 420{
 421	struct r5conf *conf = log->rdev->mddev->private;
 422	sector_t free_space;
 423	sector_t reclaim_space;
 424	bool wake_reclaim = false;
 425
 426	if (!r5c_is_writeback(log))
 427		return;
 428
 429	free_space = r5l_ring_distance(log, log->log_start,
 430				       log->last_checkpoint);
 431	reclaim_space = r5c_log_required_to_flush_cache(conf);
 432	if (free_space < 2 * reclaim_space)
 433		set_bit(R5C_LOG_CRITICAL, &conf->cache_state);
 434	else {
 435		if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state))
 436			wake_reclaim = true;
 437		clear_bit(R5C_LOG_CRITICAL, &conf->cache_state);
 438	}
 439	if (free_space < 3 * reclaim_space)
 440		set_bit(R5C_LOG_TIGHT, &conf->cache_state);
 441	else
 442		clear_bit(R5C_LOG_TIGHT, &conf->cache_state);
 443
 444	if (wake_reclaim)
 445		r5l_wake_reclaim(log, 0);
 446}
 447
 448/*
 449 * Put the stripe into writing-out phase by clearing STRIPE_R5C_CACHING.
 450 * This function should only be called in write-back mode.
 451 */
 452void r5c_make_stripe_write_out(struct stripe_head *sh)
 453{
 454	struct r5conf *conf = sh->raid_conf;
 455	struct r5l_log *log = READ_ONCE(conf->log);
 456
 457	BUG_ON(!r5c_is_writeback(log));
 458
 459	WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
 460	clear_bit(STRIPE_R5C_CACHING, &sh->state);
 461
 462	if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 463		atomic_inc(&conf->preread_active_stripes);
 464}
 465
 466static void r5c_handle_data_cached(struct stripe_head *sh)
 467{
 468	int i;
 469
 470	for (i = sh->disks; i--; )
 471		if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
 472			set_bit(R5_InJournal, &sh->dev[i].flags);
 473			clear_bit(R5_LOCKED, &sh->dev[i].flags);
 474		}
 475	clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
 476}
 477
 478/*
 479 * this journal write must contain full parity,
 480 * it may also contain some data pages
 481 */
 482static void r5c_handle_parity_cached(struct stripe_head *sh)
 483{
 484	int i;
 485
 486	for (i = sh->disks; i--; )
 487		if (test_bit(R5_InJournal, &sh->dev[i].flags))
 488			set_bit(R5_Wantwrite, &sh->dev[i].flags);
 489}
 490
 491/*
 492 * Setting proper flags after writing (or flushing) data and/or parity to the
 493 * log device. This is called from r5l_log_endio() or r5l_log_flush_endio().
 494 */
 495static void r5c_finish_cache_stripe(struct stripe_head *sh)
 496{
 497	struct r5l_log *log = READ_ONCE(sh->raid_conf->log);
 498
 499	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
 500		BUG_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
 501		/*
 502		 * Set R5_InJournal for parity dev[pd_idx]. This means
 503		 * all data AND parity in the journal. For RAID 6, it is
 504		 * NOT necessary to set the flag for dev[qd_idx], as the
 505		 * two parities are written out together.
 506		 */
 507		set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
 508	} else if (test_bit(STRIPE_R5C_CACHING, &sh->state)) {
 509		r5c_handle_data_cached(sh);
 510	} else {
 511		r5c_handle_parity_cached(sh);
 512		set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
 513	}
 514}
 515
 516static void r5l_io_run_stripes(struct r5l_io_unit *io)
 517{
 518	struct stripe_head *sh, *next;
 519
 520	list_for_each_entry_safe(sh, next, &io->stripe_list, log_list) {
 521		list_del_init(&sh->log_list);
 522
 523		r5c_finish_cache_stripe(sh);
 524
 525		set_bit(STRIPE_HANDLE, &sh->state);
 526		raid5_release_stripe(sh);
 527	}
 528}
 529
 530static void r5l_log_run_stripes(struct r5l_log *log)
 531{
 532	struct r5l_io_unit *io, *next;
 533
 534	lockdep_assert_held(&log->io_list_lock);
 535
 536	list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) {
 537		/* don't change list order */
 538		if (io->state < IO_UNIT_IO_END)
 539			break;
 540
 541		list_move_tail(&io->log_sibling, &log->finished_ios);
 542		r5l_io_run_stripes(io);
 543	}
 544}
 545
 546static void r5l_move_to_end_ios(struct r5l_log *log)
 547{
 548	struct r5l_io_unit *io, *next;
 549
 550	lockdep_assert_held(&log->io_list_lock);
 551
 552	list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) {
 553		/* don't change list order */
 554		if (io->state < IO_UNIT_IO_END)
 555			break;
 556		list_move_tail(&io->log_sibling, &log->io_end_ios);
 557	}
 558}
 559
 560static void __r5l_stripe_write_finished(struct r5l_io_unit *io);
 561static void r5l_log_endio(struct bio *bio)
 562{
 563	struct r5l_io_unit *io = bio->bi_private;
 564	struct r5l_io_unit *io_deferred;
 565	struct r5l_log *log = io->log;
 566	unsigned long flags;
 567	bool has_null_flush;
 568	bool has_flush_payload;
 569
 570	if (bio->bi_status)
 571		md_error(log->rdev->mddev, log->rdev);
 572
 573	bio_put(bio);
 574	mempool_free(io->meta_page, &log->meta_pool);
 575
 576	spin_lock_irqsave(&log->io_list_lock, flags);
 577	__r5l_set_io_unit_state(io, IO_UNIT_IO_END);
 578
 579	/*
 580	 * if the io doesn't not have null_flush or flush payload,
 581	 * it is not safe to access it after releasing io_list_lock.
 582	 * Therefore, it is necessary to check the condition with
 583	 * the lock held.
 584	 */
 585	has_null_flush = io->has_null_flush;
 586	has_flush_payload = io->has_flush_payload;
 587
 588	if (log->need_cache_flush && !list_empty(&io->stripe_list))
 589		r5l_move_to_end_ios(log);
 590	else
 591		r5l_log_run_stripes(log);
 592	if (!list_empty(&log->running_ios)) {
 593		/*
 594		 * FLUSH/FUA io_unit is deferred because of ordering, now we
 595		 * can dispatch it
 596		 */
 597		io_deferred = list_first_entry(&log->running_ios,
 598					       struct r5l_io_unit, log_sibling);
 599		if (io_deferred->io_deferred)
 600			schedule_work(&log->deferred_io_work);
 601	}
 602
 603	spin_unlock_irqrestore(&log->io_list_lock, flags);
 604
 605	if (log->need_cache_flush)
 606		md_wakeup_thread(log->rdev->mddev->thread);
 607
 608	/* finish flush only io_unit and PAYLOAD_FLUSH only io_unit */
 609	if (has_null_flush) {
 610		struct bio *bi;
 611
 612		WARN_ON(bio_list_empty(&io->flush_barriers));
 613		while ((bi = bio_list_pop(&io->flush_barriers)) != NULL) {
 614			bio_endio(bi);
 615			if (atomic_dec_and_test(&io->pending_stripe)) {
 616				__r5l_stripe_write_finished(io);
 617				return;
 618			}
 619		}
 620	}
 621	/* decrease pending_stripe for flush payload */
 622	if (has_flush_payload)
 623		if (atomic_dec_and_test(&io->pending_stripe))
 624			__r5l_stripe_write_finished(io);
 625}
 626
 627static void r5l_do_submit_io(struct r5l_log *log, struct r5l_io_unit *io)
 628{
 629	unsigned long flags;
 630
 631	spin_lock_irqsave(&log->io_list_lock, flags);
 632	__r5l_set_io_unit_state(io, IO_UNIT_IO_START);
 633	spin_unlock_irqrestore(&log->io_list_lock, flags);
 634
 635	/*
 636	 * In case of journal device failures, submit_bio will get error
 637	 * and calls endio, then active stripes will continue write
 638	 * process. Therefore, it is not necessary to check Faulty bit
 639	 * of journal device here.
 640	 *
 641	 * We can't check split_bio after current_bio is submitted. If
 642	 * io->split_bio is null, after current_bio is submitted, current_bio
 643	 * might already be completed and the io_unit is freed. We submit
 644	 * split_bio first to avoid the issue.
 645	 */
 646	if (io->split_bio) {
 647		if (io->has_flush)
 648			io->split_bio->bi_opf |= REQ_PREFLUSH;
 649		if (io->has_fua)
 650			io->split_bio->bi_opf |= REQ_FUA;
 651		submit_bio(io->split_bio);
 652	}
 653
 654	if (io->has_flush)
 655		io->current_bio->bi_opf |= REQ_PREFLUSH;
 656	if (io->has_fua)
 657		io->current_bio->bi_opf |= REQ_FUA;
 658	submit_bio(io->current_bio);
 659}
 660
 661/* deferred io_unit will be dispatched here */
 662static void r5l_submit_io_async(struct work_struct *work)
 663{
 664	struct r5l_log *log = container_of(work, struct r5l_log,
 665					   deferred_io_work);
 666	struct r5l_io_unit *io = NULL;
 667	unsigned long flags;
 668
 669	spin_lock_irqsave(&log->io_list_lock, flags);
 670	if (!list_empty(&log->running_ios)) {
 671		io = list_first_entry(&log->running_ios, struct r5l_io_unit,
 672				      log_sibling);
 673		if (!io->io_deferred)
 674			io = NULL;
 675		else
 676			io->io_deferred = 0;
 677	}
 678	spin_unlock_irqrestore(&log->io_list_lock, flags);
 679	if (io)
 680		r5l_do_submit_io(log, io);
 681}
 682
 683static void r5c_disable_writeback_async(struct work_struct *work)
 684{
 685	struct r5l_log *log = container_of(work, struct r5l_log,
 686					   disable_writeback_work);
 687	struct mddev *mddev = log->rdev->mddev;
 688	struct r5conf *conf = mddev->private;
 
 689
 690	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
 691		return;
 692	pr_info("md/raid:%s: Disabling writeback cache for degraded array.\n",
 693		mdname(mddev));
 694
 695	/* wait superblock change before suspend */
 696	wait_event(mddev->sb_wait,
 697		   !READ_ONCE(conf->log) ||
 698		   !test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags));
 699
 700	log = READ_ONCE(conf->log);
 701	if (log) {
 702		mddev_suspend(mddev, false);
 703		log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
 704		mddev_resume(mddev);
 
 705	}
 706}
 707
 708static void r5l_submit_current_io(struct r5l_log *log)
 709{
 710	struct r5l_io_unit *io = log->current_io;
 
 711	struct r5l_meta_block *block;
 712	unsigned long flags;
 713	u32 crc;
 714	bool do_submit = true;
 715
 716	if (!io)
 717		return;
 718
 719	block = page_address(io->meta_page);
 720	block->meta_size = cpu_to_le32(io->meta_offset);
 721	crc = crc32c_le(log->uuid_checksum, block, PAGE_SIZE);
 722	block->checksum = cpu_to_le32(crc);
 
 723
 724	log->current_io = NULL;
 725	spin_lock_irqsave(&log->io_list_lock, flags);
 726	if (io->has_flush || io->has_fua) {
 727		if (io != list_first_entry(&log->running_ios,
 728					   struct r5l_io_unit, log_sibling)) {
 729			io->io_deferred = 1;
 730			do_submit = false;
 731		}
 732	}
 733	spin_unlock_irqrestore(&log->io_list_lock, flags);
 734	if (do_submit)
 735		r5l_do_submit_io(log, io);
 736}
 737
 738static struct bio *r5l_bio_alloc(struct r5l_log *log)
 739{
 740	struct bio *bio = bio_alloc_bioset(log->rdev->bdev, BIO_MAX_VECS,
 741					   REQ_OP_WRITE, GFP_NOIO, &log->bs);
 742
 
 
 743	bio->bi_iter.bi_sector = log->rdev->data_offset + log->log_start;
 744
 745	return bio;
 746}
 747
 748static void r5_reserve_log_entry(struct r5l_log *log, struct r5l_io_unit *io)
 749{
 750	log->log_start = r5l_ring_add(log, log->log_start, BLOCK_SECTORS);
 751
 752	r5c_update_log_state(log);
 753	/*
 754	 * If we filled up the log device start from the beginning again,
 755	 * which will require a new bio.
 756	 *
 757	 * Note: for this to work properly the log size needs to me a multiple
 758	 * of BLOCK_SECTORS.
 759	 */
 760	if (log->log_start == 0)
 761		io->need_split_bio = true;
 762
 763	io->log_end = log->log_start;
 764}
 765
 766static struct r5l_io_unit *r5l_new_meta(struct r5l_log *log)
 767{
 768	struct r5l_io_unit *io;
 769	struct r5l_meta_block *block;
 770
 771	io = mempool_alloc(&log->io_pool, GFP_ATOMIC);
 772	if (!io)
 773		return NULL;
 774	memset(io, 0, sizeof(*io));
 775
 776	io->log = log;
 777	INIT_LIST_HEAD(&io->log_sibling);
 778	INIT_LIST_HEAD(&io->stripe_list);
 779	bio_list_init(&io->flush_barriers);
 780	io->state = IO_UNIT_RUNNING;
 781
 782	io->meta_page = mempool_alloc(&log->meta_pool, GFP_NOIO);
 783	block = page_address(io->meta_page);
 784	clear_page(block);
 785	block->magic = cpu_to_le32(R5LOG_MAGIC);
 786	block->version = R5LOG_VERSION;
 787	block->seq = cpu_to_le64(log->seq);
 788	block->position = cpu_to_le64(log->log_start);
 789
 790	io->log_start = log->log_start;
 791	io->meta_offset = sizeof(struct r5l_meta_block);
 792	io->seq = log->seq++;
 793
 794	io->current_bio = r5l_bio_alloc(log);
 795	io->current_bio->bi_end_io = r5l_log_endio;
 796	io->current_bio->bi_private = io;
 797	__bio_add_page(io->current_bio, io->meta_page, PAGE_SIZE, 0);
 798
 799	r5_reserve_log_entry(log, io);
 800
 801	spin_lock_irq(&log->io_list_lock);
 802	list_add_tail(&io->log_sibling, &log->running_ios);
 803	spin_unlock_irq(&log->io_list_lock);
 804
 805	return io;
 806}
 807
 808static int r5l_get_meta(struct r5l_log *log, unsigned int payload_size)
 809{
 810	if (log->current_io &&
 811	    log->current_io->meta_offset + payload_size > PAGE_SIZE)
 812		r5l_submit_current_io(log);
 813
 814	if (!log->current_io) {
 815		log->current_io = r5l_new_meta(log);
 816		if (!log->current_io)
 817			return -ENOMEM;
 818	}
 819
 820	return 0;
 821}
 822
 823static void r5l_append_payload_meta(struct r5l_log *log, u16 type,
 824				    sector_t location,
 825				    u32 checksum1, u32 checksum2,
 826				    bool checksum2_valid)
 827{
 828	struct r5l_io_unit *io = log->current_io;
 829	struct r5l_payload_data_parity *payload;
 830
 831	payload = page_address(io->meta_page) + io->meta_offset;
 832	payload->header.type = cpu_to_le16(type);
 833	payload->header.flags = cpu_to_le16(0);
 834	payload->size = cpu_to_le32((1 + !!checksum2_valid) <<
 835				    (PAGE_SHIFT - 9));
 836	payload->location = cpu_to_le64(location);
 837	payload->checksum[0] = cpu_to_le32(checksum1);
 838	if (checksum2_valid)
 839		payload->checksum[1] = cpu_to_le32(checksum2);
 840
 841	io->meta_offset += sizeof(struct r5l_payload_data_parity) +
 842		sizeof(__le32) * (1 + !!checksum2_valid);
 843}
 844
 845static void r5l_append_payload_page(struct r5l_log *log, struct page *page)
 846{
 847	struct r5l_io_unit *io = log->current_io;
 848
 849	if (io->need_split_bio) {
 850		BUG_ON(io->split_bio);
 851		io->split_bio = io->current_bio;
 852		io->current_bio = r5l_bio_alloc(log);
 853		bio_chain(io->current_bio, io->split_bio);
 854		io->need_split_bio = false;
 855	}
 856
 857	if (!bio_add_page(io->current_bio, page, PAGE_SIZE, 0))
 858		BUG();
 859
 860	r5_reserve_log_entry(log, io);
 861}
 862
 863static void r5l_append_flush_payload(struct r5l_log *log, sector_t sect)
 864{
 865	struct mddev *mddev = log->rdev->mddev;
 866	struct r5conf *conf = mddev->private;
 867	struct r5l_io_unit *io;
 868	struct r5l_payload_flush *payload;
 869	int meta_size;
 870
 871	/*
 872	 * payload_flush requires extra writes to the journal.
 873	 * To avoid handling the extra IO in quiesce, just skip
 874	 * flush_payload
 875	 */
 876	if (conf->quiesce)
 877		return;
 878
 879	mutex_lock(&log->io_mutex);
 880	meta_size = sizeof(struct r5l_payload_flush) + sizeof(__le64);
 881
 882	if (r5l_get_meta(log, meta_size)) {
 883		mutex_unlock(&log->io_mutex);
 884		return;
 885	}
 886
 887	/* current implementation is one stripe per flush payload */
 888	io = log->current_io;
 889	payload = page_address(io->meta_page) + io->meta_offset;
 890	payload->header.type = cpu_to_le16(R5LOG_PAYLOAD_FLUSH);
 891	payload->header.flags = cpu_to_le16(0);
 892	payload->size = cpu_to_le32(sizeof(__le64));
 893	payload->flush_stripes[0] = cpu_to_le64(sect);
 894	io->meta_offset += meta_size;
 895	/* multiple flush payloads count as one pending_stripe */
 896	if (!io->has_flush_payload) {
 897		io->has_flush_payload = 1;
 898		atomic_inc(&io->pending_stripe);
 899	}
 900	mutex_unlock(&log->io_mutex);
 901}
 902
 903static int r5l_log_stripe(struct r5l_log *log, struct stripe_head *sh,
 904			   int data_pages, int parity_pages)
 905{
 906	int i;
 907	int meta_size;
 908	int ret;
 909	struct r5l_io_unit *io;
 910
 911	meta_size =
 912		((sizeof(struct r5l_payload_data_parity) + sizeof(__le32))
 913		 * data_pages) +
 914		sizeof(struct r5l_payload_data_parity) +
 915		sizeof(__le32) * parity_pages;
 916
 917	ret = r5l_get_meta(log, meta_size);
 918	if (ret)
 919		return ret;
 920
 921	io = log->current_io;
 922
 923	if (test_and_clear_bit(STRIPE_R5C_PREFLUSH, &sh->state))
 924		io->has_flush = 1;
 925
 926	for (i = 0; i < sh->disks; i++) {
 927		if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) ||
 928		    test_bit(R5_InJournal, &sh->dev[i].flags))
 929			continue;
 930		if (i == sh->pd_idx || i == sh->qd_idx)
 931			continue;
 932		if (test_bit(R5_WantFUA, &sh->dev[i].flags) &&
 933		    log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK) {
 934			io->has_fua = 1;
 935			/*
 936			 * we need to flush journal to make sure recovery can
 937			 * reach the data with fua flag
 938			 */
 939			io->has_flush = 1;
 940		}
 941		r5l_append_payload_meta(log, R5LOG_PAYLOAD_DATA,
 942					raid5_compute_blocknr(sh, i, 0),
 943					sh->dev[i].log_checksum, 0, false);
 944		r5l_append_payload_page(log, sh->dev[i].page);
 945	}
 946
 947	if (parity_pages == 2) {
 948		r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
 949					sh->sector, sh->dev[sh->pd_idx].log_checksum,
 950					sh->dev[sh->qd_idx].log_checksum, true);
 951		r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
 952		r5l_append_payload_page(log, sh->dev[sh->qd_idx].page);
 953	} else if (parity_pages == 1) {
 954		r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
 955					sh->sector, sh->dev[sh->pd_idx].log_checksum,
 956					0, false);
 957		r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
 958	} else  /* Just writing data, not parity, in caching phase */
 959		BUG_ON(parity_pages != 0);
 960
 961	list_add_tail(&sh->log_list, &io->stripe_list);
 962	atomic_inc(&io->pending_stripe);
 963	sh->log_io = io;
 964
 965	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
 966		return 0;
 967
 968	if (sh->log_start == MaxSector) {
 969		BUG_ON(!list_empty(&sh->r5c));
 970		sh->log_start = io->log_start;
 971		spin_lock_irq(&log->stripe_in_journal_lock);
 972		list_add_tail(&sh->r5c,
 973			      &log->stripe_in_journal_list);
 974		spin_unlock_irq(&log->stripe_in_journal_lock);
 975		atomic_inc(&log->stripe_in_journal_count);
 976	}
 977	return 0;
 978}
 979
 980/* add stripe to no_space_stripes, and then wake up reclaim */
 981static inline void r5l_add_no_space_stripe(struct r5l_log *log,
 982					   struct stripe_head *sh)
 983{
 984	spin_lock(&log->no_space_stripes_lock);
 985	list_add_tail(&sh->log_list, &log->no_space_stripes);
 986	spin_unlock(&log->no_space_stripes_lock);
 987}
 988
 989/*
 990 * running in raid5d, where reclaim could wait for raid5d too (when it flushes
 991 * data from log to raid disks), so we shouldn't wait for reclaim here
 992 */
 993int r5l_write_stripe(struct r5l_log *log, struct stripe_head *sh)
 994{
 995	struct r5conf *conf = sh->raid_conf;
 996	int write_disks = 0;
 997	int data_pages, parity_pages;
 998	int reserve;
 999	int i;
1000	int ret = 0;
1001	bool wake_reclaim = false;
1002
1003	if (!log)
1004		return -EAGAIN;
1005	/* Don't support stripe batch */
1006	if (sh->log_io || !test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags) ||
1007	    test_bit(STRIPE_SYNCING, &sh->state)) {
1008		/* the stripe is written to log, we start writing it to raid */
1009		clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
1010		return -EAGAIN;
1011	}
1012
1013	WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
1014
1015	for (i = 0; i < sh->disks; i++) {
1016		void *addr;
1017
1018		if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) ||
1019		    test_bit(R5_InJournal, &sh->dev[i].flags))
1020			continue;
1021
1022		write_disks++;
1023		/* checksum is already calculated in last run */
1024		if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
1025			continue;
1026		addr = kmap_atomic(sh->dev[i].page);
1027		sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum,
1028						    addr, PAGE_SIZE);
1029		kunmap_atomic(addr);
1030	}
1031	parity_pages = 1 + !!(sh->qd_idx >= 0);
1032	data_pages = write_disks - parity_pages;
1033
1034	set_bit(STRIPE_LOG_TRAPPED, &sh->state);
1035	/*
1036	 * The stripe must enter state machine again to finish the write, so
1037	 * don't delay.
1038	 */
1039	clear_bit(STRIPE_DELAYED, &sh->state);
1040	atomic_inc(&sh->count);
1041
1042	mutex_lock(&log->io_mutex);
1043	/* meta + data */
1044	reserve = (1 + write_disks) << (PAGE_SHIFT - 9);
1045
1046	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
1047		if (!r5l_has_free_space(log, reserve)) {
1048			r5l_add_no_space_stripe(log, sh);
1049			wake_reclaim = true;
1050		} else {
1051			ret = r5l_log_stripe(log, sh, data_pages, parity_pages);
1052			if (ret) {
1053				spin_lock_irq(&log->io_list_lock);
1054				list_add_tail(&sh->log_list,
1055					      &log->no_mem_stripes);
1056				spin_unlock_irq(&log->io_list_lock);
1057			}
1058		}
1059	} else {  /* R5C_JOURNAL_MODE_WRITE_BACK */
1060		/*
1061		 * log space critical, do not process stripes that are
1062		 * not in cache yet (sh->log_start == MaxSector).
1063		 */
1064		if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
1065		    sh->log_start == MaxSector) {
1066			r5l_add_no_space_stripe(log, sh);
1067			wake_reclaim = true;
1068			reserve = 0;
1069		} else if (!r5l_has_free_space(log, reserve)) {
1070			if (sh->log_start == log->last_checkpoint)
1071				BUG();
1072			else
1073				r5l_add_no_space_stripe(log, sh);
1074		} else {
1075			ret = r5l_log_stripe(log, sh, data_pages, parity_pages);
1076			if (ret) {
1077				spin_lock_irq(&log->io_list_lock);
1078				list_add_tail(&sh->log_list,
1079					      &log->no_mem_stripes);
1080				spin_unlock_irq(&log->io_list_lock);
1081			}
1082		}
1083	}
1084
1085	mutex_unlock(&log->io_mutex);
1086	if (wake_reclaim)
1087		r5l_wake_reclaim(log, reserve);
1088	return 0;
1089}
1090
1091void r5l_write_stripe_run(struct r5l_log *log)
1092{
1093	if (!log)
1094		return;
1095	mutex_lock(&log->io_mutex);
1096	r5l_submit_current_io(log);
1097	mutex_unlock(&log->io_mutex);
1098}
1099
1100int r5l_handle_flush_request(struct r5l_log *log, struct bio *bio)
1101{
1102	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
1103		/*
1104		 * in write through (journal only)
1105		 * we flush log disk cache first, then write stripe data to
1106		 * raid disks. So if bio is finished, the log disk cache is
1107		 * flushed already. The recovery guarantees we can recovery
1108		 * the bio from log disk, so we don't need to flush again
1109		 */
1110		if (bio->bi_iter.bi_size == 0) {
1111			bio_endio(bio);
1112			return 0;
1113		}
1114		bio->bi_opf &= ~REQ_PREFLUSH;
1115	} else {
1116		/* write back (with cache) */
1117		if (bio->bi_iter.bi_size == 0) {
1118			mutex_lock(&log->io_mutex);
1119			r5l_get_meta(log, 0);
1120			bio_list_add(&log->current_io->flush_barriers, bio);
1121			log->current_io->has_flush = 1;
1122			log->current_io->has_null_flush = 1;
1123			atomic_inc(&log->current_io->pending_stripe);
1124			r5l_submit_current_io(log);
1125			mutex_unlock(&log->io_mutex);
1126			return 0;
1127		}
1128	}
1129	return -EAGAIN;
1130}
1131
1132/* This will run after log space is reclaimed */
1133static void r5l_run_no_space_stripes(struct r5l_log *log)
1134{
1135	struct stripe_head *sh;
1136
1137	spin_lock(&log->no_space_stripes_lock);
1138	while (!list_empty(&log->no_space_stripes)) {
1139		sh = list_first_entry(&log->no_space_stripes,
1140				      struct stripe_head, log_list);
1141		list_del_init(&sh->log_list);
1142		set_bit(STRIPE_HANDLE, &sh->state);
1143		raid5_release_stripe(sh);
1144	}
1145	spin_unlock(&log->no_space_stripes_lock);
1146}
1147
1148/*
1149 * calculate new last_checkpoint
1150 * for write through mode, returns log->next_checkpoint
1151 * for write back, returns log_start of first sh in stripe_in_journal_list
1152 */
1153static sector_t r5c_calculate_new_cp(struct r5conf *conf)
1154{
1155	struct stripe_head *sh;
1156	struct r5l_log *log = READ_ONCE(conf->log);
1157	sector_t new_cp;
1158	unsigned long flags;
1159
1160	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
1161		return log->next_checkpoint;
1162
1163	spin_lock_irqsave(&log->stripe_in_journal_lock, flags);
1164	if (list_empty(&log->stripe_in_journal_list)) {
1165		/* all stripes flushed */
1166		spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1167		return log->next_checkpoint;
1168	}
1169	sh = list_first_entry(&log->stripe_in_journal_list,
1170			      struct stripe_head, r5c);
1171	new_cp = sh->log_start;
1172	spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1173	return new_cp;
1174}
1175
1176static sector_t r5l_reclaimable_space(struct r5l_log *log)
1177{
1178	struct r5conf *conf = log->rdev->mddev->private;
1179
1180	return r5l_ring_distance(log, log->last_checkpoint,
1181				 r5c_calculate_new_cp(conf));
1182}
1183
1184static void r5l_run_no_mem_stripe(struct r5l_log *log)
1185{
1186	struct stripe_head *sh;
1187
1188	lockdep_assert_held(&log->io_list_lock);
1189
1190	if (!list_empty(&log->no_mem_stripes)) {
1191		sh = list_first_entry(&log->no_mem_stripes,
1192				      struct stripe_head, log_list);
1193		list_del_init(&sh->log_list);
1194		set_bit(STRIPE_HANDLE, &sh->state);
1195		raid5_release_stripe(sh);
1196	}
1197}
1198
1199static bool r5l_complete_finished_ios(struct r5l_log *log)
1200{
1201	struct r5l_io_unit *io, *next;
1202	bool found = false;
1203
1204	lockdep_assert_held(&log->io_list_lock);
1205
1206	list_for_each_entry_safe(io, next, &log->finished_ios, log_sibling) {
1207		/* don't change list order */
1208		if (io->state < IO_UNIT_STRIPE_END)
1209			break;
1210
1211		log->next_checkpoint = io->log_start;
1212
1213		list_del(&io->log_sibling);
1214		mempool_free(io, &log->io_pool);
1215		r5l_run_no_mem_stripe(log);
1216
1217		found = true;
1218	}
1219
1220	return found;
1221}
1222
1223static void __r5l_stripe_write_finished(struct r5l_io_unit *io)
1224{
1225	struct r5l_log *log = io->log;
1226	struct r5conf *conf = log->rdev->mddev->private;
1227	unsigned long flags;
1228
1229	spin_lock_irqsave(&log->io_list_lock, flags);
1230	__r5l_set_io_unit_state(io, IO_UNIT_STRIPE_END);
1231
1232	if (!r5l_complete_finished_ios(log)) {
1233		spin_unlock_irqrestore(&log->io_list_lock, flags);
1234		return;
1235	}
1236
1237	if (r5l_reclaimable_space(log) > log->max_free_space ||
1238	    test_bit(R5C_LOG_TIGHT, &conf->cache_state))
1239		r5l_wake_reclaim(log, 0);
1240
1241	spin_unlock_irqrestore(&log->io_list_lock, flags);
1242	wake_up(&log->iounit_wait);
1243}
1244
1245void r5l_stripe_write_finished(struct stripe_head *sh)
1246{
1247	struct r5l_io_unit *io;
1248
1249	io = sh->log_io;
1250	sh->log_io = NULL;
1251
1252	if (io && atomic_dec_and_test(&io->pending_stripe))
1253		__r5l_stripe_write_finished(io);
1254}
1255
1256static void r5l_log_flush_endio(struct bio *bio)
1257{
1258	struct r5l_log *log = container_of(bio, struct r5l_log,
1259		flush_bio);
1260	unsigned long flags;
1261	struct r5l_io_unit *io;
1262
1263	if (bio->bi_status)
1264		md_error(log->rdev->mddev, log->rdev);
1265	bio_uninit(bio);
1266
1267	spin_lock_irqsave(&log->io_list_lock, flags);
1268	list_for_each_entry(io, &log->flushing_ios, log_sibling)
1269		r5l_io_run_stripes(io);
1270	list_splice_tail_init(&log->flushing_ios, &log->finished_ios);
1271	spin_unlock_irqrestore(&log->io_list_lock, flags);
1272}
1273
1274/*
1275 * Starting dispatch IO to raid.
1276 * io_unit(meta) consists of a log. There is one situation we want to avoid. A
1277 * broken meta in the middle of a log causes recovery can't find meta at the
1278 * head of log. If operations require meta at the head persistent in log, we
1279 * must make sure meta before it persistent in log too. A case is:
1280 *
1281 * stripe data/parity is in log, we start write stripe to raid disks. stripe
1282 * data/parity must be persistent in log before we do the write to raid disks.
1283 *
1284 * The solution is we restrictly maintain io_unit list order. In this case, we
1285 * only write stripes of an io_unit to raid disks till the io_unit is the first
1286 * one whose data/parity is in log.
1287 */
1288void r5l_flush_stripe_to_raid(struct r5l_log *log)
1289{
1290	bool do_flush;
1291
1292	if (!log || !log->need_cache_flush)
1293		return;
1294
1295	spin_lock_irq(&log->io_list_lock);
1296	/* flush bio is running */
1297	if (!list_empty(&log->flushing_ios)) {
1298		spin_unlock_irq(&log->io_list_lock);
1299		return;
1300	}
1301	list_splice_tail_init(&log->io_end_ios, &log->flushing_ios);
1302	do_flush = !list_empty(&log->flushing_ios);
1303	spin_unlock_irq(&log->io_list_lock);
1304
1305	if (!do_flush)
1306		return;
1307	bio_init(&log->flush_bio, log->rdev->bdev, NULL, 0,
1308		  REQ_OP_WRITE | REQ_PREFLUSH);
1309	log->flush_bio.bi_end_io = r5l_log_flush_endio;
 
1310	submit_bio(&log->flush_bio);
1311}
1312
1313static void r5l_write_super(struct r5l_log *log, sector_t cp);
1314static void r5l_write_super_and_discard_space(struct r5l_log *log,
1315	sector_t end)
1316{
1317	struct block_device *bdev = log->rdev->bdev;
1318	struct mddev *mddev;
1319
1320	r5l_write_super(log, end);
1321
1322	if (!bdev_max_discard_sectors(bdev))
1323		return;
1324
1325	mddev = log->rdev->mddev;
1326	/*
1327	 * Discard could zero data, so before discard we must make sure
1328	 * superblock is updated to new log tail. Updating superblock (either
1329	 * directly call md_update_sb() or depend on md thread) must hold
1330	 * reconfig mutex. On the other hand, raid5_quiesce is called with
1331	 * reconfig_mutex hold. The first step of raid5_quiesce() is waiting
1332	 * for all IO finish, hence waiting for reclaim thread, while reclaim
1333	 * thread is calling this function and waiting for reconfig mutex. So
1334	 * there is a deadlock. We workaround this issue with a trylock.
1335	 * FIXME: we could miss discard if we can't take reconfig mutex
1336	 */
1337	set_mask_bits(&mddev->sb_flags, 0,
1338		BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1339	if (!mddev_trylock(mddev))
1340		return;
1341	md_update_sb(mddev, 1);
1342	mddev_unlock(mddev);
1343
1344	/* discard IO error really doesn't matter, ignore it */
1345	if (log->last_checkpoint < end) {
1346		blkdev_issue_discard(bdev,
1347				log->last_checkpoint + log->rdev->data_offset,
1348				end - log->last_checkpoint, GFP_NOIO);
1349	} else {
1350		blkdev_issue_discard(bdev,
1351				log->last_checkpoint + log->rdev->data_offset,
1352				log->device_size - log->last_checkpoint,
1353				GFP_NOIO);
1354		blkdev_issue_discard(bdev, log->rdev->data_offset, end,
1355				GFP_NOIO);
1356	}
1357}
1358
1359/*
1360 * r5c_flush_stripe moves stripe from cached list to handle_list. When called,
1361 * the stripe must be on r5c_cached_full_stripes or r5c_cached_partial_stripes.
1362 *
1363 * must hold conf->device_lock
1364 */
1365static void r5c_flush_stripe(struct r5conf *conf, struct stripe_head *sh)
1366{
1367	BUG_ON(list_empty(&sh->lru));
1368	BUG_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
1369	BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
1370
1371	/*
1372	 * The stripe is not ON_RELEASE_LIST, so it is safe to call
1373	 * raid5_release_stripe() while holding conf->device_lock
1374	 */
1375	BUG_ON(test_bit(STRIPE_ON_RELEASE_LIST, &sh->state));
1376	lockdep_assert_held(&conf->device_lock);
1377
1378	list_del_init(&sh->lru);
1379	atomic_inc(&sh->count);
1380
1381	set_bit(STRIPE_HANDLE, &sh->state);
1382	atomic_inc(&conf->active_stripes);
1383	r5c_make_stripe_write_out(sh);
1384
1385	if (test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state))
1386		atomic_inc(&conf->r5c_flushing_partial_stripes);
1387	else
1388		atomic_inc(&conf->r5c_flushing_full_stripes);
1389	raid5_release_stripe(sh);
1390}
1391
1392/*
1393 * if num == 0, flush all full stripes
1394 * if num > 0, flush all full stripes. If less than num full stripes are
1395 *             flushed, flush some partial stripes until totally num stripes are
1396 *             flushed or there is no more cached stripes.
1397 */
1398void r5c_flush_cache(struct r5conf *conf, int num)
1399{
1400	int count;
1401	struct stripe_head *sh, *next;
1402
1403	lockdep_assert_held(&conf->device_lock);
1404	if (!READ_ONCE(conf->log))
1405		return;
1406
1407	count = 0;
1408	list_for_each_entry_safe(sh, next, &conf->r5c_full_stripe_list, lru) {
1409		r5c_flush_stripe(conf, sh);
1410		count++;
1411	}
1412
1413	if (count >= num)
1414		return;
1415	list_for_each_entry_safe(sh, next,
1416				 &conf->r5c_partial_stripe_list, lru) {
1417		r5c_flush_stripe(conf, sh);
1418		if (++count >= num)
1419			break;
1420	}
1421}
1422
1423static void r5c_do_reclaim(struct r5conf *conf)
1424{
1425	struct r5l_log *log = READ_ONCE(conf->log);
1426	struct stripe_head *sh;
1427	int count = 0;
1428	unsigned long flags;
1429	int total_cached;
1430	int stripes_to_flush;
1431	int flushing_partial, flushing_full;
1432
1433	if (!r5c_is_writeback(log))
1434		return;
1435
1436	flushing_partial = atomic_read(&conf->r5c_flushing_partial_stripes);
1437	flushing_full = atomic_read(&conf->r5c_flushing_full_stripes);
1438	total_cached = atomic_read(&conf->r5c_cached_partial_stripes) +
1439		atomic_read(&conf->r5c_cached_full_stripes) -
1440		flushing_full - flushing_partial;
1441
1442	if (total_cached > conf->min_nr_stripes * 3 / 4 ||
1443	    atomic_read(&conf->empty_inactive_list_nr) > 0)
1444		/*
1445		 * if stripe cache pressure high, flush all full stripes and
1446		 * some partial stripes
1447		 */
1448		stripes_to_flush = R5C_RECLAIM_STRIPE_GROUP;
1449	else if (total_cached > conf->min_nr_stripes * 1 / 2 ||
1450		 atomic_read(&conf->r5c_cached_full_stripes) - flushing_full >
1451		 R5C_FULL_STRIPE_FLUSH_BATCH(conf))
1452		/*
1453		 * if stripe cache pressure moderate, or if there is many full
1454		 * stripes,flush all full stripes
1455		 */
1456		stripes_to_flush = 0;
1457	else
1458		/* no need to flush */
1459		stripes_to_flush = -1;
1460
1461	if (stripes_to_flush >= 0) {
1462		spin_lock_irqsave(&conf->device_lock, flags);
1463		r5c_flush_cache(conf, stripes_to_flush);
1464		spin_unlock_irqrestore(&conf->device_lock, flags);
1465	}
1466
1467	/* if log space is tight, flush stripes on stripe_in_journal_list */
1468	if (test_bit(R5C_LOG_TIGHT, &conf->cache_state)) {
1469		spin_lock_irqsave(&log->stripe_in_journal_lock, flags);
1470		spin_lock(&conf->device_lock);
1471		list_for_each_entry(sh, &log->stripe_in_journal_list, r5c) {
1472			/*
1473			 * stripes on stripe_in_journal_list could be in any
1474			 * state of the stripe_cache state machine. In this
1475			 * case, we only want to flush stripe on
1476			 * r5c_cached_full/partial_stripes. The following
1477			 * condition makes sure the stripe is on one of the
1478			 * two lists.
1479			 */
1480			if (!list_empty(&sh->lru) &&
1481			    !test_bit(STRIPE_HANDLE, &sh->state) &&
1482			    atomic_read(&sh->count) == 0) {
1483				r5c_flush_stripe(conf, sh);
1484				if (count++ >= R5C_RECLAIM_STRIPE_GROUP)
1485					break;
1486			}
1487		}
1488		spin_unlock(&conf->device_lock);
1489		spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1490	}
1491
1492	if (!test_bit(R5C_LOG_CRITICAL, &conf->cache_state))
1493		r5l_run_no_space_stripes(log);
1494
1495	md_wakeup_thread(conf->mddev->thread);
1496}
1497
1498static void r5l_do_reclaim(struct r5l_log *log)
1499{
1500	struct r5conf *conf = log->rdev->mddev->private;
1501	sector_t reclaim_target = xchg(&log->reclaim_target, 0);
1502	sector_t reclaimable;
1503	sector_t next_checkpoint;
1504	bool write_super;
1505
1506	spin_lock_irq(&log->io_list_lock);
1507	write_super = r5l_reclaimable_space(log) > log->max_free_space ||
1508		reclaim_target != 0 || !list_empty(&log->no_space_stripes);
1509	/*
1510	 * move proper io_unit to reclaim list. We should not change the order.
1511	 * reclaimable/unreclaimable io_unit can be mixed in the list, we
1512	 * shouldn't reuse space of an unreclaimable io_unit
1513	 */
1514	while (1) {
1515		reclaimable = r5l_reclaimable_space(log);
1516		if (reclaimable >= reclaim_target ||
1517		    (list_empty(&log->running_ios) &&
1518		     list_empty(&log->io_end_ios) &&
1519		     list_empty(&log->flushing_ios) &&
1520		     list_empty(&log->finished_ios)))
1521			break;
1522
1523		md_wakeup_thread(log->rdev->mddev->thread);
1524		wait_event_lock_irq(log->iounit_wait,
1525				    r5l_reclaimable_space(log) > reclaimable,
1526				    log->io_list_lock);
1527	}
1528
1529	next_checkpoint = r5c_calculate_new_cp(conf);
1530	spin_unlock_irq(&log->io_list_lock);
1531
1532	if (reclaimable == 0 || !write_super)
1533		return;
1534
1535	/*
1536	 * write_super will flush cache of each raid disk. We must write super
1537	 * here, because the log area might be reused soon and we don't want to
1538	 * confuse recovery
1539	 */
1540	r5l_write_super_and_discard_space(log, next_checkpoint);
1541
1542	mutex_lock(&log->io_mutex);
1543	log->last_checkpoint = next_checkpoint;
1544	r5c_update_log_state(log);
1545	mutex_unlock(&log->io_mutex);
1546
1547	r5l_run_no_space_stripes(log);
1548}
1549
1550static void r5l_reclaim_thread(struct md_thread *thread)
1551{
1552	struct mddev *mddev = thread->mddev;
1553	struct r5conf *conf = mddev->private;
1554	struct r5l_log *log = READ_ONCE(conf->log);
1555
1556	if (!log)
1557		return;
1558	r5c_do_reclaim(conf);
1559	r5l_do_reclaim(log);
1560}
1561
1562void r5l_wake_reclaim(struct r5l_log *log, sector_t space)
1563{
1564	unsigned long target;
1565	unsigned long new = (unsigned long)space; /* overflow in theory */
1566
1567	if (!log)
1568		return;
1569
1570	target = READ_ONCE(log->reclaim_target);
1571	do {
 
1572		if (new < target)
1573			return;
1574	} while (!try_cmpxchg(&log->reclaim_target, &target, new));
1575	md_wakeup_thread(log->reclaim_thread);
1576}
1577
1578void r5l_quiesce(struct r5l_log *log, int quiesce)
1579{
1580	struct mddev *mddev = log->rdev->mddev;
1581	struct md_thread *thread = rcu_dereference_protected(
1582		log->reclaim_thread, lockdep_is_held(&mddev->reconfig_mutex));
1583
1584	if (quiesce) {
1585		/* make sure r5l_write_super_and_discard_space exits */
 
1586		wake_up(&mddev->sb_wait);
1587		kthread_park(thread->tsk);
1588		r5l_wake_reclaim(log, MaxSector);
1589		r5l_do_reclaim(log);
1590	} else
1591		kthread_unpark(thread->tsk);
1592}
1593
1594bool r5l_log_disk_error(struct r5conf *conf)
1595{
1596	struct r5l_log *log = READ_ONCE(conf->log);
1597
1598	/* don't allow write if journal disk is missing */
 
 
 
1599	if (!log)
1600		return test_bit(MD_HAS_JOURNAL, &conf->mddev->flags);
1601	else
1602		return test_bit(Faulty, &log->rdev->flags);
 
 
1603}
1604
1605#define R5L_RECOVERY_PAGE_POOL_SIZE 256
1606
1607struct r5l_recovery_ctx {
1608	struct page *meta_page;		/* current meta */
1609	sector_t meta_total_blocks;	/* total size of current meta and data */
1610	sector_t pos;			/* recovery position */
1611	u64 seq;			/* recovery position seq */
1612	int data_parity_stripes;	/* number of data_parity stripes */
1613	int data_only_stripes;		/* number of data_only stripes */
1614	struct list_head cached_list;
1615
1616	/*
1617	 * read ahead page pool (ra_pool)
1618	 * in recovery, log is read sequentially. It is not efficient to
1619	 * read every page with sync_page_io(). The read ahead page pool
1620	 * reads multiple pages with one IO, so further log read can
1621	 * just copy data from the pool.
1622	 */
1623	struct page *ra_pool[R5L_RECOVERY_PAGE_POOL_SIZE];
1624	struct bio_vec ra_bvec[R5L_RECOVERY_PAGE_POOL_SIZE];
1625	sector_t pool_offset;	/* offset of first page in the pool */
1626	int total_pages;	/* total allocated pages */
1627	int valid_pages;	/* pages with valid data */
 
1628};
1629
1630static int r5l_recovery_allocate_ra_pool(struct r5l_log *log,
1631					    struct r5l_recovery_ctx *ctx)
1632{
1633	struct page *page;
1634
 
 
 
 
1635	ctx->valid_pages = 0;
1636	ctx->total_pages = 0;
1637	while (ctx->total_pages < R5L_RECOVERY_PAGE_POOL_SIZE) {
1638		page = alloc_page(GFP_KERNEL);
1639
1640		if (!page)
1641			break;
1642		ctx->ra_pool[ctx->total_pages] = page;
1643		ctx->total_pages += 1;
1644	}
1645
1646	if (ctx->total_pages == 0)
 
1647		return -ENOMEM;
 
1648
1649	ctx->pool_offset = 0;
1650	return 0;
1651}
1652
1653static void r5l_recovery_free_ra_pool(struct r5l_log *log,
1654					struct r5l_recovery_ctx *ctx)
1655{
1656	int i;
1657
1658	for (i = 0; i < ctx->total_pages; ++i)
1659		put_page(ctx->ra_pool[i]);
 
1660}
1661
1662/*
1663 * fetch ctx->valid_pages pages from offset
1664 * In normal cases, ctx->valid_pages == ctx->total_pages after the call.
1665 * However, if the offset is close to the end of the journal device,
1666 * ctx->valid_pages could be smaller than ctx->total_pages
1667 */
1668static int r5l_recovery_fetch_ra_pool(struct r5l_log *log,
1669				      struct r5l_recovery_ctx *ctx,
1670				      sector_t offset)
1671{
1672	struct bio bio;
1673	int ret;
1674
1675	bio_init(&bio, log->rdev->bdev, ctx->ra_bvec,
1676		 R5L_RECOVERY_PAGE_POOL_SIZE, REQ_OP_READ);
1677	bio.bi_iter.bi_sector = log->rdev->data_offset + offset;
1678
1679	ctx->valid_pages = 0;
1680	ctx->pool_offset = offset;
1681
1682	while (ctx->valid_pages < ctx->total_pages) {
1683		__bio_add_page(&bio, ctx->ra_pool[ctx->valid_pages], PAGE_SIZE,
1684			       0);
1685		ctx->valid_pages += 1;
1686
1687		offset = r5l_ring_add(log, offset, BLOCK_SECTORS);
1688
1689		if (offset == 0)  /* reached end of the device */
1690			break;
1691	}
1692
1693	ret = submit_bio_wait(&bio);
1694	bio_uninit(&bio);
1695	return ret;
1696}
1697
1698/*
1699 * try read a page from the read ahead page pool, if the page is not in the
1700 * pool, call r5l_recovery_fetch_ra_pool
1701 */
1702static int r5l_recovery_read_page(struct r5l_log *log,
1703				  struct r5l_recovery_ctx *ctx,
1704				  struct page *page,
1705				  sector_t offset)
1706{
1707	int ret;
1708
1709	if (offset < ctx->pool_offset ||
1710	    offset >= ctx->pool_offset + ctx->valid_pages * BLOCK_SECTORS) {
1711		ret = r5l_recovery_fetch_ra_pool(log, ctx, offset);
1712		if (ret)
1713			return ret;
1714	}
1715
1716	BUG_ON(offset < ctx->pool_offset ||
1717	       offset >= ctx->pool_offset + ctx->valid_pages * BLOCK_SECTORS);
1718
1719	memcpy(page_address(page),
1720	       page_address(ctx->ra_pool[(offset - ctx->pool_offset) >>
1721					 BLOCK_SECTOR_SHIFT]),
1722	       PAGE_SIZE);
1723	return 0;
1724}
1725
1726static int r5l_recovery_read_meta_block(struct r5l_log *log,
1727					struct r5l_recovery_ctx *ctx)
1728{
1729	struct page *page = ctx->meta_page;
1730	struct r5l_meta_block *mb;
1731	u32 crc, stored_crc;
1732	int ret;
1733
1734	ret = r5l_recovery_read_page(log, ctx, page, ctx->pos);
1735	if (ret != 0)
1736		return ret;
1737
1738	mb = page_address(page);
1739	stored_crc = le32_to_cpu(mb->checksum);
1740	mb->checksum = 0;
1741
1742	if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
1743	    le64_to_cpu(mb->seq) != ctx->seq ||
1744	    mb->version != R5LOG_VERSION ||
1745	    le64_to_cpu(mb->position) != ctx->pos)
1746		return -EINVAL;
1747
1748	crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
1749	if (stored_crc != crc)
1750		return -EINVAL;
1751
1752	if (le32_to_cpu(mb->meta_size) > PAGE_SIZE)
1753		return -EINVAL;
1754
1755	ctx->meta_total_blocks = BLOCK_SECTORS;
1756
1757	return 0;
1758}
1759
1760static void
1761r5l_recovery_create_empty_meta_block(struct r5l_log *log,
1762				     struct page *page,
1763				     sector_t pos, u64 seq)
1764{
1765	struct r5l_meta_block *mb;
1766
1767	mb = page_address(page);
1768	clear_page(mb);
1769	mb->magic = cpu_to_le32(R5LOG_MAGIC);
1770	mb->version = R5LOG_VERSION;
1771	mb->meta_size = cpu_to_le32(sizeof(struct r5l_meta_block));
1772	mb->seq = cpu_to_le64(seq);
1773	mb->position = cpu_to_le64(pos);
1774}
1775
1776static int r5l_log_write_empty_meta_block(struct r5l_log *log, sector_t pos,
1777					  u64 seq)
1778{
1779	struct page *page;
1780	struct r5l_meta_block *mb;
1781
1782	page = alloc_page(GFP_KERNEL);
1783	if (!page)
1784		return -ENOMEM;
1785	r5l_recovery_create_empty_meta_block(log, page, pos, seq);
1786	mb = page_address(page);
1787	mb->checksum = cpu_to_le32(crc32c_le(log->uuid_checksum,
1788					     mb, PAGE_SIZE));
1789	if (!sync_page_io(log->rdev, pos, PAGE_SIZE, page, REQ_OP_WRITE |
1790			  REQ_SYNC | REQ_FUA, false)) {
1791		__free_page(page);
1792		return -EIO;
1793	}
1794	__free_page(page);
1795	return 0;
1796}
1797
1798/*
1799 * r5l_recovery_load_data and r5l_recovery_load_parity uses flag R5_Wantwrite
1800 * to mark valid (potentially not flushed) data in the journal.
1801 *
1802 * We already verified checksum in r5l_recovery_verify_data_checksum_for_mb,
1803 * so there should not be any mismatch here.
1804 */
1805static void r5l_recovery_load_data(struct r5l_log *log,
1806				   struct stripe_head *sh,
1807				   struct r5l_recovery_ctx *ctx,
1808				   struct r5l_payload_data_parity *payload,
1809				   sector_t log_offset)
1810{
1811	struct mddev *mddev = log->rdev->mddev;
1812	struct r5conf *conf = mddev->private;
1813	int dd_idx;
1814
1815	raid5_compute_sector(conf,
1816			     le64_to_cpu(payload->location), 0,
1817			     &dd_idx, sh);
1818	r5l_recovery_read_page(log, ctx, sh->dev[dd_idx].page, log_offset);
1819	sh->dev[dd_idx].log_checksum =
1820		le32_to_cpu(payload->checksum[0]);
1821	ctx->meta_total_blocks += BLOCK_SECTORS;
1822
1823	set_bit(R5_Wantwrite, &sh->dev[dd_idx].flags);
1824	set_bit(STRIPE_R5C_CACHING, &sh->state);
1825}
1826
1827static void r5l_recovery_load_parity(struct r5l_log *log,
1828				     struct stripe_head *sh,
1829				     struct r5l_recovery_ctx *ctx,
1830				     struct r5l_payload_data_parity *payload,
1831				     sector_t log_offset)
1832{
1833	struct mddev *mddev = log->rdev->mddev;
1834	struct r5conf *conf = mddev->private;
1835
1836	ctx->meta_total_blocks += BLOCK_SECTORS * conf->max_degraded;
1837	r5l_recovery_read_page(log, ctx, sh->dev[sh->pd_idx].page, log_offset);
1838	sh->dev[sh->pd_idx].log_checksum =
1839		le32_to_cpu(payload->checksum[0]);
1840	set_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags);
1841
1842	if (sh->qd_idx >= 0) {
1843		r5l_recovery_read_page(
1844			log, ctx, sh->dev[sh->qd_idx].page,
1845			r5l_ring_add(log, log_offset, BLOCK_SECTORS));
1846		sh->dev[sh->qd_idx].log_checksum =
1847			le32_to_cpu(payload->checksum[1]);
1848		set_bit(R5_Wantwrite, &sh->dev[sh->qd_idx].flags);
1849	}
1850	clear_bit(STRIPE_R5C_CACHING, &sh->state);
1851}
1852
1853static void r5l_recovery_reset_stripe(struct stripe_head *sh)
1854{
1855	int i;
1856
1857	sh->state = 0;
1858	sh->log_start = MaxSector;
1859	for (i = sh->disks; i--; )
1860		sh->dev[i].flags = 0;
1861}
1862
1863static void
1864r5l_recovery_replay_one_stripe(struct r5conf *conf,
1865			       struct stripe_head *sh,
1866			       struct r5l_recovery_ctx *ctx)
1867{
1868	struct md_rdev *rdev, *rrdev;
1869	int disk_index;
1870	int data_count = 0;
1871
1872	for (disk_index = 0; disk_index < sh->disks; disk_index++) {
1873		if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
1874			continue;
1875		if (disk_index == sh->qd_idx || disk_index == sh->pd_idx)
1876			continue;
1877		data_count++;
1878	}
1879
1880	/*
1881	 * stripes that only have parity must have been flushed
1882	 * before the crash that we are now recovering from, so
1883	 * there is nothing more to recovery.
1884	 */
1885	if (data_count == 0)
1886		goto out;
1887
1888	for (disk_index = 0; disk_index < sh->disks; disk_index++) {
1889		if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
1890			continue;
1891
1892		/* in case device is broken */
1893		rdev = conf->disks[disk_index].rdev;
 
1894		if (rdev) {
1895			atomic_inc(&rdev->nr_pending);
 
1896			sync_page_io(rdev, sh->sector, PAGE_SIZE,
1897				     sh->dev[disk_index].page, REQ_OP_WRITE,
1898				     false);
1899			rdev_dec_pending(rdev, rdev->mddev);
 
1900		}
1901		rrdev = conf->disks[disk_index].replacement;
1902		if (rrdev) {
1903			atomic_inc(&rrdev->nr_pending);
 
1904			sync_page_io(rrdev, sh->sector, PAGE_SIZE,
1905				     sh->dev[disk_index].page, REQ_OP_WRITE,
1906				     false);
1907			rdev_dec_pending(rrdev, rrdev->mddev);
 
1908		}
 
1909	}
1910	ctx->data_parity_stripes++;
1911out:
1912	r5l_recovery_reset_stripe(sh);
1913}
1914
1915static struct stripe_head *
1916r5c_recovery_alloc_stripe(
1917		struct r5conf *conf,
1918		sector_t stripe_sect,
1919		int noblock)
1920{
1921	struct stripe_head *sh;
1922
1923	sh = raid5_get_active_stripe(conf, NULL, stripe_sect,
1924				     noblock ? R5_GAS_NOBLOCK : 0);
1925	if (!sh)
1926		return NULL;  /* no more stripe available */
1927
1928	r5l_recovery_reset_stripe(sh);
1929
1930	return sh;
1931}
1932
1933static struct stripe_head *
1934r5c_recovery_lookup_stripe(struct list_head *list, sector_t sect)
1935{
1936	struct stripe_head *sh;
1937
1938	list_for_each_entry(sh, list, lru)
1939		if (sh->sector == sect)
1940			return sh;
1941	return NULL;
1942}
1943
1944static void
1945r5c_recovery_drop_stripes(struct list_head *cached_stripe_list,
1946			  struct r5l_recovery_ctx *ctx)
1947{
1948	struct stripe_head *sh, *next;
1949
1950	list_for_each_entry_safe(sh, next, cached_stripe_list, lru) {
1951		r5l_recovery_reset_stripe(sh);
1952		list_del_init(&sh->lru);
1953		raid5_release_stripe(sh);
1954	}
1955}
1956
1957static void
1958r5c_recovery_replay_stripes(struct list_head *cached_stripe_list,
1959			    struct r5l_recovery_ctx *ctx)
1960{
1961	struct stripe_head *sh, *next;
1962
1963	list_for_each_entry_safe(sh, next, cached_stripe_list, lru)
1964		if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) {
1965			r5l_recovery_replay_one_stripe(sh->raid_conf, sh, ctx);
1966			list_del_init(&sh->lru);
1967			raid5_release_stripe(sh);
1968		}
1969}
1970
1971/* if matches return 0; otherwise return -EINVAL */
1972static int
1973r5l_recovery_verify_data_checksum(struct r5l_log *log,
1974				  struct r5l_recovery_ctx *ctx,
1975				  struct page *page,
1976				  sector_t log_offset, __le32 log_checksum)
1977{
1978	void *addr;
1979	u32 checksum;
1980
1981	r5l_recovery_read_page(log, ctx, page, log_offset);
1982	addr = kmap_atomic(page);
1983	checksum = crc32c_le(log->uuid_checksum, addr, PAGE_SIZE);
1984	kunmap_atomic(addr);
1985	return (le32_to_cpu(log_checksum) == checksum) ? 0 : -EINVAL;
1986}
1987
1988/*
1989 * before loading data to stripe cache, we need verify checksum for all data,
1990 * if there is mismatch for any data page, we drop all data in the mata block
1991 */
1992static int
1993r5l_recovery_verify_data_checksum_for_mb(struct r5l_log *log,
1994					 struct r5l_recovery_ctx *ctx)
1995{
1996	struct mddev *mddev = log->rdev->mddev;
1997	struct r5conf *conf = mddev->private;
1998	struct r5l_meta_block *mb = page_address(ctx->meta_page);
1999	sector_t mb_offset = sizeof(struct r5l_meta_block);
2000	sector_t log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2001	struct page *page;
2002	struct r5l_payload_data_parity *payload;
2003	struct r5l_payload_flush *payload_flush;
2004
2005	page = alloc_page(GFP_KERNEL);
2006	if (!page)
2007		return -ENOMEM;
2008
2009	while (mb_offset < le32_to_cpu(mb->meta_size)) {
2010		payload = (void *)mb + mb_offset;
2011		payload_flush = (void *)mb + mb_offset;
2012
2013		if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) {
2014			if (r5l_recovery_verify_data_checksum(
2015				    log, ctx, page, log_offset,
2016				    payload->checksum[0]) < 0)
2017				goto mismatch;
2018		} else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_PARITY) {
2019			if (r5l_recovery_verify_data_checksum(
2020				    log, ctx, page, log_offset,
2021				    payload->checksum[0]) < 0)
2022				goto mismatch;
2023			if (conf->max_degraded == 2 && /* q for RAID 6 */
2024			    r5l_recovery_verify_data_checksum(
2025				    log, ctx, page,
2026				    r5l_ring_add(log, log_offset,
2027						 BLOCK_SECTORS),
2028				    payload->checksum[1]) < 0)
2029				goto mismatch;
2030		} else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
2031			/* nothing to do for R5LOG_PAYLOAD_FLUSH here */
2032		} else /* not R5LOG_PAYLOAD_DATA/PARITY/FLUSH */
2033			goto mismatch;
2034
2035		if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
2036			mb_offset += sizeof(struct r5l_payload_flush) +
2037				le32_to_cpu(payload_flush->size);
2038		} else {
2039			/* DATA or PARITY payload */
2040			log_offset = r5l_ring_add(log, log_offset,
2041						  le32_to_cpu(payload->size));
2042			mb_offset += sizeof(struct r5l_payload_data_parity) +
2043				sizeof(__le32) *
2044				(le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
2045		}
2046
2047	}
2048
2049	put_page(page);
2050	return 0;
2051
2052mismatch:
2053	put_page(page);
2054	return -EINVAL;
2055}
2056
2057/*
2058 * Analyze all data/parity pages in one meta block
2059 * Returns:
2060 * 0 for success
2061 * -EINVAL for unknown playload type
2062 * -EAGAIN for checksum mismatch of data page
2063 * -ENOMEM for run out of memory (alloc_page failed or run out of stripes)
2064 */
2065static int
2066r5c_recovery_analyze_meta_block(struct r5l_log *log,
2067				struct r5l_recovery_ctx *ctx,
2068				struct list_head *cached_stripe_list)
2069{
2070	struct mddev *mddev = log->rdev->mddev;
2071	struct r5conf *conf = mddev->private;
2072	struct r5l_meta_block *mb;
2073	struct r5l_payload_data_parity *payload;
2074	struct r5l_payload_flush *payload_flush;
2075	int mb_offset;
2076	sector_t log_offset;
2077	sector_t stripe_sect;
2078	struct stripe_head *sh;
2079	int ret;
2080
2081	/*
2082	 * for mismatch in data blocks, we will drop all data in this mb, but
2083	 * we will still read next mb for other data with FLUSH flag, as
2084	 * io_unit could finish out of order.
2085	 */
2086	ret = r5l_recovery_verify_data_checksum_for_mb(log, ctx);
2087	if (ret == -EINVAL)
2088		return -EAGAIN;
2089	else if (ret)
2090		return ret;   /* -ENOMEM duo to alloc_page() failed */
2091
2092	mb = page_address(ctx->meta_page);
2093	mb_offset = sizeof(struct r5l_meta_block);
2094	log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2095
2096	while (mb_offset < le32_to_cpu(mb->meta_size)) {
2097		int dd;
2098
2099		payload = (void *)mb + mb_offset;
2100		payload_flush = (void *)mb + mb_offset;
2101
2102		if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
2103			int i, count;
2104
2105			count = le32_to_cpu(payload_flush->size) / sizeof(__le64);
2106			for (i = 0; i < count; ++i) {
2107				stripe_sect = le64_to_cpu(payload_flush->flush_stripes[i]);
2108				sh = r5c_recovery_lookup_stripe(cached_stripe_list,
2109								stripe_sect);
2110				if (sh) {
2111					WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
2112					r5l_recovery_reset_stripe(sh);
2113					list_del_init(&sh->lru);
2114					raid5_release_stripe(sh);
2115				}
2116			}
2117
2118			mb_offset += sizeof(struct r5l_payload_flush) +
2119				le32_to_cpu(payload_flush->size);
2120			continue;
2121		}
2122
2123		/* DATA or PARITY payload */
2124		stripe_sect = (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) ?
2125			raid5_compute_sector(
2126				conf, le64_to_cpu(payload->location), 0, &dd,
2127				NULL)
2128			: le64_to_cpu(payload->location);
2129
2130		sh = r5c_recovery_lookup_stripe(cached_stripe_list,
2131						stripe_sect);
2132
2133		if (!sh) {
2134			sh = r5c_recovery_alloc_stripe(conf, stripe_sect, 1);
2135			/*
2136			 * cannot get stripe from raid5_get_active_stripe
2137			 * try replay some stripes
2138			 */
2139			if (!sh) {
2140				r5c_recovery_replay_stripes(
2141					cached_stripe_list, ctx);
2142				sh = r5c_recovery_alloc_stripe(
2143					conf, stripe_sect, 1);
2144			}
2145			if (!sh) {
2146				int new_size = conf->min_nr_stripes * 2;
2147				pr_debug("md/raid:%s: Increasing stripe cache size to %d to recovery data on journal.\n",
2148					mdname(mddev),
2149					new_size);
2150				ret = raid5_set_cache_size(mddev, new_size);
2151				if (conf->min_nr_stripes <= new_size / 2) {
2152					pr_err("md/raid:%s: Cannot increase cache size, ret=%d, new_size=%d, min_nr_stripes=%d, max_nr_stripes=%d\n",
2153						mdname(mddev),
2154						ret,
2155						new_size,
2156						conf->min_nr_stripes,
2157						conf->max_nr_stripes);
2158					return -ENOMEM;
2159				}
2160				sh = r5c_recovery_alloc_stripe(
2161					conf, stripe_sect, 0);
2162			}
2163			if (!sh) {
2164				pr_err("md/raid:%s: Cannot get enough stripes due to memory pressure. Recovery failed.\n",
2165					mdname(mddev));
2166				return -ENOMEM;
2167			}
2168			list_add_tail(&sh->lru, cached_stripe_list);
2169		}
2170
2171		if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) {
2172			if (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
2173			    test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags)) {
2174				r5l_recovery_replay_one_stripe(conf, sh, ctx);
2175				list_move_tail(&sh->lru, cached_stripe_list);
2176			}
2177			r5l_recovery_load_data(log, sh, ctx, payload,
2178					       log_offset);
2179		} else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_PARITY)
2180			r5l_recovery_load_parity(log, sh, ctx, payload,
2181						 log_offset);
2182		else
2183			return -EINVAL;
2184
2185		log_offset = r5l_ring_add(log, log_offset,
2186					  le32_to_cpu(payload->size));
2187
2188		mb_offset += sizeof(struct r5l_payload_data_parity) +
2189			sizeof(__le32) *
2190			(le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
2191	}
2192
2193	return 0;
2194}
2195
2196/*
2197 * Load the stripe into cache. The stripe will be written out later by
2198 * the stripe cache state machine.
2199 */
2200static void r5c_recovery_load_one_stripe(struct r5l_log *log,
2201					 struct stripe_head *sh)
2202{
2203	struct r5dev *dev;
2204	int i;
2205
2206	for (i = sh->disks; i--; ) {
2207		dev = sh->dev + i;
2208		if (test_and_clear_bit(R5_Wantwrite, &dev->flags)) {
2209			set_bit(R5_InJournal, &dev->flags);
2210			set_bit(R5_UPTODATE, &dev->flags);
2211		}
2212	}
2213}
2214
2215/*
2216 * Scan through the log for all to-be-flushed data
2217 *
2218 * For stripes with data and parity, namely Data-Parity stripe
2219 * (STRIPE_R5C_CACHING == 0), we simply replay all the writes.
2220 *
2221 * For stripes with only data, namely Data-Only stripe
2222 * (STRIPE_R5C_CACHING == 1), we load them to stripe cache state machine.
2223 *
2224 * For a stripe, if we see data after parity, we should discard all previous
2225 * data and parity for this stripe, as these data are already flushed to
2226 * the array.
2227 *
2228 * At the end of the scan, we return the new journal_tail, which points to
2229 * first data-only stripe on the journal device, or next invalid meta block.
2230 */
2231static int r5c_recovery_flush_log(struct r5l_log *log,
2232				  struct r5l_recovery_ctx *ctx)
2233{
2234	struct stripe_head *sh;
2235	int ret = 0;
2236
2237	/* scan through the log */
2238	while (1) {
2239		if (r5l_recovery_read_meta_block(log, ctx))
2240			break;
2241
2242		ret = r5c_recovery_analyze_meta_block(log, ctx,
2243						      &ctx->cached_list);
2244		/*
2245		 * -EAGAIN means mismatch in data block, in this case, we still
2246		 * try scan the next metablock
2247		 */
2248		if (ret && ret != -EAGAIN)
2249			break;   /* ret == -EINVAL or -ENOMEM */
2250		ctx->seq++;
2251		ctx->pos = r5l_ring_add(log, ctx->pos, ctx->meta_total_blocks);
2252	}
2253
2254	if (ret == -ENOMEM) {
2255		r5c_recovery_drop_stripes(&ctx->cached_list, ctx);
2256		return ret;
2257	}
2258
2259	/* replay data-parity stripes */
2260	r5c_recovery_replay_stripes(&ctx->cached_list, ctx);
2261
2262	/* load data-only stripes to stripe cache */
2263	list_for_each_entry(sh, &ctx->cached_list, lru) {
2264		WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
2265		r5c_recovery_load_one_stripe(log, sh);
2266		ctx->data_only_stripes++;
2267	}
2268
2269	return 0;
2270}
2271
2272/*
2273 * we did a recovery. Now ctx.pos points to an invalid meta block. New
2274 * log will start here. but we can't let superblock point to last valid
2275 * meta block. The log might looks like:
2276 * | meta 1| meta 2| meta 3|
2277 * meta 1 is valid, meta 2 is invalid. meta 3 could be valid. If
2278 * superblock points to meta 1, we write a new valid meta 2n.  if crash
2279 * happens again, new recovery will start from meta 1. Since meta 2n is
2280 * valid now, recovery will think meta 3 is valid, which is wrong.
2281 * The solution is we create a new meta in meta2 with its seq == meta
2282 * 1's seq + 10000 and let superblock points to meta2. The same recovery
2283 * will not think meta 3 is a valid meta, because its seq doesn't match
2284 */
2285
2286/*
2287 * Before recovery, the log looks like the following
2288 *
2289 *   ---------------------------------------------
2290 *   |           valid log        | invalid log  |
2291 *   ---------------------------------------------
2292 *   ^
2293 *   |- log->last_checkpoint
2294 *   |- log->last_cp_seq
2295 *
2296 * Now we scan through the log until we see invalid entry
2297 *
2298 *   ---------------------------------------------
2299 *   |           valid log        | invalid log  |
2300 *   ---------------------------------------------
2301 *   ^                            ^
2302 *   |- log->last_checkpoint      |- ctx->pos
2303 *   |- log->last_cp_seq          |- ctx->seq
2304 *
2305 * From this point, we need to increase seq number by 10 to avoid
2306 * confusing next recovery.
2307 *
2308 *   ---------------------------------------------
2309 *   |           valid log        | invalid log  |
2310 *   ---------------------------------------------
2311 *   ^                              ^
2312 *   |- log->last_checkpoint        |- ctx->pos+1
2313 *   |- log->last_cp_seq            |- ctx->seq+10001
2314 *
2315 * However, it is not safe to start the state machine yet, because data only
2316 * parities are not yet secured in RAID. To save these data only parities, we
2317 * rewrite them from seq+11.
2318 *
2319 *   -----------------------------------------------------------------
2320 *   |           valid log        | data only stripes | invalid log  |
2321 *   -----------------------------------------------------------------
2322 *   ^                                                ^
2323 *   |- log->last_checkpoint                          |- ctx->pos+n
2324 *   |- log->last_cp_seq                              |- ctx->seq+10000+n
2325 *
2326 * If failure happens again during this process, the recovery can safe start
2327 * again from log->last_checkpoint.
2328 *
2329 * Once data only stripes are rewritten to journal, we move log_tail
2330 *
2331 *   -----------------------------------------------------------------
2332 *   |     old log        |    data only stripes    | invalid log  |
2333 *   -----------------------------------------------------------------
2334 *                        ^                         ^
2335 *                        |- log->last_checkpoint   |- ctx->pos+n
2336 *                        |- log->last_cp_seq       |- ctx->seq+10000+n
2337 *
2338 * Then we can safely start the state machine. If failure happens from this
2339 * point on, the recovery will start from new log->last_checkpoint.
2340 */
2341static int
2342r5c_recovery_rewrite_data_only_stripes(struct r5l_log *log,
2343				       struct r5l_recovery_ctx *ctx)
2344{
2345	struct stripe_head *sh;
2346	struct mddev *mddev = log->rdev->mddev;
2347	struct page *page;
2348	sector_t next_checkpoint = MaxSector;
2349
2350	page = alloc_page(GFP_KERNEL);
2351	if (!page) {
2352		pr_err("md/raid:%s: cannot allocate memory to rewrite data only stripes\n",
2353		       mdname(mddev));
2354		return -ENOMEM;
2355	}
2356
2357	WARN_ON(list_empty(&ctx->cached_list));
2358
2359	list_for_each_entry(sh, &ctx->cached_list, lru) {
2360		struct r5l_meta_block *mb;
2361		int i;
2362		int offset;
2363		sector_t write_pos;
2364
2365		WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
2366		r5l_recovery_create_empty_meta_block(log, page,
2367						     ctx->pos, ctx->seq);
2368		mb = page_address(page);
2369		offset = le32_to_cpu(mb->meta_size);
2370		write_pos = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2371
2372		for (i = sh->disks; i--; ) {
2373			struct r5dev *dev = &sh->dev[i];
2374			struct r5l_payload_data_parity *payload;
2375			void *addr;
2376
2377			if (test_bit(R5_InJournal, &dev->flags)) {
2378				payload = (void *)mb + offset;
2379				payload->header.type = cpu_to_le16(
2380					R5LOG_PAYLOAD_DATA);
2381				payload->size = cpu_to_le32(BLOCK_SECTORS);
2382				payload->location = cpu_to_le64(
2383					raid5_compute_blocknr(sh, i, 0));
2384				addr = kmap_atomic(dev->page);
2385				payload->checksum[0] = cpu_to_le32(
2386					crc32c_le(log->uuid_checksum, addr,
2387						  PAGE_SIZE));
2388				kunmap_atomic(addr);
2389				sync_page_io(log->rdev, write_pos, PAGE_SIZE,
2390					     dev->page, REQ_OP_WRITE, false);
2391				write_pos = r5l_ring_add(log, write_pos,
2392							 BLOCK_SECTORS);
2393				offset += sizeof(__le32) +
2394					sizeof(struct r5l_payload_data_parity);
2395
2396			}
2397		}
2398		mb->meta_size = cpu_to_le32(offset);
2399		mb->checksum = cpu_to_le32(crc32c_le(log->uuid_checksum,
2400						     mb, PAGE_SIZE));
2401		sync_page_io(log->rdev, ctx->pos, PAGE_SIZE, page,
2402			     REQ_OP_WRITE | REQ_SYNC | REQ_FUA, false);
2403		sh->log_start = ctx->pos;
2404		list_add_tail(&sh->r5c, &log->stripe_in_journal_list);
2405		atomic_inc(&log->stripe_in_journal_count);
2406		ctx->pos = write_pos;
2407		ctx->seq += 1;
2408		next_checkpoint = sh->log_start;
2409	}
2410	log->next_checkpoint = next_checkpoint;
2411	__free_page(page);
2412	return 0;
2413}
2414
2415static void r5c_recovery_flush_data_only_stripes(struct r5l_log *log,
2416						 struct r5l_recovery_ctx *ctx)
2417{
2418	struct mddev *mddev = log->rdev->mddev;
2419	struct r5conf *conf = mddev->private;
2420	struct stripe_head *sh, *next;
2421	bool cleared_pending = false;
2422
2423	if (ctx->data_only_stripes == 0)
2424		return;
2425
2426	if (test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags)) {
2427		cleared_pending = true;
2428		clear_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags);
2429	}
2430	log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_BACK;
2431
2432	list_for_each_entry_safe(sh, next, &ctx->cached_list, lru) {
2433		r5c_make_stripe_write_out(sh);
2434		set_bit(STRIPE_HANDLE, &sh->state);
2435		list_del_init(&sh->lru);
2436		raid5_release_stripe(sh);
2437	}
2438
2439	/* reuse conf->wait_for_quiescent in recovery */
2440	wait_event(conf->wait_for_quiescent,
2441		   atomic_read(&conf->active_stripes) == 0);
2442
2443	log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
2444	if (cleared_pending)
2445		set_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags);
2446}
2447
2448static int r5l_recovery_log(struct r5l_log *log)
2449{
2450	struct mddev *mddev = log->rdev->mddev;
2451	struct r5l_recovery_ctx *ctx;
2452	int ret;
2453	sector_t pos;
2454
2455	ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
2456	if (!ctx)
2457		return -ENOMEM;
2458
2459	ctx->pos = log->last_checkpoint;
2460	ctx->seq = log->last_cp_seq;
2461	INIT_LIST_HEAD(&ctx->cached_list);
2462	ctx->meta_page = alloc_page(GFP_KERNEL);
2463
2464	if (!ctx->meta_page) {
2465		ret =  -ENOMEM;
2466		goto meta_page;
2467	}
2468
2469	if (r5l_recovery_allocate_ra_pool(log, ctx) != 0) {
2470		ret = -ENOMEM;
2471		goto ra_pool;
2472	}
2473
2474	ret = r5c_recovery_flush_log(log, ctx);
2475
2476	if (ret)
2477		goto error;
2478
2479	pos = ctx->pos;
2480	ctx->seq += 10000;
2481
2482	if ((ctx->data_only_stripes == 0) && (ctx->data_parity_stripes == 0))
2483		pr_info("md/raid:%s: starting from clean shutdown\n",
2484			 mdname(mddev));
2485	else
2486		pr_info("md/raid:%s: recovering %d data-only stripes and %d data-parity stripes\n",
2487			 mdname(mddev), ctx->data_only_stripes,
2488			 ctx->data_parity_stripes);
2489
2490	if (ctx->data_only_stripes == 0) {
2491		log->next_checkpoint = ctx->pos;
2492		r5l_log_write_empty_meta_block(log, ctx->pos, ctx->seq++);
2493		ctx->pos = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2494	} else if (r5c_recovery_rewrite_data_only_stripes(log, ctx)) {
2495		pr_err("md/raid:%s: failed to rewrite stripes to journal\n",
2496		       mdname(mddev));
2497		ret =  -EIO;
2498		goto error;
2499	}
2500
2501	log->log_start = ctx->pos;
2502	log->seq = ctx->seq;
2503	log->last_checkpoint = pos;
2504	r5l_write_super(log, pos);
2505
2506	r5c_recovery_flush_data_only_stripes(log, ctx);
2507	ret = 0;
2508error:
2509	r5l_recovery_free_ra_pool(log, ctx);
2510ra_pool:
2511	__free_page(ctx->meta_page);
2512meta_page:
2513	kfree(ctx);
2514	return ret;
2515}
2516
2517static void r5l_write_super(struct r5l_log *log, sector_t cp)
2518{
2519	struct mddev *mddev = log->rdev->mddev;
2520
2521	log->rdev->journal_tail = cp;
2522	set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2523}
2524
2525static ssize_t r5c_journal_mode_show(struct mddev *mddev, char *page)
2526{
2527	struct r5conf *conf;
2528	int ret;
2529
2530	ret = mddev_lock(mddev);
2531	if (ret)
2532		return ret;
2533
2534	conf = mddev->private;
2535	if (!conf || !conf->log)
2536		goto out_unlock;
 
 
2537
2538	switch (conf->log->r5c_journal_mode) {
2539	case R5C_JOURNAL_MODE_WRITE_THROUGH:
2540		ret = snprintf(
2541			page, PAGE_SIZE, "[%s] %s\n",
2542			r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH],
2543			r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]);
2544		break;
2545	case R5C_JOURNAL_MODE_WRITE_BACK:
2546		ret = snprintf(
2547			page, PAGE_SIZE, "%s [%s]\n",
2548			r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH],
2549			r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]);
2550		break;
2551	default:
2552		ret = 0;
2553	}
2554
2555out_unlock:
2556	mddev_unlock(mddev);
2557	return ret;
2558}
2559
2560/*
2561 * Set journal cache mode on @mddev (external API initially needed by dm-raid).
2562 *
2563 * @mode as defined in 'enum r5c_journal_mode'.
2564 *
2565 */
2566int r5c_journal_mode_set(struct mddev *mddev, int mode)
2567{
2568	struct r5conf *conf;
2569
2570	if (mode < R5C_JOURNAL_MODE_WRITE_THROUGH ||
2571	    mode > R5C_JOURNAL_MODE_WRITE_BACK)
2572		return -EINVAL;
2573
2574	conf = mddev->private;
2575	if (!conf || !conf->log)
2576		return -ENODEV;
2577
2578	if (raid5_calc_degraded(conf) > 0 &&
2579	    mode == R5C_JOURNAL_MODE_WRITE_BACK)
2580		return -EINVAL;
2581
 
2582	conf->log->r5c_journal_mode = mode;
 
2583
2584	pr_debug("md/raid:%s: setting r5c cache mode to %d: %s\n",
2585		 mdname(mddev), mode, r5c_journal_mode_str[mode]);
2586	return 0;
2587}
2588EXPORT_SYMBOL(r5c_journal_mode_set);
2589
2590static ssize_t r5c_journal_mode_store(struct mddev *mddev,
2591				      const char *page, size_t length)
2592{
2593	int mode = ARRAY_SIZE(r5c_journal_mode_str);
2594	size_t len = length;
2595	int ret;
2596
2597	if (len < 2)
2598		return -EINVAL;
2599
2600	if (page[len - 1] == '\n')
2601		len--;
2602
2603	while (mode--)
2604		if (strlen(r5c_journal_mode_str[mode]) == len &&
2605		    !strncmp(page, r5c_journal_mode_str[mode], len))
2606			break;
2607	ret = mddev_suspend_and_lock(mddev);
2608	if (ret)
2609		return ret;
2610	ret = r5c_journal_mode_set(mddev, mode);
2611	mddev_unlock_and_resume(mddev);
2612	return ret ?: length;
2613}
2614
2615struct md_sysfs_entry
2616r5c_journal_mode = __ATTR(journal_mode, 0644,
2617			  r5c_journal_mode_show, r5c_journal_mode_store);
2618
2619/*
2620 * Try handle write operation in caching phase. This function should only
2621 * be called in write-back mode.
2622 *
2623 * If all outstanding writes can be handled in caching phase, returns 0
2624 * If writes requires write-out phase, call r5c_make_stripe_write_out()
2625 * and returns -EAGAIN
2626 */
2627int r5c_try_caching_write(struct r5conf *conf,
2628			  struct stripe_head *sh,
2629			  struct stripe_head_state *s,
2630			  int disks)
2631{
2632	struct r5l_log *log = READ_ONCE(conf->log);
2633	int i;
2634	struct r5dev *dev;
2635	int to_cache = 0;
2636	void __rcu **pslot;
2637	sector_t tree_index;
2638	int ret;
2639	uintptr_t refcount;
2640
2641	BUG_ON(!r5c_is_writeback(log));
2642
2643	if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) {
2644		/*
2645		 * There are two different scenarios here:
2646		 *  1. The stripe has some data cached, and it is sent to
2647		 *     write-out phase for reclaim
2648		 *  2. The stripe is clean, and this is the first write
2649		 *
2650		 * For 1, return -EAGAIN, so we continue with
2651		 * handle_stripe_dirtying().
2652		 *
2653		 * For 2, set STRIPE_R5C_CACHING and continue with caching
2654		 * write.
2655		 */
2656
2657		/* case 1: anything injournal or anything in written */
2658		if (s->injournal > 0 || s->written > 0)
2659			return -EAGAIN;
2660		/* case 2 */
2661		set_bit(STRIPE_R5C_CACHING, &sh->state);
2662	}
2663
2664	/*
2665	 * When run in degraded mode, array is set to write-through mode.
2666	 * This check helps drain pending write safely in the transition to
2667	 * write-through mode.
2668	 *
2669	 * When a stripe is syncing, the write is also handled in write
2670	 * through mode.
2671	 */
2672	if (s->failed || test_bit(STRIPE_SYNCING, &sh->state)) {
2673		r5c_make_stripe_write_out(sh);
2674		return -EAGAIN;
2675	}
2676
2677	for (i = disks; i--; ) {
2678		dev = &sh->dev[i];
2679		/* if non-overwrite, use writing-out phase */
2680		if (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags) &&
2681		    !test_bit(R5_InJournal, &dev->flags)) {
2682			r5c_make_stripe_write_out(sh);
2683			return -EAGAIN;
2684		}
2685	}
2686
2687	/* if the stripe is not counted in big_stripe_tree, add it now */
2688	if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
2689	    !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2690		tree_index = r5c_tree_index(conf, sh->sector);
2691		spin_lock(&log->tree_lock);
2692		pslot = radix_tree_lookup_slot(&log->big_stripe_tree,
2693					       tree_index);
2694		if (pslot) {
2695			refcount = (uintptr_t)radix_tree_deref_slot_protected(
2696				pslot, &log->tree_lock) >>
2697				R5C_RADIX_COUNT_SHIFT;
2698			radix_tree_replace_slot(
2699				&log->big_stripe_tree, pslot,
2700				(void *)((refcount + 1) << R5C_RADIX_COUNT_SHIFT));
2701		} else {
2702			/*
2703			 * this radix_tree_insert can fail safely, so no
2704			 * need to call radix_tree_preload()
2705			 */
2706			ret = radix_tree_insert(
2707				&log->big_stripe_tree, tree_index,
2708				(void *)(1 << R5C_RADIX_COUNT_SHIFT));
2709			if (ret) {
2710				spin_unlock(&log->tree_lock);
2711				r5c_make_stripe_write_out(sh);
2712				return -EAGAIN;
2713			}
2714		}
2715		spin_unlock(&log->tree_lock);
2716
2717		/*
2718		 * set STRIPE_R5C_PARTIAL_STRIPE, this shows the stripe is
2719		 * counted in the radix tree
2720		 */
2721		set_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state);
2722		atomic_inc(&conf->r5c_cached_partial_stripes);
2723	}
2724
2725	for (i = disks; i--; ) {
2726		dev = &sh->dev[i];
2727		if (dev->towrite) {
2728			set_bit(R5_Wantwrite, &dev->flags);
2729			set_bit(R5_Wantdrain, &dev->flags);
2730			set_bit(R5_LOCKED, &dev->flags);
2731			to_cache++;
2732		}
2733	}
2734
2735	if (to_cache) {
2736		set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2737		/*
2738		 * set STRIPE_LOG_TRAPPED, which triggers r5c_cache_data()
2739		 * in ops_run_io(). STRIPE_LOG_TRAPPED will be cleared in
2740		 * r5c_handle_data_cached()
2741		 */
2742		set_bit(STRIPE_LOG_TRAPPED, &sh->state);
2743	}
2744
2745	return 0;
2746}
2747
2748/*
2749 * free extra pages (orig_page) we allocated for prexor
2750 */
2751void r5c_release_extra_page(struct stripe_head *sh)
2752{
2753	struct r5conf *conf = sh->raid_conf;
2754	int i;
2755	bool using_disk_info_extra_page;
2756
2757	using_disk_info_extra_page =
2758		sh->dev[0].orig_page == conf->disks[0].extra_page;
2759
2760	for (i = sh->disks; i--; )
2761		if (sh->dev[i].page != sh->dev[i].orig_page) {
2762			struct page *p = sh->dev[i].orig_page;
2763
2764			sh->dev[i].orig_page = sh->dev[i].page;
2765			clear_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
2766
2767			if (!using_disk_info_extra_page)
2768				put_page(p);
2769		}
2770
2771	if (using_disk_info_extra_page) {
2772		clear_bit(R5C_EXTRA_PAGE_IN_USE, &conf->cache_state);
2773		md_wakeup_thread(conf->mddev->thread);
2774	}
2775}
2776
2777void r5c_use_extra_page(struct stripe_head *sh)
2778{
2779	struct r5conf *conf = sh->raid_conf;
2780	int i;
2781	struct r5dev *dev;
2782
2783	for (i = sh->disks; i--; ) {
2784		dev = &sh->dev[i];
2785		if (dev->orig_page != dev->page)
2786			put_page(dev->orig_page);
2787		dev->orig_page = conf->disks[i].extra_page;
2788	}
2789}
2790
2791/*
2792 * clean up the stripe (clear R5_InJournal for dev[pd_idx] etc.) after the
2793 * stripe is committed to RAID disks.
2794 */
2795void r5c_finish_stripe_write_out(struct r5conf *conf,
2796				 struct stripe_head *sh,
2797				 struct stripe_head_state *s)
2798{
2799	struct r5l_log *log = READ_ONCE(conf->log);
2800	int i;
2801	int do_wakeup = 0;
2802	sector_t tree_index;
2803	void __rcu **pslot;
2804	uintptr_t refcount;
2805
2806	if (!log || !test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags))
2807		return;
2808
2809	WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
2810	clear_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
2811
2812	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
2813		return;
2814
2815	for (i = sh->disks; i--; ) {
2816		clear_bit(R5_InJournal, &sh->dev[i].flags);
2817		if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2818			do_wakeup = 1;
2819	}
2820
2821	/*
2822	 * analyse_stripe() runs before r5c_finish_stripe_write_out(),
2823	 * We updated R5_InJournal, so we also update s->injournal.
2824	 */
2825	s->injournal = 0;
2826
2827	if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2828		if (atomic_dec_and_test(&conf->pending_full_writes))
2829			md_wakeup_thread(conf->mddev->thread);
2830
2831	if (do_wakeup)
2832		wake_up(&conf->wait_for_overlap);
2833
2834	spin_lock_irq(&log->stripe_in_journal_lock);
2835	list_del_init(&sh->r5c);
2836	spin_unlock_irq(&log->stripe_in_journal_lock);
2837	sh->log_start = MaxSector;
2838
2839	atomic_dec(&log->stripe_in_journal_count);
2840	r5c_update_log_state(log);
2841
2842	/* stop counting this stripe in big_stripe_tree */
2843	if (test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) ||
2844	    test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2845		tree_index = r5c_tree_index(conf, sh->sector);
2846		spin_lock(&log->tree_lock);
2847		pslot = radix_tree_lookup_slot(&log->big_stripe_tree,
2848					       tree_index);
2849		BUG_ON(pslot == NULL);
2850		refcount = (uintptr_t)radix_tree_deref_slot_protected(
2851			pslot, &log->tree_lock) >>
2852			R5C_RADIX_COUNT_SHIFT;
2853		if (refcount == 1)
2854			radix_tree_delete(&log->big_stripe_tree, tree_index);
2855		else
2856			radix_tree_replace_slot(
2857				&log->big_stripe_tree, pslot,
2858				(void *)((refcount - 1) << R5C_RADIX_COUNT_SHIFT));
2859		spin_unlock(&log->tree_lock);
2860	}
2861
2862	if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) {
2863		BUG_ON(atomic_read(&conf->r5c_cached_partial_stripes) == 0);
2864		atomic_dec(&conf->r5c_flushing_partial_stripes);
2865		atomic_dec(&conf->r5c_cached_partial_stripes);
2866	}
2867
2868	if (test_and_clear_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2869		BUG_ON(atomic_read(&conf->r5c_cached_full_stripes) == 0);
2870		atomic_dec(&conf->r5c_flushing_full_stripes);
2871		atomic_dec(&conf->r5c_cached_full_stripes);
2872	}
2873
2874	r5l_append_flush_payload(log, sh->sector);
2875	/* stripe is flused to raid disks, we can do resync now */
2876	if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
2877		set_bit(STRIPE_HANDLE, &sh->state);
2878}
2879
2880int r5c_cache_data(struct r5l_log *log, struct stripe_head *sh)
2881{
2882	struct r5conf *conf = sh->raid_conf;
2883	int pages = 0;
2884	int reserve;
2885	int i;
2886	int ret = 0;
2887
2888	BUG_ON(!log);
2889
2890	for (i = 0; i < sh->disks; i++) {
2891		void *addr;
2892
2893		if (!test_bit(R5_Wantwrite, &sh->dev[i].flags))
2894			continue;
2895		addr = kmap_atomic(sh->dev[i].page);
2896		sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum,
2897						    addr, PAGE_SIZE);
2898		kunmap_atomic(addr);
2899		pages++;
2900	}
2901	WARN_ON(pages == 0);
2902
2903	/*
2904	 * The stripe must enter state machine again to call endio, so
2905	 * don't delay.
2906	 */
2907	clear_bit(STRIPE_DELAYED, &sh->state);
2908	atomic_inc(&sh->count);
2909
2910	mutex_lock(&log->io_mutex);
2911	/* meta + data */
2912	reserve = (1 + pages) << (PAGE_SHIFT - 9);
2913
2914	if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
2915	    sh->log_start == MaxSector)
2916		r5l_add_no_space_stripe(log, sh);
2917	else if (!r5l_has_free_space(log, reserve)) {
2918		if (sh->log_start == log->last_checkpoint)
2919			BUG();
2920		else
2921			r5l_add_no_space_stripe(log, sh);
2922	} else {
2923		ret = r5l_log_stripe(log, sh, pages, 0);
2924		if (ret) {
2925			spin_lock_irq(&log->io_list_lock);
2926			list_add_tail(&sh->log_list, &log->no_mem_stripes);
2927			spin_unlock_irq(&log->io_list_lock);
2928		}
2929	}
2930
2931	mutex_unlock(&log->io_mutex);
2932	return 0;
2933}
2934
2935/* check whether this big stripe is in write back cache. */
2936bool r5c_big_stripe_cached(struct r5conf *conf, sector_t sect)
2937{
2938	struct r5l_log *log = READ_ONCE(conf->log);
2939	sector_t tree_index;
2940	void *slot;
2941
2942	if (!log)
2943		return false;
2944
 
2945	tree_index = r5c_tree_index(conf, sect);
2946	slot = radix_tree_lookup(&log->big_stripe_tree, tree_index);
2947	return slot != NULL;
2948}
2949
2950static int r5l_load_log(struct r5l_log *log)
2951{
2952	struct md_rdev *rdev = log->rdev;
2953	struct page *page;
2954	struct r5l_meta_block *mb;
2955	sector_t cp = log->rdev->journal_tail;
2956	u32 stored_crc, expected_crc;
2957	bool create_super = false;
2958	int ret = 0;
2959
2960	/* Make sure it's valid */
2961	if (cp >= rdev->sectors || round_down(cp, BLOCK_SECTORS) != cp)
2962		cp = 0;
2963	page = alloc_page(GFP_KERNEL);
2964	if (!page)
2965		return -ENOMEM;
2966
2967	if (!sync_page_io(rdev, cp, PAGE_SIZE, page, REQ_OP_READ, false)) {
2968		ret = -EIO;
2969		goto ioerr;
2970	}
2971	mb = page_address(page);
2972
2973	if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
2974	    mb->version != R5LOG_VERSION) {
2975		create_super = true;
2976		goto create;
2977	}
2978	stored_crc = le32_to_cpu(mb->checksum);
2979	mb->checksum = 0;
2980	expected_crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
2981	if (stored_crc != expected_crc) {
2982		create_super = true;
2983		goto create;
2984	}
2985	if (le64_to_cpu(mb->position) != cp) {
2986		create_super = true;
2987		goto create;
2988	}
2989create:
2990	if (create_super) {
2991		log->last_cp_seq = get_random_u32();
2992		cp = 0;
2993		r5l_log_write_empty_meta_block(log, cp, log->last_cp_seq);
2994		/*
2995		 * Make sure super points to correct address. Log might have
2996		 * data very soon. If super hasn't correct log tail address,
2997		 * recovery can't find the log
2998		 */
2999		r5l_write_super(log, cp);
3000	} else
3001		log->last_cp_seq = le64_to_cpu(mb->seq);
3002
3003	log->device_size = round_down(rdev->sectors, BLOCK_SECTORS);
3004	log->max_free_space = log->device_size >> RECLAIM_MAX_FREE_SPACE_SHIFT;
3005	if (log->max_free_space > RECLAIM_MAX_FREE_SPACE)
3006		log->max_free_space = RECLAIM_MAX_FREE_SPACE;
3007	log->last_checkpoint = cp;
3008
3009	__free_page(page);
3010
3011	if (create_super) {
3012		log->log_start = r5l_ring_add(log, cp, BLOCK_SECTORS);
3013		log->seq = log->last_cp_seq + 1;
3014		log->next_checkpoint = cp;
3015	} else
3016		ret = r5l_recovery_log(log);
3017
3018	r5c_update_log_state(log);
3019	return ret;
3020ioerr:
3021	__free_page(page);
3022	return ret;
3023}
3024
3025int r5l_start(struct r5l_log *log)
3026{
3027	int ret;
3028
3029	if (!log)
3030		return 0;
3031
3032	ret = r5l_load_log(log);
3033	if (ret) {
3034		struct mddev *mddev = log->rdev->mddev;
3035		struct r5conf *conf = mddev->private;
3036
3037		r5l_exit_log(conf);
3038	}
3039	return ret;
3040}
3041
3042void r5c_update_on_rdev_error(struct mddev *mddev, struct md_rdev *rdev)
3043{
3044	struct r5conf *conf = mddev->private;
3045	struct r5l_log *log = READ_ONCE(conf->log);
3046
3047	if (!log)
3048		return;
3049
3050	if ((raid5_calc_degraded(conf) > 0 ||
3051	     test_bit(Journal, &rdev->flags)) &&
3052	    log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK)
3053		schedule_work(&log->disable_writeback_work);
3054}
3055
3056int r5l_init_log(struct r5conf *conf, struct md_rdev *rdev)
3057{
 
3058	struct r5l_log *log;
3059	struct md_thread *thread;
3060	int ret;
3061
3062	pr_debug("md/raid:%s: using device %pg as journal\n",
3063		 mdname(conf->mddev), rdev->bdev);
3064
3065	if (PAGE_SIZE != 4096)
3066		return -EINVAL;
3067
3068	/*
3069	 * The PAGE_SIZE must be big enough to hold 1 r5l_meta_block and
3070	 * raid_disks r5l_payload_data_parity.
3071	 *
3072	 * Write journal and cache does not work for very big array
3073	 * (raid_disks > 203)
3074	 */
3075	if (sizeof(struct r5l_meta_block) +
3076	    ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32)) *
3077	     conf->raid_disks) > PAGE_SIZE) {
3078		pr_err("md/raid:%s: write journal/cache doesn't work for array with %d disks\n",
3079		       mdname(conf->mddev), conf->raid_disks);
3080		return -EINVAL;
3081	}
3082
3083	log = kzalloc(sizeof(*log), GFP_KERNEL);
3084	if (!log)
3085		return -ENOMEM;
3086	log->rdev = rdev;
3087	log->need_cache_flush = bdev_write_cache(rdev->bdev);
 
 
3088	log->uuid_checksum = crc32c_le(~0, rdev->mddev->uuid,
3089				       sizeof(rdev->mddev->uuid));
3090
3091	mutex_init(&log->io_mutex);
3092
3093	spin_lock_init(&log->io_list_lock);
3094	INIT_LIST_HEAD(&log->running_ios);
3095	INIT_LIST_HEAD(&log->io_end_ios);
3096	INIT_LIST_HEAD(&log->flushing_ios);
3097	INIT_LIST_HEAD(&log->finished_ios);
 
3098
3099	log->io_kc = KMEM_CACHE(r5l_io_unit, 0);
3100	if (!log->io_kc)
3101		goto io_kc;
3102
3103	ret = mempool_init_slab_pool(&log->io_pool, R5L_POOL_SIZE, log->io_kc);
3104	if (ret)
3105		goto io_pool;
3106
3107	ret = bioset_init(&log->bs, R5L_POOL_SIZE, 0, BIOSET_NEED_BVECS);
3108	if (ret)
3109		goto io_bs;
3110
3111	ret = mempool_init_page_pool(&log->meta_pool, R5L_POOL_SIZE, 0);
3112	if (ret)
3113		goto out_mempool;
3114
3115	spin_lock_init(&log->tree_lock);
3116	INIT_RADIX_TREE(&log->big_stripe_tree, GFP_NOWAIT | __GFP_NOWARN);
3117
3118	thread = md_register_thread(r5l_reclaim_thread, log->rdev->mddev,
3119				    "reclaim");
3120	if (!thread)
3121		goto reclaim_thread;
3122
3123	thread->timeout = R5C_RECLAIM_WAKEUP_INTERVAL;
3124	rcu_assign_pointer(log->reclaim_thread, thread);
3125
3126	init_waitqueue_head(&log->iounit_wait);
3127
3128	INIT_LIST_HEAD(&log->no_mem_stripes);
3129
3130	INIT_LIST_HEAD(&log->no_space_stripes);
3131	spin_lock_init(&log->no_space_stripes_lock);
3132
3133	INIT_WORK(&log->deferred_io_work, r5l_submit_io_async);
3134	INIT_WORK(&log->disable_writeback_work, r5c_disable_writeback_async);
3135
3136	log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
3137	INIT_LIST_HEAD(&log->stripe_in_journal_list);
3138	spin_lock_init(&log->stripe_in_journal_lock);
3139	atomic_set(&log->stripe_in_journal_count, 0);
3140
3141	WRITE_ONCE(conf->log, log);
3142
3143	set_bit(MD_HAS_JOURNAL, &conf->mddev->flags);
3144	return 0;
3145
 
 
3146reclaim_thread:
3147	mempool_exit(&log->meta_pool);
3148out_mempool:
3149	bioset_exit(&log->bs);
3150io_bs:
3151	mempool_exit(&log->io_pool);
3152io_pool:
3153	kmem_cache_destroy(log->io_kc);
3154io_kc:
3155	kfree(log);
3156	return -EINVAL;
3157}
3158
3159void r5l_exit_log(struct r5conf *conf)
3160{
3161	struct r5l_log *log = conf->log;
3162
3163	md_unregister_thread(conf->mddev, &log->reclaim_thread);
 
3164
3165	/*
3166	 * 'reconfig_mutex' is held by caller, set 'confg->log' to NULL to
3167	 * ensure disable_writeback_work wakes up and exits.
3168	 */
3169	WRITE_ONCE(conf->log, NULL);
3170	wake_up(&conf->mddev->sb_wait);
3171	flush_work(&log->disable_writeback_work);
3172
3173	mempool_exit(&log->meta_pool);
3174	bioset_exit(&log->bs);
3175	mempool_exit(&log->io_pool);
3176	kmem_cache_destroy(log->io_kc);
3177	kfree(log);
3178}
v4.17
 
   1/*
   2 * Copyright (C) 2015 Shaohua Li <shli@fb.com>
   3 * Copyright (C) 2016 Song Liu <songliubraving@fb.com>
   4 *
   5 * This program is free software; you can redistribute it and/or modify it
   6 * under the terms and conditions of the GNU General Public License,
   7 * version 2, as published by the Free Software Foundation.
   8 *
   9 * This program is distributed in the hope it will be useful, but WITHOUT
  10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11 * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  12 * more details.
  13 *
  14 */
  15#include <linux/kernel.h>
  16#include <linux/wait.h>
  17#include <linux/blkdev.h>
  18#include <linux/slab.h>
  19#include <linux/raid/md_p.h>
  20#include <linux/crc32c.h>
  21#include <linux/random.h>
  22#include <linux/kthread.h>
  23#include <linux/types.h>
  24#include "md.h"
  25#include "raid5.h"
  26#include "md-bitmap.h"
  27#include "raid5-log.h"
  28
  29/*
  30 * metadata/data stored in disk with 4k size unit (a block) regardless
  31 * underneath hardware sector size. only works with PAGE_SIZE == 4096
  32 */
  33#define BLOCK_SECTORS (8)
  34#define BLOCK_SECTOR_SHIFT (3)
  35
  36/*
  37 * log->max_free_space is min(1/4 disk size, 10G reclaimable space).
  38 *
  39 * In write through mode, the reclaim runs every log->max_free_space.
  40 * This can prevent the recovery scans for too long
  41 */
  42#define RECLAIM_MAX_FREE_SPACE (10 * 1024 * 1024 * 2) /* sector */
  43#define RECLAIM_MAX_FREE_SPACE_SHIFT (2)
  44
  45/* wake up reclaim thread periodically */
  46#define R5C_RECLAIM_WAKEUP_INTERVAL (30 * HZ)
  47/* start flush with these full stripes */
  48#define R5C_FULL_STRIPE_FLUSH_BATCH(conf) (conf->max_nr_stripes / 4)
  49/* reclaim stripes in groups */
  50#define R5C_RECLAIM_STRIPE_GROUP (NR_STRIPE_HASH_LOCKS * 2)
  51
  52/*
  53 * We only need 2 bios per I/O unit to make progress, but ensure we
  54 * have a few more available to not get too tight.
  55 */
  56#define R5L_POOL_SIZE	4
  57
  58static char *r5c_journal_mode_str[] = {"write-through",
  59				       "write-back"};
  60/*
  61 * raid5 cache state machine
  62 *
  63 * With the RAID cache, each stripe works in two phases:
  64 *	- caching phase
  65 *	- writing-out phase
  66 *
  67 * These two phases are controlled by bit STRIPE_R5C_CACHING:
  68 *   if STRIPE_R5C_CACHING == 0, the stripe is in writing-out phase
  69 *   if STRIPE_R5C_CACHING == 1, the stripe is in caching phase
  70 *
  71 * When there is no journal, or the journal is in write-through mode,
  72 * the stripe is always in writing-out phase.
  73 *
  74 * For write-back journal, the stripe is sent to caching phase on write
  75 * (r5c_try_caching_write). r5c_make_stripe_write_out() kicks off
  76 * the write-out phase by clearing STRIPE_R5C_CACHING.
  77 *
  78 * Stripes in caching phase do not write the raid disks. Instead, all
  79 * writes are committed from the log device. Therefore, a stripe in
  80 * caching phase handles writes as:
  81 *	- write to log device
  82 *	- return IO
  83 *
  84 * Stripes in writing-out phase handle writes as:
  85 *	- calculate parity
  86 *	- write pending data and parity to journal
  87 *	- write data and parity to raid disks
  88 *	- return IO for pending writes
  89 */
  90
  91struct r5l_log {
  92	struct md_rdev *rdev;
  93
  94	u32 uuid_checksum;
  95
  96	sector_t device_size;		/* log device size, round to
  97					 * BLOCK_SECTORS */
  98	sector_t max_free_space;	/* reclaim run if free space is at
  99					 * this size */
 100
 101	sector_t last_checkpoint;	/* log tail. where recovery scan
 102					 * starts from */
 103	u64 last_cp_seq;		/* log tail sequence */
 104
 105	sector_t log_start;		/* log head. where new data appends */
 106	u64 seq;			/* log head sequence */
 107
 108	sector_t next_checkpoint;
 109
 110	struct mutex io_mutex;
 111	struct r5l_io_unit *current_io;	/* current io_unit accepting new data */
 112
 113	spinlock_t io_list_lock;
 114	struct list_head running_ios;	/* io_units which are still running,
 115					 * and have not yet been completely
 116					 * written to the log */
 117	struct list_head io_end_ios;	/* io_units which have been completely
 118					 * written to the log but not yet written
 119					 * to the RAID */
 120	struct list_head flushing_ios;	/* io_units which are waiting for log
 121					 * cache flush */
 122	struct list_head finished_ios;	/* io_units which settle down in log disk */
 123	struct bio flush_bio;
 124
 125	struct list_head no_mem_stripes;   /* pending stripes, -ENOMEM */
 126
 127	struct kmem_cache *io_kc;
 128	mempool_t *io_pool;
 129	struct bio_set *bs;
 130	mempool_t *meta_pool;
 131
 132	struct md_thread *reclaim_thread;
 133	unsigned long reclaim_target;	/* number of space that need to be
 134					 * reclaimed.  if it's 0, reclaim spaces
 135					 * used by io_units which are in
 136					 * IO_UNIT_STRIPE_END state (eg, reclaim
 137					 * dones't wait for specific io_unit
 138					 * switching to IO_UNIT_STRIPE_END
 139					 * state) */
 140	wait_queue_head_t iounit_wait;
 141
 142	struct list_head no_space_stripes; /* pending stripes, log has no space */
 143	spinlock_t no_space_stripes_lock;
 144
 145	bool need_cache_flush;
 146
 147	/* for r5c_cache */
 148	enum r5c_journal_mode r5c_journal_mode;
 149
 150	/* all stripes in r5cache, in the order of seq at sh->log_start */
 151	struct list_head stripe_in_journal_list;
 152
 153	spinlock_t stripe_in_journal_lock;
 154	atomic_t stripe_in_journal_count;
 155
 156	/* to submit async io_units, to fulfill ordering of flush */
 157	struct work_struct deferred_io_work;
 158	/* to disable write back during in degraded mode */
 159	struct work_struct disable_writeback_work;
 160
 161	/* to for chunk_aligned_read in writeback mode, details below */
 162	spinlock_t tree_lock;
 163	struct radix_tree_root big_stripe_tree;
 164};
 165
 166/*
 167 * Enable chunk_aligned_read() with write back cache.
 168 *
 169 * Each chunk may contain more than one stripe (for example, a 256kB
 170 * chunk contains 64 4kB-page, so this chunk contain 64 stripes). For
 171 * chunk_aligned_read, these stripes are grouped into one "big_stripe".
 172 * For each big_stripe, we count how many stripes of this big_stripe
 173 * are in the write back cache. These data are tracked in a radix tree
 174 * (big_stripe_tree). We use radix_tree item pointer as the counter.
 175 * r5c_tree_index() is used to calculate keys for the radix tree.
 176 *
 177 * chunk_aligned_read() calls r5c_big_stripe_cached() to look up
 178 * big_stripe of each chunk in the tree. If this big_stripe is in the
 179 * tree, chunk_aligned_read() aborts. This look up is protected by
 180 * rcu_read_lock().
 181 *
 182 * It is necessary to remember whether a stripe is counted in
 183 * big_stripe_tree. Instead of adding new flag, we reuses existing flags:
 184 * STRIPE_R5C_PARTIAL_STRIPE and STRIPE_R5C_FULL_STRIPE. If either of these
 185 * two flags are set, the stripe is counted in big_stripe_tree. This
 186 * requires moving set_bit(STRIPE_R5C_PARTIAL_STRIPE) to
 187 * r5c_try_caching_write(); and moving clear_bit of
 188 * STRIPE_R5C_PARTIAL_STRIPE and STRIPE_R5C_FULL_STRIPE to
 189 * r5c_finish_stripe_write_out().
 190 */
 191
 192/*
 193 * radix tree requests lowest 2 bits of data pointer to be 2b'00.
 194 * So it is necessary to left shift the counter by 2 bits before using it
 195 * as data pointer of the tree.
 196 */
 197#define R5C_RADIX_COUNT_SHIFT 2
 198
 199/*
 200 * calculate key for big_stripe_tree
 201 *
 202 * sect: align_bi->bi_iter.bi_sector or sh->sector
 203 */
 204static inline sector_t r5c_tree_index(struct r5conf *conf,
 205				      sector_t sect)
 206{
 207	sector_t offset;
 208
 209	offset = sector_div(sect, conf->chunk_sectors);
 210	return sect;
 211}
 212
 213/*
 214 * an IO range starts from a meta data block and end at the next meta data
 215 * block. The io unit's the meta data block tracks data/parity followed it. io
 216 * unit is written to log disk with normal write, as we always flush log disk
 217 * first and then start move data to raid disks, there is no requirement to
 218 * write io unit with FLUSH/FUA
 219 */
 220struct r5l_io_unit {
 221	struct r5l_log *log;
 222
 223	struct page *meta_page;	/* store meta block */
 224	int meta_offset;	/* current offset in meta_page */
 225
 226	struct bio *current_bio;/* current_bio accepting new data */
 227
 228	atomic_t pending_stripe;/* how many stripes not flushed to raid */
 229	u64 seq;		/* seq number of the metablock */
 230	sector_t log_start;	/* where the io_unit starts */
 231	sector_t log_end;	/* where the io_unit ends */
 232	struct list_head log_sibling; /* log->running_ios */
 233	struct list_head stripe_list; /* stripes added to the io_unit */
 234
 235	int state;
 236	bool need_split_bio;
 237	struct bio *split_bio;
 238
 239	unsigned int has_flush:1;		/* include flush request */
 240	unsigned int has_fua:1;			/* include fua request */
 241	unsigned int has_null_flush:1;		/* include null flush request */
 242	unsigned int has_flush_payload:1;	/* include flush payload  */
 243	/*
 244	 * io isn't sent yet, flush/fua request can only be submitted till it's
 245	 * the first IO in running_ios list
 246	 */
 247	unsigned int io_deferred:1;
 248
 249	struct bio_list flush_barriers;   /* size == 0 flush bios */
 250};
 251
 252/* r5l_io_unit state */
 253enum r5l_io_unit_state {
 254	IO_UNIT_RUNNING = 0,	/* accepting new IO */
 255	IO_UNIT_IO_START = 1,	/* io_unit bio start writing to log,
 256				 * don't accepting new bio */
 257	IO_UNIT_IO_END = 2,	/* io_unit bio finish writing to log */
 258	IO_UNIT_STRIPE_END = 3,	/* stripes data finished writing to raid */
 259};
 260
 261bool r5c_is_writeback(struct r5l_log *log)
 262{
 263	return (log != NULL &&
 264		log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK);
 265}
 266
 267static sector_t r5l_ring_add(struct r5l_log *log, sector_t start, sector_t inc)
 268{
 269	start += inc;
 270	if (start >= log->device_size)
 271		start = start - log->device_size;
 272	return start;
 273}
 274
 275static sector_t r5l_ring_distance(struct r5l_log *log, sector_t start,
 276				  sector_t end)
 277{
 278	if (end >= start)
 279		return end - start;
 280	else
 281		return end + log->device_size - start;
 282}
 283
 284static bool r5l_has_free_space(struct r5l_log *log, sector_t size)
 285{
 286	sector_t used_size;
 287
 288	used_size = r5l_ring_distance(log, log->last_checkpoint,
 289					log->log_start);
 290
 291	return log->device_size > used_size + size;
 292}
 293
 294static void __r5l_set_io_unit_state(struct r5l_io_unit *io,
 295				    enum r5l_io_unit_state state)
 296{
 297	if (WARN_ON(io->state >= state))
 298		return;
 299	io->state = state;
 300}
 301
 302static void
 303r5c_return_dev_pending_writes(struct r5conf *conf, struct r5dev *dev)
 304{
 305	struct bio *wbi, *wbi2;
 306
 307	wbi = dev->written;
 308	dev->written = NULL;
 309	while (wbi && wbi->bi_iter.bi_sector <
 310	       dev->sector + STRIPE_SECTORS) {
 311		wbi2 = r5_next_bio(wbi, dev->sector);
 312		md_write_end(conf->mddev);
 313		bio_endio(wbi);
 314		wbi = wbi2;
 315	}
 316}
 317
 318void r5c_handle_cached_data_endio(struct r5conf *conf,
 319				  struct stripe_head *sh, int disks)
 320{
 321	int i;
 322
 323	for (i = sh->disks; i--; ) {
 324		if (sh->dev[i].written) {
 325			set_bit(R5_UPTODATE, &sh->dev[i].flags);
 326			r5c_return_dev_pending_writes(conf, &sh->dev[i]);
 327			bitmap_endwrite(conf->mddev->bitmap, sh->sector,
 328					STRIPE_SECTORS,
 329					!test_bit(STRIPE_DEGRADED, &sh->state),
 330					0);
 331		}
 332	}
 333}
 334
 335void r5l_wake_reclaim(struct r5l_log *log, sector_t space);
 336
 337/* Check whether we should flush some stripes to free up stripe cache */
 338void r5c_check_stripe_cache_usage(struct r5conf *conf)
 339{
 340	int total_cached;
 
 341
 342	if (!r5c_is_writeback(conf->log))
 343		return;
 344
 345	total_cached = atomic_read(&conf->r5c_cached_partial_stripes) +
 346		atomic_read(&conf->r5c_cached_full_stripes);
 347
 348	/*
 349	 * The following condition is true for either of the following:
 350	 *   - stripe cache pressure high:
 351	 *          total_cached > 3/4 min_nr_stripes ||
 352	 *          empty_inactive_list_nr > 0
 353	 *   - stripe cache pressure moderate:
 354	 *          total_cached > 1/2 min_nr_stripes
 355	 */
 356	if (total_cached > conf->min_nr_stripes * 1 / 2 ||
 357	    atomic_read(&conf->empty_inactive_list_nr) > 0)
 358		r5l_wake_reclaim(conf->log, 0);
 359}
 360
 361/*
 362 * flush cache when there are R5C_FULL_STRIPE_FLUSH_BATCH or more full
 363 * stripes in the cache
 364 */
 365void r5c_check_cached_full_stripe(struct r5conf *conf)
 366{
 367	if (!r5c_is_writeback(conf->log))
 
 
 368		return;
 369
 370	/*
 371	 * wake up reclaim for R5C_FULL_STRIPE_FLUSH_BATCH cached stripes
 372	 * or a full stripe (chunk size / 4k stripes).
 373	 */
 374	if (atomic_read(&conf->r5c_cached_full_stripes) >=
 375	    min(R5C_FULL_STRIPE_FLUSH_BATCH(conf),
 376		conf->chunk_sectors >> STRIPE_SHIFT))
 377		r5l_wake_reclaim(conf->log, 0);
 378}
 379
 380/*
 381 * Total log space (in sectors) needed to flush all data in cache
 382 *
 383 * To avoid deadlock due to log space, it is necessary to reserve log
 384 * space to flush critical stripes (stripes that occupying log space near
 385 * last_checkpoint). This function helps check how much log space is
 386 * required to flush all cached stripes.
 387 *
 388 * To reduce log space requirements, two mechanisms are used to give cache
 389 * flush higher priorities:
 390 *    1. In handle_stripe_dirtying() and schedule_reconstruction(),
 391 *       stripes ALREADY in journal can be flushed w/o pending writes;
 392 *    2. In r5l_write_stripe() and r5c_cache_data(), stripes NOT in journal
 393 *       can be delayed (r5l_add_no_space_stripe).
 394 *
 395 * In cache flush, the stripe goes through 1 and then 2. For a stripe that
 396 * already passed 1, flushing it requires at most (conf->max_degraded + 1)
 397 * pages of journal space. For stripes that has not passed 1, flushing it
 398 * requires (conf->raid_disks + 1) pages of journal space. There are at
 399 * most (conf->group_cnt + 1) stripe that passed 1. So total journal space
 400 * required to flush all cached stripes (in pages) is:
 401 *
 402 *     (stripe_in_journal_count - group_cnt - 1) * (max_degraded + 1) +
 403 *     (group_cnt + 1) * (raid_disks + 1)
 404 * or
 405 *     (stripe_in_journal_count) * (max_degraded + 1) +
 406 *     (group_cnt + 1) * (raid_disks - max_degraded)
 407 */
 408static sector_t r5c_log_required_to_flush_cache(struct r5conf *conf)
 409{
 410	struct r5l_log *log = conf->log;
 411
 412	if (!r5c_is_writeback(log))
 413		return 0;
 414
 415	return BLOCK_SECTORS *
 416		((conf->max_degraded + 1) * atomic_read(&log->stripe_in_journal_count) +
 417		 (conf->raid_disks - conf->max_degraded) * (conf->group_cnt + 1));
 418}
 419
 420/*
 421 * evaluate log space usage and update R5C_LOG_TIGHT and R5C_LOG_CRITICAL
 422 *
 423 * R5C_LOG_TIGHT is set when free space on the log device is less than 3x of
 424 * reclaim_required_space. R5C_LOG_CRITICAL is set when free space on the log
 425 * device is less than 2x of reclaim_required_space.
 426 */
 427static inline void r5c_update_log_state(struct r5l_log *log)
 428{
 429	struct r5conf *conf = log->rdev->mddev->private;
 430	sector_t free_space;
 431	sector_t reclaim_space;
 432	bool wake_reclaim = false;
 433
 434	if (!r5c_is_writeback(log))
 435		return;
 436
 437	free_space = r5l_ring_distance(log, log->log_start,
 438				       log->last_checkpoint);
 439	reclaim_space = r5c_log_required_to_flush_cache(conf);
 440	if (free_space < 2 * reclaim_space)
 441		set_bit(R5C_LOG_CRITICAL, &conf->cache_state);
 442	else {
 443		if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state))
 444			wake_reclaim = true;
 445		clear_bit(R5C_LOG_CRITICAL, &conf->cache_state);
 446	}
 447	if (free_space < 3 * reclaim_space)
 448		set_bit(R5C_LOG_TIGHT, &conf->cache_state);
 449	else
 450		clear_bit(R5C_LOG_TIGHT, &conf->cache_state);
 451
 452	if (wake_reclaim)
 453		r5l_wake_reclaim(log, 0);
 454}
 455
 456/*
 457 * Put the stripe into writing-out phase by clearing STRIPE_R5C_CACHING.
 458 * This function should only be called in write-back mode.
 459 */
 460void r5c_make_stripe_write_out(struct stripe_head *sh)
 461{
 462	struct r5conf *conf = sh->raid_conf;
 463	struct r5l_log *log = conf->log;
 464
 465	BUG_ON(!r5c_is_writeback(log));
 466
 467	WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
 468	clear_bit(STRIPE_R5C_CACHING, &sh->state);
 469
 470	if (!test_and_set_bit(STRIPE_PREREAD_ACTIVE, &sh->state))
 471		atomic_inc(&conf->preread_active_stripes);
 472}
 473
 474static void r5c_handle_data_cached(struct stripe_head *sh)
 475{
 476	int i;
 477
 478	for (i = sh->disks; i--; )
 479		if (test_and_clear_bit(R5_Wantwrite, &sh->dev[i].flags)) {
 480			set_bit(R5_InJournal, &sh->dev[i].flags);
 481			clear_bit(R5_LOCKED, &sh->dev[i].flags);
 482		}
 483	clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
 484}
 485
 486/*
 487 * this journal write must contain full parity,
 488 * it may also contain some data pages
 489 */
 490static void r5c_handle_parity_cached(struct stripe_head *sh)
 491{
 492	int i;
 493
 494	for (i = sh->disks; i--; )
 495		if (test_bit(R5_InJournal, &sh->dev[i].flags))
 496			set_bit(R5_Wantwrite, &sh->dev[i].flags);
 497}
 498
 499/*
 500 * Setting proper flags after writing (or flushing) data and/or parity to the
 501 * log device. This is called from r5l_log_endio() or r5l_log_flush_endio().
 502 */
 503static void r5c_finish_cache_stripe(struct stripe_head *sh)
 504{
 505	struct r5l_log *log = sh->raid_conf->log;
 506
 507	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
 508		BUG_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
 509		/*
 510		 * Set R5_InJournal for parity dev[pd_idx]. This means
 511		 * all data AND parity in the journal. For RAID 6, it is
 512		 * NOT necessary to set the flag for dev[qd_idx], as the
 513		 * two parities are written out together.
 514		 */
 515		set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
 516	} else if (test_bit(STRIPE_R5C_CACHING, &sh->state)) {
 517		r5c_handle_data_cached(sh);
 518	} else {
 519		r5c_handle_parity_cached(sh);
 520		set_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
 521	}
 522}
 523
 524static void r5l_io_run_stripes(struct r5l_io_unit *io)
 525{
 526	struct stripe_head *sh, *next;
 527
 528	list_for_each_entry_safe(sh, next, &io->stripe_list, log_list) {
 529		list_del_init(&sh->log_list);
 530
 531		r5c_finish_cache_stripe(sh);
 532
 533		set_bit(STRIPE_HANDLE, &sh->state);
 534		raid5_release_stripe(sh);
 535	}
 536}
 537
 538static void r5l_log_run_stripes(struct r5l_log *log)
 539{
 540	struct r5l_io_unit *io, *next;
 541
 542	lockdep_assert_held(&log->io_list_lock);
 543
 544	list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) {
 545		/* don't change list order */
 546		if (io->state < IO_UNIT_IO_END)
 547			break;
 548
 549		list_move_tail(&io->log_sibling, &log->finished_ios);
 550		r5l_io_run_stripes(io);
 551	}
 552}
 553
 554static void r5l_move_to_end_ios(struct r5l_log *log)
 555{
 556	struct r5l_io_unit *io, *next;
 557
 558	lockdep_assert_held(&log->io_list_lock);
 559
 560	list_for_each_entry_safe(io, next, &log->running_ios, log_sibling) {
 561		/* don't change list order */
 562		if (io->state < IO_UNIT_IO_END)
 563			break;
 564		list_move_tail(&io->log_sibling, &log->io_end_ios);
 565	}
 566}
 567
 568static void __r5l_stripe_write_finished(struct r5l_io_unit *io);
 569static void r5l_log_endio(struct bio *bio)
 570{
 571	struct r5l_io_unit *io = bio->bi_private;
 572	struct r5l_io_unit *io_deferred;
 573	struct r5l_log *log = io->log;
 574	unsigned long flags;
 575	bool has_null_flush;
 576	bool has_flush_payload;
 577
 578	if (bio->bi_status)
 579		md_error(log->rdev->mddev, log->rdev);
 580
 581	bio_put(bio);
 582	mempool_free(io->meta_page, log->meta_pool);
 583
 584	spin_lock_irqsave(&log->io_list_lock, flags);
 585	__r5l_set_io_unit_state(io, IO_UNIT_IO_END);
 586
 587	/*
 588	 * if the io doesn't not have null_flush or flush payload,
 589	 * it is not safe to access it after releasing io_list_lock.
 590	 * Therefore, it is necessary to check the condition with
 591	 * the lock held.
 592	 */
 593	has_null_flush = io->has_null_flush;
 594	has_flush_payload = io->has_flush_payload;
 595
 596	if (log->need_cache_flush && !list_empty(&io->stripe_list))
 597		r5l_move_to_end_ios(log);
 598	else
 599		r5l_log_run_stripes(log);
 600	if (!list_empty(&log->running_ios)) {
 601		/*
 602		 * FLUSH/FUA io_unit is deferred because of ordering, now we
 603		 * can dispatch it
 604		 */
 605		io_deferred = list_first_entry(&log->running_ios,
 606					       struct r5l_io_unit, log_sibling);
 607		if (io_deferred->io_deferred)
 608			schedule_work(&log->deferred_io_work);
 609	}
 610
 611	spin_unlock_irqrestore(&log->io_list_lock, flags);
 612
 613	if (log->need_cache_flush)
 614		md_wakeup_thread(log->rdev->mddev->thread);
 615
 616	/* finish flush only io_unit and PAYLOAD_FLUSH only io_unit */
 617	if (has_null_flush) {
 618		struct bio *bi;
 619
 620		WARN_ON(bio_list_empty(&io->flush_barriers));
 621		while ((bi = bio_list_pop(&io->flush_barriers)) != NULL) {
 622			bio_endio(bi);
 623			if (atomic_dec_and_test(&io->pending_stripe)) {
 624				__r5l_stripe_write_finished(io);
 625				return;
 626			}
 627		}
 628	}
 629	/* decrease pending_stripe for flush payload */
 630	if (has_flush_payload)
 631		if (atomic_dec_and_test(&io->pending_stripe))
 632			__r5l_stripe_write_finished(io);
 633}
 634
 635static void r5l_do_submit_io(struct r5l_log *log, struct r5l_io_unit *io)
 636{
 637	unsigned long flags;
 638
 639	spin_lock_irqsave(&log->io_list_lock, flags);
 640	__r5l_set_io_unit_state(io, IO_UNIT_IO_START);
 641	spin_unlock_irqrestore(&log->io_list_lock, flags);
 642
 643	/*
 644	 * In case of journal device failures, submit_bio will get error
 645	 * and calls endio, then active stripes will continue write
 646	 * process. Therefore, it is not necessary to check Faulty bit
 647	 * of journal device here.
 648	 *
 649	 * We can't check split_bio after current_bio is submitted. If
 650	 * io->split_bio is null, after current_bio is submitted, current_bio
 651	 * might already be completed and the io_unit is freed. We submit
 652	 * split_bio first to avoid the issue.
 653	 */
 654	if (io->split_bio) {
 655		if (io->has_flush)
 656			io->split_bio->bi_opf |= REQ_PREFLUSH;
 657		if (io->has_fua)
 658			io->split_bio->bi_opf |= REQ_FUA;
 659		submit_bio(io->split_bio);
 660	}
 661
 662	if (io->has_flush)
 663		io->current_bio->bi_opf |= REQ_PREFLUSH;
 664	if (io->has_fua)
 665		io->current_bio->bi_opf |= REQ_FUA;
 666	submit_bio(io->current_bio);
 667}
 668
 669/* deferred io_unit will be dispatched here */
 670static void r5l_submit_io_async(struct work_struct *work)
 671{
 672	struct r5l_log *log = container_of(work, struct r5l_log,
 673					   deferred_io_work);
 674	struct r5l_io_unit *io = NULL;
 675	unsigned long flags;
 676
 677	spin_lock_irqsave(&log->io_list_lock, flags);
 678	if (!list_empty(&log->running_ios)) {
 679		io = list_first_entry(&log->running_ios, struct r5l_io_unit,
 680				      log_sibling);
 681		if (!io->io_deferred)
 682			io = NULL;
 683		else
 684			io->io_deferred = 0;
 685	}
 686	spin_unlock_irqrestore(&log->io_list_lock, flags);
 687	if (io)
 688		r5l_do_submit_io(log, io);
 689}
 690
 691static void r5c_disable_writeback_async(struct work_struct *work)
 692{
 693	struct r5l_log *log = container_of(work, struct r5l_log,
 694					   disable_writeback_work);
 695	struct mddev *mddev = log->rdev->mddev;
 696	struct r5conf *conf = mddev->private;
 697	int locked = 0;
 698
 699	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
 700		return;
 701	pr_info("md/raid:%s: Disabling writeback cache for degraded array.\n",
 702		mdname(mddev));
 703
 704	/* wait superblock change before suspend */
 705	wait_event(mddev->sb_wait,
 706		   conf->log == NULL ||
 707		   (!test_bit(MD_SB_CHANGE_PENDING, &mddev->sb_flags) &&
 708		    (locked = mddev_trylock(mddev))));
 709	if (locked) {
 710		mddev_suspend(mddev);
 
 711		log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
 712		mddev_resume(mddev);
 713		mddev_unlock(mddev);
 714	}
 715}
 716
 717static void r5l_submit_current_io(struct r5l_log *log)
 718{
 719	struct r5l_io_unit *io = log->current_io;
 720	struct bio *bio;
 721	struct r5l_meta_block *block;
 722	unsigned long flags;
 723	u32 crc;
 724	bool do_submit = true;
 725
 726	if (!io)
 727		return;
 728
 729	block = page_address(io->meta_page);
 730	block->meta_size = cpu_to_le32(io->meta_offset);
 731	crc = crc32c_le(log->uuid_checksum, block, PAGE_SIZE);
 732	block->checksum = cpu_to_le32(crc);
 733	bio = io->current_bio;
 734
 735	log->current_io = NULL;
 736	spin_lock_irqsave(&log->io_list_lock, flags);
 737	if (io->has_flush || io->has_fua) {
 738		if (io != list_first_entry(&log->running_ios,
 739					   struct r5l_io_unit, log_sibling)) {
 740			io->io_deferred = 1;
 741			do_submit = false;
 742		}
 743	}
 744	spin_unlock_irqrestore(&log->io_list_lock, flags);
 745	if (do_submit)
 746		r5l_do_submit_io(log, io);
 747}
 748
 749static struct bio *r5l_bio_alloc(struct r5l_log *log)
 750{
 751	struct bio *bio = bio_alloc_bioset(GFP_NOIO, BIO_MAX_PAGES, log->bs);
 
 752
 753	bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
 754	bio_set_dev(bio, log->rdev->bdev);
 755	bio->bi_iter.bi_sector = log->rdev->data_offset + log->log_start;
 756
 757	return bio;
 758}
 759
 760static void r5_reserve_log_entry(struct r5l_log *log, struct r5l_io_unit *io)
 761{
 762	log->log_start = r5l_ring_add(log, log->log_start, BLOCK_SECTORS);
 763
 764	r5c_update_log_state(log);
 765	/*
 766	 * If we filled up the log device start from the beginning again,
 767	 * which will require a new bio.
 768	 *
 769	 * Note: for this to work properly the log size needs to me a multiple
 770	 * of BLOCK_SECTORS.
 771	 */
 772	if (log->log_start == 0)
 773		io->need_split_bio = true;
 774
 775	io->log_end = log->log_start;
 776}
 777
 778static struct r5l_io_unit *r5l_new_meta(struct r5l_log *log)
 779{
 780	struct r5l_io_unit *io;
 781	struct r5l_meta_block *block;
 782
 783	io = mempool_alloc(log->io_pool, GFP_ATOMIC);
 784	if (!io)
 785		return NULL;
 786	memset(io, 0, sizeof(*io));
 787
 788	io->log = log;
 789	INIT_LIST_HEAD(&io->log_sibling);
 790	INIT_LIST_HEAD(&io->stripe_list);
 791	bio_list_init(&io->flush_barriers);
 792	io->state = IO_UNIT_RUNNING;
 793
 794	io->meta_page = mempool_alloc(log->meta_pool, GFP_NOIO);
 795	block = page_address(io->meta_page);
 796	clear_page(block);
 797	block->magic = cpu_to_le32(R5LOG_MAGIC);
 798	block->version = R5LOG_VERSION;
 799	block->seq = cpu_to_le64(log->seq);
 800	block->position = cpu_to_le64(log->log_start);
 801
 802	io->log_start = log->log_start;
 803	io->meta_offset = sizeof(struct r5l_meta_block);
 804	io->seq = log->seq++;
 805
 806	io->current_bio = r5l_bio_alloc(log);
 807	io->current_bio->bi_end_io = r5l_log_endio;
 808	io->current_bio->bi_private = io;
 809	bio_add_page(io->current_bio, io->meta_page, PAGE_SIZE, 0);
 810
 811	r5_reserve_log_entry(log, io);
 812
 813	spin_lock_irq(&log->io_list_lock);
 814	list_add_tail(&io->log_sibling, &log->running_ios);
 815	spin_unlock_irq(&log->io_list_lock);
 816
 817	return io;
 818}
 819
 820static int r5l_get_meta(struct r5l_log *log, unsigned int payload_size)
 821{
 822	if (log->current_io &&
 823	    log->current_io->meta_offset + payload_size > PAGE_SIZE)
 824		r5l_submit_current_io(log);
 825
 826	if (!log->current_io) {
 827		log->current_io = r5l_new_meta(log);
 828		if (!log->current_io)
 829			return -ENOMEM;
 830	}
 831
 832	return 0;
 833}
 834
 835static void r5l_append_payload_meta(struct r5l_log *log, u16 type,
 836				    sector_t location,
 837				    u32 checksum1, u32 checksum2,
 838				    bool checksum2_valid)
 839{
 840	struct r5l_io_unit *io = log->current_io;
 841	struct r5l_payload_data_parity *payload;
 842
 843	payload = page_address(io->meta_page) + io->meta_offset;
 844	payload->header.type = cpu_to_le16(type);
 845	payload->header.flags = cpu_to_le16(0);
 846	payload->size = cpu_to_le32((1 + !!checksum2_valid) <<
 847				    (PAGE_SHIFT - 9));
 848	payload->location = cpu_to_le64(location);
 849	payload->checksum[0] = cpu_to_le32(checksum1);
 850	if (checksum2_valid)
 851		payload->checksum[1] = cpu_to_le32(checksum2);
 852
 853	io->meta_offset += sizeof(struct r5l_payload_data_parity) +
 854		sizeof(__le32) * (1 + !!checksum2_valid);
 855}
 856
 857static void r5l_append_payload_page(struct r5l_log *log, struct page *page)
 858{
 859	struct r5l_io_unit *io = log->current_io;
 860
 861	if (io->need_split_bio) {
 862		BUG_ON(io->split_bio);
 863		io->split_bio = io->current_bio;
 864		io->current_bio = r5l_bio_alloc(log);
 865		bio_chain(io->current_bio, io->split_bio);
 866		io->need_split_bio = false;
 867	}
 868
 869	if (!bio_add_page(io->current_bio, page, PAGE_SIZE, 0))
 870		BUG();
 871
 872	r5_reserve_log_entry(log, io);
 873}
 874
 875static void r5l_append_flush_payload(struct r5l_log *log, sector_t sect)
 876{
 877	struct mddev *mddev = log->rdev->mddev;
 878	struct r5conf *conf = mddev->private;
 879	struct r5l_io_unit *io;
 880	struct r5l_payload_flush *payload;
 881	int meta_size;
 882
 883	/*
 884	 * payload_flush requires extra writes to the journal.
 885	 * To avoid handling the extra IO in quiesce, just skip
 886	 * flush_payload
 887	 */
 888	if (conf->quiesce)
 889		return;
 890
 891	mutex_lock(&log->io_mutex);
 892	meta_size = sizeof(struct r5l_payload_flush) + sizeof(__le64);
 893
 894	if (r5l_get_meta(log, meta_size)) {
 895		mutex_unlock(&log->io_mutex);
 896		return;
 897	}
 898
 899	/* current implementation is one stripe per flush payload */
 900	io = log->current_io;
 901	payload = page_address(io->meta_page) + io->meta_offset;
 902	payload->header.type = cpu_to_le16(R5LOG_PAYLOAD_FLUSH);
 903	payload->header.flags = cpu_to_le16(0);
 904	payload->size = cpu_to_le32(sizeof(__le64));
 905	payload->flush_stripes[0] = cpu_to_le64(sect);
 906	io->meta_offset += meta_size;
 907	/* multiple flush payloads count as one pending_stripe */
 908	if (!io->has_flush_payload) {
 909		io->has_flush_payload = 1;
 910		atomic_inc(&io->pending_stripe);
 911	}
 912	mutex_unlock(&log->io_mutex);
 913}
 914
 915static int r5l_log_stripe(struct r5l_log *log, struct stripe_head *sh,
 916			   int data_pages, int parity_pages)
 917{
 918	int i;
 919	int meta_size;
 920	int ret;
 921	struct r5l_io_unit *io;
 922
 923	meta_size =
 924		((sizeof(struct r5l_payload_data_parity) + sizeof(__le32))
 925		 * data_pages) +
 926		sizeof(struct r5l_payload_data_parity) +
 927		sizeof(__le32) * parity_pages;
 928
 929	ret = r5l_get_meta(log, meta_size);
 930	if (ret)
 931		return ret;
 932
 933	io = log->current_io;
 934
 935	if (test_and_clear_bit(STRIPE_R5C_PREFLUSH, &sh->state))
 936		io->has_flush = 1;
 937
 938	for (i = 0; i < sh->disks; i++) {
 939		if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) ||
 940		    test_bit(R5_InJournal, &sh->dev[i].flags))
 941			continue;
 942		if (i == sh->pd_idx || i == sh->qd_idx)
 943			continue;
 944		if (test_bit(R5_WantFUA, &sh->dev[i].flags) &&
 945		    log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK) {
 946			io->has_fua = 1;
 947			/*
 948			 * we need to flush journal to make sure recovery can
 949			 * reach the data with fua flag
 950			 */
 951			io->has_flush = 1;
 952		}
 953		r5l_append_payload_meta(log, R5LOG_PAYLOAD_DATA,
 954					raid5_compute_blocknr(sh, i, 0),
 955					sh->dev[i].log_checksum, 0, false);
 956		r5l_append_payload_page(log, sh->dev[i].page);
 957	}
 958
 959	if (parity_pages == 2) {
 960		r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
 961					sh->sector, sh->dev[sh->pd_idx].log_checksum,
 962					sh->dev[sh->qd_idx].log_checksum, true);
 963		r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
 964		r5l_append_payload_page(log, sh->dev[sh->qd_idx].page);
 965	} else if (parity_pages == 1) {
 966		r5l_append_payload_meta(log, R5LOG_PAYLOAD_PARITY,
 967					sh->sector, sh->dev[sh->pd_idx].log_checksum,
 968					0, false);
 969		r5l_append_payload_page(log, sh->dev[sh->pd_idx].page);
 970	} else  /* Just writing data, not parity, in caching phase */
 971		BUG_ON(parity_pages != 0);
 972
 973	list_add_tail(&sh->log_list, &io->stripe_list);
 974	atomic_inc(&io->pending_stripe);
 975	sh->log_io = io;
 976
 977	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
 978		return 0;
 979
 980	if (sh->log_start == MaxSector) {
 981		BUG_ON(!list_empty(&sh->r5c));
 982		sh->log_start = io->log_start;
 983		spin_lock_irq(&log->stripe_in_journal_lock);
 984		list_add_tail(&sh->r5c,
 985			      &log->stripe_in_journal_list);
 986		spin_unlock_irq(&log->stripe_in_journal_lock);
 987		atomic_inc(&log->stripe_in_journal_count);
 988	}
 989	return 0;
 990}
 991
 992/* add stripe to no_space_stripes, and then wake up reclaim */
 993static inline void r5l_add_no_space_stripe(struct r5l_log *log,
 994					   struct stripe_head *sh)
 995{
 996	spin_lock(&log->no_space_stripes_lock);
 997	list_add_tail(&sh->log_list, &log->no_space_stripes);
 998	spin_unlock(&log->no_space_stripes_lock);
 999}
1000
1001/*
1002 * running in raid5d, where reclaim could wait for raid5d too (when it flushes
1003 * data from log to raid disks), so we shouldn't wait for reclaim here
1004 */
1005int r5l_write_stripe(struct r5l_log *log, struct stripe_head *sh)
1006{
1007	struct r5conf *conf = sh->raid_conf;
1008	int write_disks = 0;
1009	int data_pages, parity_pages;
1010	int reserve;
1011	int i;
1012	int ret = 0;
1013	bool wake_reclaim = false;
1014
1015	if (!log)
1016		return -EAGAIN;
1017	/* Don't support stripe batch */
1018	if (sh->log_io || !test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags) ||
1019	    test_bit(STRIPE_SYNCING, &sh->state)) {
1020		/* the stripe is written to log, we start writing it to raid */
1021		clear_bit(STRIPE_LOG_TRAPPED, &sh->state);
1022		return -EAGAIN;
1023	}
1024
1025	WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
1026
1027	for (i = 0; i < sh->disks; i++) {
1028		void *addr;
1029
1030		if (!test_bit(R5_Wantwrite, &sh->dev[i].flags) ||
1031		    test_bit(R5_InJournal, &sh->dev[i].flags))
1032			continue;
1033
1034		write_disks++;
1035		/* checksum is already calculated in last run */
1036		if (test_bit(STRIPE_LOG_TRAPPED, &sh->state))
1037			continue;
1038		addr = kmap_atomic(sh->dev[i].page);
1039		sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum,
1040						    addr, PAGE_SIZE);
1041		kunmap_atomic(addr);
1042	}
1043	parity_pages = 1 + !!(sh->qd_idx >= 0);
1044	data_pages = write_disks - parity_pages;
1045
1046	set_bit(STRIPE_LOG_TRAPPED, &sh->state);
1047	/*
1048	 * The stripe must enter state machine again to finish the write, so
1049	 * don't delay.
1050	 */
1051	clear_bit(STRIPE_DELAYED, &sh->state);
1052	atomic_inc(&sh->count);
1053
1054	mutex_lock(&log->io_mutex);
1055	/* meta + data */
1056	reserve = (1 + write_disks) << (PAGE_SHIFT - 9);
1057
1058	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
1059		if (!r5l_has_free_space(log, reserve)) {
1060			r5l_add_no_space_stripe(log, sh);
1061			wake_reclaim = true;
1062		} else {
1063			ret = r5l_log_stripe(log, sh, data_pages, parity_pages);
1064			if (ret) {
1065				spin_lock_irq(&log->io_list_lock);
1066				list_add_tail(&sh->log_list,
1067					      &log->no_mem_stripes);
1068				spin_unlock_irq(&log->io_list_lock);
1069			}
1070		}
1071	} else {  /* R5C_JOURNAL_MODE_WRITE_BACK */
1072		/*
1073		 * log space critical, do not process stripes that are
1074		 * not in cache yet (sh->log_start == MaxSector).
1075		 */
1076		if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
1077		    sh->log_start == MaxSector) {
1078			r5l_add_no_space_stripe(log, sh);
1079			wake_reclaim = true;
1080			reserve = 0;
1081		} else if (!r5l_has_free_space(log, reserve)) {
1082			if (sh->log_start == log->last_checkpoint)
1083				BUG();
1084			else
1085				r5l_add_no_space_stripe(log, sh);
1086		} else {
1087			ret = r5l_log_stripe(log, sh, data_pages, parity_pages);
1088			if (ret) {
1089				spin_lock_irq(&log->io_list_lock);
1090				list_add_tail(&sh->log_list,
1091					      &log->no_mem_stripes);
1092				spin_unlock_irq(&log->io_list_lock);
1093			}
1094		}
1095	}
1096
1097	mutex_unlock(&log->io_mutex);
1098	if (wake_reclaim)
1099		r5l_wake_reclaim(log, reserve);
1100	return 0;
1101}
1102
1103void r5l_write_stripe_run(struct r5l_log *log)
1104{
1105	if (!log)
1106		return;
1107	mutex_lock(&log->io_mutex);
1108	r5l_submit_current_io(log);
1109	mutex_unlock(&log->io_mutex);
1110}
1111
1112int r5l_handle_flush_request(struct r5l_log *log, struct bio *bio)
1113{
1114	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH) {
1115		/*
1116		 * in write through (journal only)
1117		 * we flush log disk cache first, then write stripe data to
1118		 * raid disks. So if bio is finished, the log disk cache is
1119		 * flushed already. The recovery guarantees we can recovery
1120		 * the bio from log disk, so we don't need to flush again
1121		 */
1122		if (bio->bi_iter.bi_size == 0) {
1123			bio_endio(bio);
1124			return 0;
1125		}
1126		bio->bi_opf &= ~REQ_PREFLUSH;
1127	} else {
1128		/* write back (with cache) */
1129		if (bio->bi_iter.bi_size == 0) {
1130			mutex_lock(&log->io_mutex);
1131			r5l_get_meta(log, 0);
1132			bio_list_add(&log->current_io->flush_barriers, bio);
1133			log->current_io->has_flush = 1;
1134			log->current_io->has_null_flush = 1;
1135			atomic_inc(&log->current_io->pending_stripe);
1136			r5l_submit_current_io(log);
1137			mutex_unlock(&log->io_mutex);
1138			return 0;
1139		}
1140	}
1141	return -EAGAIN;
1142}
1143
1144/* This will run after log space is reclaimed */
1145static void r5l_run_no_space_stripes(struct r5l_log *log)
1146{
1147	struct stripe_head *sh;
1148
1149	spin_lock(&log->no_space_stripes_lock);
1150	while (!list_empty(&log->no_space_stripes)) {
1151		sh = list_first_entry(&log->no_space_stripes,
1152				      struct stripe_head, log_list);
1153		list_del_init(&sh->log_list);
1154		set_bit(STRIPE_HANDLE, &sh->state);
1155		raid5_release_stripe(sh);
1156	}
1157	spin_unlock(&log->no_space_stripes_lock);
1158}
1159
1160/*
1161 * calculate new last_checkpoint
1162 * for write through mode, returns log->next_checkpoint
1163 * for write back, returns log_start of first sh in stripe_in_journal_list
1164 */
1165static sector_t r5c_calculate_new_cp(struct r5conf *conf)
1166{
1167	struct stripe_head *sh;
1168	struct r5l_log *log = conf->log;
1169	sector_t new_cp;
1170	unsigned long flags;
1171
1172	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
1173		return log->next_checkpoint;
1174
1175	spin_lock_irqsave(&log->stripe_in_journal_lock, flags);
1176	if (list_empty(&conf->log->stripe_in_journal_list)) {
1177		/* all stripes flushed */
1178		spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1179		return log->next_checkpoint;
1180	}
1181	sh = list_first_entry(&conf->log->stripe_in_journal_list,
1182			      struct stripe_head, r5c);
1183	new_cp = sh->log_start;
1184	spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1185	return new_cp;
1186}
1187
1188static sector_t r5l_reclaimable_space(struct r5l_log *log)
1189{
1190	struct r5conf *conf = log->rdev->mddev->private;
1191
1192	return r5l_ring_distance(log, log->last_checkpoint,
1193				 r5c_calculate_new_cp(conf));
1194}
1195
1196static void r5l_run_no_mem_stripe(struct r5l_log *log)
1197{
1198	struct stripe_head *sh;
1199
1200	lockdep_assert_held(&log->io_list_lock);
1201
1202	if (!list_empty(&log->no_mem_stripes)) {
1203		sh = list_first_entry(&log->no_mem_stripes,
1204				      struct stripe_head, log_list);
1205		list_del_init(&sh->log_list);
1206		set_bit(STRIPE_HANDLE, &sh->state);
1207		raid5_release_stripe(sh);
1208	}
1209}
1210
1211static bool r5l_complete_finished_ios(struct r5l_log *log)
1212{
1213	struct r5l_io_unit *io, *next;
1214	bool found = false;
1215
1216	lockdep_assert_held(&log->io_list_lock);
1217
1218	list_for_each_entry_safe(io, next, &log->finished_ios, log_sibling) {
1219		/* don't change list order */
1220		if (io->state < IO_UNIT_STRIPE_END)
1221			break;
1222
1223		log->next_checkpoint = io->log_start;
1224
1225		list_del(&io->log_sibling);
1226		mempool_free(io, log->io_pool);
1227		r5l_run_no_mem_stripe(log);
1228
1229		found = true;
1230	}
1231
1232	return found;
1233}
1234
1235static void __r5l_stripe_write_finished(struct r5l_io_unit *io)
1236{
1237	struct r5l_log *log = io->log;
1238	struct r5conf *conf = log->rdev->mddev->private;
1239	unsigned long flags;
1240
1241	spin_lock_irqsave(&log->io_list_lock, flags);
1242	__r5l_set_io_unit_state(io, IO_UNIT_STRIPE_END);
1243
1244	if (!r5l_complete_finished_ios(log)) {
1245		spin_unlock_irqrestore(&log->io_list_lock, flags);
1246		return;
1247	}
1248
1249	if (r5l_reclaimable_space(log) > log->max_free_space ||
1250	    test_bit(R5C_LOG_TIGHT, &conf->cache_state))
1251		r5l_wake_reclaim(log, 0);
1252
1253	spin_unlock_irqrestore(&log->io_list_lock, flags);
1254	wake_up(&log->iounit_wait);
1255}
1256
1257void r5l_stripe_write_finished(struct stripe_head *sh)
1258{
1259	struct r5l_io_unit *io;
1260
1261	io = sh->log_io;
1262	sh->log_io = NULL;
1263
1264	if (io && atomic_dec_and_test(&io->pending_stripe))
1265		__r5l_stripe_write_finished(io);
1266}
1267
1268static void r5l_log_flush_endio(struct bio *bio)
1269{
1270	struct r5l_log *log = container_of(bio, struct r5l_log,
1271		flush_bio);
1272	unsigned long flags;
1273	struct r5l_io_unit *io;
1274
1275	if (bio->bi_status)
1276		md_error(log->rdev->mddev, log->rdev);
 
1277
1278	spin_lock_irqsave(&log->io_list_lock, flags);
1279	list_for_each_entry(io, &log->flushing_ios, log_sibling)
1280		r5l_io_run_stripes(io);
1281	list_splice_tail_init(&log->flushing_ios, &log->finished_ios);
1282	spin_unlock_irqrestore(&log->io_list_lock, flags);
1283}
1284
1285/*
1286 * Starting dispatch IO to raid.
1287 * io_unit(meta) consists of a log. There is one situation we want to avoid. A
1288 * broken meta in the middle of a log causes recovery can't find meta at the
1289 * head of log. If operations require meta at the head persistent in log, we
1290 * must make sure meta before it persistent in log too. A case is:
1291 *
1292 * stripe data/parity is in log, we start write stripe to raid disks. stripe
1293 * data/parity must be persistent in log before we do the write to raid disks.
1294 *
1295 * The solution is we restrictly maintain io_unit list order. In this case, we
1296 * only write stripes of an io_unit to raid disks till the io_unit is the first
1297 * one whose data/parity is in log.
1298 */
1299void r5l_flush_stripe_to_raid(struct r5l_log *log)
1300{
1301	bool do_flush;
1302
1303	if (!log || !log->need_cache_flush)
1304		return;
1305
1306	spin_lock_irq(&log->io_list_lock);
1307	/* flush bio is running */
1308	if (!list_empty(&log->flushing_ios)) {
1309		spin_unlock_irq(&log->io_list_lock);
1310		return;
1311	}
1312	list_splice_tail_init(&log->io_end_ios, &log->flushing_ios);
1313	do_flush = !list_empty(&log->flushing_ios);
1314	spin_unlock_irq(&log->io_list_lock);
1315
1316	if (!do_flush)
1317		return;
1318	bio_reset(&log->flush_bio);
1319	bio_set_dev(&log->flush_bio, log->rdev->bdev);
1320	log->flush_bio.bi_end_io = r5l_log_flush_endio;
1321	log->flush_bio.bi_opf = REQ_OP_WRITE | REQ_PREFLUSH;
1322	submit_bio(&log->flush_bio);
1323}
1324
1325static void r5l_write_super(struct r5l_log *log, sector_t cp);
1326static void r5l_write_super_and_discard_space(struct r5l_log *log,
1327	sector_t end)
1328{
1329	struct block_device *bdev = log->rdev->bdev;
1330	struct mddev *mddev;
1331
1332	r5l_write_super(log, end);
1333
1334	if (!blk_queue_discard(bdev_get_queue(bdev)))
1335		return;
1336
1337	mddev = log->rdev->mddev;
1338	/*
1339	 * Discard could zero data, so before discard we must make sure
1340	 * superblock is updated to new log tail. Updating superblock (either
1341	 * directly call md_update_sb() or depend on md thread) must hold
1342	 * reconfig mutex. On the other hand, raid5_quiesce is called with
1343	 * reconfig_mutex hold. The first step of raid5_quiesce() is waitting
1344	 * for all IO finish, hence waitting for reclaim thread, while reclaim
1345	 * thread is calling this function and waitting for reconfig mutex. So
1346	 * there is a deadlock. We workaround this issue with a trylock.
1347	 * FIXME: we could miss discard if we can't take reconfig mutex
1348	 */
1349	set_mask_bits(&mddev->sb_flags, 0,
1350		BIT(MD_SB_CHANGE_DEVS) | BIT(MD_SB_CHANGE_PENDING));
1351	if (!mddev_trylock(mddev))
1352		return;
1353	md_update_sb(mddev, 1);
1354	mddev_unlock(mddev);
1355
1356	/* discard IO error really doesn't matter, ignore it */
1357	if (log->last_checkpoint < end) {
1358		blkdev_issue_discard(bdev,
1359				log->last_checkpoint + log->rdev->data_offset,
1360				end - log->last_checkpoint, GFP_NOIO, 0);
1361	} else {
1362		blkdev_issue_discard(bdev,
1363				log->last_checkpoint + log->rdev->data_offset,
1364				log->device_size - log->last_checkpoint,
1365				GFP_NOIO, 0);
1366		blkdev_issue_discard(bdev, log->rdev->data_offset, end,
1367				GFP_NOIO, 0);
1368	}
1369}
1370
1371/*
1372 * r5c_flush_stripe moves stripe from cached list to handle_list. When called,
1373 * the stripe must be on r5c_cached_full_stripes or r5c_cached_partial_stripes.
1374 *
1375 * must hold conf->device_lock
1376 */
1377static void r5c_flush_stripe(struct r5conf *conf, struct stripe_head *sh)
1378{
1379	BUG_ON(list_empty(&sh->lru));
1380	BUG_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
1381	BUG_ON(test_bit(STRIPE_HANDLE, &sh->state));
1382
1383	/*
1384	 * The stripe is not ON_RELEASE_LIST, so it is safe to call
1385	 * raid5_release_stripe() while holding conf->device_lock
1386	 */
1387	BUG_ON(test_bit(STRIPE_ON_RELEASE_LIST, &sh->state));
1388	lockdep_assert_held(&conf->device_lock);
1389
1390	list_del_init(&sh->lru);
1391	atomic_inc(&sh->count);
1392
1393	set_bit(STRIPE_HANDLE, &sh->state);
1394	atomic_inc(&conf->active_stripes);
1395	r5c_make_stripe_write_out(sh);
1396
1397	if (test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state))
1398		atomic_inc(&conf->r5c_flushing_partial_stripes);
1399	else
1400		atomic_inc(&conf->r5c_flushing_full_stripes);
1401	raid5_release_stripe(sh);
1402}
1403
1404/*
1405 * if num == 0, flush all full stripes
1406 * if num > 0, flush all full stripes. If less than num full stripes are
1407 *             flushed, flush some partial stripes until totally num stripes are
1408 *             flushed or there is no more cached stripes.
1409 */
1410void r5c_flush_cache(struct r5conf *conf, int num)
1411{
1412	int count;
1413	struct stripe_head *sh, *next;
1414
1415	lockdep_assert_held(&conf->device_lock);
1416	if (!conf->log)
1417		return;
1418
1419	count = 0;
1420	list_for_each_entry_safe(sh, next, &conf->r5c_full_stripe_list, lru) {
1421		r5c_flush_stripe(conf, sh);
1422		count++;
1423	}
1424
1425	if (count >= num)
1426		return;
1427	list_for_each_entry_safe(sh, next,
1428				 &conf->r5c_partial_stripe_list, lru) {
1429		r5c_flush_stripe(conf, sh);
1430		if (++count >= num)
1431			break;
1432	}
1433}
1434
1435static void r5c_do_reclaim(struct r5conf *conf)
1436{
1437	struct r5l_log *log = conf->log;
1438	struct stripe_head *sh;
1439	int count = 0;
1440	unsigned long flags;
1441	int total_cached;
1442	int stripes_to_flush;
1443	int flushing_partial, flushing_full;
1444
1445	if (!r5c_is_writeback(log))
1446		return;
1447
1448	flushing_partial = atomic_read(&conf->r5c_flushing_partial_stripes);
1449	flushing_full = atomic_read(&conf->r5c_flushing_full_stripes);
1450	total_cached = atomic_read(&conf->r5c_cached_partial_stripes) +
1451		atomic_read(&conf->r5c_cached_full_stripes) -
1452		flushing_full - flushing_partial;
1453
1454	if (total_cached > conf->min_nr_stripes * 3 / 4 ||
1455	    atomic_read(&conf->empty_inactive_list_nr) > 0)
1456		/*
1457		 * if stripe cache pressure high, flush all full stripes and
1458		 * some partial stripes
1459		 */
1460		stripes_to_flush = R5C_RECLAIM_STRIPE_GROUP;
1461	else if (total_cached > conf->min_nr_stripes * 1 / 2 ||
1462		 atomic_read(&conf->r5c_cached_full_stripes) - flushing_full >
1463		 R5C_FULL_STRIPE_FLUSH_BATCH(conf))
1464		/*
1465		 * if stripe cache pressure moderate, or if there is many full
1466		 * stripes,flush all full stripes
1467		 */
1468		stripes_to_flush = 0;
1469	else
1470		/* no need to flush */
1471		stripes_to_flush = -1;
1472
1473	if (stripes_to_flush >= 0) {
1474		spin_lock_irqsave(&conf->device_lock, flags);
1475		r5c_flush_cache(conf, stripes_to_flush);
1476		spin_unlock_irqrestore(&conf->device_lock, flags);
1477	}
1478
1479	/* if log space is tight, flush stripes on stripe_in_journal_list */
1480	if (test_bit(R5C_LOG_TIGHT, &conf->cache_state)) {
1481		spin_lock_irqsave(&log->stripe_in_journal_lock, flags);
1482		spin_lock(&conf->device_lock);
1483		list_for_each_entry(sh, &log->stripe_in_journal_list, r5c) {
1484			/*
1485			 * stripes on stripe_in_journal_list could be in any
1486			 * state of the stripe_cache state machine. In this
1487			 * case, we only want to flush stripe on
1488			 * r5c_cached_full/partial_stripes. The following
1489			 * condition makes sure the stripe is on one of the
1490			 * two lists.
1491			 */
1492			if (!list_empty(&sh->lru) &&
1493			    !test_bit(STRIPE_HANDLE, &sh->state) &&
1494			    atomic_read(&sh->count) == 0) {
1495				r5c_flush_stripe(conf, sh);
1496				if (count++ >= R5C_RECLAIM_STRIPE_GROUP)
1497					break;
1498			}
1499		}
1500		spin_unlock(&conf->device_lock);
1501		spin_unlock_irqrestore(&log->stripe_in_journal_lock, flags);
1502	}
1503
1504	if (!test_bit(R5C_LOG_CRITICAL, &conf->cache_state))
1505		r5l_run_no_space_stripes(log);
1506
1507	md_wakeup_thread(conf->mddev->thread);
1508}
1509
1510static void r5l_do_reclaim(struct r5l_log *log)
1511{
1512	struct r5conf *conf = log->rdev->mddev->private;
1513	sector_t reclaim_target = xchg(&log->reclaim_target, 0);
1514	sector_t reclaimable;
1515	sector_t next_checkpoint;
1516	bool write_super;
1517
1518	spin_lock_irq(&log->io_list_lock);
1519	write_super = r5l_reclaimable_space(log) > log->max_free_space ||
1520		reclaim_target != 0 || !list_empty(&log->no_space_stripes);
1521	/*
1522	 * move proper io_unit to reclaim list. We should not change the order.
1523	 * reclaimable/unreclaimable io_unit can be mixed in the list, we
1524	 * shouldn't reuse space of an unreclaimable io_unit
1525	 */
1526	while (1) {
1527		reclaimable = r5l_reclaimable_space(log);
1528		if (reclaimable >= reclaim_target ||
1529		    (list_empty(&log->running_ios) &&
1530		     list_empty(&log->io_end_ios) &&
1531		     list_empty(&log->flushing_ios) &&
1532		     list_empty(&log->finished_ios)))
1533			break;
1534
1535		md_wakeup_thread(log->rdev->mddev->thread);
1536		wait_event_lock_irq(log->iounit_wait,
1537				    r5l_reclaimable_space(log) > reclaimable,
1538				    log->io_list_lock);
1539	}
1540
1541	next_checkpoint = r5c_calculate_new_cp(conf);
1542	spin_unlock_irq(&log->io_list_lock);
1543
1544	if (reclaimable == 0 || !write_super)
1545		return;
1546
1547	/*
1548	 * write_super will flush cache of each raid disk. We must write super
1549	 * here, because the log area might be reused soon and we don't want to
1550	 * confuse recovery
1551	 */
1552	r5l_write_super_and_discard_space(log, next_checkpoint);
1553
1554	mutex_lock(&log->io_mutex);
1555	log->last_checkpoint = next_checkpoint;
1556	r5c_update_log_state(log);
1557	mutex_unlock(&log->io_mutex);
1558
1559	r5l_run_no_space_stripes(log);
1560}
1561
1562static void r5l_reclaim_thread(struct md_thread *thread)
1563{
1564	struct mddev *mddev = thread->mddev;
1565	struct r5conf *conf = mddev->private;
1566	struct r5l_log *log = conf->log;
1567
1568	if (!log)
1569		return;
1570	r5c_do_reclaim(conf);
1571	r5l_do_reclaim(log);
1572}
1573
1574void r5l_wake_reclaim(struct r5l_log *log, sector_t space)
1575{
1576	unsigned long target;
1577	unsigned long new = (unsigned long)space; /* overflow in theory */
1578
1579	if (!log)
1580		return;
 
 
1581	do {
1582		target = log->reclaim_target;
1583		if (new < target)
1584			return;
1585	} while (cmpxchg(&log->reclaim_target, target, new) != target);
1586	md_wakeup_thread(log->reclaim_thread);
1587}
1588
1589void r5l_quiesce(struct r5l_log *log, int quiesce)
1590{
1591	struct mddev *mddev;
 
 
1592
1593	if (quiesce) {
1594		/* make sure r5l_write_super_and_discard_space exits */
1595		mddev = log->rdev->mddev;
1596		wake_up(&mddev->sb_wait);
1597		kthread_park(log->reclaim_thread->tsk);
1598		r5l_wake_reclaim(log, MaxSector);
1599		r5l_do_reclaim(log);
1600	} else
1601		kthread_unpark(log->reclaim_thread->tsk);
1602}
1603
1604bool r5l_log_disk_error(struct r5conf *conf)
1605{
1606	struct r5l_log *log;
1607	bool ret;
1608	/* don't allow write if journal disk is missing */
1609	rcu_read_lock();
1610	log = rcu_dereference(conf->log);
1611
1612	if (!log)
1613		ret = test_bit(MD_HAS_JOURNAL, &conf->mddev->flags);
1614	else
1615		ret = test_bit(Faulty, &log->rdev->flags);
1616	rcu_read_unlock();
1617	return ret;
1618}
1619
1620#define R5L_RECOVERY_PAGE_POOL_SIZE 256
1621
1622struct r5l_recovery_ctx {
1623	struct page *meta_page;		/* current meta */
1624	sector_t meta_total_blocks;	/* total size of current meta and data */
1625	sector_t pos;			/* recovery position */
1626	u64 seq;			/* recovery position seq */
1627	int data_parity_stripes;	/* number of data_parity stripes */
1628	int data_only_stripes;		/* number of data_only stripes */
1629	struct list_head cached_list;
1630
1631	/*
1632	 * read ahead page pool (ra_pool)
1633	 * in recovery, log is read sequentially. It is not efficient to
1634	 * read every page with sync_page_io(). The read ahead page pool
1635	 * reads multiple pages with one IO, so further log read can
1636	 * just copy data from the pool.
1637	 */
1638	struct page *ra_pool[R5L_RECOVERY_PAGE_POOL_SIZE];
 
1639	sector_t pool_offset;	/* offset of first page in the pool */
1640	int total_pages;	/* total allocated pages */
1641	int valid_pages;	/* pages with valid data */
1642	struct bio *ra_bio;	/* bio to do the read ahead */
1643};
1644
1645static int r5l_recovery_allocate_ra_pool(struct r5l_log *log,
1646					    struct r5l_recovery_ctx *ctx)
1647{
1648	struct page *page;
1649
1650	ctx->ra_bio = bio_alloc_bioset(GFP_KERNEL, BIO_MAX_PAGES, log->bs);
1651	if (!ctx->ra_bio)
1652		return -ENOMEM;
1653
1654	ctx->valid_pages = 0;
1655	ctx->total_pages = 0;
1656	while (ctx->total_pages < R5L_RECOVERY_PAGE_POOL_SIZE) {
1657		page = alloc_page(GFP_KERNEL);
1658
1659		if (!page)
1660			break;
1661		ctx->ra_pool[ctx->total_pages] = page;
1662		ctx->total_pages += 1;
1663	}
1664
1665	if (ctx->total_pages == 0) {
1666		bio_put(ctx->ra_bio);
1667		return -ENOMEM;
1668	}
1669
1670	ctx->pool_offset = 0;
1671	return 0;
1672}
1673
1674static void r5l_recovery_free_ra_pool(struct r5l_log *log,
1675					struct r5l_recovery_ctx *ctx)
1676{
1677	int i;
1678
1679	for (i = 0; i < ctx->total_pages; ++i)
1680		put_page(ctx->ra_pool[i]);
1681	bio_put(ctx->ra_bio);
1682}
1683
1684/*
1685 * fetch ctx->valid_pages pages from offset
1686 * In normal cases, ctx->valid_pages == ctx->total_pages after the call.
1687 * However, if the offset is close to the end of the journal device,
1688 * ctx->valid_pages could be smaller than ctx->total_pages
1689 */
1690static int r5l_recovery_fetch_ra_pool(struct r5l_log *log,
1691				      struct r5l_recovery_ctx *ctx,
1692				      sector_t offset)
1693{
1694	bio_reset(ctx->ra_bio);
1695	bio_set_dev(ctx->ra_bio, log->rdev->bdev);
1696	bio_set_op_attrs(ctx->ra_bio, REQ_OP_READ, 0);
1697	ctx->ra_bio->bi_iter.bi_sector = log->rdev->data_offset + offset;
 
 
1698
1699	ctx->valid_pages = 0;
1700	ctx->pool_offset = offset;
1701
1702	while (ctx->valid_pages < ctx->total_pages) {
1703		bio_add_page(ctx->ra_bio,
1704			     ctx->ra_pool[ctx->valid_pages], PAGE_SIZE, 0);
1705		ctx->valid_pages += 1;
1706
1707		offset = r5l_ring_add(log, offset, BLOCK_SECTORS);
1708
1709		if (offset == 0)  /* reached end of the device */
1710			break;
1711	}
1712
1713	return submit_bio_wait(ctx->ra_bio);
 
 
1714}
1715
1716/*
1717 * try read a page from the read ahead page pool, if the page is not in the
1718 * pool, call r5l_recovery_fetch_ra_pool
1719 */
1720static int r5l_recovery_read_page(struct r5l_log *log,
1721				  struct r5l_recovery_ctx *ctx,
1722				  struct page *page,
1723				  sector_t offset)
1724{
1725	int ret;
1726
1727	if (offset < ctx->pool_offset ||
1728	    offset >= ctx->pool_offset + ctx->valid_pages * BLOCK_SECTORS) {
1729		ret = r5l_recovery_fetch_ra_pool(log, ctx, offset);
1730		if (ret)
1731			return ret;
1732	}
1733
1734	BUG_ON(offset < ctx->pool_offset ||
1735	       offset >= ctx->pool_offset + ctx->valid_pages * BLOCK_SECTORS);
1736
1737	memcpy(page_address(page),
1738	       page_address(ctx->ra_pool[(offset - ctx->pool_offset) >>
1739					 BLOCK_SECTOR_SHIFT]),
1740	       PAGE_SIZE);
1741	return 0;
1742}
1743
1744static int r5l_recovery_read_meta_block(struct r5l_log *log,
1745					struct r5l_recovery_ctx *ctx)
1746{
1747	struct page *page = ctx->meta_page;
1748	struct r5l_meta_block *mb;
1749	u32 crc, stored_crc;
1750	int ret;
1751
1752	ret = r5l_recovery_read_page(log, ctx, page, ctx->pos);
1753	if (ret != 0)
1754		return ret;
1755
1756	mb = page_address(page);
1757	stored_crc = le32_to_cpu(mb->checksum);
1758	mb->checksum = 0;
1759
1760	if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
1761	    le64_to_cpu(mb->seq) != ctx->seq ||
1762	    mb->version != R5LOG_VERSION ||
1763	    le64_to_cpu(mb->position) != ctx->pos)
1764		return -EINVAL;
1765
1766	crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
1767	if (stored_crc != crc)
1768		return -EINVAL;
1769
1770	if (le32_to_cpu(mb->meta_size) > PAGE_SIZE)
1771		return -EINVAL;
1772
1773	ctx->meta_total_blocks = BLOCK_SECTORS;
1774
1775	return 0;
1776}
1777
1778static void
1779r5l_recovery_create_empty_meta_block(struct r5l_log *log,
1780				     struct page *page,
1781				     sector_t pos, u64 seq)
1782{
1783	struct r5l_meta_block *mb;
1784
1785	mb = page_address(page);
1786	clear_page(mb);
1787	mb->magic = cpu_to_le32(R5LOG_MAGIC);
1788	mb->version = R5LOG_VERSION;
1789	mb->meta_size = cpu_to_le32(sizeof(struct r5l_meta_block));
1790	mb->seq = cpu_to_le64(seq);
1791	mb->position = cpu_to_le64(pos);
1792}
1793
1794static int r5l_log_write_empty_meta_block(struct r5l_log *log, sector_t pos,
1795					  u64 seq)
1796{
1797	struct page *page;
1798	struct r5l_meta_block *mb;
1799
1800	page = alloc_page(GFP_KERNEL);
1801	if (!page)
1802		return -ENOMEM;
1803	r5l_recovery_create_empty_meta_block(log, page, pos, seq);
1804	mb = page_address(page);
1805	mb->checksum = cpu_to_le32(crc32c_le(log->uuid_checksum,
1806					     mb, PAGE_SIZE));
1807	if (!sync_page_io(log->rdev, pos, PAGE_SIZE, page, REQ_OP_WRITE,
1808			  REQ_SYNC | REQ_FUA, false)) {
1809		__free_page(page);
1810		return -EIO;
1811	}
1812	__free_page(page);
1813	return 0;
1814}
1815
1816/*
1817 * r5l_recovery_load_data and r5l_recovery_load_parity uses flag R5_Wantwrite
1818 * to mark valid (potentially not flushed) data in the journal.
1819 *
1820 * We already verified checksum in r5l_recovery_verify_data_checksum_for_mb,
1821 * so there should not be any mismatch here.
1822 */
1823static void r5l_recovery_load_data(struct r5l_log *log,
1824				   struct stripe_head *sh,
1825				   struct r5l_recovery_ctx *ctx,
1826				   struct r5l_payload_data_parity *payload,
1827				   sector_t log_offset)
1828{
1829	struct mddev *mddev = log->rdev->mddev;
1830	struct r5conf *conf = mddev->private;
1831	int dd_idx;
1832
1833	raid5_compute_sector(conf,
1834			     le64_to_cpu(payload->location), 0,
1835			     &dd_idx, sh);
1836	r5l_recovery_read_page(log, ctx, sh->dev[dd_idx].page, log_offset);
1837	sh->dev[dd_idx].log_checksum =
1838		le32_to_cpu(payload->checksum[0]);
1839	ctx->meta_total_blocks += BLOCK_SECTORS;
1840
1841	set_bit(R5_Wantwrite, &sh->dev[dd_idx].flags);
1842	set_bit(STRIPE_R5C_CACHING, &sh->state);
1843}
1844
1845static void r5l_recovery_load_parity(struct r5l_log *log,
1846				     struct stripe_head *sh,
1847				     struct r5l_recovery_ctx *ctx,
1848				     struct r5l_payload_data_parity *payload,
1849				     sector_t log_offset)
1850{
1851	struct mddev *mddev = log->rdev->mddev;
1852	struct r5conf *conf = mddev->private;
1853
1854	ctx->meta_total_blocks += BLOCK_SECTORS * conf->max_degraded;
1855	r5l_recovery_read_page(log, ctx, sh->dev[sh->pd_idx].page, log_offset);
1856	sh->dev[sh->pd_idx].log_checksum =
1857		le32_to_cpu(payload->checksum[0]);
1858	set_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags);
1859
1860	if (sh->qd_idx >= 0) {
1861		r5l_recovery_read_page(
1862			log, ctx, sh->dev[sh->qd_idx].page,
1863			r5l_ring_add(log, log_offset, BLOCK_SECTORS));
1864		sh->dev[sh->qd_idx].log_checksum =
1865			le32_to_cpu(payload->checksum[1]);
1866		set_bit(R5_Wantwrite, &sh->dev[sh->qd_idx].flags);
1867	}
1868	clear_bit(STRIPE_R5C_CACHING, &sh->state);
1869}
1870
1871static void r5l_recovery_reset_stripe(struct stripe_head *sh)
1872{
1873	int i;
1874
1875	sh->state = 0;
1876	sh->log_start = MaxSector;
1877	for (i = sh->disks; i--; )
1878		sh->dev[i].flags = 0;
1879}
1880
1881static void
1882r5l_recovery_replay_one_stripe(struct r5conf *conf,
1883			       struct stripe_head *sh,
1884			       struct r5l_recovery_ctx *ctx)
1885{
1886	struct md_rdev *rdev, *rrdev;
1887	int disk_index;
1888	int data_count = 0;
1889
1890	for (disk_index = 0; disk_index < sh->disks; disk_index++) {
1891		if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
1892			continue;
1893		if (disk_index == sh->qd_idx || disk_index == sh->pd_idx)
1894			continue;
1895		data_count++;
1896	}
1897
1898	/*
1899	 * stripes that only have parity must have been flushed
1900	 * before the crash that we are now recovering from, so
1901	 * there is nothing more to recovery.
1902	 */
1903	if (data_count == 0)
1904		goto out;
1905
1906	for (disk_index = 0; disk_index < sh->disks; disk_index++) {
1907		if (!test_bit(R5_Wantwrite, &sh->dev[disk_index].flags))
1908			continue;
1909
1910		/* in case device is broken */
1911		rcu_read_lock();
1912		rdev = rcu_dereference(conf->disks[disk_index].rdev);
1913		if (rdev) {
1914			atomic_inc(&rdev->nr_pending);
1915			rcu_read_unlock();
1916			sync_page_io(rdev, sh->sector, PAGE_SIZE,
1917				     sh->dev[disk_index].page, REQ_OP_WRITE, 0,
1918				     false);
1919			rdev_dec_pending(rdev, rdev->mddev);
1920			rcu_read_lock();
1921		}
1922		rrdev = rcu_dereference(conf->disks[disk_index].replacement);
1923		if (rrdev) {
1924			atomic_inc(&rrdev->nr_pending);
1925			rcu_read_unlock();
1926			sync_page_io(rrdev, sh->sector, PAGE_SIZE,
1927				     sh->dev[disk_index].page, REQ_OP_WRITE, 0,
1928				     false);
1929			rdev_dec_pending(rrdev, rrdev->mddev);
1930			rcu_read_lock();
1931		}
1932		rcu_read_unlock();
1933	}
1934	ctx->data_parity_stripes++;
1935out:
1936	r5l_recovery_reset_stripe(sh);
1937}
1938
1939static struct stripe_head *
1940r5c_recovery_alloc_stripe(struct r5conf *conf,
1941			  sector_t stripe_sect)
 
 
1942{
1943	struct stripe_head *sh;
1944
1945	sh = raid5_get_active_stripe(conf, stripe_sect, 0, 1, 0);
 
1946	if (!sh)
1947		return NULL;  /* no more stripe available */
1948
1949	r5l_recovery_reset_stripe(sh);
1950
1951	return sh;
1952}
1953
1954static struct stripe_head *
1955r5c_recovery_lookup_stripe(struct list_head *list, sector_t sect)
1956{
1957	struct stripe_head *sh;
1958
1959	list_for_each_entry(sh, list, lru)
1960		if (sh->sector == sect)
1961			return sh;
1962	return NULL;
1963}
1964
1965static void
1966r5c_recovery_drop_stripes(struct list_head *cached_stripe_list,
1967			  struct r5l_recovery_ctx *ctx)
1968{
1969	struct stripe_head *sh, *next;
1970
1971	list_for_each_entry_safe(sh, next, cached_stripe_list, lru) {
1972		r5l_recovery_reset_stripe(sh);
1973		list_del_init(&sh->lru);
1974		raid5_release_stripe(sh);
1975	}
1976}
1977
1978static void
1979r5c_recovery_replay_stripes(struct list_head *cached_stripe_list,
1980			    struct r5l_recovery_ctx *ctx)
1981{
1982	struct stripe_head *sh, *next;
1983
1984	list_for_each_entry_safe(sh, next, cached_stripe_list, lru)
1985		if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) {
1986			r5l_recovery_replay_one_stripe(sh->raid_conf, sh, ctx);
1987			list_del_init(&sh->lru);
1988			raid5_release_stripe(sh);
1989		}
1990}
1991
1992/* if matches return 0; otherwise return -EINVAL */
1993static int
1994r5l_recovery_verify_data_checksum(struct r5l_log *log,
1995				  struct r5l_recovery_ctx *ctx,
1996				  struct page *page,
1997				  sector_t log_offset, __le32 log_checksum)
1998{
1999	void *addr;
2000	u32 checksum;
2001
2002	r5l_recovery_read_page(log, ctx, page, log_offset);
2003	addr = kmap_atomic(page);
2004	checksum = crc32c_le(log->uuid_checksum, addr, PAGE_SIZE);
2005	kunmap_atomic(addr);
2006	return (le32_to_cpu(log_checksum) == checksum) ? 0 : -EINVAL;
2007}
2008
2009/*
2010 * before loading data to stripe cache, we need verify checksum for all data,
2011 * if there is mismatch for any data page, we drop all data in the mata block
2012 */
2013static int
2014r5l_recovery_verify_data_checksum_for_mb(struct r5l_log *log,
2015					 struct r5l_recovery_ctx *ctx)
2016{
2017	struct mddev *mddev = log->rdev->mddev;
2018	struct r5conf *conf = mddev->private;
2019	struct r5l_meta_block *mb = page_address(ctx->meta_page);
2020	sector_t mb_offset = sizeof(struct r5l_meta_block);
2021	sector_t log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2022	struct page *page;
2023	struct r5l_payload_data_parity *payload;
2024	struct r5l_payload_flush *payload_flush;
2025
2026	page = alloc_page(GFP_KERNEL);
2027	if (!page)
2028		return -ENOMEM;
2029
2030	while (mb_offset < le32_to_cpu(mb->meta_size)) {
2031		payload = (void *)mb + mb_offset;
2032		payload_flush = (void *)mb + mb_offset;
2033
2034		if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) {
2035			if (r5l_recovery_verify_data_checksum(
2036				    log, ctx, page, log_offset,
2037				    payload->checksum[0]) < 0)
2038				goto mismatch;
2039		} else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_PARITY) {
2040			if (r5l_recovery_verify_data_checksum(
2041				    log, ctx, page, log_offset,
2042				    payload->checksum[0]) < 0)
2043				goto mismatch;
2044			if (conf->max_degraded == 2 && /* q for RAID 6 */
2045			    r5l_recovery_verify_data_checksum(
2046				    log, ctx, page,
2047				    r5l_ring_add(log, log_offset,
2048						 BLOCK_SECTORS),
2049				    payload->checksum[1]) < 0)
2050				goto mismatch;
2051		} else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
2052			/* nothing to do for R5LOG_PAYLOAD_FLUSH here */
2053		} else /* not R5LOG_PAYLOAD_DATA/PARITY/FLUSH */
2054			goto mismatch;
2055
2056		if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
2057			mb_offset += sizeof(struct r5l_payload_flush) +
2058				le32_to_cpu(payload_flush->size);
2059		} else {
2060			/* DATA or PARITY payload */
2061			log_offset = r5l_ring_add(log, log_offset,
2062						  le32_to_cpu(payload->size));
2063			mb_offset += sizeof(struct r5l_payload_data_parity) +
2064				sizeof(__le32) *
2065				(le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
2066		}
2067
2068	}
2069
2070	put_page(page);
2071	return 0;
2072
2073mismatch:
2074	put_page(page);
2075	return -EINVAL;
2076}
2077
2078/*
2079 * Analyze all data/parity pages in one meta block
2080 * Returns:
2081 * 0 for success
2082 * -EINVAL for unknown playload type
2083 * -EAGAIN for checksum mismatch of data page
2084 * -ENOMEM for run out of memory (alloc_page failed or run out of stripes)
2085 */
2086static int
2087r5c_recovery_analyze_meta_block(struct r5l_log *log,
2088				struct r5l_recovery_ctx *ctx,
2089				struct list_head *cached_stripe_list)
2090{
2091	struct mddev *mddev = log->rdev->mddev;
2092	struct r5conf *conf = mddev->private;
2093	struct r5l_meta_block *mb;
2094	struct r5l_payload_data_parity *payload;
2095	struct r5l_payload_flush *payload_flush;
2096	int mb_offset;
2097	sector_t log_offset;
2098	sector_t stripe_sect;
2099	struct stripe_head *sh;
2100	int ret;
2101
2102	/*
2103	 * for mismatch in data blocks, we will drop all data in this mb, but
2104	 * we will still read next mb for other data with FLUSH flag, as
2105	 * io_unit could finish out of order.
2106	 */
2107	ret = r5l_recovery_verify_data_checksum_for_mb(log, ctx);
2108	if (ret == -EINVAL)
2109		return -EAGAIN;
2110	else if (ret)
2111		return ret;   /* -ENOMEM duo to alloc_page() failed */
2112
2113	mb = page_address(ctx->meta_page);
2114	mb_offset = sizeof(struct r5l_meta_block);
2115	log_offset = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2116
2117	while (mb_offset < le32_to_cpu(mb->meta_size)) {
2118		int dd;
2119
2120		payload = (void *)mb + mb_offset;
2121		payload_flush = (void *)mb + mb_offset;
2122
2123		if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_FLUSH) {
2124			int i, count;
2125
2126			count = le32_to_cpu(payload_flush->size) / sizeof(__le64);
2127			for (i = 0; i < count; ++i) {
2128				stripe_sect = le64_to_cpu(payload_flush->flush_stripes[i]);
2129				sh = r5c_recovery_lookup_stripe(cached_stripe_list,
2130								stripe_sect);
2131				if (sh) {
2132					WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
2133					r5l_recovery_reset_stripe(sh);
2134					list_del_init(&sh->lru);
2135					raid5_release_stripe(sh);
2136				}
2137			}
2138
2139			mb_offset += sizeof(struct r5l_payload_flush) +
2140				le32_to_cpu(payload_flush->size);
2141			continue;
2142		}
2143
2144		/* DATA or PARITY payload */
2145		stripe_sect = (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) ?
2146			raid5_compute_sector(
2147				conf, le64_to_cpu(payload->location), 0, &dd,
2148				NULL)
2149			: le64_to_cpu(payload->location);
2150
2151		sh = r5c_recovery_lookup_stripe(cached_stripe_list,
2152						stripe_sect);
2153
2154		if (!sh) {
2155			sh = r5c_recovery_alloc_stripe(conf, stripe_sect);
2156			/*
2157			 * cannot get stripe from raid5_get_active_stripe
2158			 * try replay some stripes
2159			 */
2160			if (!sh) {
2161				r5c_recovery_replay_stripes(
2162					cached_stripe_list, ctx);
2163				sh = r5c_recovery_alloc_stripe(
2164					conf, stripe_sect);
2165			}
2166			if (!sh) {
 
2167				pr_debug("md/raid:%s: Increasing stripe cache size to %d to recovery data on journal.\n",
2168					mdname(mddev),
2169					conf->min_nr_stripes * 2);
2170				raid5_set_cache_size(mddev,
2171						     conf->min_nr_stripes * 2);
2172				sh = r5c_recovery_alloc_stripe(conf,
2173							       stripe_sect);
 
 
 
 
 
 
 
 
2174			}
2175			if (!sh) {
2176				pr_err("md/raid:%s: Cannot get enough stripes due to memory pressure. Recovery failed.\n",
2177				       mdname(mddev));
2178				return -ENOMEM;
2179			}
2180			list_add_tail(&sh->lru, cached_stripe_list);
2181		}
2182
2183		if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_DATA) {
2184			if (!test_bit(STRIPE_R5C_CACHING, &sh->state) &&
2185			    test_bit(R5_Wantwrite, &sh->dev[sh->pd_idx].flags)) {
2186				r5l_recovery_replay_one_stripe(conf, sh, ctx);
2187				list_move_tail(&sh->lru, cached_stripe_list);
2188			}
2189			r5l_recovery_load_data(log, sh, ctx, payload,
2190					       log_offset);
2191		} else if (le16_to_cpu(payload->header.type) == R5LOG_PAYLOAD_PARITY)
2192			r5l_recovery_load_parity(log, sh, ctx, payload,
2193						 log_offset);
2194		else
2195			return -EINVAL;
2196
2197		log_offset = r5l_ring_add(log, log_offset,
2198					  le32_to_cpu(payload->size));
2199
2200		mb_offset += sizeof(struct r5l_payload_data_parity) +
2201			sizeof(__le32) *
2202			(le32_to_cpu(payload->size) >> (PAGE_SHIFT - 9));
2203	}
2204
2205	return 0;
2206}
2207
2208/*
2209 * Load the stripe into cache. The stripe will be written out later by
2210 * the stripe cache state machine.
2211 */
2212static void r5c_recovery_load_one_stripe(struct r5l_log *log,
2213					 struct stripe_head *sh)
2214{
2215	struct r5dev *dev;
2216	int i;
2217
2218	for (i = sh->disks; i--; ) {
2219		dev = sh->dev + i;
2220		if (test_and_clear_bit(R5_Wantwrite, &dev->flags)) {
2221			set_bit(R5_InJournal, &dev->flags);
2222			set_bit(R5_UPTODATE, &dev->flags);
2223		}
2224	}
2225}
2226
2227/*
2228 * Scan through the log for all to-be-flushed data
2229 *
2230 * For stripes with data and parity, namely Data-Parity stripe
2231 * (STRIPE_R5C_CACHING == 0), we simply replay all the writes.
2232 *
2233 * For stripes with only data, namely Data-Only stripe
2234 * (STRIPE_R5C_CACHING == 1), we load them to stripe cache state machine.
2235 *
2236 * For a stripe, if we see data after parity, we should discard all previous
2237 * data and parity for this stripe, as these data are already flushed to
2238 * the array.
2239 *
2240 * At the end of the scan, we return the new journal_tail, which points to
2241 * first data-only stripe on the journal device, or next invalid meta block.
2242 */
2243static int r5c_recovery_flush_log(struct r5l_log *log,
2244				  struct r5l_recovery_ctx *ctx)
2245{
2246	struct stripe_head *sh;
2247	int ret = 0;
2248
2249	/* scan through the log */
2250	while (1) {
2251		if (r5l_recovery_read_meta_block(log, ctx))
2252			break;
2253
2254		ret = r5c_recovery_analyze_meta_block(log, ctx,
2255						      &ctx->cached_list);
2256		/*
2257		 * -EAGAIN means mismatch in data block, in this case, we still
2258		 * try scan the next metablock
2259		 */
2260		if (ret && ret != -EAGAIN)
2261			break;   /* ret == -EINVAL or -ENOMEM */
2262		ctx->seq++;
2263		ctx->pos = r5l_ring_add(log, ctx->pos, ctx->meta_total_blocks);
2264	}
2265
2266	if (ret == -ENOMEM) {
2267		r5c_recovery_drop_stripes(&ctx->cached_list, ctx);
2268		return ret;
2269	}
2270
2271	/* replay data-parity stripes */
2272	r5c_recovery_replay_stripes(&ctx->cached_list, ctx);
2273
2274	/* load data-only stripes to stripe cache */
2275	list_for_each_entry(sh, &ctx->cached_list, lru) {
2276		WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
2277		r5c_recovery_load_one_stripe(log, sh);
2278		ctx->data_only_stripes++;
2279	}
2280
2281	return 0;
2282}
2283
2284/*
2285 * we did a recovery. Now ctx.pos points to an invalid meta block. New
2286 * log will start here. but we can't let superblock point to last valid
2287 * meta block. The log might looks like:
2288 * | meta 1| meta 2| meta 3|
2289 * meta 1 is valid, meta 2 is invalid. meta 3 could be valid. If
2290 * superblock points to meta 1, we write a new valid meta 2n.  if crash
2291 * happens again, new recovery will start from meta 1. Since meta 2n is
2292 * valid now, recovery will think meta 3 is valid, which is wrong.
2293 * The solution is we create a new meta in meta2 with its seq == meta
2294 * 1's seq + 10000 and let superblock points to meta2. The same recovery
2295 * will not think meta 3 is a valid meta, because its seq doesn't match
2296 */
2297
2298/*
2299 * Before recovery, the log looks like the following
2300 *
2301 *   ---------------------------------------------
2302 *   |           valid log        | invalid log  |
2303 *   ---------------------------------------------
2304 *   ^
2305 *   |- log->last_checkpoint
2306 *   |- log->last_cp_seq
2307 *
2308 * Now we scan through the log until we see invalid entry
2309 *
2310 *   ---------------------------------------------
2311 *   |           valid log        | invalid log  |
2312 *   ---------------------------------------------
2313 *   ^                            ^
2314 *   |- log->last_checkpoint      |- ctx->pos
2315 *   |- log->last_cp_seq          |- ctx->seq
2316 *
2317 * From this point, we need to increase seq number by 10 to avoid
2318 * confusing next recovery.
2319 *
2320 *   ---------------------------------------------
2321 *   |           valid log        | invalid log  |
2322 *   ---------------------------------------------
2323 *   ^                              ^
2324 *   |- log->last_checkpoint        |- ctx->pos+1
2325 *   |- log->last_cp_seq            |- ctx->seq+10001
2326 *
2327 * However, it is not safe to start the state machine yet, because data only
2328 * parities are not yet secured in RAID. To save these data only parities, we
2329 * rewrite them from seq+11.
2330 *
2331 *   -----------------------------------------------------------------
2332 *   |           valid log        | data only stripes | invalid log  |
2333 *   -----------------------------------------------------------------
2334 *   ^                                                ^
2335 *   |- log->last_checkpoint                          |- ctx->pos+n
2336 *   |- log->last_cp_seq                              |- ctx->seq+10000+n
2337 *
2338 * If failure happens again during this process, the recovery can safe start
2339 * again from log->last_checkpoint.
2340 *
2341 * Once data only stripes are rewritten to journal, we move log_tail
2342 *
2343 *   -----------------------------------------------------------------
2344 *   |     old log        |    data only stripes    | invalid log  |
2345 *   -----------------------------------------------------------------
2346 *                        ^                         ^
2347 *                        |- log->last_checkpoint   |- ctx->pos+n
2348 *                        |- log->last_cp_seq       |- ctx->seq+10000+n
2349 *
2350 * Then we can safely start the state machine. If failure happens from this
2351 * point on, the recovery will start from new log->last_checkpoint.
2352 */
2353static int
2354r5c_recovery_rewrite_data_only_stripes(struct r5l_log *log,
2355				       struct r5l_recovery_ctx *ctx)
2356{
2357	struct stripe_head *sh;
2358	struct mddev *mddev = log->rdev->mddev;
2359	struct page *page;
2360	sector_t next_checkpoint = MaxSector;
2361
2362	page = alloc_page(GFP_KERNEL);
2363	if (!page) {
2364		pr_err("md/raid:%s: cannot allocate memory to rewrite data only stripes\n",
2365		       mdname(mddev));
2366		return -ENOMEM;
2367	}
2368
2369	WARN_ON(list_empty(&ctx->cached_list));
2370
2371	list_for_each_entry(sh, &ctx->cached_list, lru) {
2372		struct r5l_meta_block *mb;
2373		int i;
2374		int offset;
2375		sector_t write_pos;
2376
2377		WARN_ON(!test_bit(STRIPE_R5C_CACHING, &sh->state));
2378		r5l_recovery_create_empty_meta_block(log, page,
2379						     ctx->pos, ctx->seq);
2380		mb = page_address(page);
2381		offset = le32_to_cpu(mb->meta_size);
2382		write_pos = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2383
2384		for (i = sh->disks; i--; ) {
2385			struct r5dev *dev = &sh->dev[i];
2386			struct r5l_payload_data_parity *payload;
2387			void *addr;
2388
2389			if (test_bit(R5_InJournal, &dev->flags)) {
2390				payload = (void *)mb + offset;
2391				payload->header.type = cpu_to_le16(
2392					R5LOG_PAYLOAD_DATA);
2393				payload->size = cpu_to_le32(BLOCK_SECTORS);
2394				payload->location = cpu_to_le64(
2395					raid5_compute_blocknr(sh, i, 0));
2396				addr = kmap_atomic(dev->page);
2397				payload->checksum[0] = cpu_to_le32(
2398					crc32c_le(log->uuid_checksum, addr,
2399						  PAGE_SIZE));
2400				kunmap_atomic(addr);
2401				sync_page_io(log->rdev, write_pos, PAGE_SIZE,
2402					     dev->page, REQ_OP_WRITE, 0, false);
2403				write_pos = r5l_ring_add(log, write_pos,
2404							 BLOCK_SECTORS);
2405				offset += sizeof(__le32) +
2406					sizeof(struct r5l_payload_data_parity);
2407
2408			}
2409		}
2410		mb->meta_size = cpu_to_le32(offset);
2411		mb->checksum = cpu_to_le32(crc32c_le(log->uuid_checksum,
2412						     mb, PAGE_SIZE));
2413		sync_page_io(log->rdev, ctx->pos, PAGE_SIZE, page,
2414			     REQ_OP_WRITE, REQ_SYNC | REQ_FUA, false);
2415		sh->log_start = ctx->pos;
2416		list_add_tail(&sh->r5c, &log->stripe_in_journal_list);
2417		atomic_inc(&log->stripe_in_journal_count);
2418		ctx->pos = write_pos;
2419		ctx->seq += 1;
2420		next_checkpoint = sh->log_start;
2421	}
2422	log->next_checkpoint = next_checkpoint;
2423	__free_page(page);
2424	return 0;
2425}
2426
2427static void r5c_recovery_flush_data_only_stripes(struct r5l_log *log,
2428						 struct r5l_recovery_ctx *ctx)
2429{
2430	struct mddev *mddev = log->rdev->mddev;
2431	struct r5conf *conf = mddev->private;
2432	struct stripe_head *sh, *next;
 
2433
2434	if (ctx->data_only_stripes == 0)
2435		return;
2436
 
 
 
 
2437	log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_BACK;
2438
2439	list_for_each_entry_safe(sh, next, &ctx->cached_list, lru) {
2440		r5c_make_stripe_write_out(sh);
2441		set_bit(STRIPE_HANDLE, &sh->state);
2442		list_del_init(&sh->lru);
2443		raid5_release_stripe(sh);
2444	}
2445
2446	/* reuse conf->wait_for_quiescent in recovery */
2447	wait_event(conf->wait_for_quiescent,
2448		   atomic_read(&conf->active_stripes) == 0);
2449
2450	log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
 
 
2451}
2452
2453static int r5l_recovery_log(struct r5l_log *log)
2454{
2455	struct mddev *mddev = log->rdev->mddev;
2456	struct r5l_recovery_ctx *ctx;
2457	int ret;
2458	sector_t pos;
2459
2460	ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
2461	if (!ctx)
2462		return -ENOMEM;
2463
2464	ctx->pos = log->last_checkpoint;
2465	ctx->seq = log->last_cp_seq;
2466	INIT_LIST_HEAD(&ctx->cached_list);
2467	ctx->meta_page = alloc_page(GFP_KERNEL);
2468
2469	if (!ctx->meta_page) {
2470		ret =  -ENOMEM;
2471		goto meta_page;
2472	}
2473
2474	if (r5l_recovery_allocate_ra_pool(log, ctx) != 0) {
2475		ret = -ENOMEM;
2476		goto ra_pool;
2477	}
2478
2479	ret = r5c_recovery_flush_log(log, ctx);
2480
2481	if (ret)
2482		goto error;
2483
2484	pos = ctx->pos;
2485	ctx->seq += 10000;
2486
2487	if ((ctx->data_only_stripes == 0) && (ctx->data_parity_stripes == 0))
2488		pr_info("md/raid:%s: starting from clean shutdown\n",
2489			 mdname(mddev));
2490	else
2491		pr_info("md/raid:%s: recovering %d data-only stripes and %d data-parity stripes\n",
2492			 mdname(mddev), ctx->data_only_stripes,
2493			 ctx->data_parity_stripes);
2494
2495	if (ctx->data_only_stripes == 0) {
2496		log->next_checkpoint = ctx->pos;
2497		r5l_log_write_empty_meta_block(log, ctx->pos, ctx->seq++);
2498		ctx->pos = r5l_ring_add(log, ctx->pos, BLOCK_SECTORS);
2499	} else if (r5c_recovery_rewrite_data_only_stripes(log, ctx)) {
2500		pr_err("md/raid:%s: failed to rewrite stripes to journal\n",
2501		       mdname(mddev));
2502		ret =  -EIO;
2503		goto error;
2504	}
2505
2506	log->log_start = ctx->pos;
2507	log->seq = ctx->seq;
2508	log->last_checkpoint = pos;
2509	r5l_write_super(log, pos);
2510
2511	r5c_recovery_flush_data_only_stripes(log, ctx);
2512	ret = 0;
2513error:
2514	r5l_recovery_free_ra_pool(log, ctx);
2515ra_pool:
2516	__free_page(ctx->meta_page);
2517meta_page:
2518	kfree(ctx);
2519	return ret;
2520}
2521
2522static void r5l_write_super(struct r5l_log *log, sector_t cp)
2523{
2524	struct mddev *mddev = log->rdev->mddev;
2525
2526	log->rdev->journal_tail = cp;
2527	set_bit(MD_SB_CHANGE_DEVS, &mddev->sb_flags);
2528}
2529
2530static ssize_t r5c_journal_mode_show(struct mddev *mddev, char *page)
2531{
2532	struct r5conf *conf;
2533	int ret;
2534
2535	ret = mddev_lock(mddev);
2536	if (ret)
2537		return ret;
2538
2539	conf = mddev->private;
2540	if (!conf || !conf->log) {
2541		mddev_unlock(mddev);
2542		return 0;
2543	}
2544
2545	switch (conf->log->r5c_journal_mode) {
2546	case R5C_JOURNAL_MODE_WRITE_THROUGH:
2547		ret = snprintf(
2548			page, PAGE_SIZE, "[%s] %s\n",
2549			r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH],
2550			r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]);
2551		break;
2552	case R5C_JOURNAL_MODE_WRITE_BACK:
2553		ret = snprintf(
2554			page, PAGE_SIZE, "%s [%s]\n",
2555			r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_THROUGH],
2556			r5c_journal_mode_str[R5C_JOURNAL_MODE_WRITE_BACK]);
2557		break;
2558	default:
2559		ret = 0;
2560	}
 
 
2561	mddev_unlock(mddev);
2562	return ret;
2563}
2564
2565/*
2566 * Set journal cache mode on @mddev (external API initially needed by dm-raid).
2567 *
2568 * @mode as defined in 'enum r5c_journal_mode'.
2569 *
2570 */
2571int r5c_journal_mode_set(struct mddev *mddev, int mode)
2572{
2573	struct r5conf *conf;
2574
2575	if (mode < R5C_JOURNAL_MODE_WRITE_THROUGH ||
2576	    mode > R5C_JOURNAL_MODE_WRITE_BACK)
2577		return -EINVAL;
2578
2579	conf = mddev->private;
2580	if (!conf || !conf->log)
2581		return -ENODEV;
2582
2583	if (raid5_calc_degraded(conf) > 0 &&
2584	    mode == R5C_JOURNAL_MODE_WRITE_BACK)
2585		return -EINVAL;
2586
2587	mddev_suspend(mddev);
2588	conf->log->r5c_journal_mode = mode;
2589	mddev_resume(mddev);
2590
2591	pr_debug("md/raid:%s: setting r5c cache mode to %d: %s\n",
2592		 mdname(mddev), mode, r5c_journal_mode_str[mode]);
2593	return 0;
2594}
2595EXPORT_SYMBOL(r5c_journal_mode_set);
2596
2597static ssize_t r5c_journal_mode_store(struct mddev *mddev,
2598				      const char *page, size_t length)
2599{
2600	int mode = ARRAY_SIZE(r5c_journal_mode_str);
2601	size_t len = length;
2602	int ret;
2603
2604	if (len < 2)
2605		return -EINVAL;
2606
2607	if (page[len - 1] == '\n')
2608		len--;
2609
2610	while (mode--)
2611		if (strlen(r5c_journal_mode_str[mode]) == len &&
2612		    !strncmp(page, r5c_journal_mode_str[mode], len))
2613			break;
2614	ret = mddev_lock(mddev);
2615	if (ret)
2616		return ret;
2617	ret = r5c_journal_mode_set(mddev, mode);
2618	mddev_unlock(mddev);
2619	return ret ?: length;
2620}
2621
2622struct md_sysfs_entry
2623r5c_journal_mode = __ATTR(journal_mode, 0644,
2624			  r5c_journal_mode_show, r5c_journal_mode_store);
2625
2626/*
2627 * Try handle write operation in caching phase. This function should only
2628 * be called in write-back mode.
2629 *
2630 * If all outstanding writes can be handled in caching phase, returns 0
2631 * If writes requires write-out phase, call r5c_make_stripe_write_out()
2632 * and returns -EAGAIN
2633 */
2634int r5c_try_caching_write(struct r5conf *conf,
2635			  struct stripe_head *sh,
2636			  struct stripe_head_state *s,
2637			  int disks)
2638{
2639	struct r5l_log *log = conf->log;
2640	int i;
2641	struct r5dev *dev;
2642	int to_cache = 0;
2643	void **pslot;
2644	sector_t tree_index;
2645	int ret;
2646	uintptr_t refcount;
2647
2648	BUG_ON(!r5c_is_writeback(log));
2649
2650	if (!test_bit(STRIPE_R5C_CACHING, &sh->state)) {
2651		/*
2652		 * There are two different scenarios here:
2653		 *  1. The stripe has some data cached, and it is sent to
2654		 *     write-out phase for reclaim
2655		 *  2. The stripe is clean, and this is the first write
2656		 *
2657		 * For 1, return -EAGAIN, so we continue with
2658		 * handle_stripe_dirtying().
2659		 *
2660		 * For 2, set STRIPE_R5C_CACHING and continue with caching
2661		 * write.
2662		 */
2663
2664		/* case 1: anything injournal or anything in written */
2665		if (s->injournal > 0 || s->written > 0)
2666			return -EAGAIN;
2667		/* case 2 */
2668		set_bit(STRIPE_R5C_CACHING, &sh->state);
2669	}
2670
2671	/*
2672	 * When run in degraded mode, array is set to write-through mode.
2673	 * This check helps drain pending write safely in the transition to
2674	 * write-through mode.
2675	 *
2676	 * When a stripe is syncing, the write is also handled in write
2677	 * through mode.
2678	 */
2679	if (s->failed || test_bit(STRIPE_SYNCING, &sh->state)) {
2680		r5c_make_stripe_write_out(sh);
2681		return -EAGAIN;
2682	}
2683
2684	for (i = disks; i--; ) {
2685		dev = &sh->dev[i];
2686		/* if non-overwrite, use writing-out phase */
2687		if (dev->towrite && !test_bit(R5_OVERWRITE, &dev->flags) &&
2688		    !test_bit(R5_InJournal, &dev->flags)) {
2689			r5c_make_stripe_write_out(sh);
2690			return -EAGAIN;
2691		}
2692	}
2693
2694	/* if the stripe is not counted in big_stripe_tree, add it now */
2695	if (!test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) &&
2696	    !test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2697		tree_index = r5c_tree_index(conf, sh->sector);
2698		spin_lock(&log->tree_lock);
2699		pslot = radix_tree_lookup_slot(&log->big_stripe_tree,
2700					       tree_index);
2701		if (pslot) {
2702			refcount = (uintptr_t)radix_tree_deref_slot_protected(
2703				pslot, &log->tree_lock) >>
2704				R5C_RADIX_COUNT_SHIFT;
2705			radix_tree_replace_slot(
2706				&log->big_stripe_tree, pslot,
2707				(void *)((refcount + 1) << R5C_RADIX_COUNT_SHIFT));
2708		} else {
2709			/*
2710			 * this radix_tree_insert can fail safely, so no
2711			 * need to call radix_tree_preload()
2712			 */
2713			ret = radix_tree_insert(
2714				&log->big_stripe_tree, tree_index,
2715				(void *)(1 << R5C_RADIX_COUNT_SHIFT));
2716			if (ret) {
2717				spin_unlock(&log->tree_lock);
2718				r5c_make_stripe_write_out(sh);
2719				return -EAGAIN;
2720			}
2721		}
2722		spin_unlock(&log->tree_lock);
2723
2724		/*
2725		 * set STRIPE_R5C_PARTIAL_STRIPE, this shows the stripe is
2726		 * counted in the radix tree
2727		 */
2728		set_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state);
2729		atomic_inc(&conf->r5c_cached_partial_stripes);
2730	}
2731
2732	for (i = disks; i--; ) {
2733		dev = &sh->dev[i];
2734		if (dev->towrite) {
2735			set_bit(R5_Wantwrite, &dev->flags);
2736			set_bit(R5_Wantdrain, &dev->flags);
2737			set_bit(R5_LOCKED, &dev->flags);
2738			to_cache++;
2739		}
2740	}
2741
2742	if (to_cache) {
2743		set_bit(STRIPE_OP_BIODRAIN, &s->ops_request);
2744		/*
2745		 * set STRIPE_LOG_TRAPPED, which triggers r5c_cache_data()
2746		 * in ops_run_io(). STRIPE_LOG_TRAPPED will be cleared in
2747		 * r5c_handle_data_cached()
2748		 */
2749		set_bit(STRIPE_LOG_TRAPPED, &sh->state);
2750	}
2751
2752	return 0;
2753}
2754
2755/*
2756 * free extra pages (orig_page) we allocated for prexor
2757 */
2758void r5c_release_extra_page(struct stripe_head *sh)
2759{
2760	struct r5conf *conf = sh->raid_conf;
2761	int i;
2762	bool using_disk_info_extra_page;
2763
2764	using_disk_info_extra_page =
2765		sh->dev[0].orig_page == conf->disks[0].extra_page;
2766
2767	for (i = sh->disks; i--; )
2768		if (sh->dev[i].page != sh->dev[i].orig_page) {
2769			struct page *p = sh->dev[i].orig_page;
2770
2771			sh->dev[i].orig_page = sh->dev[i].page;
2772			clear_bit(R5_OrigPageUPTDODATE, &sh->dev[i].flags);
2773
2774			if (!using_disk_info_extra_page)
2775				put_page(p);
2776		}
2777
2778	if (using_disk_info_extra_page) {
2779		clear_bit(R5C_EXTRA_PAGE_IN_USE, &conf->cache_state);
2780		md_wakeup_thread(conf->mddev->thread);
2781	}
2782}
2783
2784void r5c_use_extra_page(struct stripe_head *sh)
2785{
2786	struct r5conf *conf = sh->raid_conf;
2787	int i;
2788	struct r5dev *dev;
2789
2790	for (i = sh->disks; i--; ) {
2791		dev = &sh->dev[i];
2792		if (dev->orig_page != dev->page)
2793			put_page(dev->orig_page);
2794		dev->orig_page = conf->disks[i].extra_page;
2795	}
2796}
2797
2798/*
2799 * clean up the stripe (clear R5_InJournal for dev[pd_idx] etc.) after the
2800 * stripe is committed to RAID disks.
2801 */
2802void r5c_finish_stripe_write_out(struct r5conf *conf,
2803				 struct stripe_head *sh,
2804				 struct stripe_head_state *s)
2805{
2806	struct r5l_log *log = conf->log;
2807	int i;
2808	int do_wakeup = 0;
2809	sector_t tree_index;
2810	void **pslot;
2811	uintptr_t refcount;
2812
2813	if (!log || !test_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags))
2814		return;
2815
2816	WARN_ON(test_bit(STRIPE_R5C_CACHING, &sh->state));
2817	clear_bit(R5_InJournal, &sh->dev[sh->pd_idx].flags);
2818
2819	if (log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_THROUGH)
2820		return;
2821
2822	for (i = sh->disks; i--; ) {
2823		clear_bit(R5_InJournal, &sh->dev[i].flags);
2824		if (test_and_clear_bit(R5_Overlap, &sh->dev[i].flags))
2825			do_wakeup = 1;
2826	}
2827
2828	/*
2829	 * analyse_stripe() runs before r5c_finish_stripe_write_out(),
2830	 * We updated R5_InJournal, so we also update s->injournal.
2831	 */
2832	s->injournal = 0;
2833
2834	if (test_and_clear_bit(STRIPE_FULL_WRITE, &sh->state))
2835		if (atomic_dec_and_test(&conf->pending_full_writes))
2836			md_wakeup_thread(conf->mddev->thread);
2837
2838	if (do_wakeup)
2839		wake_up(&conf->wait_for_overlap);
2840
2841	spin_lock_irq(&log->stripe_in_journal_lock);
2842	list_del_init(&sh->r5c);
2843	spin_unlock_irq(&log->stripe_in_journal_lock);
2844	sh->log_start = MaxSector;
2845
2846	atomic_dec(&log->stripe_in_journal_count);
2847	r5c_update_log_state(log);
2848
2849	/* stop counting this stripe in big_stripe_tree */
2850	if (test_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state) ||
2851	    test_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2852		tree_index = r5c_tree_index(conf, sh->sector);
2853		spin_lock(&log->tree_lock);
2854		pslot = radix_tree_lookup_slot(&log->big_stripe_tree,
2855					       tree_index);
2856		BUG_ON(pslot == NULL);
2857		refcount = (uintptr_t)radix_tree_deref_slot_protected(
2858			pslot, &log->tree_lock) >>
2859			R5C_RADIX_COUNT_SHIFT;
2860		if (refcount == 1)
2861			radix_tree_delete(&log->big_stripe_tree, tree_index);
2862		else
2863			radix_tree_replace_slot(
2864				&log->big_stripe_tree, pslot,
2865				(void *)((refcount - 1) << R5C_RADIX_COUNT_SHIFT));
2866		spin_unlock(&log->tree_lock);
2867	}
2868
2869	if (test_and_clear_bit(STRIPE_R5C_PARTIAL_STRIPE, &sh->state)) {
2870		BUG_ON(atomic_read(&conf->r5c_cached_partial_stripes) == 0);
2871		atomic_dec(&conf->r5c_flushing_partial_stripes);
2872		atomic_dec(&conf->r5c_cached_partial_stripes);
2873	}
2874
2875	if (test_and_clear_bit(STRIPE_R5C_FULL_STRIPE, &sh->state)) {
2876		BUG_ON(atomic_read(&conf->r5c_cached_full_stripes) == 0);
2877		atomic_dec(&conf->r5c_flushing_full_stripes);
2878		atomic_dec(&conf->r5c_cached_full_stripes);
2879	}
2880
2881	r5l_append_flush_payload(log, sh->sector);
2882	/* stripe is flused to raid disks, we can do resync now */
2883	if (test_bit(STRIPE_SYNC_REQUESTED, &sh->state))
2884		set_bit(STRIPE_HANDLE, &sh->state);
2885}
2886
2887int r5c_cache_data(struct r5l_log *log, struct stripe_head *sh)
2888{
2889	struct r5conf *conf = sh->raid_conf;
2890	int pages = 0;
2891	int reserve;
2892	int i;
2893	int ret = 0;
2894
2895	BUG_ON(!log);
2896
2897	for (i = 0; i < sh->disks; i++) {
2898		void *addr;
2899
2900		if (!test_bit(R5_Wantwrite, &sh->dev[i].flags))
2901			continue;
2902		addr = kmap_atomic(sh->dev[i].page);
2903		sh->dev[i].log_checksum = crc32c_le(log->uuid_checksum,
2904						    addr, PAGE_SIZE);
2905		kunmap_atomic(addr);
2906		pages++;
2907	}
2908	WARN_ON(pages == 0);
2909
2910	/*
2911	 * The stripe must enter state machine again to call endio, so
2912	 * don't delay.
2913	 */
2914	clear_bit(STRIPE_DELAYED, &sh->state);
2915	atomic_inc(&sh->count);
2916
2917	mutex_lock(&log->io_mutex);
2918	/* meta + data */
2919	reserve = (1 + pages) << (PAGE_SHIFT - 9);
2920
2921	if (test_bit(R5C_LOG_CRITICAL, &conf->cache_state) &&
2922	    sh->log_start == MaxSector)
2923		r5l_add_no_space_stripe(log, sh);
2924	else if (!r5l_has_free_space(log, reserve)) {
2925		if (sh->log_start == log->last_checkpoint)
2926			BUG();
2927		else
2928			r5l_add_no_space_stripe(log, sh);
2929	} else {
2930		ret = r5l_log_stripe(log, sh, pages, 0);
2931		if (ret) {
2932			spin_lock_irq(&log->io_list_lock);
2933			list_add_tail(&sh->log_list, &log->no_mem_stripes);
2934			spin_unlock_irq(&log->io_list_lock);
2935		}
2936	}
2937
2938	mutex_unlock(&log->io_mutex);
2939	return 0;
2940}
2941
2942/* check whether this big stripe is in write back cache. */
2943bool r5c_big_stripe_cached(struct r5conf *conf, sector_t sect)
2944{
2945	struct r5l_log *log = conf->log;
2946	sector_t tree_index;
2947	void *slot;
2948
2949	if (!log)
2950		return false;
2951
2952	WARN_ON_ONCE(!rcu_read_lock_held());
2953	tree_index = r5c_tree_index(conf, sect);
2954	slot = radix_tree_lookup(&log->big_stripe_tree, tree_index);
2955	return slot != NULL;
2956}
2957
2958static int r5l_load_log(struct r5l_log *log)
2959{
2960	struct md_rdev *rdev = log->rdev;
2961	struct page *page;
2962	struct r5l_meta_block *mb;
2963	sector_t cp = log->rdev->journal_tail;
2964	u32 stored_crc, expected_crc;
2965	bool create_super = false;
2966	int ret = 0;
2967
2968	/* Make sure it's valid */
2969	if (cp >= rdev->sectors || round_down(cp, BLOCK_SECTORS) != cp)
2970		cp = 0;
2971	page = alloc_page(GFP_KERNEL);
2972	if (!page)
2973		return -ENOMEM;
2974
2975	if (!sync_page_io(rdev, cp, PAGE_SIZE, page, REQ_OP_READ, 0, false)) {
2976		ret = -EIO;
2977		goto ioerr;
2978	}
2979	mb = page_address(page);
2980
2981	if (le32_to_cpu(mb->magic) != R5LOG_MAGIC ||
2982	    mb->version != R5LOG_VERSION) {
2983		create_super = true;
2984		goto create;
2985	}
2986	stored_crc = le32_to_cpu(mb->checksum);
2987	mb->checksum = 0;
2988	expected_crc = crc32c_le(log->uuid_checksum, mb, PAGE_SIZE);
2989	if (stored_crc != expected_crc) {
2990		create_super = true;
2991		goto create;
2992	}
2993	if (le64_to_cpu(mb->position) != cp) {
2994		create_super = true;
2995		goto create;
2996	}
2997create:
2998	if (create_super) {
2999		log->last_cp_seq = prandom_u32();
3000		cp = 0;
3001		r5l_log_write_empty_meta_block(log, cp, log->last_cp_seq);
3002		/*
3003		 * Make sure super points to correct address. Log might have
3004		 * data very soon. If super hasn't correct log tail address,
3005		 * recovery can't find the log
3006		 */
3007		r5l_write_super(log, cp);
3008	} else
3009		log->last_cp_seq = le64_to_cpu(mb->seq);
3010
3011	log->device_size = round_down(rdev->sectors, BLOCK_SECTORS);
3012	log->max_free_space = log->device_size >> RECLAIM_MAX_FREE_SPACE_SHIFT;
3013	if (log->max_free_space > RECLAIM_MAX_FREE_SPACE)
3014		log->max_free_space = RECLAIM_MAX_FREE_SPACE;
3015	log->last_checkpoint = cp;
3016
3017	__free_page(page);
3018
3019	if (create_super) {
3020		log->log_start = r5l_ring_add(log, cp, BLOCK_SECTORS);
3021		log->seq = log->last_cp_seq + 1;
3022		log->next_checkpoint = cp;
3023	} else
3024		ret = r5l_recovery_log(log);
3025
3026	r5c_update_log_state(log);
3027	return ret;
3028ioerr:
3029	__free_page(page);
3030	return ret;
3031}
3032
3033int r5l_start(struct r5l_log *log)
3034{
3035	int ret;
3036
3037	if (!log)
3038		return 0;
3039
3040	ret = r5l_load_log(log);
3041	if (ret) {
3042		struct mddev *mddev = log->rdev->mddev;
3043		struct r5conf *conf = mddev->private;
3044
3045		r5l_exit_log(conf);
3046	}
3047	return ret;
3048}
3049
3050void r5c_update_on_rdev_error(struct mddev *mddev, struct md_rdev *rdev)
3051{
3052	struct r5conf *conf = mddev->private;
3053	struct r5l_log *log = conf->log;
3054
3055	if (!log)
3056		return;
3057
3058	if ((raid5_calc_degraded(conf) > 0 ||
3059	     test_bit(Journal, &rdev->flags)) &&
3060	    conf->log->r5c_journal_mode == R5C_JOURNAL_MODE_WRITE_BACK)
3061		schedule_work(&log->disable_writeback_work);
3062}
3063
3064int r5l_init_log(struct r5conf *conf, struct md_rdev *rdev)
3065{
3066	struct request_queue *q = bdev_get_queue(rdev->bdev);
3067	struct r5l_log *log;
3068	char b[BDEVNAME_SIZE];
 
3069
3070	pr_debug("md/raid:%s: using device %s as journal\n",
3071		 mdname(conf->mddev), bdevname(rdev->bdev, b));
3072
3073	if (PAGE_SIZE != 4096)
3074		return -EINVAL;
3075
3076	/*
3077	 * The PAGE_SIZE must be big enough to hold 1 r5l_meta_block and
3078	 * raid_disks r5l_payload_data_parity.
3079	 *
3080	 * Write journal and cache does not work for very big array
3081	 * (raid_disks > 203)
3082	 */
3083	if (sizeof(struct r5l_meta_block) +
3084	    ((sizeof(struct r5l_payload_data_parity) + sizeof(__le32)) *
3085	     conf->raid_disks) > PAGE_SIZE) {
3086		pr_err("md/raid:%s: write journal/cache doesn't work for array with %d disks\n",
3087		       mdname(conf->mddev), conf->raid_disks);
3088		return -EINVAL;
3089	}
3090
3091	log = kzalloc(sizeof(*log), GFP_KERNEL);
3092	if (!log)
3093		return -ENOMEM;
3094	log->rdev = rdev;
3095
3096	log->need_cache_flush = test_bit(QUEUE_FLAG_WC, &q->queue_flags) != 0;
3097
3098	log->uuid_checksum = crc32c_le(~0, rdev->mddev->uuid,
3099				       sizeof(rdev->mddev->uuid));
3100
3101	mutex_init(&log->io_mutex);
3102
3103	spin_lock_init(&log->io_list_lock);
3104	INIT_LIST_HEAD(&log->running_ios);
3105	INIT_LIST_HEAD(&log->io_end_ios);
3106	INIT_LIST_HEAD(&log->flushing_ios);
3107	INIT_LIST_HEAD(&log->finished_ios);
3108	bio_init(&log->flush_bio, NULL, 0);
3109
3110	log->io_kc = KMEM_CACHE(r5l_io_unit, 0);
3111	if (!log->io_kc)
3112		goto io_kc;
3113
3114	log->io_pool = mempool_create_slab_pool(R5L_POOL_SIZE, log->io_kc);
3115	if (!log->io_pool)
3116		goto io_pool;
3117
3118	log->bs = bioset_create(R5L_POOL_SIZE, 0, BIOSET_NEED_BVECS);
3119	if (!log->bs)
3120		goto io_bs;
3121
3122	log->meta_pool = mempool_create_page_pool(R5L_POOL_SIZE, 0);
3123	if (!log->meta_pool)
3124		goto out_mempool;
3125
3126	spin_lock_init(&log->tree_lock);
3127	INIT_RADIX_TREE(&log->big_stripe_tree, GFP_NOWAIT | __GFP_NOWARN);
3128
3129	log->reclaim_thread = md_register_thread(r5l_reclaim_thread,
3130						 log->rdev->mddev, "reclaim");
3131	if (!log->reclaim_thread)
3132		goto reclaim_thread;
3133	log->reclaim_thread->timeout = R5C_RECLAIM_WAKEUP_INTERVAL;
 
 
3134
3135	init_waitqueue_head(&log->iounit_wait);
3136
3137	INIT_LIST_HEAD(&log->no_mem_stripes);
3138
3139	INIT_LIST_HEAD(&log->no_space_stripes);
3140	spin_lock_init(&log->no_space_stripes_lock);
3141
3142	INIT_WORK(&log->deferred_io_work, r5l_submit_io_async);
3143	INIT_WORK(&log->disable_writeback_work, r5c_disable_writeback_async);
3144
3145	log->r5c_journal_mode = R5C_JOURNAL_MODE_WRITE_THROUGH;
3146	INIT_LIST_HEAD(&log->stripe_in_journal_list);
3147	spin_lock_init(&log->stripe_in_journal_lock);
3148	atomic_set(&log->stripe_in_journal_count, 0);
3149
3150	rcu_assign_pointer(conf->log, log);
3151
3152	set_bit(MD_HAS_JOURNAL, &conf->mddev->flags);
3153	return 0;
3154
3155	rcu_assign_pointer(conf->log, NULL);
3156	md_unregister_thread(&log->reclaim_thread);
3157reclaim_thread:
3158	mempool_destroy(log->meta_pool);
3159out_mempool:
3160	bioset_free(log->bs);
3161io_bs:
3162	mempool_destroy(log->io_pool);
3163io_pool:
3164	kmem_cache_destroy(log->io_kc);
3165io_kc:
3166	kfree(log);
3167	return -EINVAL;
3168}
3169
3170void r5l_exit_log(struct r5conf *conf)
3171{
3172	struct r5l_log *log = conf->log;
3173
3174	conf->log = NULL;
3175	synchronize_rcu();
3176
3177	/* Ensure disable_writeback_work wakes up and exits */
 
 
 
 
3178	wake_up(&conf->mddev->sb_wait);
3179	flush_work(&log->disable_writeback_work);
3180	md_unregister_thread(&log->reclaim_thread);
3181	mempool_destroy(log->meta_pool);
3182	bioset_free(log->bs);
3183	mempool_destroy(log->io_pool);
3184	kmem_cache_destroy(log->io_kc);
3185	kfree(log);
3186}